Four-dimensional agent-based approaches for

modeling spatial dynamic phenomena using complex

systems and GIScience theories

by
Alexander Smith

M.Sc., University of Waterloo, 2017

B.E.S., University of Waterloo, 2015

Thesis Submitted in Partial Fulfillment of the

Requirements for the Degree of
Doctor of Philosophy

in the
Department of Geography
Faculty of Environment

© Alexander Smith 2022

SIMON FRASER UNIVERSITY
Fall 2022

Copyright in this work is held by the author. Please ensure that any reproduction
or re-use is done in accordance with the relevant national copyright legislation.
Declaration of Committee

Name: Alexander Smith

Degree: Doctor of Philosophy

Title: Four-dimensional agent-based approaches for

modeling spatial dynamic phenomena using
complex systems and GIScience theories

Committee: Chair: Bing Lu

Assistant Professor, Geography

Suzana Dragićević
Supervisor
Professor, Geography

Margaret Schmidt
Committee Member
Associate Professor, Geography

Hao (Richard) Zhang

Committee Member
Professor, Computing Science

Nicholas Hedley
Examiner
Associate Professor, Geography

Andrew Crooks
External Examiner
Professor, Geography
University at Buffalo

ii
Abstract

Spatial phenomena are often perceived as complex systems (CS) that cannot be fully
represented through equation or statistical modeling as they are spatial, dynamic, and
nonlinear. This requires alternate methods such as geographic automata systems (GAS)
including cellular automata (CA) and agent-based modeling (ABM), based on
geographic information science (GIScience), geographic information systems (GIS), and
CS theories, to achieve a bottom-up representation of spatial processes that generate
spatial patterns. However, most GAS use two-dimensional space over time to represent
the dynamics of change of spatial systems, while in reality these systems exist in three-
dimensional (3D) space that is perpetually changing over time, thus are four-dimensional
(4D). Developments have advanced data representation into 3D space using several
GIS data formats, LIDAR data and voxel automata, the 4D equivalent of CA. However,
the ABM frameworks have not experienced the same progress, and therefore, the main
objective of this dissertation includes the development and implementation of a novel 4D
ABM theoretical framework applied on two case studies. The first case study implements
the 4D ABM approach to represent the propagation of forest-fire smoke in British
Columbia, Canada. The second case study implements the 4D ABM approach to
represent the predator-prey dynamics of the Southern Resident Killer Whale and
Chinook Salmon in the Salish Sea, Canada and USA, and identifies locations of whale
presence for marine environment preservation. One challenge of 4D ABM approaches is
the model testing methods, specifically the availability of comparison techniques for 3D
map and simulation outcomes. Therefore, in addition, this dissertation proposes new 3D
and 4D map comparison techniques based on several 3D and 4D Kappa measures,
using probability and fuzzy logic to calculate the similarity between two 3D maps.
Obtained results indicate that the 4D ABM theoretical framework can effectively be
implemented to represent 4D spatial dynamic phenomena for multiple proposed
scenarios. The results of this dissertation have potential for use in decision-making,
spatial planning and policy testing for conservation purposes and environmental
protection. This thesis contributes to the advancements of the field of GIScience and
modeling with multidimensional geographic automata systems.

iii
Keywords: geographic information science; complex systems theory; 4D agent-
based modeling; 3D map comparison; geographic automata systems;
four dimensions (4D)

iv
Dedication

This thesis is dedicated to my family and friends, without their support I would have
never been able to get to this point

v
Acknowledgements

I would like to acknowledge the support from the Natural Science and
Engineering Research Council (NSERC) of Canada through the Canada Discover Grant
awarded to Dr. Suzana Dragićević. I would also like to acknowledge the support from the
Department of Geography and Simon Fraser University through various awards, grants,
and scholarships, including two Graduate Fellowships, three Travel & Research Awards,
the Presidents PhD Scholarship, the Canadian Pacific/Teck Resources Award, and the
Wayne Goeson Memorial Award.

I would like to thank the numerous people through my academic career that have
been mentors, colleagues, co-workers, and friends. I would like to specifically thank Dr.
Suzana Dragićević for her support throughout my Ph.D. program, she has been vital in
my growth as an academic and scholar. Her mentorship has been immeasurable to me.
I would also like to thank my supervisory committee, Drs. Margaret Schmidt and Richard
Zhang, for their support and feedback as I progressed through my program. To my
fellow SAM Lab members and Geography graduate students, thank you for making my
years at SFU the wonder they were. I would like to thank the Department of Geography
staff that keep the department running smooth and helping me solve all my
administrative and IT problems, including Anke Baker, Joyce Chen, Erin Rodgers, Erin
Huddleston, Curtis Platson, B-Jae Kelly, Justin Song, John Ng, and Chris Au-Yeung,
and anyone else I may have missed.

vi
Table of Contents

Declaration of Committee.................................................................................................. ii
Abstract .............................................................................................................................iii
Dedication ......................................................................................................................... v
Acknowledgements .......................................................................................................... vi
Table of Contents .............................................................................................................vii
List of Tables .................................................................................................................... ix
List of Figures.................................................................................................................... x
List of Acronyms...............................................................................................................xii

Chapter 1. Introduction.............................................................................................. 1
1.1. Brief Literature Overview on Multidimensional Models ........................................... 5
1.2. Research Problems and Questions ........................................................................ 8
1.3. Research Objectives ............................................................................................... 9
1.4. Dissertation Overview ............................................................................................. 9
1.5. References ............................................................................................................ 12

Chapter 2. A Four-Dimensional Agent-Based Model: A Case Study of Forest-

Fire Smoke Propagation ..................................................................................... 20
2.1. Introduction ........................................................................................................... 20
2.2. Theoretical Framework of a 4D Agent Based Model ............................................ 24
2.3. Smoke Propagation Model .................................................................................... 26
2.3.1. Study Area and Data ............................................................................... 26
2.3.2. Model Overview ....................................................................................... 26
2.3.3. Model Components ................................................................................. 29
2.4. Results and Discussion ......................................................................................... 33
2.4.1. Geosimulation Outputs ............................................................................ 33
2.4.2. Model Validation ...................................................................................... 39
2.5. Conclusions ........................................................................................................... 41
2.6. References ............................................................................................................ 42

Chapter 3. Assessing the similarities of 3D simulation model outcomes .......... 47

3.1. Introduction ........................................................................................................... 47
3.2. Methodology .......................................................................................................... 49
3.3. Implementation ...................................................................................................... 50
3.4. Results and Discussion ......................................................................................... 54
3.5. Conclusions ........................................................................................................... 57
3.6. References ............................................................................................................ 58

Chapter 4. Map comparison methods for four-dimensional space and time

voxel data ............................................................................................................. 60
4.1. Introduction ........................................................................................................... 60
4.2. Mathematical background ..................................................................................... 62
4.2.1. Cohen’s Kappa ........................................................................................ 62

vii
 4.2.2. Fuzzy Kappa............................................................................................ 63
4.3. Methods ................................................................................................................ 65
4.3.1. Geospatial Data Representation in 3D and 4D ....................................... 65
4.3.2. Approaches for 3D Map Comparisons .................................................... 66
3D and 4D Accuracy............................................................................................ 66
3D Kappa Coefficient........................................................................................... 66
3D Fuzzy Kappa Coefficient ................................................................................ 67
4D Fuzzy Kappa .................................................................................................. 71
4.4. Implementation ...................................................................................................... 74
4.5. Results and Discussion ......................................................................................... 81
4.6. General Discussion ............................................................................................... 85
4.7. Conclusions ........................................................................................................... 86
4.8. References ............................................................................................................ 87

Chapter 5. Representing the Southern Resident Killer Whales in the Salish Sea
using a four-dimensional agent-based model .................................................. 93
5.1. Introduction ........................................................................................................... 93
5.2. Methods ................................................................................................................ 97
5.2.1. Study Area and Datasets......................................................................... 97
5.2.2. Agent Based Model of Southern Resident Killer Whales ........................ 98
Agents................................................................................................................ 102
Whale Agents (WAs) ......................................................................................... 103
Salmon Agents (SAs) ........................................................................................ 105
Model Computing Environment ......................................................................... 105
5.3. Simulation Results .............................................................................................. 106
5.3.1. Model Testing ........................................................................................ 111
5.4. Discussion ........................................................................................................... 112
5.5. Conclusions ......................................................................................................... 115
5.6. References .......................................................................................................... 116

Chapter 6. Conclusions ......................................................................................... 122

6.1. General Conclusions ........................................................................................... 122
6.2. Summary of Findings .......................................................................................... 122
6.3. Future Research ................................................................................................. 125
6.4. Research Contributions ....................................................................................... 128
6.5. References .......................................................................................................... 128

viii
List of Tables

Table 2.1 The ODD protocol of the 4D ABM ........................................................... 30

Table 3.1 Values obtained for 3D Accuracy and 3D Kappa for the six comparisons
of the forest fire smoke propagation ABM simulation outputs. ................ 54
Table 3.2 Values obtained for 3D Accuracy and 3D Kappa for the 3D comparisons
of four forest-fire smoke model simulation outputs for two runs S1 and S2.
................................................................................................................. 56
Table 3.3 Values obtained for 3D Kappa for the comparisons of the forest-fire
smoke propagation model simulation outputs with varying diffusion rates
(DF). ........................................................................................................ 57
Table 4.1 Allocation of shell numbers, voxel distances, and similarity of 3D distance
for the 5x5x5 3D voxel window. ............................................................... 68
Table 4.2 Hypothetical 3D maps with two and four classes and their characteristics.
................................................................................................................. 75
Table 4.3 Voxelized LiDAR 3D data representing different types and heights of
trees......................................................................................................... 78
Table 4.4 Two soil horizon 3D maps with depths and thicknesses. ........................ 79
Table 4.5 Results from the comparisons of hypothetical 3D voxel data. ................. 82
Table 4.6 Results from the comparisons of 3D tree maps. ..................................... 83
Table 4.7 3D Fuzzy Kappa values for 3D soil data by 3D Voxel Window Size. ...... 84
Table 4.8 Values for 4D Fuzzy Kappa obtained for comparisons of forest-fire smoke
simulation outputs.................................................................................... 84
Table 5.1 The Overview, Design, and Details (ODD) of the 4D ABM. .................. 100
Table 5.2 Results of simulation outcome comparisons for Scenarios A, B, C, and D.
............................................................................................................... 112

ix
List of Figures

Figure 1.1 Integration of CS and GIScience theories with GAS approaches

organized by space, time and topology ..................................................... 7
Figure 2.1 Schematic representation of 4D ABM with the agents’ states S, input data
I, and transitions rules R from initial time t and constitutive iterations at t+1
to t+n. ...................................................................................................... 25
Figure 2.2 Flow chart of 4D ABM of smoke propagation process with information on
each agent type. ...................................................................................... 28
Figure 2.3 4D ABM simulation smoke outputs presented in 2D for every 2 days up to
12 days for (A) Scenario 1 with static fire, and (B) Scenario 2 with moving
fire, (C) comparison of Day 12 for both scenarios, and (D) the path of the
Fire Agent for Scenario 2......................................................................... 35
Figure 2.4 Geosimulation outputs for Scenario 2 with moving Fire Agent depicting
(A) an oblique view of Smoke Agents after 24 hours from a low view
point; oblique views (left of pair) and vertical views (right of pair) of Smoke
Agents forming a plume from the moving fire location with outputs after
(B) 3 hours, (C) 6 hours, (D) 9 hours and (E) 12 hours from the
initialization of the model. Vertical distances scaled 5x to exaggerate
elevation differences................................................................................ 36
Figure 2.5 The quantification of simulation distribution of all Smoke Agents in 100 m
bins of vertical distance at a daily time step for Scenario 2. .................... 38
Figure 2.6 Examples of smoke propagation patterns for British Columbia and
Alberta, Canada, on (A) August 11th, 2017, (B) August 14th, 2017, (C)
July 18th, 2017, and (D) July 6th, 2015 (NASA, 2019). ........................... 40
Figure 3.1 Hierarchy of Kappa approaches originating from Cohen’s Kappa ........... 48
Figure 3.2 Agent-based model forest-fire smoke propagation simulation outcomes in
voxel data format used for six comparisons. ........................................... 52
Figure 3.3 Simulation outputs S1 and S2 at initial stage and after three hours, six
hours, and nine hours of smoke propagation. ......................................... 53
Figure 3.4 Randomly distributed smoke voxels highlighted in red from ABM outputs
for simulation S1 and S2. ........................................................................ 55
Figure 4.1 Evolution of GIS data representation from 2D map containing rasters to
3D map containing voxels in 3D space or 4D as space-time defined by
x,y,z coordinates and time t with three map classes. ............................... 66
Figure 4.2 The sliced 3D voxel window extracted from a 3D map, consisting of 5x5x5
voxels with the shell numbers for each voxel. ......................................... 67
Figure 4.3 Shell numbers in the sliced space-time 4D voxel window of 5×5×5×3 size
with time steps ts = 3 and time offset o = 2 for t-1, t, and t + 1. ............... 73
Figure 4.4 Voxelized model outputs at t = 3 for the three scenarios of forest-fire
smoke geosimulations. ............................................................................ 80
Figure 5.1 Study area Salish Sea, Canada and USA, as study area with key
datasets: (A) Bathymetry, (B) Shore Distance, and (C) Vessel Density. . 98
Figure 5.2 Flowchart of the 4D ABM approach for spatial simulations of Southern
Resident Killer Whales (SRKW). ............................................................. 99

x
Figure 5.3 Locations of whale and salmon agents in 3D for three consecutive output
intervals at time step t=0, t=60 and t=120 and for a selected section of the
study area. ............................................................................................. 107
Figure 5.4 Simulation outputs for density of locations for whale agents for Scenarios
A, B, C, and D for all iterations of each scenario representing 365 days.
............................................................................................................... 108
Figure 5.5 Simulation outputs for accumulated density of locations for whale agents
for the subsection around the San Juan Islands for Scenarios A, B, C,
and D. .................................................................................................... 110

xi
List of Acronyms

2.5D Two-and-a-half dimensional

2D Two-dimensional
3D Three-dimensional
4D Four-dimensional
ABM Agent-based model
AIS Automatic identification systems
AQHI Air Quality Health Index
BC British Columbia
BIM Building information modeling
CA Cellular automata
CS Complex system
DEM Digital elevation model
DF Diffusion factor
EMD Earth movers distance
GAS Geographic automata systems
GIS Geographic information system
GIScience Geographic information science
LiDAR Light detection and ranging
NOAA National Oceanic and Atmospheric Administration
NPCs Non-player characters
ODD Overview, Design, and Details
SA Salmon agent
SRKW Southern Resident Killer Whale
USA United States of America
USGS United States Geological Survey
VA Voxel automata
WA Whale agent

xii
Chapter 1.

Introduction

Geographical or geospatial systems can be categorized as complex systems

(CS) as they are process-based and they operate in space and time. They are
characterized by their aggregate complexity, where the local interactions and behaviors
of individuals work from the bottom-up to cause the properties of the system to emerge
from local to larger spatial scales (Messina et al., 2008). Based on complexity theory, the
CS have several characteristics such as non-linearity, evolution, emergence, self-
organization, adaptation, and feedback, which make CS difficult to represent (Manson,
2001) as they can create multiple continuous states (Burger et al., 2021). These
characteristics give the system various dynamic properties, such as changes to system
structure, elements’ interactions and behaviors as a result of changing environments and
internal forces, with processes that reinforce or repress these changes (Crooks et al.,
2008).

The characteristics of CS and their spatial and dynamic nature that is close to
chaotic make them difficult to represent and predict (Batty and Torrens, 2005). Some
modeling methods are not capable of including the internal mechanisms and instead
represent the system-level spatial process. Models can be categorized into two general
types, equation models (Rahmandad and Sterman, 2008) and process models (Alam
and Dutta, 2012). While equation models may be sufficient in some cases, they
represent the effects of the processes within the system through mathematical equations
(Paraunak et al., 1998), and as such, do not represent the processes directly. This can
cause the attributes of the spatial CS to be ignored and only have their outputs
represented in a top-down manner (Murray-Rust et al., 2014; Rounsevell et al., 2012).
Process models operate in the opposite direction, bottom-up, where they represent the
individual-level processes and allow for the system-level patterns to emerge over space
and time (Luus et al., 2011; Parunak et al., 1998). Process models represent a spatial
CS in a way that emulates how the spatial CS itself operates and can more closely act in
a similar manner (Brown et al., 2005).

1
 Spatial processes of a CS can rely on geospatial data and thus the concepts of
geographic information science (GIScience) theory can be combined for their
mathematical representations with approaches such as cellular automata (CA) and
agent-based models (ABM). This integration of CS with GIScience has been achieved
with the concept of geographic automata systems (GAS) (Torrens and Benenson, 2005)
that are allowing representation of dynamic spatial processes with the use of geospatial
data in two-dimensional (2D) space and time. In a CA approach, 2D space is
represented by cells that are typically homogeneous squares, such as rasters which are
typical for geographic information systems (GIS) data formats and remote sensing (RS)
data. Sets of transition rules define the process to be modeled and are applied to each
cell (Batty et al., 1999; White and Engelen, 2000). The transition rules update the cell
state, while considering the states of the surrounding cells, for each time step of the
model. ABMs approaches work with a similar concept to CAs, with some important
differences. Transition rules in an ABM rather represent agents’ individual behavior,
interaction or decision-making to name a few (Crooks et al., 2008). Each agent type has
their own set of transition rules that incorporate an agent’s individual characteristics and
environment in the decision-making process to determine their behavior as they interact
with other agents and the environment (Ligmann-Zielinska and Sun, 2010). Agents’
individual characteristics can change as they learn from their actions and adapt to a
changing environment, changing their behavior (Parker et al., 2003). Instead of being
spatially static like cells in a CA, agents move through the environment as dictated by
their decision-making processes (Heppenstall et al., 2016).

The GAS approaches have been widely used in geographical applications. CAs,
for example, have been used to represent many physical systems on the Earth’s surface
such as landslides (Segre and Deangeli, 1995), soil erosion (Heung et al., 2013), or
sand dunes (Narteau et al., 2009), and land systems such as urban growth (Batty et al.,
1999; White et al., 1997) and land use change (Mitsova et al., 2011) or ecological
systems such as predator-prey systems (Chen et al., 2012) or insect infestations (Bone
et al., 2006), to name a few. These CS systems have been represented as 2D CA
changing with time. Further, ABMs are typically applied to represent systems involving
several individuals with decision-making processes, such as land use change (Castella
et al., 2005; Mialhe et al., 2012; Murray-Rust et al., 2014; Parker et al., 2003), insect
infestations (Anderson and Dragićević, 2015), disease spread (Perez and Dragicevic,

2
2009), sand dunes (Genois et al., 2013), chemical reactions (Agusdinata et al., 2015),
and ballistic fragment impact (Bova et al., 2016). A limitation of these GAS models is
they represent space as 2D, when all real-world geographic systems are operating in
three-dimensional (3D) space and time, thus are four-dimensional (4D) (Goodchild,
2009). The advance of GIScience to 4D is also one of the five topics identified by
Goodchild (2010) for research in GIScience in the next decade. While this advancement
has begun, ABMs have not experienced it to the same extent.

The integration of CS theory with GIScience theory is operationalized through the

concept of GAS and geosimulation approaches to better represent spatial dynamic
phenomena and with use of geospatial data. CS theory considers the approach to
defining and representing systems from their individual components and relationships
instead of reducing the system to its generalized behaviour (Batty and Torrens, 2005).
This theory is represented in the bottom-up approach of GAS modeling. GIScience
theory considers the representation of the various aspects of the geographical world,
such as space, time, attributes, and relationships (Cova and Goodchild, 2002). Due to
inherited cartographic principles that are transposed into the digital geospatial world,
geographic information systems (GIS) have been traditionally focused on 2D space
(Dragicevic and Marceau, 2000; Zlatanova et al., 2002). The main aspect of GIScience
theory still relied upon are the uses of geospatial data in that are in vector and raster
data formats that fall under the object and field perspectives, respectfully. Time is
generally represented iteratively with attributes changing with each iteration, or as an
attribute of space (Langran and Chrisman, 1988). GAS use these theories of GIScience
in the field and object representations within CAs and ABMs, and time applies through
iterative applications of transition rules, changing the state and attributes of the model
and its components. A GAS can also represent relationships of GIScience through
network theory using nodes and links to represent individuals and their relationships
(Anderson and Dragićević, 2018). It also uses Tobler’s First Law of geography where
near things are more similar than distant things (Tobler 1970) which is operationalized
through the concept of neighborhood where local interactions are influencing the
process that spread on larger scale.

There are alternatives to the object and field perspectives for representing
individuals of a system, however their integration with GAS is still in its infancy. These
alternatives include concepts based off fields and objects such as the multi-

3
representational perspective, network-based field, and corridor object field (Cova and
Goodchild, 2002). More recently, another alternative is the geo-atom which has been
introduced as an abstract representation of geographic information (Goodchild et al.,
2007) that can be integrated to modeling GAS (Jjumba and Dragićević, 2015).

Compressing 3D reality into 2D space creates difficulties in representation when

multiple objects interfere and overlap, causing a single location to belong to more than
one object (Arroyo Ohori et al., 2015). Some data can be considered between full
dimensions, called 2.5D, and includes digital elevation models (DEM). Data in 2.5D
resides on a non-overlapping 3D surface with defined height extruded above a 2D plane
(Lenk and Heipke, 2006). Advancements in 3D space representation have resulted in
point datatypes such as point clouds used for Light Detection and Ranging (LiDAR)
(Hosoi et al., 2013), polyhedrons such as tetrahedra (Shi et al., 2003), in vector
datatypes such as Building Information Modeling (BIM) (Jung and Joo, 2011), and field
datatypes such as voxels, the 3D counterpart to the raster, (Jjumba and Dragićević,
2016a). With these representations, the real world can be shown in 3D, however
advancements in the representations of systems as processes that change over 3D and
time must also be addressed.

Model testing is a vital part in the development of all types of models and has
several key steps. These include model verification, sensitivity analysis, calibration,
validation, and qualification (Pérez et al. 2013). Verification of a model is the assessment
of the integrity of the implemented code and processes to ensure the model is free of
errors and bugs and that the model operates as expected (Parker et al. 2003). The
sensitivity analysis can be a part of verification, and is the quantification of the impacts of
incremental changes to variables (Ligmann-Zielinska, 2013). Calibration and validation
are highly connected steps of model testing done to ensure a verified model can
accurately represent the phenomena (van Vliet et al., 2016). Calibration involves
comparing model outputs to a portion of the available reference data to set parameters
to fit the model to the specific case being represented. Validation then uses the
remaining reference data to assess the ability of the model to represent reality. Both
calibration and validation can be conducted on portions of the model and the model in
entirety. Qualification of a model involves determining its domain of applicability to set
limits on its use (Rykiel, 1996). Together, the model testing steps ensure a model works
as intended with a known performance for an identified domain so that results can be

4
trusted and used in decision-making processes. Several difficulties can be encountered
when conducting model testing on models of CS. Properties of CS such as sensitivity
and path dependence can lead to nonlinearity and positive feedback (Manson, 2007).
Model testing techniques for models such as ABMs must account for these properties in
the system while also dealing with the increased runtime due to the level of complexity
being represented (Luus et al., 2011). This reduces the amount of simulation available
for model testing and contributes to the struggle to calibrate and validate a model.

1.1. Brief Literature Overview on Multidimensional Models

The use of 3D and 4D models in general is not a novel concept as they have
been used in other fields including medical imaging, computer graphics and vision,
aerospace, and civil engineering. However, their use in geospatial applications has been
limited, and the applications in other fields are either not appropriate for geospatial
applications or have not been adapted or migrated. Within the field of medical imaging,
3D representation and 4D models are typically used with various forms of scans, such
as magnetic resonance imaging (MRI) (Fritz et al., 2014; Wright et al., 2017) and CT
imaging (Li et al., 2008). However, these models are typically not simulations and are
instead used to display observations or for image visualization, analysis, and
comparison similar to procedures in GIScience for processing geospatial data for maps
and classification of RS images. The fields of aerospace and civil engineering use 3D
and 4D modeling in ways that are dissimilar to the geospatial applications and are
generally less applicable to GIScience. Some models such as computation fluid
dynamics of aircraft and wind effects on buildings (Baghaei Daemei et al., 2019; Russo
et al., 2021) may have applications for atmospheric and hydrological modeling, however
other models such as deformation models (Che Osmi et al., 2020) are difficult to link to
GIScience concepts and geospatial data.

Computer graphics and vision have many applications that are useful in
geospatial applications. The crossover of methods has already occurred, especially for
uses in image classification for remote sensing such as LiDAR (Livny et al., 2010;
Maturana and Scherer, 2015; Seiferling et al., 2017) and soil classification (Srivastava et
al., 2021). In computer graphics, perhaps the most notable use is in applications such as
video games, with 3D particle physics, ray tracing, object interactions, and 3D
environments, all being interacted with by the player and rendered for display on a 2D

5
computer screen. Some video games with non-player characters (NPCs) can also be
viewed as similar to ABMs. An NPC is like an agent, where they follow a set of rules
depending on their character type and interact with other NPCs, the player, and the
environment. Role-playing games are also sometimes used in the description of ABMs
(Castella et al., 2005). Several computer vision methods may be beneficial to the field of
GIScience for map comparison and analysis, such as Earth Mover’s Distance (EMD)
(Rubner et al., 2000). EMD, also called the Wasserstein metric, calculates the difference
between probability distributions using the distance the distributions need to move to
match and can be useful in map comparison for model testing (Potts et al., 2014). Forms
of EMD can also be applied to contour tracking, showing potential for object detection in
satellite images (Li, 2013). Computer vision methods can also be used for the
conversion of 3D data between datatypes, such as voxelization (Nourian et al., 2016). A
major difference between computer vision and geospatial applications is the scale, with
computer vision generally focused on smaller scale applications such as buildings (Nan
et al., 2010), compared to the use at any scale from local to global in geospatial
applications. However, this should not be difficult to overcome with either using new
training data or a redesign using the computer vision concepts alongside georeferencing
of the data. Additionally, one could argue that many of the methods used in image
classification are already a form of computer vision (Poomani et al., 2021).

CA can be extended into 4D by using voxel geographic automata (VA) (Figure

1.1) (Greene, 1989; Jjumba and Dragićević, 2016c). However, ABMs have not yet been
extended into the 4D realm by postulating a 4D ABM theoretical framework and its
applications thus this represents a gap in the literature. The integration of concepts of
CS and GIScience theories via GAS approaches in the space, time and topology cube
has been presented in Figure 1.1. In a VA, the cells of a CA are extended into 3D with a
3D grid forming voxels. Voxels can be uniform, regular, irregular, or structured,
depending on the field topology required (Fritsch, 1996). Each voxel contains the values
that defines the volume enclosed in the same way a cell defines the enclosed area.
While the voxels of a VA are typically uniform cubes, other shapes can be used
(Karssenberg and De Jong, 2005). Similarly, a 4D ABM can use voxel data, along with
3D vector data such as points, lines, and polyhedrons. The integration of various data
representation types allows ABMs to utilize object field theory, such as where the object

6
of an agent can have a neighborhood it that is a field as a medium to interact with the
space or volume around it.

Figure 1.1 Integration of CS and GIScience theories with GAS approaches

organized by space, time and topology

In both VA and 4D ABM, the transition rules need to be updated to account for
3D space. While typical CA transition rules may depend on the nine cells of a three-by-
three Moore neighborhood in 2D space, VA use the 27 voxels of an equivalent 3D
Moore neighborhood (Jjumba and Dragićević, 2016c). This exponential growth in the
number of neighbors greatly increases the amount of data needed to properly represent
the multi-dimensional spatial process and thus augments computational requirements to
run 4D agent-based models. The transition rules also vary between the models in how
time is represented, based on Goodchild’s (1992) five options for representing time.
Time in a CA and VA, can be considered as option 1, with “a finite number of discrete
time slices, each views as a field” (Goodchild, 1992, p. 407) while in ABMs it can be

7
considered as option 4, with “discrete time slices, with objects identified in each slide
and linkages between corresponding objects at different times” (Goodchild 1992, p. 407)

One of the challenges of geographic automata systems approaches is

performing full model testing (Crooks et al., 2008; Dragicevic and Marceau, 2000; Kang
et al., 2022). The advancement of map comparison methods into 3D space and 4D is
used for model testing such as calibration, validation, sensitivity analysis, and
verification. Many of the methods used for model testing are not compatible with 3D or
4D data (Jjumba and Dragićević, 2016b), a modern continuation of the trend of a lack of
spatial analysis methods already identified for 2D data by Openshaw (1998) in the late
1990s. This lack of analytical methods that operate with 3D or 4D data forms a second
gap in the literature. With model testing already a difficult process for GAS, these
methods must also advance if VA or 4D ABM are going to be useful. Some map
comparison methods are trivial to advance into 3D or 4D because they do not rely on
spatial relationships or neighborhoods. These include accuracy and Cohen’s Kappa
(Cohen, 1960) which only require pairing voxels between maps. However, spatially fuzzy
methods such as Fuzzy Kappa (Hagen-Zanker, 2009; Hagen, 2003) require major
advancement in order to be adequate and functional for the use with 3D and 4D data
and geosimulation models.

1.2. Research Problems and Questions

There are several gaps in the scientific literature on 4D spatial CS modeling and
4D map comparison that are of concern. The first gap is the lack of 4D ABMs to
complement the extension of CA into VA within GAS. This will bring GAS and CS
modeling further into the realm of 4D, which will assist in the representation of the
inherently 4D spatial CS. This gap is composed of two parts, the theoretical and the
practical, of which both parts must be addressed. The second gap is the lack of
appropriate 3D and 4D model testing methods caused by the need for 3D and 4D map
comparison and analysis. Addressing this gap will enable more robust representations of
4D spatial CS with GAS 4D agent-based models (Figure 1.1). Based on the presented
gaps in the scientific literature, this dissertation is addressing the following research
questions:

8
 1. How can theories of complex systems and GISciences, particularly
geographic automata systems be used as a theoretical framework to
represent spatial processes that operate in four-dimensions (4D),
composed of three spatial dimensions (3D) and time?

2. How can a four-dimensional (4D) geographic automata system,

particularly 4D agent-based modeling (ABM) approaches, be
leveraged for the representation and analysis of real world multi-
dimensional dynamic spatial complex systems (CS)?

3. To what extent can existing spatial CS model testing approaches be

used to perform a comparison of 3D maps and simulation outcomes
as part of model testing of the developed four-dimensional (4D) agent-
based modeling approaches?

1.3. Research Objectives

To answer the identified research questions, this dissertation works towards

integration of complex system and GIScience theories to develop a novel 4D modeling
framework to represent dynamic spatial multi-dimensional phenomena. Therefore, the
main objectives of this thesis research are as follows:

1. To develop a theoretical 4D GAS, specifically a 4D ABM framework,

for the representation of spatial phenomena in 4D

2. To implement the 4D ABM framework on two case studies:

a. 4D ABM of the propagation of forest-fire smoke

b. 4D ABM of marine species interactions

3. To develop approaches for comparing 3D voxel data for the purposes

of comparing:

a. simulation modeling outputs in 3D and

b. comparing 3D voxel maps and testing 4D ABM models

1.4. Dissertation Overview

This thesis is composed of five chapters, starting with this Introduction chapter,
followed by four chapters addressing the three research objectives, and a concluding
chapter. Chapter 2 introduces the novel theoretical framework of 4D ABMs to extend the
field of ABMs into 4D. This framework is used to develop a 4D ABM with relatively basic
agent functions informed by physical interactions without complex decision-making

9
abilities. The empirical equations are used for representing agents’ interactions within
the model as one end of a spectrum of agents’ characteristics with no decision-making
abilities as useful starting point for implementing the developed framework. At the other
end other characteristics and decision-making abilities of the agents have been
deployed at later stage for second case study. For the case study in this chapter, the
spatial phenomenon of forest-fire smoke propagation is used as an example to apply the
modeling framework. Fire agents represent the source of smoke, such as forest fires or
point source emitters, and smoke agents represent the emitted smoke as it propagates
around the study area. The smoke is discretized into agents that represent the
movement of individual smoke particles in order to understand the movement of the
larger plumes of smoke. The smoke agents’ movements are determined by smoke
concentration, the terrain, and atmospheric conditions using physics-based calculations
and calibrations. The 4D ABM is implemented on the datasets for the study area of
British Columbia (BC), and western Alberta, Canada, and northern parts of Washington
and Montana, USA. Two datasets were used, atmospheric data from the National Center
for Atmospheric Research, North American Regional Reanalysis (National Centers for
Environmental Prediction et al., 2005), and digital elevation data from the United States
Geological Survey (USGS) Global Digital Elevation Model (GTOPO30) (U.S. Geological
Survey, 2012). While the framework sets the theoretical space for the 4D ABM, the
developed 4D ABM identifies the advancements needed for the future use of 4D ABMs.
This primarily includes the need for model testing methods that operate in 3D and 4D,
and the need for further advancement of 4D ABMs with more complex agent behavior.

Chapter 3 presents new 3D voxel map comparison methods, 3D accuracy and

3D Kappa. These methods are novel extensions of existing common 2D map
comparisons as there are no readily available 3D methods for use with 3D geospatial
data. These methods are applied to outputs from the smoke propagation 4D ABM to
determine the similarity between scenarios. Both 3D accuracy and 3D Kappa lack any
spatial logic beyond the matching of voxels between 3D maps, therefore the next step of
the thesis research is to develop 3D and 4D Fuzzy Kappa methods.

The 3D Kappa method is then further expanded in Chapter 4 where 3D and 4D

model testing methods are presented in greater detail including comparisons of 3D map
and other 3D spatial objects with developed methods of 3D and 4D Fuzzy Kappa
adopted to compare 4D ABM simulation outputs. When conducting model testing for

10
ABMs in 2D space, there are many metrics and map comparisons to select from to find
the best fit, however 3D space and 4D do not have this benefit. In order to assist in the
model testing of 4D ABMs, and other 4D models or comparing 3D maps, new methods
must be developed. Cohen’s Kappa and Fuzzy Kappa techniques were selected to
advance into 3D and 4D as they have been used in geospatial applications and account
for chance agreement. The fuzzy logic of the 3D and 4D Fuzzy Kappa methods allows
for partial similarity between two points in space instead of the binary logic of similar and
dissimilar. Fuzzy Kappa was advanced into 3D and 4D Fuzzy Kappa through the
extension of the 2D neighborhood composed of rings of equidistant cells into a 3D
neighborhood composed of shells of equidistant voxels. The advancement into time was
introduced through the inclusion of a time offset, allowing the user to make voxels from
adjacent time intervals to be seen at a greater distance than the same locations from the
same time interval. 3D and 4D Fuzzy Kappa was demonstrated through the application
of various 3D datasets and the simulation outputs of the 4D ABM of smoke propagation.

With the methodological abilities to compare simulation outputs, Chapter 5

focuses on the development of another more complex 4D ABM. The developed 4D ABM
contains multiple interacting agents with decision-making abilities above and beyond that
of the 4D ABM developed in Chapter 2. This model also changes medium from the air of
Chapter 2 to water, demonstrating the broad range in useful applications of the
framework, while still operating in a 3D fluid that can be moved through. The model
represents the predator-prey system of the Southern Resident Killer Whale (SRKW) and
their primary prey, Chinook Salmon, within the Salish Sea between Vancouver Island,
Canada and the mainland, including waters of BC, Canada, and Washington, USA.
These two agent types use their own parameters, the environment, and nearby agents
as they live, interact, and navigate within the study area. The model uses three datasets,
obtained from two sources. The first is the bathymetry data from the Western
Washington University Salish Sea Atlas (Western Washington University, 2022), and this
was also used to generate the distance from shore dataset. The model also used vessel
density from 2013 obtained from the Office for Coastal Management at the National
Oceanic and Atmospheric Administration (NOAA) (Office for Coastal Management,
2022). This study highlights areas and conditions that may be important for the
protection of both species. It also identifies the next steps needed in order to improve
model testing of 4D ABMs. This includes the further advancement of geospatial specific

11
map analysis methods, such as 3D kernel density, to enhance the inputs to map
comparison methods.

Both 4D ABMs were developed using Repast Simphony (North et al., 2013,
2007), a Java based platform that uses the Eclipse integrated development environment
(IDE). While Repast Simphony works with 2D geospatial data, 3D and 4D compatibility is
limited, and most processes required custom programming. 3D and 4D Fuzzy Kappa
were developed using the Python (The Python Software Foundation, 2020)
programming language. The visualizations were made using a variety of software
options, including ArcMap, ArcScene, ArcGlobe, and ArcGIS Pro from the ESRI product
line (ESRI, 2018), Autodesk Fusion 360 (Autodesk, 2019), and MagicaVoxel (Ephtracy,
2019).

The final chapter concludes this doctoral dissertation with a short summary of the
obtained research results, explores future directions of the research, and situates the
contributions within the fields of research.

1.5. References

Agusdinata, D. B., Amouie, M., & Xu, T. (2015). Diffusion dynamics and concentration of
toxic materials from quantum dots-based nanotechnologies: an agent-based
modeling simulation framework. Journal of Nanoparticle Research, 17(1), 26.
https://doi.org/10.1007/s11051-014-2844-x

Alam, M. J., & Dutta, D. (2012). A process-based and distributed model for nutrient
dynamics in river basin: Development, testing and applications. Ecological
Modelling, 247, 112–124. https://doi.org/10.1016/j.ecolmodel.2012.07.031

Anderson, T., & Dragićević, S. (2015). An agent-based modeling approach to represent

infestation dynamics of the emerald ash borer beetle. Ecological Informatics, 30,
97–109. https://doi.org/10.1016/j.ecoinf.2015.09.003

Anderson, T. M., & Dragićević, S. (2018). Network-agent based model for simulating the
dynamic spatial network structure of complex ecological systems. Ecological
Modelling, 389(October), 19–32. https://doi.org/10.1016/j.ecolmodel.2018.10.008
Arroyo Ohori, K., Ledoux, H., & Stoter, J. (2015). An evaluation and classification of nD
topological data structures for the representation of objects in a higher-
dimensional GIS. International Journal of Geographical Information Science,
29(5), 825–849. https://doi.org/10.1080/13658816.2014.999683

Autodesk. (2019). Fusion 360. Retrieved from

https://www.autodesk.com/products/fusion-360/overview

12
Baghaei Daemei, A., Khotbehsara, E. M., Nobarani, E. M., & Bahrami, P. (2019). Study
on wind aerodynamic and flow characteristics of triangular-shaped tall buildings
and CFD simulation in order to assess drag coefficient. Ain Shams Engineering
Journal, 10(3), 541–548. https://doi.org/10.1016/j.asej.2018.08.008

Batty, M., & Torrens, P. M. (2005). Modelling and prediction in a complex world. Futures,
37(7 SPEC.ISS.), 745–766. https://doi.org/10.1016/j.futures.2004.11.003

Batty, M., Xie, Y., & Sun, Z. (1999). Modeling urban dynamics through GIS-based
cellular automata. Computers, Environment and Urban Systems, 23(3), 205–233.
https://doi.org/10.1016/S0198-9715(99)00015-0

Bone, C., Dragicevic, S., & Roberts, A. (2006). A fuzzy-constrained cellular automata
model of forest insect infestations. Ecological Modelling, 192(1–2), 107–125.
https://doi.org/10.1016/j.ecolmodel.2005.09.013

Bova, M. J., Ciarallo, F. W., & Hill, R. R. (2016). Development of an agent-based model
for the secondary threat resulting from a ballistic impact event. Journal of
Simulation, 10(1), 24–35. https://doi.org/10.1057/jos.2015.1

Brown, D. G., Riolo, R., Robinson, D. T., North, M., & Rand, W. (2005). Spatial process
and data models: Toward integration of agent-based models and GIS. Journal of
Geographical Systems, 7(1), 25–47. https://doi.org/10.1007/s10109-005-0148-5

Burger, A., Kennedy, W. G., & Crooks, A. (2021). Organizing Theories for Disasters into
a Complex Adaptive System Framework. Urban Science, 5(3), 61.
https://doi.org/10.3390/urbansci5030061

Castella, J.-C., Trung, T. N., & Boissau, S. (2005). Participatory simulation of land-use
changes in the northern mountains of vietnam: the combined use of an agent-
based model, a role-playing game, and a geographic information system.
Ecology and Society, 10(1), 27.

Che Osmi, S. K., Rosdi, A., Husen, H., Sojipto, S., Misnon, N. A., & Khairuddin, F. H.
(2020). Numerical simulation and full-scale field testing of steel parapet railing
under impact of heavy truck loads - Comparative study. IOP Conference Series:
Earth and Environmental Science, 476(1). https://doi.org/10.1088/1755-
1315/476/1/012056

Chen, K., Chan, K. S., Bailey, K. M., Aydin, K., & Ciannelli, L. (2012). A probabilistic
cellular automata approach for predator-prey interactions of arrowtooth flounder
(Atheresthes stomias) and walleye pollock (Theragra chalcogramma) in the
eastern bering sea. Canadian Journal of Fisheries and Aquatic Sciences, 69(2),
259–272. https://doi.org/10.1139/F2011-160

Cohen, J. (1960). A Coefficient of Agreement for Nominal Scales. Educational and

Psychological Measurement, 20(1), 37–46.

Cova, T. J., & Goodchild, M. F. (2002). Extending geographical representation to include

fields of spatial objects. International Journal of Geographical Information
Science, 16(6), 509–532. https://doi.org/10.1080/13658810210137040

13
Crooks, A., Castle, C., & Batty, M. (2008). Key challenges in agent-based modelling for
geo-spatial simulation. Computers, Environment and Urban Systems, 32(6),
417–430. https://doi.org/10.1016/j.compenvurbsys.2008.09.004

Dragicevic, S., & Marceau, D. J. (2000). A fuzzy set approach for modelling time in GIS.
International Journal of Geographical Information Science, 14(3), 225–245.
https://doi.org/10.1080/136588100240822

Ephtracy. (2019). MagicaVoxel. Retrieved from https://ephtracy.github.io/

ESRI. (2018). ArcGIS Desktop: Release 10.5. Redlands, CA, CA: Environmental System
Research Institute. Retrieved from https://www.esri.com/en-
us/arcgis/products/arcgis-desktop/overview

Fritsch, D. (1996). Three-dimensional geographic information systems-status and

prospects. International Archives of Photogrammetry and Remote Sensins, 31,
215–221.
Fritz, H. C., Wittfeld, K., Schmidt, C. O., Domin, M., Grabe, H. J., Hegenscheid, K., …
Lotze, M. (2014). Current smoking and reduced gray matter volume - A voxel-
based morphometry study. Neuropsychopharmacology, 39(11), 2594–2600.
https://doi.org/10.1038/npp.2014.112

Genois, M., Du Pont, S. C., Hersen, P., & Gregoire, G. (2013). An agent-based model of
dune interactions produces the emergence of patterns in deserts. Geophysical
Research Letters, 40(15), 3909–3914. https://doi.org/10.1002/grl.50757

Goodchild, M. F. (1992). Geographical data modeling. Computers and Geosciences,

18(4), 401–408. https://doi.org/10.1016/0098-3004(92)90069-4
Goodchild, M. F. (2009). Geographic information systems and science: today and
tomorrow. Procedia Earth and Planetary Science, 1(1), 1037–1043.
https://doi.org/10.1016/j.proeps.2009.09.160
Goodchild, M. F., Yuan, M., & Cova, T. J. (2007). Towards a general theory of
geographic representation in GIS. International Journal of Geographical
Information Science, 21(3), 239–260.
https://doi.org/10.1080/13658810600965271
Greene, N. (1989). Voxel space automata: Modeling with stochastic growth processes in
voxel space. Computer Graphics, 23(3), 175–184.
https://doi.org/10.1145/74334.74351

Hagen-Zanker, A. (2009). An improved Fuzzy Kappa statistic that accounts for spatial
autocorrelation. International Journal of Geographical Information Science, 23(1),
61–73. https://doi.org/10.1080/13658810802570317

Hagen, A. (2003). Fuzzy set approach to assessing similarity of categorical maps.

International Journal of Geographical Information Science, 17(3), 235–249.
https://doi.org/10.1080/13658810210157822

14
Heppenstall, A., Malleson, N., & Crooks, A. (2016). “Space, the Final Frontier”: How
good are agent-based models at simulating individuals and space in cities?
Systems, 4(1), 9. https://doi.org/10.3390/systems4010009

Heung, B., Bakker, L., Schmidt, M. G., & Dragićević, S. (2013). Modelling the dynamics
of soil redistribution induced by sheet erosion using the Universal Soil Loss
Equation and cellular automata. Geoderma, 202–203, 112–125.
https://doi.org/10.1016/j.geoderma.2013.03.019

Hosoi, F., Nakai, Y., & Omasa, K. (2013). 3-D voxel-based solid modeling of a broad-
leaved tree for accurate volume estimation using portable scanning lidar. ISPRS
Journal of Photogrammetry and Remote Sensing, 82, 41–48.
https://doi.org/10.1016/j.isprsjprs.2013.04.011

Jjumba, A., & Dragićević, S. (2015). Integrating GIS-Based Geo-Atom Theory and Voxel
Automata to Simulate the Dispersal of Airborne Pollutants. Transactions in GIS,
19(4), 582–603. https://doi.org/10.1111/tgis.12113
Jjumba, A., & Dragićević, S. (2016a). A development of spatiotemporal queries to
analyze the simulation outcomes from a voxel automata model. Earth Science
Informatics, 9(3), 343–353. https://doi.org/10.1007/s12145-016-0260-8

Jjumba, A., & Dragićević, S. (2016b). Spatial indices for measuring three-dimensional
patterns in a voxel-based space. Journal of Geographical Systems, 18(3), 183–
204. https://doi.org/10.1007/s10109-016-0231-0
Jjumba, A., & Dragićević, S. (2016c). Towards a voxel-based geographic automata for
the simulation of geospatial processes. ISPRS Journal of Photogrammetry and
Remote Sensing, 117, 206–216. https://doi.org/10.1016/j.isprsjprs.2016.01.017

Jung, Y., & Joo, M. (2011). Building information modelling ( BIM ) framework for practical
implementation. Automation in Construction, 20(2), 126–133.
https://doi.org/10.1016/j.autcon.2010.09.010

Kang, J. Y., Michels, A., Crooks, A., Aldstadt, J., & Wang, S. (2022). An integrated
framework of global sensitivity analysis and calibration for spatially explicit agent-
based models. Transactions in GIS, 26(1), 100–128.
https://doi.org/10.1111/tgis.12837

Karssenberg, D., & De Jong, K. (2005). Dynamic environmental modelling in GIS: 1.

Modelling in three spatial dimensions. International Journal of Geographical
Information Science, 19(5), 559–579.
https://doi.org/10.1080/13658810500032362

Langran, G., & Chrisman, N. R. (1988). A framework for temporal geographic

information. Cartographica, 25(3), 1–14. https://doi.org/10.3138/K877-7273-
2238-5Q6V

Lenk, U., & Heipke, C. (2006). The radial topology algorithm - A new approach for
deriving 2.5D GIS data models. GeoInformatica, 10(4), 447–468.
https://doi.org/10.1007/s10707-006-0342-8

15
Li, G., Citrin, D., Camphausen, K., Mueller, B., Burman, C., Mychalczak, B., … Song, Y.
(2008). Advances in 4D medical imaging and 4D radiation therapy. Technology in
Cancer Research and Treatment, 7(1), 67–81.
https://doi.org/10.1177/153303460800700109

Li, P. (2013). Tensor-SIFT based earth mover’s distance for contour tracking. Journal of
Mathematical Imaging and Vision, 46(1), 44–65. https://doi.org/10.1007/s10851-
012-0369-4

Ligmann-Zielinska, A., & Sun, L. (2010). Applying time-dependent variance-based global

sensitivity analysis to represent the dynamics of an agent-based model of land
use change. International Journal of Geographical Information Science, 24(12),
1829–1850. https://doi.org/10.1080/13658816.2010.490533

Ligmann-Zielinska, A. (2013). Spatially-explicit sensitivity analysis of an agent-based

model of land use change. International Journal of Geographical Information
Science, 27(9), 1764–1781. https://doi.org/10.1080/13658816.2013.782613
Livny, Y., Yan, F., Olson, M., Chen, B., Zhang, H., & El-Sana, J. (2010). Automatic
reconstruction of tree skeletal structures from point clouds. ACM Transactions on
Graphics, 29(6), 1. https://doi.org/10.1145/1882261.1866177

Luus, K. a., Robinson, D. T., & Deadman, P. J. (2011). Representing ecological

processes in agent-based models of land use and cover change. Journal of Land
Use Science, 8(2), 175–198. https://doi.org/10.1080/1747423X.2011.640357

Manson, S. M. (2001). Simplifying complexity: a review of complexity theory. Geoforum,

32, 405–414. https://doi.org/10.1016/S0016-7185(00)00035-X

Manson, S. M. (2007). Challenges in evaluating models of geographic complexity.

Environment and Planning B: Planning and Design, 34(2), 245–260.
https://doi.org/10.1068/b31179
Maturana, D., & Scherer, S. (2015). VoxNet: A 3D Convolutional Neural Network for
real-time object recognition. IEEE International Conference on Intelligent Robots
and Systems, 2015–Decem, 922–928.
https://doi.org/10.1109/IROS.2015.7353481

Messina, J. P., Evans, T. P., Manson, S. M., Shortridge, A. M., Deadman, P. J., &
Verburg, P. H. (2008). Complex systems models and the management of error
and uncertainty. Journal of Land Use Science, 3(1), 11–25.
https://doi.org/10.1080/17474230802047989

Mialhe, F., Becu, N., & Gunnell, Y. (2012). An agent-based model for analyzing land use
dynamics in response to farmer behaviour and environmental change in the
Pampanga delta (Philippines). Agriculture, Ecosystems and Environment, 161,
55–69. https://doi.org/10.1016/j.agee.2012.07.016

Mitsova, D., Shuster, W., & Wang, X. (2011). A cellular automata model of land cover
change to integrate urban growth with open space conservation. Landscape and
Urban Planning, 99(2), 141–153.
https://doi.org/10.1016/j.landurbplan.2010.10.001

16
Murray-Rust, D., Brown, C., van Vliet, J., Alam, S. J., Robinson, D. T., Verburg, P. H., &
Rounsevell, M. (2014). Combining agent functional types, capitals and services
to model land use dynamics. Environmental Modelling and Software, 59, 187–
201. https://doi.org/10.1016/j.envsoft.2014.05.019

Nan, L., Sharf, A., Zhang, H., Cohen-Or, D., & Chen, B. (2010). SmartBoxes for
interactive urban reconstruction. ACM Transactions on Graphics, 29(4), 1.
https://doi.org/10.1145/1833351.1778830

Narteau, C., Zhang, D., Rozier, O., & Claudin, P. (2009). Setting the length and time
scales of a cellular automaton dune model from the analysis of superimposed
bed forms. Journal of Geophysical Research: Earth Surface, 114(3), 1–18.
https://doi.org/10.1029/2008JF001127

National Centers for Environmental Prediction, National Weather Service, NOAA, & U.S.
Department of Commerce. (2005). NCEP North American Regional Reanalysis
(NARR), Research Data Archive at the National Center for Atmospheric
Research, Computational and information Systems Labratory. Retrieved March
11, 2017, from http://rda.ucar.edu/datasets/ds608.0/

North, M. J., Collier, N. T., Ozik, J., Tatara, E. R., Macal, C. M., Bragen, M., & Sydelko,
P. (2013). Complex adaptive systems modeling with Repast Simphony. Complex
Adaptive Systems Modeling, 1(1), 3. https://doi.org/10.1186/2194-3206-1-3

North, M. J., Tatara, E., Collier, N. T., & Ozik, J. (2007). Visual Agent-based Model
Development with Repast Simphony. Proceedings of the Agent 2007 Conference
on Complex Interaction and Social Emergence, (1), 1–20. Retrieved from
http://www.osti.gov/energycitations/product.biblio.jsp?osti_id=973007

Nourian, P., Gonçalves, R., Zlatanova, S., Ohori, K. A., & Vu Vo, A. (2016). Voxelization
algorithms for geospatial applications: Computational methods for voxelating
spatial datasets of 3D city models containing 3D surface, curve and point data
models. MethodsX, 3, 69–86. https://doi.org/10.1016/j.mex.2016.01.001

Office for Coastal Management. (2022). 2013 Vessel Density. Retrieved from
https://www.fisheries.noaa.gov/inport/item/48932

Openshaw, S. (1998). Some trends and future perspectives for spatial analysis in gis.
Geographic Information Sciences, 4(1–2), 5–13.
https://doi.org/10.1080/10824009809480497
Paraunak, H. V. D., Savit, R., & Riolo, R. L. (1998). Agent-Based Modeling vs. Equation-
Based Modeling: A Case Study and Users’ Guide. Multi-Agent Systems and
Agent-Based Simulation: First International Workshop, MABS’98.
https://doi.org/10.1007/10692956_2

Parker, D. C., Manson, S. M., Janssen, M. A., Hoffmann, M. J., & Deadman, P. (2003).
Multi-agent systems for the simulation of land-use and land-cover change: A
review. Annals of the Association of American Geographers, 93(2), 314–337.
https://doi.org/10.1111/1467-8306.9302004

17
Parunak, H. V. D., Savit, R., & Riolo, R. L. (1998). Agent-Based Modeling vs. Equation-
Based Modeling: A Case Study and Users’ Guide. In Multi-Agent Systems and
Agent-Based Simulation: First International Workshop, MABS’981 (pp. 10–25).
Lecture Notes in Computer Science.

Perez, L., & Dragicevic, S. (2009). An agent-based approach for modeling dynamics of
contagious disease spread. International Journal of Health Geographics, 8(1), 1–
17. https://doi.org/10.1186/1476-072X-8-50

Pérez, L., Dragićević, S., & White, R. (2013). Model testing and assessment:
Perspectives from a swarm intelligence, agent-based model of forest insect
infestations. Computers, Environment and Urban Systems, 39, 121–135.
https://doi.org/10.1016/j.compenvurbsys.2012.10.004
Poomani, M., Sutha, J., & Soundar, K. R. (2021). Wiener filter based deep convolutional
network approach for classification of satellite images. Journal of Ambient
Intelligence and Humanized Computing, 12(7), 7343–7351.
https://doi.org/10.1007/s12652-020-02410-3

Potts, J. R., Auger-Méthé, M., Mokross, K., & Lewis, M. A. (2014). A generalized
residual technique for analysing complex movement models using earth mover’s
distance. Methods in Ecology and Evolution, 5(10), 1012–1022.
https://doi.org/10.1111/2041-210X.12253

Rahmandad, H., & Sterman, J. (2008). Heterogeneity and Network Structure in the
Dynamics of Diffusion : Comparing Agent-Based and Differential Equation
Models, 54(5), 998–1014.

Rounsevell, M. D. a, Robinson, D. T., & Murray-Rust, D. (2012). From actors to agents

in socio-ecological systems models. Philosophical Transactions of the Royal
Society of London. Series B, Biological Sciences, 367(1586), 259–69.
https://doi.org/10.1098/rstb.2011.0187

Rubner, Y., Tomasi, C., & Guibas, L. J. (2000). Earth mover’s distance as a metric for
image retrieval. International Journal of Computer Vision, 40(2), 99–121.
https://doi.org/10.1023/A:1026543900054

Russo, S., Müller, J., Alderman, J., Paletta, N., Adden, S., & Ruiz-Calavera, L. P. (2021).
A CFD study on the strut interference on a regional aircraft wind-tunnel model.
IOP Conference Series: Materials Science and Engineering, 1024(1), 0–8.
https://doi.org/10.1088/1757-899X/1024/1/012047

Rykiel, E. J. (1996). Testing ecological models: the meaning of validation. Ecological

Modelling, 90, 229–244.
Segre, E., & Deangeli, C. (1995). Cellular automaton for realistic modelling of landslides.
Nonlinear Processes in Geophysics, 2(1), 1–15. https://doi.org/10.5194/npg-2-1-
1995

Seiferling, I., Naik, N., Ratti, C., & Proulx, R. (2017). Green streets − Quantifying and
mapping urban trees with street-level imagery and computer vision. Landscape

18
 and Urban Planning, 165(May), 93–101.
https://doi.org/10.1016/j.landurbplan.2017.05.010

Shi, W., Yang, B., & Li, Q. (2003). An object-oriented data model for complex objects in
three-dimensional geographical information systems. International Journal of
Geographical Information Science, 17(5), 411–430.
https://doi.org/10.1080/1365881031000086974

Srivastava, P., Shukla, A., & Bansal, A. (2021). A comprehensive review on soil
classification using deep learning and computer vision techniques. Multimedia
Tools and Applications, 80(10), 14887–14914. https://doi.org/10.1007/s11042-
021-10544-5

The Python Software Foundation. (2020). Python. Retrieved September 29, 2020, from
python.org

Tobler, W. R. (1970). A Computer Movie Simulating Urban Growth in the Detroit Region.
Economic Geography, 46, 234–240.
Torrens, P. M., & Benenson, I. (2005). Geographic Automata Systems. International
Journal of Geographical Information Science, 19(4), 385–412.
https://doi.org/10.1080/13658810512331325139

U.S. Geological Survey. (2012). Global Digital Elevation Model (GTOPO30). U.S.
Geological Survey (USGS) Earth Resources Observation and Science (EROS)
Center.

van Vliet, J., Bregt, A. K., Brown, D. G., van Delden, H., Heckbert, S., & Verburg, P. H.
(2016). A review of current calibration and validation practices in land-change
modeling. Environmental Modelling and Software, 82, 174–182.
https://doi.org/10.1016/j.envsoft.2016.04.017
Western Washington University. (2022). Salish Sea Atlas. Retrieved March 22, 2022,
from https://wp.wwu.edu/salishseaatlas/

White, R., & Engelen, G. (2000). High-resolution Integrated Modelling of the Spatial
Dynamics of Urban and Regional Systems. Computers, Environment and Urban
Systems, 24(5), 383–400. https://doi.org/10.1016/S0198-9715(00)00012-0

White, R., Engelen, G., & Uljee, I. (1997). The use of constrained cellular automata for
high-resolution modelling of urban land-use dynamics. Environment and Planning
B: Planning and Design, 24(3), 323–343. https://doi.org/10.1068/b240323

Wright, A., Scadeng, M., Stec, D., Dubowitz, R., Ridgway, S., & Leger, J. S. (2017).
Neuroanatomy of the killer whale (Orcinus orca): a magnetic resonance imaging
investigation of structure with insights on function and evolution. Brain Structure
and Function, 222(1), 417–436. https://doi.org/10.1007/s00429-016-1225-x

Zlatanova, S., Rahman, A., & Pilouk, M. (2002). Trends in 3D GIS development. Journal
of Geospatial Engineering, 4(2), 71–80.

19
Chapter 2.

A Four-Dimensional Agent-Based Model: A Case

Study of Forest-Fire Smoke Propagation1

2.1. Introduction

Many geographical and environmental systems are constantly evolving in three-

dimensional space and time and thus by nature they are four-dimensional (4D).
Examples of such systems include changes of geological structures, dune formation,
landslides, pollutant propagation through the atmosphere or in water environments,
nutrient cycles in the soil, forest fires, change of high-rise buildings in the cites or
evacuation in urban settings. However, these phenomena are often analyzed and
represented with modeling approaches that consider only two spatial dimensions (2D)
and time. As examples, some spatio-temporal models that can represent the dynamics
of change over 2D and time are those representing land use change (Luus et al., 2011),
human movement in a city (Crooks et al., 2008; Zhu and Timmermans, 2011), soil
erosion (Heung et al., 2013), or wildfire spread (Clarke et al., 1994). This simplification is
largely determined by the nature of the data collection and availability that are primarily
related to two spatial dimensions such as satellite imagery, soil samples, land use or
land cover states, all captured at different temporal snapshots.

Furthermore, there are many different methods of spatio-temporal

modeling including statistical and numerical models that can also be grouped into
the two general categories of equation models (Rahmandad and Sterman, 2008)
and process models (Alam and Dutta, 2012). Equation models are based on
mathematical equations that are developed to represent the effects of processes
on the system (Parunak et al., 1998). With this method, the inner mechanisms of
the system are not well understood and only the output patterns are analyzed
(Murray-Rust et al., 2014). In addition, these are top-down modeling approaches

1
A version of this chapter has been published, coauthored with S. Dragićević, under the title “A
Four-Dimensional Agent-Based Model: A Case Study of Forest-Fire Smoke Propagation” in the
journal Transactions in GIS, 2019, 23, 417-434

20
that represent aggregate behavior (Rounsevell et al., 2012). Process models
represent micro-level interactions of system elements to determine the macro-
level behavior (Luus et al., 2011) making them a bottom-up approach
(Heppenstall et al., 2016). The processes within a system are represented and
are directly studied to determine how they create spatial patterns that change
over space and time (Parunak et al., 1998). In this context, the observed patterns
of the system emerge from the components of the system allowing for
unexpected relationships and interactions among system parts and the
environment to form the change (Heppenstall et al., 2016). For example, an
equation model of land use change may use regression analysis to determine the
impacts of specific variables such as economic growth, then the equations can
be applied to various regions to determine how each will affect the land use
change over time (He et al., 2013). On the other hand, the process model will
determine how the decisions and interactions of different elements of the system
change the land under expected conditions and then forecast several possible
scenarios of change process for the region (Sun et al., 2014).

Given the bottom-up local interactions and behavior of elements of the

systems, most of the geospatial dynamic systems can be viewed as complex
systems (CS). They behave in a nonlinear manner, and are characterized by
evolution, emergence, self-organization, adaptation, feedback loops, and
bifurcation, which make them difficult to represent and model (Manson, 2001).
The most common mathematical approaches for CS modeling of geospatial
dynamic phenomena are Geographic Automata Systems (GAS) (Torrens and
Benenson, 2005) particularly cellular automata (CA), voxel automata (VA), and
agent-based modeling (ABM). These are bottom-up methods that account of the
interactions at local or individual level thereby allowing for the evolution of the
spatial pattern of the system at larger spatial scale.

Geographic space is typically represented in a CA by a regular grid of

square cells also called rasters and can be extended to 3D cubes called voxels
(Greene, 1989), while irregular shapes or volumetric elements can also be

21
considered (Jjumba and Dragićević, 2015). The process of change is
represented by a set of transition rules applied to each cell or voxel at a discrete
time increment and within a defined neighborhood. When representing individual
behavior and the movement and interactions between elements of the system or
the environment, ABM approaches are more appropriate (Huang et al., 2014). An
ABM operates using a series of mobile agents that are software routines moving
and interacting with the landscape using their own rulesets (Torrens and
Benenson, 2005), allowing for higher degrees of heterogeneity than CA or VA
models.

While there are many uses of ABMs, one of the major geospatial
applications is the modeling of land use change where agents represent various
stakeholders impacting the process such as households or homeowners
(Castella et al., 2005; Krebs, 2017; Rounsevell et al., 2012), and land developers
(Dragićević and Hatch, 2018). These agents make decisions to change the land
to meet their demand for urban resources such as employment, and natural
resources such as farmland (Castella et al., 2005). Ecological processes are also
often modeled using an ABM such as the spread of invasive insects (Anderson
and Dragićević, 2015; Pérez and Dragićević, 2011), killer whales (Testa et al.,
2012), animal movements (Tang and Bennett, 2010), and human disease
propagation to name a few (Crooks and Hailegiorgis, 2014; Perez and
Dragicevic, 2009).

ABM have been used to represent physical systems where agents

represent objects or particles such as sand dunes (Genois et al., 2013), overland
flow of water (Reaney, 2008), chemical reaction and diffusion (Agusdinata et al.,
2015), or even ballistic fragment impacts (Bova et al., 2016). The agent decision-
making is based on concepts from physics and empirical observations to
calculate the correct action (Azimi et al., 2011; Reaney, 2008) and affects the
system state by changing attributes such as particle location, volume or size
(Genois et al., 2013).

22
 However, many of these geospatial processes are inherently 4-
dimensional thus entailing a need for advancement of the GAS models in 4D for
more realistic representation of geographical and environmental phenomena.
Recently, a VA approach (Jjumba and Dragićević, 2015) was developed to model
phenomena such as pollution propagation in 4D. The use of 4D ABM has been
developed to a lesser extent and focused on systems such as building
evacuations (Macatulad and Blanco, 2014) or urban earthquake evacuation
(Torrens, 2014). Consequently, these models are not modeling natural CS from a
4D perspective which demands an advancement of the development of 4D ABM
into natural systems that are not necessarily based on human decision-making
processes. Consequently, the main objectives of this study are to explore the
further advancement of GAS models in 4D by: 1) proposing a theoretical
modeling framework for 4D ABM, and 2) implementing the developed approach
to the 4D case study for forest-fire smoke propagation and using two scenarios.
The study area used in this research is Central and Southern British Columbia
(BC) and the western parts of Alberta, Canada for the forest fires that occurred in
the summer season 2017.

Smoke propagation was selected as the focus of the case study because
of its inherent four-dimensionality and also due to the need to mitigate adverse
impacts on human health and the environment. Human exposure to smoke has
led to increased demand on medical services for respiratory problems
(Dennekamp and Abramson, 2011). Advanced warning of smoke events through
simulation of smoke propagation could help reduce exposure to smoke as
recommended for example by the BC Centre for Disease Control (Elliott, 2014).
Two models of smoke propagation are available, WRF-Chem (Grell et al., 2005),
based on the Eulerian perspective of motion, and HYSPLIT (Rolph et al., 2009)
based on the Lagrangian perspective of motion, both for space-time smoke
propagation. The Lagrangian motion improves the representation of movement
over the Eulerian motion perspective by directly modeling the movement of
smoke particles instead of the movement of smoke concentrations past static
points. The representation of smoke concentration as a continuous gradient in

23
WRF-Chem is better than the parcel representation of smoke used in HYSPLIT
model. However, the ABM of smoke propagation can comprise the advantages of
both existing models by including moving agents with Lagrangian perspective of
motion and by calculating a gradient of smoke concentration. An existing ABM of
smoke propagation has been proposed (Smith and Dragicevic 2018) however it
is operating only in 2D space and time with a static fire emitting the smoke.
Therefore, in this study, the proposed 4D ABM builds upon limitations of the
exiting models by representing the dynamics of smoke propagation as a 4D
process, considering moving forest fire and smoke across the 3D landscape.

2.2. Theoretical Framework of a 4D Agent Based Model

An ABM is a GAS model consisting of mobile agents that transmit information as

they move and interact (Torrens and Benenson, 2005) and can be formulated as:

𝐴𝐵𝑀 ~ 𝑆 , 𝐼 , 𝑅 1

where ABM at time t represents the automaton composed of agents

states’ S and transition rules R used by the agents to change the state of the
model using input data I. In a 3D ABM, the states S can be defined by:

𝑆 :𝑆 , 2

where (x, y) represents the horizontal position of the agent 2D state S2D. In
a 4D ABM, the state must include the third spatial dimension z and is
represented by:

𝑆 :𝑆 , , 3

where (x, y, z) represents the horizontal and vertical positions the 3D

agent with state S3D. Similarly, the input data I are also 3D and represented by:

𝐼 :𝐼 , , 4

24
 where (x, y, z) represents the horizontal and vertical positions of the data
available to the agents. The 3D representation of both S3D and I3D can be seen
on the left side of Figure 2.1 for time t. While not all input data need to be 3D, 2D
data can be given z coordinates or can be stretched over the entire third
dimension. Together with the transition rules R, this results in a 4D ABM being
defined at the current time t by:

𝐴𝐵𝑀 ~ 𝑆 , 𝐼 , 𝑅 5

where the four dimensions are located in the three spatial dimensions of
the 3D state S3D, and the one temporal dimension of time t. This is represented
twice in Figure 2.1, once at time t, and once at time t+1. As time t progresses, the
4D ABM moves from the current time t to the next time step t+1:

𝐴𝐵𝑀 ~ 𝑆 𝐼 𝑅 → 𝐴𝐵𝑀 ~ 𝑆 𝐼 𝑅 → ⋯ → 𝐴𝐵𝑀 ~ 𝑆 𝐼 𝑅

6

where the agents use their current state S3Dt, current input data I3Dt, and
the current transition rules Rt to determine their state in the next time step S3Dt+1.
At time t+1, the agents’ have access to the data I3Dt+1 for use with transition rules
Rt+1 that may or may not be different than the previous transition rules Rt. This
progression of time in the model is displayed in Figure 2.1 where the model
evolves from initial time t through the next time t+1 and generally for times t+n.

Figure 2.1 Schematic representation of 4D ABM with the agents’ states S, input
data I, and transitions rules R from initial time t and constitutive
iterations at t+1 to t+n.

25
2.3. Smoke Propagation Model

2.3.1. Study Area and Data

The study area for this case study is British Columbia (BC) and the western half
of Alberta, Canada including the metropolitan regions of Edmonton and Calgary. This
area experiences an annual forest-fire season that fills the skies with smoke that
disperses to major urban centers such as Vancouver, Calgary, Edmonton, Canada and
Seattle, USA. Forest fires in BC burned a record-breaking area of 12,000 km2 in 2017
(BC Wildfire Service, 2018a), and the record was broken again in 2018 with more than
13,000 km2 burned (BC Wildfire Service, 2018b). The location of the hypothetical fire
used in the case study to initialize the geosimulation is on the west coast of BC, near
Khartoum Lake, approximately 100km northwest of Vancouver (49.86°N, -124.19°E).
This location was chosen because of the westward wind patterns during the selected
dates between August 10th and 25th, 2017. On the modeled dates, real fires were
burning across BC with a majority of the fires further inland, near Kamloops.

Two datasets are used for the model input: a digital elevation model (DEM) (U.S.
Geological Survey, 2012) and an atmospheric conditions dataset (National Centers for
Environmental Prediction et al., 2005). The DEM represents the landscape of the study
area and is not updated during model processing. It has a horizontal and vertical
resolution of 1km and 1m, respectively. The atmospheric conditions data have wind
speeds and directions at a horizontal resolution of 32km, and a variable vertical
resolution based of 29 pressure levels with pressure differences of 25 or 50 pascals.
Because elevation in the atmospheric conditions data is based on pressure levels, there
is no fixed distance for vertical resolution. In some areas these pressure levels can be
closer together while in others they may be further apart.

2.3.2. Model Overview

The 4D ABM represents smoke as a 4D process by modelling in the three spatial

dimensions (x, y, z) and the temporal dimension (t). By including three spatial
dimensions, the model has the ability to propagate smoke vertically through the
atmosphere and over terrain where different wind directions and velocities may be
present.

26
 Using the theoretical formulation of the 4D ABM, the model of smoke propagation
in the 4D ABMsmoke can be represented as follows:

𝐴𝐵𝑀 ~ 𝐴; 𝑆 , 𝐼 , 𝑅 ; 𝑆 , 𝐼 , 𝑅 7

where the types of agents A are represented in two agent sets. The first set
represents the Fire Agents F with their states SF, input data IF, and transition rules RF.
Smoke Agents K also have their own states SK, input data IK, and transition rules RK.

The flowchart presented in Figure 2.2 shows the processes the two agent types
and model control use to model the smoke propagation. Fire Agents (Figure 2.2)
represent the location of the fires being modeled. The Fire Agents release smoke by
creating a specified number of Smoke Agents every iteration. This number is determined
through calibration for the available processing power and relative size of fires. Two
scenarios were developed for Fire Agents, Scenario 1 representing the baseline model
with static fire and Scenario 2 representing moving Fire Agent starting at the same
location as the first one. While full modeling of the forest fires is a large and complicated
undertaking, some basic movement of Fire Agents has been included. The Scenario 2
adopts that the Fire Agents move at 2% of the speed of the ground-level wind that
results in an average speed of 3.8m/min, within the range of observed forest fire speeds
(Perrakis et al., 2014).

Smoke Agents (Figure 2.2) are based on two methods of movement through the
atmosphere, the processes of passive diffusion and active dispersion by wind. The
movement from these two methods are calculated into separate vectors which are then
added together to form a movement vector. This reduces the number of times the agents
move, increasing the efficiency of the model. The diffusion process occurs as the Smoke
Agents move from areas of high concentration of Smoke Agents to areas of low
concentrations to realistically represent smoke behavior in the environment. In order to
model the diffusion process, the Smoke Agents use a 3D cube to find the neighboring
voxel with the most agents and move directly away from it. The agents randomize the
order in which they search for the highest concentration to avoid a bias in direction when
two voxels have an equal amount of agents, forming the first of three sources of
heterogeneity in the model. When the agent is in the most populated voxel, it will move
in a random direction, representing the second source of heterogeneity.

27
Figure 2.2 Flow chart of 4D ABM of smoke propagation process with
information on each agent type.

28
 Movement by the wind process is operationalized by the Smoke Agents reading
the horizontal wind speed, horizontal direction, and vertical movement from the
atmospheric data at their locations. The wind movement vector is determined by the
distance the wind would blow in the specified horizontal direction and repeating the
process for the vertical component. When both movements are done the agent’s new
location is checked to ensure it is above ground, and the agent is moved to the surface if
it is below ground. The order in which the agents are processed is determined randomly
by the modeling environment and represents the third source of heterogeneity in the
model.

The Control Module (Figure 2.2) dictates data update and model outputs.
Atmospheric data are updated periodically depending on the frequency of the available
data. For example, if new atmospheric data are available every 3 hours, the model
performs an update that corresponds to 3 hours of real time. Output from the model at
regular intervals is done when the Smoke Agents and Fire Agents output their locations
to external files containing all the agent locations for the current output. These files can
be used to view agent locations or for post-processing in a suitable geographic
information system software.

2.3.3. Model Components

The variation of the Overview, Design, and Details (ODD) protocol (Grimm et al.,
2010, 2006) adapted for this study was used to describe the 4D ABM of smoke
propagation to improve communication of the model development process. The ODD
protocol has been modified for this study to better represent the different components of
the developed 4D ABM by omitting unused sections and adding a section for output
(Table 2.1).

29
Table 2.1 The ODD protocol of the 4D ABM

Model Overview
Purpose The purpose of the model is to represent smoke propagation in 4D by
providing an environment in which to run customized scenarios. The model
can generate various scenarios to forecast possible 3D locations of smoke
from existing fires and to estimate the risks from fire emissions. The model
also aims to demonstrate the use of ABM in the representation of a physical
CS to mitigate difficulties of process representation with often non-spatial
equation models.

Entities, state variables, The model contains two agent types:

and scales Fire Agents: These agents represent the source of the smoke, often forest
fires. Fire Agents have two variables. One variable determines the number of
Smoke Agents released every model iteration, and for this study set to five
agents per one-minute iteration. The other variable determines how fast the
Fire Agent moves compared to the wind and it is set to 0% for Scenario 1 to
represent no influence by wind, and 2% of wind speed for Scenario 2 to
represent moving location for fire.

Smoke Agents: These agents represent the smoke released by the Fire
Agents and are uniform in characteristics and transition rules. Smoke Agents
move through the atmosphere by diffusion and wind processes.

Diffusion process: Smoke Agents search a 3D neighborhood for the voxel with
the most Smoke Agents. The agent will then move directly away from this
voxel or in a random direction if it is detected in the central voxel. Diffusion
has two variables, cubic neighborhood size, and movement distance.

Wind process: Smoke Agents move with the wind by reading the atmospheric
conditions for their location and moving in the same direction as the wind by
the distance the wind moves in the time of one iteration. There are no agent
variables in this process, simply the agent location and input data.

Ground Check: If the Smoke Agent is below ground level as reported by the
digital elevation model, the agent will move to the surface.

The model also contains a Control Module with three components:

 Main Control: This part of the module sets the model parameters
such as scale, extent, temporal resolution, and the model time and
date, and defines the Fire Agents’ starting locations. This module
also manages the location of the input data for Data Control.

 Data Control: This part of the module reads input data into the model
for use by the agents and converts data into alternate formats if
required.

 Time Control: This part of the module dictates when it is time to

change to the next atmospheric dataset and when to write agent
outputs.

30
 The user is able to modify several key variables that effect agents (current
values in parentheses): voxel size or spatial resolution (100 m3), diffusion
distance inner shell (0.36 voxels) outer shell (0.18 voxels), neighborhood size
(5x5x5 voxels), output frequency (180 minutes), data update frequency (180
minutes), iteration length or temporal resolution (1 minute), model duration (15
days), as well as many other variables found in the “Details” section of this
table.

Process overview and The model begins with a setup phase where input data are loaded into the
scheduling appropriate location for access. Then the model creates the desired Fire
Agents.

Each time step or iteration follows the same series of steps:

1. Fire Agents emit the desired Smoke Agents
2. Fire Agents move considering the wind
3. Smoke Agents move using diffusion and wind
4. Smoke Agents checks if output is required. Writes location if required
5. Control Module checks if a data update is required. Updates if
required

Design Concepts
Emergence The emergence of smoke propagation is represented by the aggregated
movement of Smoke Agents. The smoke patterns are mainly affected by the
measured atmospheric conditions.

Adaptation The Smoke Agents adapt to changing atmospheric conditions and nearby
agents through their movement by moving with the current wind and moving
away from other Smoke Agents. The rules Smoke Agents use to determine
movement are static for the entire run of the model.

Interaction Smoke Agents interact through the process of diffusion where agents move
from areas of high smoke concentration to areas of low concentration. To do
this the Smoke Agents search their surrounding area and move directly away
from high concentrations of other Smoke Agents.

Stochasticity Three random processes occur within the model:

 The order in which agents are processed is random and controlled
by the modeling environment.
 During the diffusion process, the Smoke Agents randomize the order
they search the voxels within the neighborhood to avoid a bias in any
one direction caused by searching in the same order every time.
 During the diffusion process, when the Smoke Agent is in the voxel
with the most Smoke Agents, it will move in a random direction.

31
Collectives Collectives are actively avoided through diffusion to prevent the Smoke
Agents from forming clumps. This allows smoke plumes to emerge when wind
patterns and terrain overpower diffusion instead of the Smoke Agents actively
trying to form plumes.

Calibration The number of agents and the diffusion process need to be calibrated for
each unique scenario. The number of agents is dependent primarily on the
computational power available. While more agents will typically produce better
results, they will also increase the processing time required linearly, meaning
that the model has a computational complexity of O(n). The number of Fire
Agents used will also be impacted by the number and size of fires being
represented. Some larger fires may require two or more agents to be
accurately represented.

The diffusion process needs to be calibrated for the movement distance and
the cubic neighborhood size. These variables are adjusted together to match
the results from Kellogg (1956) and is mainly dependent on the voxel size and
number of Smoke Agents.

Outputs Smoke Agents’ locations are collected at regular intervals and saved to
external files where each agent writes its (x, y, z) coordinate to one line of the
file. These files can be opened for viewing and post-processing using GIS
software.
Details
Initialization Modeling Environment: Repast Simphony 2.4
Language: Java

Scenario Settings:
Projection: Lambert Conformal Conic
Origin Lower Left (x,y,z) [m]: -1785776.725475, -376124.151374, 0
Voxel Size (x,y,z) [m]: 100, 100, 100
Extent (x,y,z) [voxels]: 14000, 14000, 200
Iteration length [minutes]: 1
Output Interval [minutes]: 180 (3 hours)
Time Extent [minutes]: 21600 (15 days)
Simulation Start Date: corresponds to August 10th, 2017, 0:00
Simulation End Date: corresponds to August 25th, 2017, 0:00

Number of Fire Agents: 1

Fire Agent location (x,y,z) [voxels]: 5723, 5090, 10
Smoke Agent spawn rate [agents/iteration]: 5
Fire Agent speed Scenario 1: 0%
Fire Agent speed Scenario 2: 2%

Smoke Agent Neighborhood Size (x,y,z) [voxels]: 5, 5, 5

Smoke Agent Diffusion Outer Shell [voxels]: 0.18
Smoke Agent Diffusion Inner Shell [voxels]: 0.36
Smoke Agent Diffusion Centre [voxels]: 0.18

32
 Input Data Atmospheric Data Source: National Center for Atmospheric Research’s North
American Regional Reanalysis, ds608.0
Link: https://rda.ucar.edu/datasets/ds608.0/
Format: GRIB1
Extent: North America, clipped to modeling extent
Horizontal Resolution [km]: 32
Vertical Resolution [pa]: variable between 25 and 50
Vertical Layers: 29
Dates Available: January 1st, 1979 to present, updated monthly
Dates Used: 00:00 August 10th, 2017 to 24:00 August 24th, 2017
Layers Used: Geopotential Height [gpm], u wind [m/s], v wind [m/s], vertical
velocity [Pa/s], Pressure (obtained from layer name)

Digital Elevation Data Source: USGS Global Digital Elevation Model

(GTOPO30)
Resolution – horizontal [km]: 1
Resolution – vertical [m]: 1
Extent: North America, clipped to modeling extent
Format: GeoTIFF

2.4. Results and Discussion

2.4.1. Geosimulation Outputs

The geosimulation model was implemented by development of coding routines in

Repast Simphony using Java (North et al., 2013). The model has a temporal resolution
of one minute represented by one single iteration. At every 180 iterations, one output is
generated that represents 180 minutes of time in reality. In total, 120 outputs are created
representing the 15 days between August 10th, 2017, and August 25th, 2017. Outputs
from the model are in form of “.txt” files but differ between the agent types. Fire Agents
output their location and time to a single file for all iterations. Smoke Agents output their
location to a file specific to the current iteration, to create a series of files where each file
contains the locations of all agents during the one iteration it represents. These files are
processed in ESRI’s ArcGIS suite of software (ESRI, 2018) to create a GIS layer of point
locations for each output. For select simulation outputs, ArcMap was used to create 2D
maps and ArcGlobe was used to create 3D representations.

The simulation results for two-day intervals are presented in Figure 2.3 A and B,
representing the first 12 days of smoke propagation for both the Scenario 1 (Figure
2.3A) and Scenario 2 (Figure 2.3B). The simulations indicate that Smoke Agents stay

33
local to the fire in the Strait of Georgia, BC prior to Day 2 before being blown north into
the interior of BC by Day 4. For the rest of the model, the smoke generally propagates
east into Alberta, reaching as far south as the Metropolitan area of Calgary between Day
4 and 6, and leaving the study area in northern Alberta after Day 8 before returning to
the Metropolitan area of Edmonton by Day 10. Day 12 for both scenarios are compared
in Figure 2.3C with two main differences. The first is the difference in the smoke near the
coast where the fires are located, mostly due to the fires being in different locations. The
second is the location of the next plume in each scenario, where the static fire (Scenario
1), in yellow, has not travelled as far north as the related plume from the moving fire
(Scenario 2) towards the north, in red. In all other locations the simulated pattern of
smoke propagation from Scenario 2 is very similar to the one obtained for Scenario 1.
The path of the moving Fire Agent is shown in Figure 2.3D with a general eastward
movement.

Outputs for Scenario 2, as more realistic with representation of moving forest fire
locations, are used to present the first 12 hours of smoke propagation on Figure 2.4.
These outputs are presented in the form of 3D perspectives with an oblique view on the
left of each pair and the matching vertical view on the right with elevation exaggerated
by a multiple of five. In addition, the output at 24 hours is presented at the top from a low
viewpoint. The results indicate that the Smoke Agents form two main plumes, with the
second forming around the second output (Figure 2.4C) when the wind patterns change.
The second and larger plume gets blown towards the Strait of Georgia, BC, the large
body of water to the west. In terms of height, the simulations indicate that the second
plume is much taller than the first, getting its height from converging and uplifting winds
occurring between 9 and 12 hours (Figure 2.4D-E).

34
Figure 2.3 4D ABM simulation smoke outputs presented in 2D for every 2 days
up to 12 days for (A) Scenario 1 with static fire, and (B) Scenario 2
with moving fire, (C) comparison of Day 12 for both scenarios, and
(D) the path of the Fire Agent for Scenario 2.

35
Figure 2.4 Geosimulation outputs for Scenario 2 with moving Fire Agent
depicting (A) an oblique view of Smoke Agents after 24 hours from a
low view point; oblique views (left of pair) and vertical views (right of
pair) of Smoke Agents forming a plume from the moving fire
location with outputs after (B) 3 hours, (C) 6 hours, (D) 9 hours
and (E) 12 hours from the initialization of the model. Vertical
distances scaled 5x to exaggerate elevation differences.

36
 The summary of the vertical distribution of the Smoke Agents in Scenario 2 is
presented in Figure 2.5. The maximum number of Smoke Agents is reached on the
output for 8100 minutes, or 5 days and 15 hours. The peak in the distribution of smoke
reaches a maximum of around 2600 m on day 7 where it remains close to this height
until day 10 where it begins to drop and rises again towards the end where another peak
of approximately 2600 m is reached on day 15. However, days 1, 2, 5, 8, and 13 show
more smoke under 1000 m than the other days. This has been presented on Figure 2.3
where the simulation outputs for these days contain the most smoke in the Strait of
Georgia, which has an elevation of 0 m. For comparison, the elevation of Vancouver
ranges from ~0 m to ~350 m, Blackcomb Peak, the alpine location of the 2010 Winter
Olympics, is ~2400 m, Kelowna, BC in the Okanagan Valley is ~350 m, Banff, Alberta is
~1400 m, and Calgary, Alberta is ~1000 m. With smoke generally higher than even
some mountain peaks, it is these days with low smoke that have the largest threat to
effecting human health.

37
Figure 2.5 The quantification of simulation distribution of all Smoke Agents in 100 m bins of vertical distance at a daily
time step for Scenario 2.

38
2.4.2. Model Validation

The lack of 3D smoke location data prevents a full model validation. Only two
sources of data were found, satellite imagery and ground reports, but both include
smoke from anthropogenic sources such as factories. Satellite imagery also compresses
all smoke into a 2D layer and has limited capabilities to determine accurate
concentration values. Ground reports, such as weather stations, only report on ground
level smoke, and do not measure smoke higher in the atmosphere.

In order to proceed with 4D ABM validation, resulting simulated spatial patterns

of smoke propagation are visually compared to satellite images from NASA (Figure 2.6),
making several key points of comparison. While similarities between the modeled smoke
propagation and satellite imagery of smoke from different fires do not fully validate the
model, the similarities in the observed patterns and distribution still forms a partial
validation. The validation of patterns can show that the model produces a realistic
representation of the smoke propagation as complex system.

The first two images (Figure 2.6A and 2.6B) are from August 11th and 14th 2017,
corresponding to approximately 1.5 days and 4.5 days from the start of the model at
midnight on the morning of August 10th. These images show similar spatial patterns as
the ones presented on Figure 2.3A and 2.3B where the smoke enters the Interior
Plateau, moves north, and on Day 4 begins to exit the plateau to the northeast, north of
the Cariboo Mountains. This shows that the model follows wind patterns and terrain
properly. The main difference with these images is that the smoke in the imagery fills the
entire plateau instead of a single plume in the model. This can be attributed to the model
containing only a single source of smoke instead of many small fires geographically
distributed.

Figure 2.6C shows the smoke moving to the east through Alberta, and while it is
approximately 20 days after the time period that was modeled, it shows similarities with
the simulated spatial patterns. This shows that the prevailing wind patterns will cause the
smoke to take a similar path. However, because the imagery is from a later date that is
not included in the simulation, the patterns are not an exact match.

39
Figure 2.6 Examples of smoke propagation patterns for British Columbia and
Alberta, Canada, on (A) August 11th, 2017, (B) August 14th, 2017, (C)
July 18th, 2017, and (D) July 6th, 2015 (NASA, 2019).

The final comparison shows the smoke trapped over the Strait of Georgia (Figure
2.6D) and can be seen occurring at smaller extent in the simulations presented in
Figures 2.3A, 2.3B, 2.4E. In the area between the mainland and Vancouver Island, the
tall mountains can trap smoke as both a physical barrier and by influencing wind
patterns. This is significant in the study area due to the urban centers of Vancouver,
Victoria, and Seattle being located in this region. This smoke pattern was visible in many
satellite images including Figures 2.6A and 2.6B, however Figure 2.6D is the most
extreme in terms of amount of smoke trapped in this location. The obtained simulation
outcomes demonstrate trapped smoke over Strait of George, indicating that the

40
proposed 4D ABM model can generate non-imposed patterns emerging from the
represented processes.

In correspondence to the actual simulation dates, it was observed that on August

14th, 2017, at 4:00 AM smoke was reported in Metropolitan area of Calgary, Alberta with
an Air Quality Health Index of 10 as reported by the Government of Canada’s real-time
air quality website (Environment and Natural Resources, 2017). The Metropolitan area of
Calgary is on the far-east side of the study area, and 4:00 AM on August 14th
corresponds to 5940 minutes of model simulation, or in reality 4 days and 3 hours, after
adjusting for time zones. The model shows smoke present in Metropolitan area of
Calgary after 7200 minutes (5 days) of model simulations, or in reality midnight of
August 14th. Due to limited reports, it is difficult to identify how long the smoke event
lasted in Metropolitan area of Calgary. The delay in the simulated smoke reaching
Calgary may be due to the distance of travel after the change in wind patterns. The real
smoke originated from the nearby Rocky Mountains while the simulated smoke traveled
from the coastal area. The overall obtained spatial patterns of smoke propagation
generated by the proposed 4D ABM are similar to the real smoke propagation patterns.

2.5. Conclusions

The 4D ABM developed for the representation of forest-fire smoke propagation is

a proof of concept demonstrating the potential of ABMs to model 4D physical complex
systems. The model produced smoke propagation patterns that correspond to the
location and distribution patterns observed in satellite imagery of forest-fire smoke. This
indicates the 4D ABM is capable of representing the geospatial complex systems of a
physical process where agents characterize matter such as particles of smoke or dirt
with interactions and relatively simple decision-making process. Agents used in a model
representing a physical process only need to use information from their surroundings
with simple equation to influence their decisions and replicate the actual geospatial
process.

There are some limitations in the current model and further research in needed to
address them. The first is the inability to fully validate the 4D ABM with actual smoke
data collected in 3D over time. One solution is to use Lidar ceilometers (van der Kamp et
al., 2008), but their use is cost prohibitive at regional scale. The second limitation is the

41
resolution of the atmospheric conditions data. At the current spatial resolution, entire
valleys and mountains can be contained within one voxel horizontally, and only a few
voxels vertically. The low spatial and temporal resolution of the atmospheric data also
causes the Smoke Agents to move in near linear patterns because they will be blown in
one constant direction until they reach the next atmospheric voxel. The solutions to this
are either obtain higher resolution data, or to create an additional atmospheric model to
artificially increase the resolution. Moreover, there is a lack of methods to validate model
outputs that fully include both 3D space and time. Consequently, the next research stage
includes the development of 4D model validation approaches that can compare
simulated patterns of voxels and inner processes of agents’ behavior in 4D. With the
proposed model enhancement and full validation, there is a potential of using the
proposed model by professionals in need for smoke propagation forecasting such as
public health sector or fire and rescue services.

In summary, the developed 4D ABM demonstrates the strengths of the proposed

geosimulation approach and provides a foundation for further development of models
and analysis methods that can operate in 4D and the ability of representing physical
processes as complex systems.

2.6. References

Agusdinata, D. B., Amouie, M., & Xu, T. (2015). Diffusion dynamics and concentration of
toxic materials from quantum dots-based nanotechnologies: an agent-based
modeling simulation framework. Journal of Nanoparticle Research, 17(1), 26.
https://doi.org/10.1007/s11051-014-2844-x

Alam, M. J., & Dutta, D. (2012). A process-based and distributed model for nutrient
dynamics in river basin: Development, testing and applications. Ecological
Modelling, 247, 112–124. https://doi.org/10.1016/j.ecolmodel.2012.07.031

Anderson, T., & Dragićević, S. (2015). An agent-based modeling approach to represent

infestation dynamics of the emerald ash borer beetle. Ecological Informatics, 30,
97–109. https://doi.org/10.1016/j.ecoinf.2015.09.003

Azimi, M., Jamali, Y., & Mofrad, M. R. K. (2011). Accounting for diffusion in agent based
models of reaction-diffusion systems with application to Cytoskeletal diffusion.
PLoS ONE, 6(9). https://doi.org/10.1371/journal.pone.0025306

BC Wildfire Service. (2018a). Current Statistics. Retrieved September 12, 2018, from
http://bcfireinfo.for.gov.bc.ca/hprScripts/WildfireNews/Statistics.asp

42
BC Wildfire Service. (2018b). Wildfire Season Summary. Retrieved April 30, 2018, from
https://www2.gov.bc.ca/gov/content/safety/wildfire-status/about-bcws/wildfire-
history/wildfire-season-summary

Bova, M. J., Ciarallo, F. W., & Hill, R. R. (2016). Development of an agent-based model
for the secondary threat resulting from a ballistic impact event. Journal of
Simulation, 10(1), 24–35. https://doi.org/10.1057/jos.2015.1

Castella, J.-C., Trung, T. N., & Boissau, S. (2005). Participatory simulation of land-use
changes in the northern mountains of vietnam: the combined use of an agent-
based model, a role-playing game, and a geographic information system.
Ecology and Society, 10(1), 27.

Clarke, K. C., Brass, J. a, & Riggan, P. J. (1994). A Cellular-Automaton Model of Wildfire

Propagation and Extinction. Photogrammetric Engineering and Remote Sensing,
60(11), 1355–1367.

Crooks, A., Castle, C., & Batty, M. (2008). Key challenges in agent-based modelling for
geo-spatial simulation. Computers, Environment and Urban Systems, 32(6),
417–430. https://doi.org/10.1016/j.compenvurbsys.2008.09.004

Crooks, A. T., & Hailegiorgis, A. B. (2014). An agent-based modeling approach applied

to the spread of cholera. Environmental Modelling and Software, 62, 164–177.
https://doi.org/10.1016/j.envsoft.2014.08.027

Dennekamp, M., & Abramson, M. J. (2011). The effects of bushfire smoke on respiratory
health. Respirology, 16(2), 198–209. https://doi.org/10.1111/j.1440-
1843.2010.01868.x

Dragićević, S., & Hatch, K. (2018). Urban geosimulations with the Logic Scoring of
Preference method for agent-based decision-making. Habitat International, 72,
3–17. https://doi.org/10.1016/j.habitatint.2017.09.006

Elliott, C. (2014). Guidance for BC Public Health Decision Makers During Wildfire Smoke
Events. Retrieved from www.bccdc.ca

Environment and Natural Resources. (2017). Calgary - Air Quality Health Index.
Retrieved August 14, 2017, from https://weather.gc.ca/airquality/pages/abaq-
002_e.html

ESRI. (2018). ArcGIS Desktop: Release 10.5. Redlands, CA, CA: Environmental System
Research Institute. Retrieved from https://www.esri.com/en-
us/arcgis/products/arcgis-desktop/overview

Genois, M., Du Pont, S. C., Hersen, P., & Gregoire, G. (2013). An agent-based model of
dune interactions produces the emergence of patterns in deserts. Geophysical
Research Letters, 40(15), 3909–3914. https://doi.org/10.1002/grl.50757

Greene, N. (1989). Voxel space automata: Modeling with stochastic growth processes in
voxel space. Computer Graphics, 23(3), 175–184.
https://doi.org/10.1145/74334.74351

43
Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., … DeAngelis, D.
L. (2006). A standard protocol for describing individual-based and agent-based
models. Ecological Modelling, 198(1–2), 115–126.
https://doi.org/10.1016/j.ecolmodel.2006.04.023

Grimm, V., Berger, U., DeAngelis, D. L., Polhill, J. G., Giske, J., & Railsback, S. F.
(2010). The ODD protocol: A review and first update. Ecological Modelling,
221(23), 2760–2768. https://doi.org/10.1016/j.ecolmodel.2010.08.019

He, C., Huang, Z., & Wang, R. (2013). Land use change and economic growth in urban
China: A structural equation analysis. Urban Studies, 51(13), 2880–2898.
https://doi.org/10.1177/0042098013513649

Heppenstall, A., Malleson, N., & Crooks, A. (2016). “Space, the Final Frontier”: How
good are agent-based models at simulating individuals and space in cities?
Systems, 4(1), 9. https://doi.org/10.3390/systems4010009

Heung, B., Bakker, L., Schmidt, M. G., & Dragićević, S. (2013). Modelling the dynamics
of soil redistribution induced by sheet erosion using the Universal Soil Loss
Equation and cellular automata. Geoderma, 202–203, 112–125.
https://doi.org/10.1016/j.geoderma.2013.03.019

Huang, Q., Parker, D. C., Filatova, T., & Sun, S. (2014). A Review of Urban Residential
Choice Models Using Agent-Based Modeling. Environment and Planning B:
Planning and Design, 41(4), 661–689. https://doi.org/10.1068/b120043p

Jjumba, A., & Dragićević, S. (2015). Integrating GIS-Based Geo-Atom Theory and Voxel
Automata to Simulate the Dispersal of Airborne Pollutants. Transactions in GIS,
19(4), 582–603. https://doi.org/10.1111/tgis.12113

Kellogg, W. W. (1956). Diffusion of Smoke in the Stratosphere. Journal of Meteorology,

13, 241–250.

Krebs, F. (2017). An empirically grounded model of green electricity adoption in

Germany: Calibration, validation and insights into patterns of diffusion. Journal of
Artificial Societies and Social Simulation, 20(2).
https://doi.org/10.18564/jasss.3429

Luus, K. a., Robinson, D. T., & Deadman, P. J. (2011). Representing ecological

processes in agent-based models of land use and cover change. Journal of Land
Use Science, 8(2), 175–198. https://doi.org/10.1080/1747423X.2011.640357

Macatulad, E. G., & Blanco, A. C. (2014). 3DGIS-based multi-agent geosimulation and

visualization of building evacuation using gama platform. International Archives
of the Photogrammetry, Remote Sensing and Spatial Information Sciences -
ISPRS Archives, 40(2), 87–91. https://doi.org/10.5194/isprsarchives-XL-2-87-
2014

Manson, S. M. (2001). Simplifying complexity: a review of complexity theory. Geoforum,

32, 405–414. https://doi.org/10.1016/S0016-7185(00)00035-X

44
Murray-Rust, D., Brown, C., van Vliet, J., Alam, S. J., Robinson, D. T., Verburg, P. H., &
Rounsevell, M. (2014). Combining agent functional types, capitals and services
to model land use dynamics. Environmental Modelling and Software, 59, 187–
201. https://doi.org/10.1016/j.envsoft.2014.05.019

NASA. (2019). World Fire Image Gallery. Retrieved January 31, 2019, from
https://www.nasa.gov/mission_pages/fires/main/world/index.html

National Centers for Environmental Prediction, National Weather Service, NOAA, & U.S.
Department of Commerce. (2005). NCEP North American Regional Reanalysis
(NARR), Research Data Archive at the National Center for Atmospheric
Research, Computational and information Systems Labratory. Retrieved March
11, 2017, from http://rda.ucar.edu/datasets/ds608.0/

North, M. J., Collier, N. T., Ozik, J., Tatara, E. R., Macal, C. M., Bragen, M., & Sydelko,
P. (2013). Complex adaptive systems modeling with Repast Simphony. Complex
Adaptive Systems Modeling, 1(1), 3. https://doi.org/10.1186/2194-3206-1-3

Parunak, H. V. D., Savit, R., & Riolo, R. L. (1998). Agent-Based Modeling vs. Equation-
Based Modeling: A Case Study and Users’ Guide. In Multi-Agent Systems and
Agent-Based Simulation: First International Workshop, MABS’981 (pp. 10–25).
Lecture Notes in Computer Science.

Perez, L., & Dragicevic, S. (2009). An agent-based approach for modeling dynamics of
contagious disease spread. International Journal of Health Geographics, 8(1), 1–
17. https://doi.org/10.1186/1476-072X-8-50

Pérez, L., & Dragićević, S. (2011). ForestSimMPB: A swarming intelligence and agent-
based modeling approach for mountain pine beetle outbreaks. Ecological
Informatics, 6(1), 62–72. https://doi.org/10.1016/j.ecoinf.2010.09.003

Perrakis, D. D. B., Lanoville, R. A., Taylor, S. W., & Hicks, D. (2014). Modeling wildfire
spread in mountain pine beetle-affected forest stands, British Columbia, Canada.
Fire Ecology, 10(2), 10–35. https://doi.org/10.4996/fireecology.1002010

Rahmandad, H., & Sterman, J. (2008). Heterogeneity and Network Structure in the
Dynamics of Diffusion : Comparing Agent-Based and Differential Equation
Models, 54(5), 998–1014.

Reaney, S. M. (2008). The use of agent based modelling techniques in hydrology:

determining the spatial and temporal origin of channel flow in semi-arid
catchments. Earth Surface Processes and Landforms, 33, 317–327.
https://doi.org/10.1002/esp

Rounsevell, M. D. a, Robinson, D. T., & Murray-Rust, D. (2012). From actors to agents

in socio-ecological systems models. Philosophical Transactions of the Royal
Society of London. Series B, Biological Sciences, 367(1586), 259–69.
https://doi.org/10.1098/rstb.2011.0187

45
Smith, A. K., & Dragicevic, S. (2018). An Agent-Based Model to Represent Space-Time
Propagation of Forest-Fire Smoke. ISPRS Annals of the Photogrammetry,
Remote Sensing and Spatial Information Sciences, IV-4(ISPRS TC IV Mid-term
Symposium “3D Spatial Information Science – The Engine of Change”, 1–5
October 2018, Delft, The Netherlands), 207–212.

Sun, S., Parker, D. C., Huang, Q., Filatova, T., Robinson, D. T., Riolo, R. L., … Brown,
D. G. (2014). Market Impacts on Land-Use Change :, 104(3), 460–484.

Tang, W., & Bennett, D. A. (2010). Agent-based modeling of animal movement: A

review. Geography Compass, 4(7), 682–700. https://doi.org/10.1111/j.1749-
8198.2010.00337.x

Testa, J. W., Mock, K. J., Taylor, C., Koyuk, H., Coyle, J. R., & Waggoner, R. (2012).
Agent-based modeling of the dynamics of mammal-eating killer whales and their
prey. Marine Ecology Progress Series, 466(2003), 275–291.
https://doi.org/10.3354/meps09845

Torrens, P. M. (2014). High-resolution space-time processes for agents at the built-

human interface of urban earthquakes. International Journal of Geographical
Information Science, 28(5), 964–986.
https://doi.org/10.1080/13658816.2013.835816

Torrens, P. M., & Benenson, I. (2005). Geographic Automata Systems. International

Journal of Geographical Information Science, 19(4), 385–412.
https://doi.org/10.1080/13658810512331325139

U.S. Geological Survey. (2012). Global Digital Elevation Model (GTOPO30). U.S.
Geological Survey (USGS) Earth Resources Observation and Science (EROS)
Center.

van der Kamp, D., McKendry, I., Wong, M., & Stull, R. (2008). Lidar ceilometer
observations and modeling of a fireworks plume in Vancouver, British Columbia.
Atmospheric Environment, 42(30), 7174–7178.
https://doi.org/10.1016/j.atmosenv.2008.06.047

Zhu, W., & Timmermans, H. (2011). Modeling pedestrian shopping behavior using
principles of bounded rationality: Model comparison and validation. Journal of
Geographical Systems, 13(2), 101–126. https://doi.org/10.1007/s10109-010-
0122-8

46
Chapter 3.

Assessing the similarities of 3D simulation model

outcomes2

3.1. Introduction

The multidimensional characteristics of geospatial phenomena and the increase

in three-dimensional (3D) and four-dimensional (4D) simulation modeling (Eshraghi et al.
2012; Gobron et al. 2011; Jjumba and Dragićević 2015; Narteau et al. 2009) indicates
the necessity to develop improved methods for simulation model output comparisons.
With spatial analysis and modeling approaches advancing from two-dimensional (2D)
into 3D space, and even 4D with space-time considerations, the methods required for
map comparisons must evolve with this advance into 3D and 4D. However, non-spatial
techniques from concepts independent of the spatial dimensions can be applied to
spatial scenarios including 3D and 4D. Potentially, the most common map comparison
method is the simple overall map accuracy where two maps are compared cell-by-cell to
calculate the percent agreement (Van Genderen et al. 1978). Accuracy does not account
for chance agreement in its calculation where the two maps being compared have some
degree of similarity expected due to the randomness of cell values. Cohen’s Kappa
coefficient (Cohen 1960, 1968), or simply Kappa, was developed to account for the
chance agreement between two datasets although it was originally developed for the
field of psychology. Previous to the Kappa, the mathematically identical Heidke Skill
Score was developed for meteorology (Heidke 1926). However, Kappa is the commonly
used name in Remote Sensing and GIScience studies when dealing with image or map
comparisons. Since its development, Kappa has been widely adopted for use in remote
sensing (Congalton and Mead 1986; Hudson 1987) and raster GIS for the purpose of
comparison of raster-based images using paired cells, also known as pixels or raster,
from two images or GIS data layers representing maps. The Kappa coefficient has been
further enhanced to multiple variants such as Fleiss’ Kappa (Fleiss 1971) which has

2
A version of this chapter has been published, coauthored with S. Dragićević, under the title
“Assessing the similarities of 3D simulation model outcomes” in the ISPRS International Archives
of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B4-
2021, 339-344

47
seen use in the comparison of more than two maps (Rogers et al. 2014), Fuzzy Kappa
which introduces spatial fuzzy logic into the calculation of Kappa (Hagen 2003), and
many others specifically for land-use change comparisons including Kappa Histogram,
Kappa Location, Kappa Simulation, Kappa Transition, and Kappa Transition Location
(van Vliet et al. 2011). The hierarchy of the various Kappa variants are shown in Figure
3.1 and includes new generations of metrics, 3D Kappa, 3D Fuzzy Kappa and 4D Fuzzy
Kappa. The techniques used by 2D map comparison methods such as Fuzzy Kappa are
incompatible with 3D applications and therefore require major redevelopment before
they can be used for comparisons of simulation outcomes in multiple spatial dimensions.

Figure 3.1 Hierarchy of Kappa approaches originating from Cohen’s Kappa

48
 Both accuracy and Kappa measures are used for the comparison of 3D maps,
typically generated from LiDAR data (Roberts et al. 2019; Wang et al. 2018) however
were not as yet applied for the needs of geosimulation modeling. Therefore, the main
objective of this study is to compare the similarity of 3D simulation modeling outcomes
using 3D Accuracy and 3D Kappa, and then to theoretically extend the analysis towards
possible spatial extension of Kappa to 3D and 4D metrics. A previously developed 4D
agent-based model (ABM) for forest fire smoke propagation (Smith and Dragićević 2019)
was used to generate the simulation outcomes used as a case study for comparison of
3D simulation outcomes using 3D Accuracy and 3D Kappa measures.

3.2. Methodology

Similar to the resel or raster concept in 2D GIS (Tobler 1995), voxels are the
smallest representation of 3D space, divided into a regular 3D lattice, typically forming
cubes (Greene 1989; Jjumba and Dragićević 2016). Similar to raster GIS data layer,
representing array of equal-size uniform square cells, a voxel data layer represents the
equal-size cubes. Each voxel contains a value, such as a class, that represents an
attribute characterizing the 3D space that the voxel occupies, empty voxels must be
represented by null values. In this study the 3D Kappa approach (Smith and Dragićević
2022) have been used to compare simulation outcomes in 3D. The 3D Accuracy and 3D
Kappa measures are designed for use with voxel representation of model simulation
outputs. The 3D Accuracy is calculated by comparing two voxel datasets voxel-by-voxel
and calculating the proportion of agreement between the pairs of voxels. The name 3D
Accuracy is given to distinguish it from the cell-by-cell methods used by Accuracy in 2D
map comparison. Cohen’s Kappa is extended to be applied to the 3D voxel datasets
through 3D Kappa (K3D) and can be defined by the following equation:

𝑃 𝑃
𝐾 1
1 𝑃

where Po represents the observed proportion of agreement, or observed

similarity, among voxels in two 3D maps or simulation outcomes S, and Pe represents
the expected proportion of agreement, or expected similarity, if voxel classes were
assigned randomly to the voxels of both datasets. Po is equal to the 3D Accuracy of the
two voxel datasets. The expected similarity Pe is defined by:

49
 𝑃 𝑃 ∗𝑃 2

where PiA and PiB represent the proportion of class i in simulation outcome SA
and outcome SB respectively, for all n classes in the two 3D maps or simulation
outcomes. To implement both 3D Accuracy and 3D Kappa for the 3D modeling outputs,
a custom program in Python 2.7 was developed. This program accepts two voxel
datasets converted into 3D arrays of the same extent. It synchronously iterates over
each entry of the arrays, recording relevant information, including the counts of each
class per array, the number of matches, and the total number of voxels with data.
Equations (1) and (2) are calculated and 3D Kappa is returned alongside Po as 3D
Accuracy.

This research study uses an existing ABM to simulate the propagation of forest
fire smoke (Smith and Dragićević 2019) based on a hypothetical fire. The ABM operates
with two primary agent types - fire agents and smoke agents. The fire agents represent
fire locations, and they emit smoke agents at a predetermined rate that represent smoke
in the atmosphere. Smoke agents move using two methods, the first being carried by the
wind, and the second through diffusion to increase dispersion. The local terrain acts as a
lower limit for the elevation of smoke agents as they travel over, around, and through
mountains and valleys. The simulation outcomes are voxelized, creating voxels with
values representing the number of agents they contain, thus the concentration of smoke.

3.3. Implementation

The study area is located in British Columbia (BC), Canada, where fire and
smoke are represented as agents as they move through the mountains using wind
patterns and terrain. The 3D spatial extent of the study area is 96km x 45km x 10km with
a southwest corner located at 70 Mile House, BC and a northeast corner of
approximately 5km northeast of Mahood Lake, where the Murtle River joins the
Clearwater River. The study area is approximately 100km north of Kamloops, BC. The
simulation uses a hypothetical forest fire started on August 10th, 2017 on the west coast
of BC and burns for a total of 15 days (Smith and Dragićević 2019). The model
simulation outputs are obtained in voxel data format with the locations of smoke agents
and at a voxel of resolution of 1km. The ABM simulate the concentration of smoke

50
agents that are represented with five voxel classes. Voxel classes were determined by
the number of smoke agents in each voxel based on the information from the
Government of BC Air Quality Health Index (Government of British Columbia 2020).
Class 0 represents no smoke when voxels contain no smoke agents. Class 1 represents
low smoke with 1-3 smoke agents, Class 2 represents moderate smoke with 4-6 smoke
agents, Class 3 represents high smoke with 7-10 smoke agents, and Class 4 represents
very high smoke with more than 10 smoke agents in the voxel.

Three sets of comparisons were made between different output times and
settings of the agent-based model. The first set of comparisons is composed of four
simulation outcomes to allow for six comparisons (Figure 3.2). The study area contains
smoke on August 23rd at 3pm, 13 days and 15 hours into the simulation. The model is
run twice to retrieve two simulation outcomes, Simulation 1 and Simulation 2. The model
also has an option for the forest fire to move under the influence of the wind and is used
to retrieve Simulation 3. A second output time was selected for Simulation 4 on the 7th
day, or 0:00am on August 17th, 2017 when smoke was also found to be present in the
study area for Simulation 4. Comparison 1 is between Simulation 1 and Simulation 2 of
the static fire scenarios, where both simulations use the same initial conditions but
different automatically generated random seeds that define the set of random values
used. This comparison is used to assess the differences caused by the randomness in
the model. The randomness originates from two sources, one is the random diffusion of
smoke and the other is due to the random order of the agents’ movement. Comparison 2
is between Simulation 1 and Simulation 3 using the scenario with a fire that moves over
space-time through the landscape of the study area. Comparison 3 is between
Simulation 1 and Simulation 4 which is obtained from the alternate time step assuming
the 3D terrain is identical, but the smoke will have experienced different wind conditions.
Comparisons 4, 5, and 6 are between all other possible combinations of these four
simulation outcomes.

51
Figure 3.2 Agent-based model forest-fire smoke propagation simulation
outcomes in voxel data format used for six comparisons.

52
 Smoke is present in the study area for an additional 9 hours after the output time
used for Simulation 1, and a total of eight additional ABM simulation outputs have been
generated from the two runs and presented on Figure 3.3. These four outputs are
compared against the outputs of the second run of the model, where the first outputs are
Simulation 1 and Simulation 2 from the first set of comparisons. The 16 comparisons
between these two sets of four simulation outputs make up the second set of
comparisons. These comparisons are done to investigate a use case of 3D Kappa for
comparing simulations to temporally match patterns from two simulations.

Figure 3.3 Simulation outputs S1 and S2 at initial stage and after three hours,
six hours, and nine hours of smoke propagation.

53
 The third and final set of simulation outcome comparisons is performed to
explore the sensitivity analysis of the diffusion rate of the smoke of the agent-based
model. The distance used by the diffusion process is multiplied by the diffusion factor
(DF), where DF = 2 is the most similar to the calibrated DF of the model. Six levels of
diffusion are used and results in DF1, DF4, DF8, DF12, DF16, and DF20. This analysis
will help assess the impact of the diffusion factor, for the selected simulation runs. Due
to the required computer processing time of the model only one run for each diffusion
factor was completed.

3.4. Results and Discussion

The obtained results in Table 3.1 show the 3D Accuracy and 3D Kappa values
for the first set of comparisons. Due to the large number of voxels with no smoke in the
outputs, the accuracy typically receives high values. The only comparison to receive a
moderate or better 3D Kappa value is Comparison 1, between the two runs with the
same initial conditions. This comparison was expected to receive much higher 3D Kappa
values due to the very similar appearance of the smoke patterns. However, the 3D
Accuracy of this comparison is only slightly higher than Comparisons 3 and 5 which
obtained the lowest 3D Kappa values. These two Comparisons involve Simulation 4
which along with Comparison 6 achieve the lowest three 3D Kappa values and with
exception of Comparison 1, the highest 3D Accuracy values. This is due to the large
number of voxels with no smoke in Simulation 4, more than the other simulations.

Table 3.1 Values obtained for 3D Accuracy and 3D Kappa for the six
comparisons of the forest fire smoke propagation ABM simulation
outputs.

Comparison Simulations 3D Accuracy 3D Kappa

C1 S1 S2 0.96 0.53
C2 S1 S3 0.87 0.15
C3 S1 S4 0.95 0.01
C4 S2 S3 0.89 0.15
C5 S2 S4 0.95 0.00
C6 S3 S4 0.90 0.03

The randomness incorporated in the ABM creates the differences between

Simulation 1 and Simulation 2 presenting itself partially in the form of randomly

54
distributed voxels within the 3D space (Figure 3.4). While this can make two 3D maps
appear visually similar, few of these voxels with smoke in one map agree with their
counterpart in the other. This leads to noticeably lower 3D Kappa values than expected
and slightly lower 3D Accuracy values. In 2D spatial analysis techniques, a moving
window is often used to incorporate nearby cells into the analysis of the central cell.
Situations like this could be more accurately assessed if this type of moving window
analysis could be developed into a more advanced 3D map comparison method. Using a
moving 3D window, nearby voxels could be used to partially substitute the agreement
between two disagreeing voxels. A similar technique with the moving window was
adopted for the calculation of Fuzzy Kappa (Hagen 2003) in 2D and can allow for further
expansion into 3D Fuzzy Kappa to accommodate the benefits of the use of 3D window
space.

Figure 3.4 Randomly distributed smoke voxels highlighted in red from ABM
outputs for simulation S1 and S2.

55
 The values for 3D Accuracy and 3D Kappa for the second set of comparisons are
presented in Table 3.2. Similar to the first set of comparisons, accuracy is very high, with
the lowest value of 0.86. The accuracy values of the comparisons of outputs of the same
time step are typically the highest in their row and column, although by small differences.
However, 3D Kappa produces greater differences between the comparisons of outputs
of the same time step and all other comparisons. Neither 3D Accuracy nor 3D Kappa
include any time component that can incorporate differences in time to the comparison
of data.

Table 3.2 Values obtained for 3D Accuracy and 3D Kappa for the 3D
comparisons of four forest-fire smoke model simulation outputs for
two runs S1 and S2.

3D Accuracy
Time Steps S2 S2+3h S2+6h S2+9h
S1 0.96 0.91 0.86 0.91
S1+3h 0.92 0.94 0.87 0.90
S1+6h 0.88 0.87 0.92 0.88
S1+9h 0.91 0.88 0.86 0.96
3D Kappa
Time Steps S2 S2+3h S2+6h S2+9h
S1 0.53 0.17 0.08 0.03
S1+3h 0.17 0.49 0.10 0.01
S1+6h 0.09 0.11 0.54 0.08
S1+9h 0.02 0.00 0.08 0.60

Table 3.3 presents the values obtained for 3D Kappa for the third set of
comparisons. These last comparisons are accomplished to investigate the effect of the
rate of the smoke diffusion on the simulation outputs generated by the ABM. Similar to
the other results, the 3D Accuracy values are very high, but were also very consistent
with all values being 0.95-0.96. While the 3D Kappa values are also somewhat similar,
they show a clearer trend. The larger the difference in DF, the lower the 3D Kappa
values. Additionally, comparing simulation outputs with higher DF values is also
associated with a lower 3D Kappa. This is most likely caused by the increased diffusion
rate increasing the amount of randomness in smoke location in the model. Because 3D
Kappa has no spatial fuzziness unlike Fuzzy Kappa, it cannot account for the increasing
amounts of randomness in smoke locations with higher diffusion factors. This indicates

56
that more advanced 3D Kappa metrics have to be developed such as 3D and 4D Fuzzy
Kappa to incorporate a variable 3D window size and 3D spatial autocorrelation to
accommodate the variation of simulation outcomes in 3D and 4D space-time.

Table 3.3 Values obtained for 3D Kappa for the comparisons of the forest-fire
smoke propagation model simulation outputs with varying diffusion
rates (DF).

Simulation
Outcomes DF1 DF4 DF8 DF12 DF16 DF20
DF1 1.00 0.49 0.40 0.37 0.33 0.28
DF4 1.00 0.46 0.42 0.39 0.35
DF8 1.00 0.43 0.40 0.37
DF12 1.00 0.39 0.36
DF16 1.00 0.34
DF20 1.00

3.5. Conclusions

This research study extends two existing methods of Accuracy and Kappa
measures into 3D by proposing 3D Accuracy and 3D Kappa for comparing model
simulation outcomes in 3D space. Critics of Kappa typically have concerns with how
chance is accounted for (Foody 2020; Pontius and Millones 2011), however for use in
assessing models such as the one used in this study that have built-in randomness, this
concern may be reduced. Kappa is also an often-required metric in many fields, such as
remote sensing and GIS, and is applied despite the criticisms. While in this research
study 3D Accuracy and 3D Kappa metrics show potential for use in comparison of model
outputs dominated by a single voxel class, it also revealed a greater potential to expand
these methods to account for spatial neighborhoods or spatial autocorrelation in 3D
domain. The existing 2D Fuzzy Kappa method can be extended into 3D and even 4D
Fuzzy Kappa counterparts to advance map and simulation outcome comparison into 3D
and 4D while incorporating fuzzy logic (Smith and Dragićević in press). The proposed 3D
Accuracy and 3D Kappa measures may be the first step towards enabling more
comprehensive comparison methods of simulation outcomes and generated patterns in
3D. While designed for 3D simulation outcomes and 3D map comparisons, such
measures may also be useful to other fields, including 3D medical imaging, and in the
training and comparison of predicted 3D AI-system behaviour with the real-world data.
Based on findings from this research study, further advance methods of 3D simulation

57
outcomes and 3D pattern comparisons should be developed that facilitate the calibration
and validation of simulation models such as 4D ABM or voxel automata operating in the
4D space-time domain.

3.6. References

Cohen, J. (1960). A Coefficient of Agreement for Nominal Scales. Educational and

Psychological Measurement, 20(1), 37–46.

Cohen, J. (1968). Weighted Kappa: Nominal scale agreement provision for scaled
disagreement or partial credit. Psychological Bulletin, 70(4), 213–220.
https://doi.org/10.1037/h0026256

Congalton, R. G., & Mead, R. A. (1986). A Review of Three Discrete Multivariate

Analysis Techniques Used in Assessing the Accuracy of Remotely Sensed Data
from Error Matrices. IEEE Transactions on Geoscience and Remote Sensing,
GE-24(1), 169–174. https://doi.org/10.1109/TGRS.1986.289546

Eshraghi, M., Felicelli, S. D., & Jelinek, B. (2012). Three dimensional simulation of
solutal dendrite growth using lattice Boltzmann and cellular automaton methods.
Journal of Crystal Growth, 354(1), 129–134.
https://doi.org/10.1016/j.jcrysgro.2012.06.002

Fleiss, J. L. (1971). Measuring Nominal Scale Agreement Among Many Raters.

Psycological Bulletin, 76(5), 378–382.

Foody, G. M. (2020). Explaining the unsuitability of the kappa coefficient in the

assessment and comparison of the accuracy of thematic maps obtained by
image classification. Remote Sensing of Environment, 239, 111630.
https://doi.org/10.1016/j.rse.2019.111630

Gobron, S., Çöltekin, A., Bonafos, H., & Thalmann, D. (2011). GPGPU computation and
visualization of three-dimensional cellular automata. Visual Computer, 27(1), 67–
81. https://doi.org/10.1007/s00371-010-0515-1

Government of British Columbia. (2020). BC Air Quality. Retrieved January 31, 2020,
from http://www.env.gov.bc.ca/epd/bcairquality/data/aqhi-table.html

Greene, N. (1989). Voxel space automata: Modeling with stochastic growth processes in
voxel space. Computer Graphics, 23(3), 175–184.
https://doi.org/10.1145/74334.74351

Hagen, A. (2003). Fuzzy set approach to assessing similarity of categorical maps.

International Journal of Geographical Information Science, 17(3), 235–249.
https://doi.org/10.1080/13658810210157822

Heidke, P. (1926). Berechnung des Erfolges und der Gute der Windstarkevorhersagen
im Sturmwarnungsdienst. Geografiska Annaler, 8, 301–349.

58
Hudson, W. D. (1987). Evaluation of several classification schemes for mapping forest
cover types in Michigan. International Journal of Remote Sensing, 8(12), 1785–
1796. https://doi.org/10.1080/01431168708954816

Jjumba, A., & Dragićević, S. (2015). Integrating GIS-Based Geo-Atom Theory and Voxel
Automata to Simulate the Dispersal of Airborne Pollutants. Transactions in GIS,
19(4), 582–603. https://doi.org/10.1111/tgis.12113

Jjumba, A., & Dragićević, S. (2016). Spatial indices for measuring three-dimensional
patterns in a voxel-based space. Journal of Geographical Systems, 18(3), 183–
204. https://doi.org/10.1007/s10109-016-0231-0

Narteau, C., Zhang, D., Rozier, O., & Claudin, P. (2009). Setting the length and time
scales of a cellular automaton dune model from the analysis of superimposed
bed forms. Journal of Geophysical Research: Earth Surface, 114(3), 1–18.
https://doi.org/10.1029/2008JF001127

Pontius, R. G., & Millones, M. (2011). Death to Kappa: Birth of quantity disagreement
and allocation disagreement for accuracy assessment. International Journal of
Remote Sensing, 32(15), 4407–4429.
https://doi.org/10.1080/01431161.2011.552923

Roberts, K. C., Lindsay, J. B., & Berg, A. A. (2019). An analysis of ground-point

classifiers for terrestrial LiDAR. Remote Sensing, 11(16).
https://doi.org/10.3390/rs11161915

Rogers, D. J., Suk, J. E., & Semenza, J. C. (2014). Using global maps to predict the risk
of dengue in Europe. Acta Tropica, 129(1), 1–14.
https://doi.org/10.1016/j.actatropica.2013.08.008

Smith, A. K., & Dragićević, S. (2019). A four-dimensional agent-based model: A case

study of forest-fire smoke propagation. Transactions in GIS, 23(3), 417–434.
https://doi.org/10.1111/tgis.12551

Smith, A. K., & Dragićević, S. (2022). Map comparison methods for three-dimensional
space and time voxel data. Geographical Analysis, 54(1), 149–172.
https://doi.org/10.1111/gean.12279

Tobler, W. (1995). The resel-based GIS. International Journal of Geographical

Information Systems, 9(1), 95–100. https://doi.org/10.1080/02693799508902027

Van Genderen, J. L., Lock, B. F., & Vass, P. A. (1978). Remote Sensing: Statistical
Testing of Thematic Map Accuracy. Remote Sensing of Environment, 7(1), 3–14.
https://doi.org/10.1016/0034-4257(78)90003-2

van Vliet, J., Bregt, A. K., & Hagen-Zanker, A. (2011). Revisiting Kappa to account for
change in the accuracy assessment of land-use change models. Ecological
Modelling, 222(8), 1367–1375. https://doi.org/10.1016/j.ecolmodel.2011.01.017

Wang, J., Lindenbergh, R., & Menenti, M. (2018). Scalable individual tree delineation in
3D point clouds. Photogrammetric Record, 33(163), 315–340.
https://doi.org/10.1111/phor.12247

59
Chapter 4.

Map comparison methods for four-dimensional

space and time voxel data3

4.1. Introduction

Map comparison in general is an important topic within the field of geographic

information science (GIScience), specifically cartography, spatial analysis and
geosimulation modeling. The purpose of map comparisons is to find similarities among
features and spatial patterns on two maps (Muller 1975; van Vliet et al. 2011), discover
possible correlations or correspondences cartographically (Power et al. 2001; Stehman
1999; Visser and De Nijs 2006; Woodcock and Gopal 2010) or with use of spatial
analysis methods (Foody 2007; Lavigne et al. 2010). With developments of
geosimulation and other modeling methods, model evaluation procedures demand
approaches that permit the comparisons of model outcomes with actual data (Brown et
al. 2005; Li et al. 2015; Samardzic-Petrovic et al. 2017).

The existing research work in GIScience over the past several decades has been
focused on map comparison or pattern analysis techniques uniquely for determining
similarity in two-dimensional (2D) space. Techniques have been developed for
comparisons of two categorical maps as raster-based geographic information system
(GIS) layers and have evolved from the simple overall map accuracy (Van Genderen et
al. 1978), Cohen’s Kappa statistics (Cohen 1960; Congalton 1991), Weighted Kappa
(Balzter et al. 2002; Cohen 1968), and Fleiss’ Kappa (Fleiss 1971), into techniques such
as Fuzzy Kappa (Hagen 2003), Kappa location (van Vliet et al. 2011) and Kappa
simulation (Hagen-Zanker 2009; van Vliet et al. 2011).

While Cohen’s Kappa (Cohen 1960, 1968) and Fleiss Kappa (Fleiss 1971) were
originally developed for the field of psychology, they were later adopted to work on
spatial data. Although Cohen’s naming of Kappa has been adopted, it is predated by the

3
A version of this chapter has been published, coauthored with S. Dragićević, under the title
“Map comparison methods for three-dimensional space and time voxel data” in the journal
Geographic Analysis, 2022, 54, 149-172

60
Heidke Skill Score that uses the same calculations for use in meteorological modelling
(Heidke 1926). Remote sensing (RS) and GIScience research communities are widely
using Kappa coefficients to compare the satellite images or as a tool for map
comparisons to account for chance agreement between maps, dating back to at least
1986 (Congalton and Mead 1986; Hudson 1987). Kappa was expanded to Weighted
Kappa and is used to account for degree of similarity between categorical maps or
different categories of image classification. Fleiss’ Kappa was further adapted to
compare agreement between multiple maps (Pringle et al. 2012; Rogers et al. 2014).
While initial versions of Kappa coefficients were not developed specifically for spatial
analysis and map comparison, Fuzzy Kappa was developed specifically to address
spatial dependencies and similarity (Hagen 2003). Fuzzy Kappa incorporates Weighted
Kappa with a fuzzy spatial component that uses a fuzzy neighborhood and spatial
autocorrelation (Hagen-Zanker 2009) to comparing raster-based cells while accounting
for the influence of nearby cells on the similarity. Furthermore to avoid cell-by-cell
comparisons, pattern-oriented map comparisons methods have been developed to
analyze and compare spatial patterns generated from simulation modeling approaches
(White 2006). Such methods include Hierarchical Fuzzy Pattern Matching (Power et al.
2001) and patch-based metrics such as Contagion (O’Neill et al. 1988).

These techniques have been useful for remote sensing image classification
procedures (Foody 2002; Hu and Wang 2013) and modeling approaches (Ahmed et al.
2013; Ding et al. 2013) specifically designed for 2D. More recent data collection and
model outputs are produced in three and four dimensions (3D and 4D) and are present
with increasing frequency. For example, Light Detection And Ranging (LiDAR)
technologies permit geospatial data collection in 3D point clouds (Woodhouse et al.
2011), Building Information Models (BIM) provide representations of built structures in
3D (Jung and Joo 2011), as well as the rise of geosimulation models that operate in 3D
(Koziatek and Dragicevic 2017; Smith and Dragićević 2019; Torrens 2014) or 4D, as 3D
changing with time (Aurbach et al. 2018; Zielinski et al. 2018). Model outputs are
generated in 3D or 4D from various applications including plant growth (Greene 1989),
weather forecasting (Grell et al. 2005), soil organic carbon stocks (Poggio and Gimona
2014), ballistic impacts (Bova et al. 2016), bird migration (Aurbach et al. 2018), airborne
pollutants (Jjumba and Dragićević 2015), or forest-fire smoke modeling (Smith and
Dragićević 2019), to name a few. There is a necessity to develop methods that allow for

61
comparison of such multidimensional data or model outputs to ensure abilities of map
comparison in 3D or 4D.

Work in other fields, such as computer vision (Gobron et al. 2011), medical
imaging (Dou et al. 2007), geology (Pellerin et al. 2015), landscape ecology (Parrott et
al. 2008), and urban applications (Chen et al. 2014; Kedron et al. 2019; Koziatek and
Dragićević 2019) have already advanced into 3D and 4D. Some work has been done in
GIScience in developing 3D metrics of patterns in voxel space including 3D spatial auto-
correlation (Jjumba and Dragićević 2016a) and urban civil engineering applications
(Shirowzhan et al. 2019). Various aspects of the 2D spatial Kappa coefficient have been
criticized (Foody 2020; Pontius and Millones 2011), including the limitations of the
Heidke Skill Score (Appleman 1960). These criticisms are typically targeted towards how
chance is modeled by the Kappa coefficient, especially for use in remote sensing.
However, there have been different debates on the usefulness of Kappa for space-time
model assessments (van Vliet et al. 2011). Even with the debates for and against the
use of the Kappa coefficient and its derivatives, they are still commonly used, and in
some fields expected metrics to use in comparing maps or model outputs. Although
these forms of Kappa statistics cannot be directly applied to multidimensional data or
processes but rather to their confusion matrices, they can be expanded to address
comparisons of data and model outputs produced in 3D and 4D. Therefore, the main
objective of this research study is to develop and implement comparison methods that
work in the 3D and 4D space-time domain and use voxel geospatial data representation.

4.2. Mathematical background

4.2.1. Cohen’s Kappa

Cohen’s Kappa Coefficient (K) is used as the basis of the later developed forms
of Kappa coefficient for spatial applications, it is firstly defined (Cohen 1960; Congalton
1991; Heidke 1926; Monserud and Leemans 1992) as follows:

𝑃 𝑃
𝛫 (1)
1 𝑃

62
 where Po is the observed similarity and Pe is the expected similarity of two
datasets, or map A and map B in spatial applications. The observed similarity Po is the
measured proportion of agreement between the two maps. The expected similarity Pe is
defined by:

𝑃 𝑃 ∙𝑃 (2)

where PiA represents the proportion of class i in map A, and PiB represents the
proportion of class i in map B, both for n classes in the two maps. For example, if class 1
in map A (P1A) represents 0.40 of the map, and class 1 in map B (P1B) represents 0.50 of
the map, their product will be 0.20, and will be added to the same process for all other
classes. By calculating Po and Pe from a cell-by-cell comparison, counting the proportion
of cells that both maps are in agreement and the proportion of each class in both maps.
Given the wide applications of the Kappa coefficient, it is a part of common raster GIS or
RS software packages as well as stand-alone tools such as the Map Comparison Kit
(Visser and De Nijs 2006).

4.2.2. Fuzzy Kappa

To develop a map comparison technique for 2D spatial patterns and

heterogeneity, Cohen’s Kappa coefficient was modified to Fuzzy Kappa coefficient
(KFuzzy) (Hagen 2003). Fuzziness is used to allow raster data with similarity to have a
degree of equality rather than the binary similarity of equal or unequal. This is
accomplished in two ways, for distances among raster cells and for their categories.
Depending on the scenario, some categories, or classes, in a map may be considered
similar to each other, such as different forms of residential land use, or agricultural land
use. This similarity is defined using a Fuzzy Category Vector that assigns a value
between 0 and 1 to each pair of classes representing the degree of similarity ranging
respectively from no to total similarity. The Fuzzy Category Vector is used along with
distance for determining the similarity of two cells. Fuzziness in distance is where cells
can be considered similar due to proximity. As with category, Fuzzy Kappa metrics
assigns a value between 0 and 1 to cells in the neighborhood of a central cell. The
assigned value depends on the user, however resulting similarity decreases with

63
increase of distance. Fuzzy Kappa is based around Equation (1) used in Cohen’s Kappa
coefficient however, Po and Pe are calculated using new methods that incorporate the
fuzzy components.

The observed similarity, Po, is calculated cell by cell with a moving window and is
averaged for the entire map. For each cell, a two-way comparison is done by taking the
minimum value of two one-way comparisons that are done in opposite directions. A one-
way comparison calculates the similarities of the central cell of one window compared to
each cell in the neighborhood of the other window to return the maximum value. The
similarity between the central cell and another cell is calculated by multiplying the class
similarity by the user-defined distance function M(di) and should result in a value of 1
when comparing the two central cells with the same class, and should decrease if the
cells are further apart or with less similar classes (Hagen 2003). As an example, if the
maximum similarity between the central cell from the window in Map A with the cells in
the window from Map B is 0.5, and from Map B to Map A is 0.2, the similarity of these
two windows is 0.2.

The expected similarity Pe, is calculated using the moving window neighborhood.
The neighborhood is segmented into rings of cells that are equidistant from the central
cell. For example, Ring 0 contains only the central cell, Ring 1 contains the four cells in
the cardinal directions, Ring 2 contains the four cells directly diagonal, Ring 3 contains
the four cells that are two cells away in the cardinal directions, and Ring 4 contains the
eight cells a knight’s move away, and so on. The probability E(i) that an average ring i
from one map contains a matching cell to the central cell of another map is multiplied by
a user-defined distance function M(di) using the distance of the ring di for each ring i
(i=0,R). The distance function M(di) is the same function used in the calculation of the
observed similarity Po. Therefore, the expected similarity Pe for a neighborhood with R
number of rings can be calculated as:

𝑃 𝐸 𝑖 𝑀 𝑑 (3)

Then the resulting value for Fuzzy Kappa coefficient for two map comparison can
be calculated with Po and Pe using equation (1).

64
4.3. Methods

4.3.1. Geospatial Data Representation in 3D and 4D

Representation of data in 3D can be accomplished in several ways, including

points defined with three spatial coordinates, irregular 3D tessellations such as
tetrahedral or Voronoi networks, and voxels. LiDAR data (Beland et al. 2019) is an
example of 3D point data that forms clouds of points outlining and representing the
measured objects in 3D. Tetrahedral networks (Shi et al. 2003) and Voronoi 3D
networks (Beni et al. 2011) were used to represent data in 3D volumes composed of
polygons with several triangular or Voronoi diagrams as sides. Finally, voxels (Greene
1989) representing the smallest cubic spatial units are used to represent 3D space
through a 3D lattice, similar to raster-based cells used in 2D maps. Some examples of
voxels can be seen in representation of trees (Hosoi et al. 2013), urban landscapes
(Nourian et al. 2016) or brain magnetic resonance imaging (MRI) scans (Fritz et al.
2014). Figure 4.1 represents the evolution of raster to voxels space from 2D to 4D that
represent 2D and 3D maps. Voxels have been used for modeling approaches such as
voxel automata (VA), the 3D equivalent of the 2D cellular automata (CA), to represent
various multidimensional processes (Gobron et al. 2011; Jjumba and Dragićević 2016b).
Voxels are iteratively updated to represent change over space and time using a set of
transition rules based on voxel states in the local 3D neighborhood to determine the new
state of a central voxel. Voxel automata have been used to simulate landslides (Segre
and Deangeli 1995), bed forms of dunes (Narteau et al. 2009), root growth (Mulia et al.
2010), procedural cave generation (Cui et al. 2011), dendrite growth (Eshraghi et al.
2012) or air pollution propagation (Jjumba and Dragićević 2015). The VA approach has
been also further extended to agent-based models (ABM) that operate in 4D (Smith and
Dragićević 2019). In this research study a 3D map is considered to be a cuboid
containing voxels that are defined with coordinates (x,y,z) and class value for each voxel
that together represent various 3D features as part of the 3D map. With map
comparisons, two 3D maps, Map A and Map B are considered, captured at the same
time ti and the same spatial extent, are considered where both maps are representing
either two sets of actual data, simulation outcomes and actual data, or two simulation
outcomes.

65
Figure 4.1 Evolution of GIS data representation from 2D map containing rasters
to 3D map containing voxels in 3D space or 4D as space-time
defined by x,y,z coordinates and time t with three map classes.

4.3.2. Approaches for 3D Map Comparisons

3D and 4D Accuracy

In this research study, 3D and 4D accuracy is calculated based off the traditional
method used by 2D map comparisons, using the proportion of agreement of two
datasets, derived from an error or confusion matrix (Congalton and Green 1993). For 3D
map comparison, the two datasets are two 3D maps, or two space-time 3D maps for a
4D comparison. The voxels are spatially paired voxel-by-voxel, so that each voxel in one
map is paired with the voxel at the same location (x,y,z) in the other map. The proportion
of agreement is calculated by dividing the total number of pairs that are equal in value by
the total number of pairs, resulting in a value between 0 and 1.

3D Kappa Coefficient

The methodology behind Cohen’s Kappa coefficient is adapted to compare two

3D maps by comparison of spatial data or model outcomes represented with 3D voxels
and specifically for the 3D space. The methodology is changed in how the values are
measured from 3D voxels instead of 2D raster cells as in spatial Kappa applications
involving raster data. For the 3D Kappa coefficient (K3D) firstly Po is calculated through a

66
voxel-by-voxel comparison for each 3D coordinate in overall map cubes for the two 3D
maps to find the proportion of voxels in agreement. Secondly, PiA and PiB from equation
(2) for map A and B are calculated to find the proportion of all classes in each 3D map.

3D Fuzzy Kappa Coefficient

The 3D Fuzzy Kappa coefficient (K3DFuzzy) method expands the existing 2D

counterpart by comparing two 3D maps using observed similarity (Po3D) and expected
similarity (Pe3D) to account for fuzziness in both distance and category of classes in 3D
maps, as follows:

𝑃 𝑃
𝛫 (4)
1 𝑃

However, in order to complete the calculations, Po3D and Pe3D and their
components must be defined in 3D space using a 3D voxel window and shells (Figure
4.2). The 3D voxel window represents the voxels forming a cube around a given central
voxel. Due to the regular lattice, multiple voxels in the 3D voxel window have the same
distance from the central voxel and are grouped into shells. Similar to the 2D rings used
in KFuzzy, shells (s) in K3DFuzzy are the 3D equivalent and are numbered as s=0,N where 0
is assigned to central voxel and increases to N for the furthest shell from the central
voxel at distance (L).

Figure 4.2 The sliced 3D voxel window extracted from a 3D map, consisting of
5x5x5 voxels with the shell numbers for each voxel.

67
 Observed 3D Similarity (Po3D) has three main components, Voxel Similarity, One-
Way 3D Similarity, and Two-Way 3D Similarity and they are defined as follows:

Voxel Similarity S(a,b) – The similarity of two voxels a and b, is the product of
the similarity of 3D distance D(a,b) and the similarity of voxel class category C(a,b). The
similarity of 3D distance D(a,b) is calculated as the 3D Euclidian distance d(a,b) between
two voxels a and b with a user-defined distance decay function M(d) and maximum
distance L, as follows:

𝑆 𝑎, 𝑏 𝐷 𝑎, 𝑏 ∙ 𝐶 𝑎, 𝑏 (5)

𝑀 𝑑 𝑎, 𝑏 , 𝑓𝑜𝑟 𝑑 𝑎, 𝑏 𝐿
𝐷 𝑎, 𝑏 (6)
0, 𝑓𝑜𝑟 𝑑 𝑎, 𝑏 𝐿

𝑀 𝑑 𝑒 2 (7)

𝑑 𝑎, 𝑏 𝑎 𝑏 𝑎 𝑏 𝑎 𝑏 (8)

where ai, aj, ak, bi, bj, and bk represent the x, y, and z coordinates of the two
voxels a and b. The distance decay function M(d) can be calculated in various ways but
in this research study an exponential distance decay function has been chosen, similar
to the KFuzzy in 2D. The values of voxel similarity of 3D distance ds for each shell (s) of a
window size (w,w,w) with w=5 voxels are presented in the example from Figure 4.2 and
Table 4.1. The user can define the maximum size w of the 3D voxel window to hold all
the desired shells and in respect to overall extent of the 3D map defining the study area.

Table 4.1 Allocation of shell numbers, voxel distances, and similarity of 3D

distance for the 5x5x5 3D voxel window.
Shell number (s) 0 1 2 3 4 5 6 7 8 9
Distance from central voxel (ds) for 0 1 √2 √3 2 √5 √6 √8 3 √12
each shell [in number of voxels]
Similarity of 3D distance D(a,b) 1 0.71 0.61 0.55 0.5 0.46 0.43 0.38 0.35 0.30

The similarity in class category C(a,b) calls upon a user generated matrix
defining the similarity between the classes categories of the two 3D maps and by

68
extracting the similarity between the classes that voxels a and b represent. The values
can be between 0 and 1, where 0 represents no similarity and 1 represent perfect
similarity among voxel class categories. For example, the classes of “tree canopy” and
“underbrush” may have a user specified similarity of 0.85 as they are quite similar while
“tree canopy” and “bedrock” may be given a similarity of 0.05 as they have very low
similarity. This would result in two voxels of “tree canopy” and “underbrush” having
C(a,b) = 0.85.

One-Way 3D Similarity Sone-way (W,V) – provides the maximum value for the
Voxel Similarity S(a,b) between each voxel Vi,j,k on one map compared to its counterpart
voxels in the second map located in the 3D voxel window Wi,I,k centered on voxel Vi,j,k,
such as:

𝑆 𝑊, 𝑉 𝑆 𝑊, , ,𝑉, , ,𝑆 𝑊 , , ,𝑉, , ,…,𝑆 𝑊 , , ,𝑉, , (9)

where w represents the dimensions of the 3D voxel window and i, j, k indicate the
location of the voxel in the map.

Two-Way 3D Similarity Stwo-way(Map AW(I,j,k), Map BW(I,j,k)) – provides the

minimum value of the two values obtained for the One-Way 3D Similarity Sone-way. One
value is calculated between a 3D voxel window WA(i,j,k) from Map A centered on voxel
VA(i,j,k) and voxel VB(i,j,k) from Map B. The second value is calculated between a 3D voxel
window WB(i,j,k) from Map B centered on voxel VB(i,j,k) and voxel VA(i,j,k) from Map A such
as:

𝑆 𝑊 , , ,𝑊 , ,
(10)
𝑆 𝑊 , , ,𝑉 , , ,𝑆 𝑊 , , ,𝑉 , ,

where i, j, and k represent the same location of a voxel in both Map A and B.

Observed 3D Similarity Po3D – is calculated between the two 3D maps, Map A

and Map B, by averaging the Two-Way 3D Similarity (Stwo-way) between all pairs of voxels
such as:

69
 1
𝑃 ∙ 𝑆 𝑊 , , ,𝑊 , , (11)
𝑝∙𝑞∙𝑟

where p, q, and r represent the 3D spatial extent of the overall voxel 3D Map.

Expected 3D Similarity (Pe3D) has two main components, the Expected Shell
Similarity for the outer shells E(s|s≥1), and the central shell E(s|s=o).

Expected Shell Similarity for outer shells E(s|s≥1) – can be calculated for shell
number s, when s≥1, by using the probability of voxels matching in various shells of the
3D voxel window except for Shell 0, s=0, which contains the central voxel. Firstly, the
probability P(n) is calculated with the probability (Ba) that a voxel from Map B contains
class a from Map A, and the probability (Ab) that a voxel from Map A contains class b
from Map B, for n voxels. Secondly, the probability P(ns) and P(ns-1) are calculated for
each shell s and s-1, as follows:

𝑃 𝑛 1 1 𝐵 ∙ 1 1 𝐴 (12)

𝐸 𝑠| 1 𝛿 ∙𝐴 ∙𝐵 ∙ 𝑃 𝑛 𝑃 𝑛 (13)

where a and b represent classes of Map A and Map B respectively, c is total

number of class categories, ns represents the number of voxels in shell s, and δab is the
Kronecker-delta of class a and b that returns a value of 1 if the classes are equal or 0 if
they are unequal.

Expected Shell Similarity for the central shell E(s|s=o) – is calculated as the
probability that the central voxels of the 3D windows from Map A and Map B have the
same class category, such as:

𝐸 𝑠| 𝐴 ∙𝐵 (14)

where c represents the total number of class categories.

70
 Expected 3D Similarity Pe3D – is the sum of the products of the Expected Shell
Similarity E(s) and the distance decay function M(di), for each shell s, either s≥1 or s=0,
of a 3D voxel Window from each map, calculated as:

𝑃 𝐸 𝑠 ∙𝑀 𝑑 (15)

where N represents the maximum shell number and ds represents the distance of
shell s from the central voxel Vi,j,k.

With all components determined, the K3DFuzzy can be calculated using Equation (4)
and has an upper limit of one while the lower limit can be less than negative one as it is
dependent on the user defined functions and variables, such as the distance decay
function and similarities of class categories. Values above zero represent a better than
random agreement, while values below zero represent a lower similarity than random
agreement.

4D Fuzzy Kappa

4D Fuzzy Kappa (K4DFuzzy) coefficient is developed to account for changes in 4D

thus incorporates time t to the three spatial dimensions and is based on the comparison
of series of 3D maps over time. K4DFuzzy can be formulated as:

𝑃 𝑃
𝛫 (16)
1 𝑃

where Observed Similarity Po4D and Expected Similarity Pe4D have been modified
from Po3D and Pe3D to account for comparisons of time series of 3D maps. In Po4D, the
Similarity of 4D Distance D4D(a,b) has been adjusted from D(a,b) to include Temporal
Distance T(d,Δt,o) as input to M(d).

Temporal Distance T(d,Δt,o) – is the distance between two voxels based on the
difference in time steps Δt. For every time step the shell number is increased by the user
defined variable time offset (o):

71
 𝑇 𝑑, ∆𝑡, 𝑜 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 𝑆ℎ𝑒𝑙𝑙 𝑑 |∆𝑡| ∗ 𝑜 (17)

where d represents the 3D Euclidean distance of the voxel from the central voxel
based on equation (5), Shell (d) returns the shell s with the distance d, and distance (s)
returns the distance of shell s.

Similarity of 4D Distance D4D(a,b) – has been expanded from equation (7) to

account for the difference in time steps Δt and the offset o by including a Temporal
Distance T(d,Δt,o) to account for voxels at different time steps such as:

𝑀 𝑇 𝑑 𝑎, 𝑏 , ∆𝑡, 𝑜 , 𝑓𝑜𝑟 𝑇 𝑑 𝑎, 𝑏 , ∆𝑡, 𝑜 𝐿

𝐷 𝑎, 𝑏 (18)
0, 𝑓𝑜𝑟 𝑇 𝑑 𝑎, 𝑏 , ∆𝑡, 𝑜 𝐿

Pe3D is modified into Pe4D to include the voxels of neighboring time steps (ts) of
the 4D space-time window that accounts for time steps (w,w,w,ts) into the shells s for the
number of voxels n in the Expected Shell Similarity for outer shells (E(s|s≥1)). The voxels
to include are determined by the difference in time steps Δt and a user-defined time
offset o where the shell s of a voxel is increased by the time offset (o) for each step in
time away from the central voxel. Modifying the shell number allows the calculation of
distance to scale from the central voxel. This means that voxels further from the center
will be affected less than those that are closer. For example, the difference between
Shell 1 and Shell 2 is greater than the difference between Shell 2 and Shell 3. Figure 4.3
presents the sliced 4D voxel window with the time offset o=2. For the 4D window defined
as 5x5x5x3, the number of voxels in each shell will vary, so that Shell 3 containing n=20
voxels would include 6 voxels at t-1, 8 voxels at t, and 6 voxels at t+1. The voxels from t-
1 and t+1 would be part of Shell 1 in a 3D voxel window of each 3D map in time series.

72
Figure 4.3 Shell numbers in the sliced space-time 4D voxel window of 5×5×5×3
size with time steps ts = 3 and time offset o = 2 for t-1, t, and t + 1.

73
4.4. Implementation

The proposed 3D Kappa coefficients are novel and have not been implemented
in any common GIS or RS software package. For the purposes of testing and
demonstration, the proposed methodology of comparisons of 3D maps and calculations
of the 3D and 4D Kappa coefficients have been implemented using Python version 2.7
(The Python Software Foundation 2020). The overall processing time of 3D Fuzzy
Kappa varied between 4 minute and 15 seconds (Asus GL502VMK laptop with an Intel
i7-7700HQ, 16GB of RAM) and 1 minute and 53 seconds (Dell Precision Tower 7810
workstation with an Intel Xeon E5-2526v4 and 64GB of RAM) depending on the
processor and the available RAM on two computers, using a 50x50x50 voxel study area.
The processing time of 4D Fuzzy Kappa is longer and varied depending on the number
of temporal iterations and the overall number of voxels contained in the study area.

The preparation of spatial data to create 3D maps and time series was
accomplished with the ArcGIS 10.5 (ESRI 2018) software. Two equal-sized voxel
datasets are stored in two 4D lists where 3D voxel datasets can be represented using
only one timestep. The program accepts both 3D and 4D datasets and returns the
values for accuracy, 3D Kappa, 3D Fuzzy Kappa or 4D Fuzzy Kappa coefficients. The
size of the 3D and 4D datasets are limited by processing time and available computer
memory, and due to computational limitations, this research study used only a maximum
of 9x9x9x9 3D or 4D voxel window. While the Expected Similarity Pe uses values
measured from the dataset, Observed Similarity Po requires multithreading to iterate the
3D or 4D voxel windows over the datasets in a reasonable time.

To implement the proposed methodology and test different 3D map comparisons,

various 3D datasets have been used ranging from hypothetical data, measured LiDAR
data of different trees, soil horizon data as well as agent-based simulation outcomes for
forest fire smoke propagation, all for 3D and 4D cases. Firstly, hypothetical datasets
were created representing different 3D maps with simplified voxel patterns. Table 4.2
lists all 13 hypothetical 3D maps with for two and four classes and with 3D spatial
patterns ranging from identical, smaller to significantly different, and also uniform,
random, and clustered. These comparisons were conducted with a maximum shell size
of 3, which creates a 3D voxel window that is a cube with dimensions 3x3x3 around the
central voxel.

74
Table 4.2 Hypothetical 3D maps with two and four classes and their
characteristics.

Case ID 3D Map A 3D Map B Characteristics

1 Identical 3D maps
containing two classes

2 Moderately different 3D
patterns of two classes

3 Different 3D patterns of
two classes

4 Simple 3D pattern and

randomly generated 3D
pattern of two classes

5 Randomly generated 3D
patterns of two classes

75
Case ID 3D Map A 3D Map B Characteristics
6 Completely different 3D
maps containing two
inverted classes

7 Uniform distribution with

offset by one voxel of
two classes in a
checkerboard 3D pattern

8 Identical 3D maps
containing four classes

9 Moderately different 3D
maps features
containing four classes

10 Different 3D maps
features containing four
classes

76
 Case ID 3D Map A 3D Map B Characteristics
11 Considerably different
3D map features of four
classes at various
locations

12 Randomly generated 3D
patterns of four classes

13 Uniform distribution with

offset by one voxel of
four classes in a
checkerboard 3D pattern

Secondly, LiDAR data was acquired from the City of Surrey Open Data
Catalogue (Surrey 2019) and voxelized using Esri’s ArcScene (ESRI 2018) for six lone
trees located on a golf course in the southeast of Surrey, BC to implement the
developed methodology on actual data. The voxel size used for these comparisons is
0.50 m to ensure feasible performance of the 32-bit Python software. Table 4.3 presents
the 3D map for each lone tree along with the tree types and tree heights. A 3D voxel
window s=3 was also used for these comparisons.

77
Table 4.3 Voxelized LiDAR 3D data representing different types and heights of
trees.

Tree ID Type Height 3D Map

C1 Coniferous 10.21 m

C2 Coniferous 9.61 m

D1 Deciduous 9.37 m

D2 Deciduous 9.37 m

T1 Tall Coniferous 22.31 m

T2 Tall Coniferous 19.65 m

78
 The developed methodology was further tested on measured soil horizon data
with specific thickness, published by Khresat (2001), using measurements from Pedon 1
and Pedon 3 as Map A and Map B respectively. Voxel datasets used a study area extent
of 50x50x32 voxels with voxel resolution of 5cm. The comparison between soil Maps A
and B was repeated for window sizes s=0 to s=6 to investigate how the window size
affects the resulting 3D Fuzzy Kappa value. Table 4.4 presents the depth and thickness
of each soil horizon along with the voxel maps of both pedons.

Table 4.4 Two soil horizon 3D maps with depths and thicknesses.

Soil horizon Pedon Map A Pedon Map B

types
Depth Thickness Depth Thickness
Ap 0-20 cm 20cm 0-20 cm 20cm
Bt 20-65 cm 45cm 20-50 cm 30cm
Btk1 65-90 cm 25cm 50-100 cm 50cm
Btk2 90-160 cm 70cm 100-160 cm 60cm

3D Maps

Finally, 3D agent-based model (ABM) simulation outcomes have been used to

implement the methodology. A previously developed smoke propagation model (Smith
and Dragićević 2019) has been used to implement the proposed coefficients on
simulation outcomes used for 3D comparisons. A specific study area was selected in
Metro Vancouver Region, Canada to run the 3D ABM for a hypothetical fire that lasted
15 days. For the purpose of comparison, simulation outcomes have been used to
represent 15 hours of smoke propagation starting from 12th day and 6 hrs. This
corresponds to six-time steps starting from 6am to 9pm. The simulations outcome
comparisons were based on 1km3 voxel resolution with total study area extent of

79
44x32x10km3 and with the southwest corner approximately corresponding to the location
of Tsawwassen Ferry Terminal in the Metro Vancouver Region. Voxels were classified in
five categories based on Air Quality Health Index (AQHI) classification values
(Government of British Columbia 2020), and based on the amount of smoke agents
simulated for each voxel.

Three simulation outcomes have been generated for calculation of 4D Fuzzy

Kappa for two comparisons presented in Figure 4.4 primarily to demonstrate the
dissimilar and similar 3D map space-time series. The first comparison is between
Simulation 1 (Map A) that represents the smoke generated from a static fire, and
Simulation 2 (Map B) that represents moving fire influenced by the wind. The second
comparison is between Simulation 1 (Map A) and Simulation 3 (Map B) of the same
initial conditions but depicting the differences due to the randomness within the model.
The size of the 3D voxel window is tested between 0 and 14 shells similar to the process
in the soil horizon comparison, but to investigate the effects of window size on 4D Fuzzy
Kappa.

Figure 4.4 Voxelized model outputs at t = 3 for the three scenarios of forest-fire
smoke geosimulations.

80
4.5. Results and Discussion

The results from the comparisons of the various 3D maps have been presented
showing values for accuracy, 3D Kappa, K3D, 3D Fuzzy Kappa, K3D Fuzzy, and 4D Fuzzy
Kappa, K4D Fuzzy, where appropriate. In two case studies the values for K3D Fuzzy and K4D
Fuzzy have been derived for different window sizes to see the test for the sensitivity
analysis and how it effects the values. It is expected that as more shells are included in
the 3D voxel window, the number of voxels increases for both Po and Pe however, the
similarity of 3D distance decreases for each shell. The effects of Po and Pe on the K3D
Fuzzy value are dependent on the composition and patterns of voxel classes of the 3D
maps being compared. With larger 3D voxel windows these effects will become
negligible and K3D Fuzzy will become stable. This approach can assist in a way to
determine the optimal 3D voxel window size to reduce the level of user bias in the 3D
map comparison.

The obtained results for hypothetical datasets (Table 4.2) are presented in Table
4.5. These tests were run to ensure the methods and program operate as expected and
to inspect the results of controlled case studies. The identical maps in Case IDs 1 and 8
show that the methods correctly calculate a perfect value for all indices, while the
comparison of similar patterns in Case IDs 2, 3, 9, 10, and 11 result in a lower K3D Fuzzy.
The accuracy in these five cases are very similar while the K3D shows less variability
than K3D Fuzzy. The most extreme example is between Case IDs 10 and 11, where the
accuracy and K3D are identical while the K3DFuzzy shows there are differences in similarity.
The K3DFuzzy value for Case ID 4 comparing a simple dataset to a random dataset is lower
than initially expected, it is due to two causes. The first is due to the large area pattern of
the same class in Map A containing no nearby voxels of the same class to the central
cell from Map B. For half the voxels of Map B this creates a low One-Way 3D Similarity,
forcing a low Two-Way 3D Similarity, and in turn lowers the Observed 3D Similarity. The
second cause is the presence of only two classes causing a very high Pe which in turn
causes a low K3DFuzzy. This effect is seen across all cases with two classes and adding
more classes will reduce this effect as seen in the cases with four classes. The random
datasets in Case IDs 5 and 12 confirm that comparing two random dataset results in a
value of zero, showing that the methods fully account for the randomness in the 4D
maps. Case ID 6 demonstrates two maps containing inverted classes thus are

81
completely different, and results are indicating extremely low values for all three
measures. The results from Case IDs 7 and 13 with the evenly distributed classes
resulted in much lower values than initially expected due to the visual similarity of the
two maps. However, a similar 2D map with the same uniform pattern of two classes was
calculated with KFuzzy and had similar results. This is caused by a high (>0.8) Pe and a
moderate (0.50) Po in both examples.

Table 4.5 Results from the comparisons of hypothetical 3D voxel data.

Case ID Accuracy K3D K3D Fuzzy

1 1.00 1.00 1.00
2 0.79 0.58 -0.41
3 0.72 0.44 -0.86
4 0.50 0.00 -2.21
5 0.50 0.00 0.00
6 0.00 -1.00 -5.41
7 0.00 -1.00 -0.98
8 1.00 1.00 1.00
9 0.86 0.82 0.74
10 0.78 0.70 0.41
11 0.78 0.70 0.27
12 0.25 0.00 0.00
13 0.00 -0.33 -0.20

The datasets representing trees (Table 4.3) have been used for all possible
comparisons among the different type of trees. Table 4.6 presents the obtained values
for accuracy, K3D and K3D Fuzzy respectively. The comparisons between trees of the same
types, such as coniferous or deciduous, resulted in the only positive K3D values, showing
that they are the most similar 3D maps. In the trees of different types, the height of the
trees seemingly played a large factor with the smaller height differences leading to
generally more similar K3D Fuzzy values. Similar to the hypothetical 3D map comparisons,
the tree 3D maps only contained two classes and resulted in overall low values, further
enhancing the concept of fewer classes leading to lower K3D Fuzzy values.

82
Table 4.6 Results from the comparisons of 3D tree maps.

Accuracy
Trees C1 C2 D1 D2 T1 T2
C1 1 0.85 0.68 0.69 0.68 0.72
C2 1 0.78 0.74 0.65 0.71
D1 1 0.87 0.63 0.66
D2 1 0.63 0.66
T1 1 0.84
T2 1

K3D
Trees C1 C2 D1 D2 T1 T2
C1 1 0.68 0.39 0.30 0.21 0.31
C2 1 0.48 0.36 0.13 0.27
D1 1 0.70 0.08 0.15
D2 1 0.06 0.12
T1 1 0.63
T2 1

K3D Fuzzy
Trees C1 C2 D1 D2 T1 T2
C1 1 0.46 -0.58 -0.94 -1.20 -0.86
C2 1 -0.26 -0.68 -1.37 -0.95
D1 1 0.60 -1.43 -1.22
D2 1 -1.20 -1.03
T1 1 0.14
T2 1

The comparison of the two soil horizon datasets (Table 4.4) resulted in a 3D
Accuracy of 0.84 and a K3D value of 0.79. The values for K3D Fuzzy, presented in Table
4.7, becomes stable with the value of 0.55 when the 3D voxel window size was set to 4.
This can be partially caused by the parallel layers of the soil horizons. Shells 2 and 3 do
not reach further in any axis than Shell 1, meaning the observed similarity Po does not
increase while the expected similarity Pe does increase. Shell 4 does extend the 3D
voxel window further in the x, y, and z axes, increasing the observed accuracy. The
impact of each additional shell after Shell 4 is too low to have an impact on the K3D Fuzzy
value in this comparison.

83
Table 4.7 3D Fuzzy Kappa values for 3D soil data by 3D Voxel Window Size.

3D Voxel Window 0 1 2 3 4 5 6
Size [shell]
K3D Fuzzy 0.79 0.62 0.40 0.37 0.55 0.55 0.55

The comparisons between the forest-fire smoke model outputs (Figure 4.4) were
conducted for various 3D voxel window sizes to further test the sensitivity of the output
results. Comparison between Scenario 1 and 2 resulted in an accuracy of 0.90 and a K3D
of 0.21, and comparison between Scenario 1 and 3 resulted in an accuracy of 0.94 and
a K3D of 0.65. The results for K4D Fuzzy values are presented in Table 4.8. The comparison
between Simulations 1 and 2 reaches a somewhat stable value close to Shell 11, only
increasing slowly afterwards, while the comparison between Simulations 1 and 3
reaches a stable value around Shell 7. Most of the similarity between the voxels with
smoke in the first comparison is from the upper right corner of the map where smoke has
collected in both simulations. The second comparison is much more similar than the first
simulation due to having the same initial conditions in both simulations. Both
comparisons have large amounts of voxels with no smoke causing the high accuracy
while the K4D Fuzzy shows the similarity much better. By including the time component of
K4D Fuzzy, small temporal differences between 4D datasets can be accounted for in a
similar way to spatial differences in K3D Fuzzy. This has the benefit of two events close in
time at a similar location being considered somewhat similar instead of dissimilar if time
was not included.

Table 4.8 Values for 4D Fuzzy Kappa obtained for comparisons of forest-fire
smoke simulation outputs.

4D Fuzzy Kappa by 3D Voxel Window Size [shell]

Simulations

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

1,2 0.21 0.22 0.19 0.18 0.24 0.19 0.20 0.32 0.35 0.37 0.40 0.41 0.42 0.42 0.43

1,3 0.65 0.82 0.82 0.84 0.86 0.86 0.87 0.88 0.88 0.88 0.88 0.88 0.88 0.88 0.88

84
4.6. General Discussion

Overall, the 3D and 4D Fuzzy Kappa results in different values than K3D but they
have similar patterns and are promising options for the use in 3D and 4D map
comparison. However, they suffer from edge effects in a similar way as the 2D
counterpart KFuzzy. This edge effect occurs because voxels at and near the edge of the
dataset have fewer nearby voxels in their 4D voxel window. This decreases the
probability of a matching voxel being in the window, lowering the observed similarity.
However, this is not accounted for in the expected similarity and with larger study area
extents or 3D Map sizes the edge effect becomes less. Accounting for the edge effect
can be operationalized by calculating the expected similarity for every voxel and finding
the average as described by Hagen (2003). Implementing this solution to the edge effect
could further enhance the ability of the proposed methodology.

Some aspects of K3D Fuzzy and K4D Fuzzy such as 3D or 4D voxel window size and
category similarity are controlled by the user and can significantly affect the results. For
example, in the 3D soil horizons case study, if a window size of s=3 was selected, the
result would be very low compared to the other window sizes. Sensitivity to both, window
size and category similarity, are not new to the Fuzzy Kappa concept and are present in
various other 2D map comparisons and analysis, however the temporal component of
fuzziness is relatively new. A direct conversion between 3D space and time is difficult
and will depend on the nature of 3D data, case studies, and interpretation by different
users. If the 4D data has a very fine temporal resolution then a small time offset should
be used, but with coarse temporal resolution, a large time offset will be more
appropriate. This offset must be determined by the user for the specific scenario,
however a default value of one can be used as a starting point where each step in time
correlates to a voxel being considered one shell further away. The user also needs to
consider the furthest in time they want the shells to reach and set the time offset and
shell size to meet that requirement. Testing of the user defined values for a specific
scenario may be beneficial prior to final implementation.

While acquiring measured 4D data can be difficult, modeling in 4D can provide

an alternative. Without a suite of map comparison methods similar to what is available
for 2D, 3D and 4D models are lacking in model evaluation, calibration and validation
techniques. The proposed 3D and 4D Fuzzy Kappa approaches are just an initial

85
research study that can fill this gap in the needs of comparing 3D maps and simulation
outcomes accounting for chance agreement, fuzzy space, time and voxel category.
While more study may be needed to determine thresholds for levels of similarity as those
existing for Cohen’s Kappa, relative comparisons of similarity should be done between
outputs of a model or outputs of multiple models for calculation of 3D Fuzzy Kappa. The
3D Fuzzy Kappa and 4D Fuzzy Kappa values can be used to calibrate or validate a
model by comparing outputs under varying initial conditions until an optimal value is
reached. Practical uses of 3D Fuzzy Kappa and 4D Fuzzy Kappa could be to assist in
the comparisons of study areas or landscape 3D features in geology, geomorphology,
hydrology, or soil science to determine the best course of action depending on similarity
to past study areas. Additionally, the comparison of various machine vision and medical
imaging 3D techniques against K3D Fuzzy, and other map comparison methods extended
into 3D, may provide valuable feedback on the usefulness of each method for GIScience
purposes.

4.7. Conclusions

The main objective of this research study is to develop and implement a new
suite of 3D methods that allow comparisons of 3D and 4D maps. The developed
program routines and the choice of various 3D data sets and case studies allowed for
various tests to be conducted to demonstrate the usefulness of proposed approaches.
Some methods, such as 3D Accuracy or 3D Kappa coefficients, can be applied to 3D
and 4D space-time data, however they are by their nature aspatial methods. To account
for spatial properties in 2D maps, windows and fuzziness have been used, and these
methods are developed further into 3D and 4D Fuzzy Kappa coefficients. By including
fuzzy location, time, and voxel class, many aspects of a 3D map can be included
according to the user’s specifications. This allows for comparisons of different types of
similarity and dissimilarity, for example where one model may be similar to reality in
location but not time, and another model may be similar to reality in time but not location.
There is a potential to modify the proposed methods to account for additional factors,
such as directional similarity and the edge effect. Incorporating spatial autocorrelation in
3D is a possible extension of the 3D Fuzzy measures, similar to the extended 2D Fuzzy
Kappa (Hagen-Zanker 2009). This would solve the bias towards lower values in 3D
maps with stronger spatial autocorrelation. In summary, the proposed suite of 3D

86
approaches is fully operational and build upon initial research efforts. This research
supports the sustained need for methods that incorporate 3D and 4D data used in the
3D space-time domain for comparison of data, patterns and model simulation outcomes.

4.8. References

Ahmed, B., Ahmed, R., and Zhu, X. (2013). “Evaluation of model validation techniques in
land cover dynamics.” ISPRS International Journal of Geo-Information, 2(3),
577–597.

Appleman, H. S. (1960). “A Fallacy in the Use of Skill Scores.” Bulletin of the American
Meteorological Society, 41(2), 64–67.

Aurbach, A., Schmid, B., Liechti, F., Chokani, N., and Abhari, R. (2018). “Complex
behaviour in complex terrain - Modelling bird migration in a high resolution wind
field across mountainous terrain to simulate observed patterns.” Journal of
Theoretical Biology, 454, 126–138.

Balzter, H., Talmon, E., Wagner, W., Gaveau, D., Plummer, S., Yu, J. J., Quegan, S.,
Davidson, M., Toan, T. Le, Gluck, M., Shvidenko, A., Nilsson, S., Tansey, K.,
Luckman, A., and Schmullius, C. (2002). “Accuracy assessment of a large-scale
forest cover map of central siberia from synthetic aperture radar.” Canadian
Journal of Remote Sensing, 28(6), 719–737.

Beland, M., Parker, G., Sparrow, B., Harding, D., Chasmer, L., Phinn, S., Antonarakis,
A., and Strahler, A. (2019). “On promoting the use of lidar systems in forest
ecosystem research.” Forest Ecology and Management, 450, 117484.

Beni, L. H., Mostafavi, M. A., Pouliot, J., and Gavrilova, M. (2011). “Toward 3D spatial
dynamic field simulation within GIS using kinetic Voronoi diagram and Delaunay
tetrahedralization.” International Journal of Geographical Information Science,
25(1), 25–80.

Bova, M. J., Ciarallo, F. W., and Hill, R. R. (2016). “Development of an agent-based

model for the secondary threat resulting from a ballistic impact event.” Journal of
Simulation, 10(1), 24–35.

Brown, D. G., Page, S., Riolo, R., Zellner, M., and Rand, W. (2005). “Path dependence
and the validation of agent-based spatial models of land use.” International
Journal of Geographical Information Science, 19(2), 153–174.

Chen, Z., Xu, B., and Devereux, B. (2014). “Urban landscape pattern analysis based on
3D landscape models.” Applied Geography, 55, 82–91.

Cohen, J. (1960). “A Coefficient of Agreement for Nominal Scales.” Educational and

Psychological Measurement, 20(1), 37–46.

Cohen, J. (1968). “Weighted Kappa: Nominal scale agreement provision for scaled
disagreement or partial credit.” Psychological Bulletin, 70(4), 213–220.

87
Congalton, R. G. (1991). “A review of assessing the accuracy of classifications of
remotely sensed data.” Remote Sensing of Environment, 37(1), 35–46.

Congalton, R. G., and Green, K. (1993). “A Practical Look at the Sources of Confusion in
Error Matrix Generation.” Photogrammetric Engineering & Remote Sensing,
59(5), 641–644.

Congalton, R. G., and Mead, R. A. (1986). “A Review of Three Discrete Multivariate

Analysis Techniques Used in Assessing the Accuracy of Remotely Sensed Data
from Error Matrices.” IEEE Transactions on Geoscience and Remote Sensing,
GE-24(1), 169–174.

Cui, J., Chow, Y., and Zhang, M. (2011). “A Voxel-based Octree Construction Approach
for Procedural Cave Generation.” Journal of Computer Science, 11(6), 160–168.

Ding, W. J., Wang, R. Q., Wu, D. Q., and Liu, J. (2013). “Cellular automata model as an
intuitive approach to simulate complex land-use changes: An evaluation of two
multi-state land-use models in the Yellow River Delta.” Stochastic Environmental
Research and Risk Assessment, 27(4), 899–907.

Dou, W., Ren, Y., Wu, Q., Ruan, S., Chen, Y., Bloyet, D., and Constans, J. M. (2007).
“Fuzzy kappa for the agreement measure of fuzzy classifications.”
Neurocomputing, 70(4–6), 726–734.

Eshraghi, M., Felicelli, S. D., and Jelinek, B. (2012). “Three dimensional simulation of
solutal dendrite growth using lattice Boltzmann and cellular automaton methods.”
Journal of Crystal Growth, 354(1), 129–134.

ESRI. (2018). “ArcGIS Desktop: Release 10.5.” Environmental System Research

Institute, Redlands, CA.

Fleiss, J. L. (1971). “Measuring Nominal Scale Agreement Among Many Raters.”

Psycological Bulletin, 76(5), 378–382.

Foody, G. M. (2002). “Status of land cover classification accuracy assessment.” Remote

Sensing of Environment, 80, 185–201.

Foody, G. M. (2007). “Map comparison in GIS.” Progress in Physical Geography, 31(4),

439–445.

Foody, G. M. (2020). “Explaining the unsuitability of the kappa coefficient in the

assessment and comparison of the accuracy of thematic maps obtained by
image classification.” Remote Sensing of Environment, 239, 111630.

Fritz, H. C., Wittfeld, K., Schmidt, C. O., Domin, M., Grabe, H. J., Hegenscheid, K.,
Hosten, N., and Lotze, M. (2014). “Current smoking and reduced gray matter
volume - A voxel-based morphometry study.” Neuropsychopharmacology,
39(11), 2594–2600.

88
Van Genderen, J. L., Lock, B. F., and Vass, P. A. (1978). “Remote Sensing: Statistical
Testing of Thematic Map Accuracy.” Remote Sensing of Environment, 7(1), 3–
14.

Gobron, S., Çöltekin, A., Bonafos, H., and Thalmann, D. (2011). “GPGPU computation
and visualization of three-dimensional cellular automata.” Visual Computer,
27(1), 67–81.

Government of British Columbia. (2020). “BC Air Quality.”

<http://www.env.gov.bc.ca/epd/bcairquality/data/aqhi-table.html> (Jan. 31, 2020).

Greene, N. (1989). “Voxel space automata: Modeling with stochastic growth processes
in voxel space.” Computer Graphics, 23(3), 175–184.

Grell, G. A., Peckham, S. E., Schmitz, R., McKeen, S. A., Frost, G., Skamarock, W. C.,
and Eder, B. (2005). “Fully coupled ‘online’ chemistry within the WRF model.”
Atmospheric Environment, 39(37), 6957–6975.

Hagen-Zanker, A. (2009). “An improved Fuzzy Kappa statistic that accounts for spatial
autocorrelation.” International Journal of Geographical Information Science,
23(1), 61–73.

Hagen, A. (2003). “Fuzzy set approach to assessing similarity of categorical maps.”

International Journal of Geographical Information Science, 17(3), 235–249.

Heidke, P. (1926). “Berechnung des Erfolges und der Gute der Windstarkevorhersagen
im Sturmwarnungsdienst.” Geografiska Annaler, 8, 301–349.

Hosoi, F., Nakai, Y., and Omasa, K. (2013). “3-D voxel-based solid modeling of a broad-
leaved tree for accurate volume estimation using portable scanning lidar.” ISPRS
Journal of Photogrammetry and Remote Sensing, 82, 41–48.

Hu, S., and Wang, L. (2013). “Automated urban land-use classification with remote
sensing.” International Journal of Remote Sensing, 34(3), 790–803.

Hudson, W. D. (1987). “Evaluation of several classification schemes for mapping forest

cover types in Michigan.” International Journal of Remote Sensing, 8(12), 1785–
1796.

Jjumba, A., and Dragićević, S. (2015). “Integrating GIS-Based Geo-Atom Theory and
Voxel Automata to Simulate the Dispersal of Airborne Pollutants.” Transactions in
GIS, 19(4), 582–603.

Jjumba, A., and Dragićević, S. (2016a). “Spatial indices for measuring three-dimensional
patterns in a voxel-based space.” Journal of Geographical Systems, 18(3), 183–
204.

Jjumba, A., and Dragićević, S. (2016b). “Towards a voxel-based geographic automata

for the simulation of geospatial processes.” ISPRS Journal of Photogrammetry
and Remote Sensing, 117, 206–216.

89
Jung, Y., and Joo, M. (2011). “Building information modelling ( BIM ) framework for
practical implementation.” Automation in Construction, 20(2), 126–133.

Kedron, P., Zhao, Y., and Frazier, A. E. (2019). “Three dimensional (3D) spatial metrics
for objects.” Landscape Ecology, 34(9), 2123–2132.

Khresat, S. A. (2001). “Calcic horizon distribution and soil classification in selected soils
of north-western jordan.” Journal of Arid Environments, 47(2), 145–152.

Koziatek, O., and Dragicevic, S. (2017). “iCity 3D : A geosimualtion method and tool for
three-dimensional modeling of vertical urban development.” Landscape and
Urban Planning, 167, 356–367.

Koziatek, O., and Dragićević, S. (2019). “A local and regional spatial index for measuring
three-dimensional urban compactness growth.” Environment and Planning B:
Urban Analytics and City Science, 46(1), 143–164.

Lavigne, C., Ricci, B., Franck, P., and Senoussi, R. (2010). “Spatial analyses of
ecological count data: A density map comparison approach.” Basic and Applied
Ecology, 11(8), 734–742.

Li, S., Colson, V., Lejeune, P., Speybroeck, N., and Vanwambeke, S. O. (2015). “Agent-
based modelling of the spatial pattern of leisure visitation in forests: A case study
in Wallonia, south Belgium.” Environmental Modelling & Software, 71, 111–125.

Monserud, R. A., and Leemans, R. (1992). “Comparing global vegetation maps with the
Kappa statistic.” Ecological Modelling, 62(4), 275–293.

Mulia, R., Dupraz, C., and van Noordwijk, M. (2010). “Reconciling root plasticity and
architectural ground rules in tree root growth models with voxel automata.” Plant
and Soil, 337(1), 77–92.

Muller, J.-C. (1975). “Associations in Choropleth Map Comparison.” Annals of the

Association of American Geographers, 65(3), 403–413.

Narteau, C., Zhang, D., Rozier, O., and Claudin, P. (2009). “Setting the length and time
scales of a cellular automaton dune model from the analysis of superimposed
bed forms.” Journal of Geophysical Research: Earth Surface, 114(3), 1–18.

Nourian, P., Gonçalves, R., Zlatanova, S., Ohori, K. A., and Vu Vo, A. (2016).
“Voxelization algorithms for geospatial applications: Computational methods for
voxelating spatial datasets of 3D city models containing 3D surface, curve and
point data models.” MethodsX, 3, 69–86.

O’Neill, R. V, Krummel, J. R., Gardner, R. H., Sugihara, G., Jackson, B., Deangelis, D.
L., Milne, B. T., Turner, M. G. I., Zygmunt, B., Christensen, S. W., Dale, V. H.,
and Graham, R. L. (1988). “Indices of landscape pattern.” Landscape Ecology,
1(3), 153–162.

90
Parrott, L., Proulx, R., and Thibert-Plante, X. (2008). “Three-dimensional metrics for the
analysis of spatiotemporal data in ecology.” Ecological Informatics, 3(6), 343–
353.

Pellerin, J., Caumon, G., Julio, C., Mejia-Herrera, P., and Botella, A. (2015). “Elements
for measuring the complexity of 3D structural models: Connectivity and
geometry.” Computers and Geosciences, Elsevier, 76, 130–140.

Poggio, L., and Gimona, A. (2014). “National scale 3D modelling of soil organic carbon
stocks with uncertainty propagation - An example from Scotland.” Geoderma,
232–234, 284–299.

Pontius, R. G., and Millones, M. (2011). “Death to Kappa: Birth of quantity disagreement
and allocation disagreement for accuracy assessment.” International Journal of
Remote Sensing, 32(15), 4407–4429.

Power, C., Simms, A., and White, R. (2001). “Hierarchical fuzzy pattern matching for the
regional comparison of land use maps.” International Journal of Geographical
Information Science, 15(1), 77–100.

Pringle, M. J., Denham, R. J., and Devadas, R. (2012). “International Journal of Applied
Earth Observation and Geoinformation Identification of cropping activity in central
and southern Queensland, Australia, with the aid of MODIS MOD13Q1 imagery.”
International Journal of Applied Earth Observations and Geoinformation, 19,
276–285.

Rogers, D. J., Suk, J. E., and Semenza, J. C. (2014). “Using global maps to predict the
risk of dengue in Europe.” Acta Tropica, 129(1), 1–14.

Samardzic-Petrovic, M., Kovacevic, M., Bajat, B., and Dragicevic, S. (2017). “Machine
Learning Techniques for Modelling Short Term Land-Use Change.” International
Journal of Geo-Information, 6(12), 387.

Segre, E., and Deangeli, C. (1995). “Cellular automaton for realistic modelling of
landslides.” Nonlinear Processes in Geophysics, 2(1), 1–15.

Shi, W., Yang, B., and Li, Q. (2003). “An object-oriented data model for complex objects
in three-dimensional geographical information systems.” International Journal of
Geographical Information Science, 17(5), 411–430.

Shirowzhan, S., Sepasgozar, S. M. E., Li, H., Trinder, J., and Tang, P. (2019).
“Comparative analysis of machine learning and point-based algorithms for
detecting 3D changes in buildings over time using bi-temporal lidar data.”
Automation in Construction, 105, 102841.

Smith, A. K., and Dragićević, S. (2019). “A four-dimensional agent-based model: A case

study of forest-fire smoke propagation.” Transactions in GIS, 23(3), 417–434.

Stehman, S. V. (1999). “Comparing thematic maps based on map value.” International

Journal of Remote Sensing, 20(12), 2347–2366.

91
Surrey, C. of. (2019). “City of Surrey Open Data Catalogue.” <https://data.surrey.ca/>
(Nov. 26, 2019).

The Python Software Foundation. (2020). “Python.” <python.org> (Sep. 29, 2020).

Torrens, P. M. (2014). “High-resolution space-time processes for agents at the built-

human interface of urban earthquakes.” International Journal of Geographical
Information Science, 28(5), 964–986.

Visser, H., and De Nijs, T. (2006). “The map comparison kit.” Environmental Modelling
and Software, 21(3), 346–358.

van Vliet, J., Bregt, A. K., and Hagen-Zanker, A. (2011). “Revisiting Kappa to account for
change in the accuracy assessment of land-use change models.” Ecological
Modelling, 222(8), 1367–1375.

White, R. (2006). “Pattern based map comparisons.” Journal of Geographical Systems,

8, 145–164.

Woodcock, C. E., and Gopal, S. (2010). “Fuzzy set theory and thematic maps : accuracy
assessment and area estimation.” 8816.

Woodhouse, I. H., Nichol, C., Sinclair, P., Jack, J., Morsdorf, F., Malthus, T. J., and
Patenaude, G. (2011). “A multispectral canopy LiDAR demonstrator project.”
IEEE Geoscience and Remote Sensing Letters, 8(5), 839–843.

Zielinski, D. P., Voller, V. R., and Sorensen, P. W. (2018). “A physiologically inspired

agent-based approach to model upstream passage of invasive fish at a lock-and-
dam.” Ecological Modelling, 382, 18–32

92
Chapter 5.

Representing the Southern Resident Killer Whales in

the Salish Sea using a four-dimensional agent-based
model4

5.1. Introduction

Many real-world systems such as the atmosphere, the lithosphere, the

hydrosphere as well as the boundaries between them are dynamic and driven by
complex processes that operate in three spatial dimensions (3D) and time thereby being
perceived as four-dimensional (4D). Marine environments being part of the hydrosphere
are as such complex systems and make up most of the surface of the Earth. They have
come under focus for conservation efforts such as reducing plastic pollution (Eriksen et
al., 2014; Rios et al., 2007) and the impact of fishing activity on biodiversity (Coro et al.,
2021). One species that has attracted attention from environmentalists and the public off
the west coast of North America is the orca (Orcinus orca), a member of the oceanic
dolphin family and known as the killer whale. One type of orca is the resident which
remains in a smaller range than the more migratory types such as the transient.
Southern Resident Killer Whales (SRKW) inhabit the northeast Pacific Ocean in the
waters off British Columbia (BC), Canada as well as the United States of America (USA)
including the Salish Sea between BC and Washington. They are considered endangered
with only 73 individuals in the wild in September 2021, a decline from 78 individuals in
2017 (Joy et al., 2019). The SRKW form three distinct social groups or pods and given
the letter designations of J, K, and L pods (Olson et al., 2018). J pod is the focus of this
research study and has 25 individuals as of 2022 following the recent births of J59 and
K45, bringing the SRKW population total up to 75 wild individuals (Orca Network,
2022a).

The Salish Sea is on the west coast of North America between Vancouver Island
and the mainland, consisting of the Strait of Georgia, the Strait of Juan de Fuca, and the
Puget Sound. Major rivers have outlets in the Salish Sea, such as the Fraser River, that

4
A version of this chapter coauthored with S. Dragićević has been submitted to the journal
Ecological Informatics.

93
are the spawning grounds of the Chinook salmon (Oncorhynchus tshawytscha), the
SRKW’s main prey, composing 79.5% of its diet (Ford et al., 2016). This area
experiences heavy marine vessel traffic from Vancouver and Victoria (Canada) and
Seattle (USA). Their ports generate large volumes of cargo and tanker vessel traffic, with
additional ferries connecting the many islands and coastal cities and towns, fishing
boats, and pleasure craft. The vessel traffic particularly the cargo and tanker vessels
create acoustic disturbances to the SRKW (Veirs et al., 2016).

The SRKW rely on sounds for many aspects of their life from communication to
hunting, and nearby vessels generate noise from their engines, propellers, and active
sonar that can interfere with their hearing (Joy et al., 2019). SRKW have three main
types of sounds; clicks for echolocation, whistles, and pulsed calls for communication
such as during socialization or travelling (Ford, 1989). While the full extent of the impact
of exposure to vessel noise is not known, it is logical to expect that exposure decreases
the ability of the SRWK to hunt and forage for Chinook salmon as vessel noise overlaps
with the clicks SRKW use for echolocation (Veirs et al., 2016). This disruption may also
cause more energy expenditure due to a lower rate of successfully catching their prey.
The impact of noise exposure affects female SRKW more than males, potentially
causing disruptions to the matriarch of a pod (Joy et al., 2019).

Marine animals have been represented in several ways with various modeling
approaches, but these typically do not consider important aspects needed for the SRKW
context. Equation models such as positive mathematical programming (Jansson and
Waldo, 2022) do not include spatial aspects except when applied to multiple regions (He
et al., 2013), and are based on top-down processes on populations as a whole. Spatial
bottom-up models such as cellar automata (Chen et al., 2012) have difficulty
representing individuals and analyzing their interactions (Parker et al., 2003), which is
important given that only few SRKW remain. Modeling methods that have the capability
to address these issues include agent-based models (ABM).

Agent-based models are a type of Geographic Automata System (GAS) (Torrens

and Benenson, 2005) with the explicit ability to represent complex spatial processes,
behavior and interactions of multiple elements, individuals, actors or species in different
contexts including the marine environment. They operate through a bottom-up process
where the local behavior and interactions of individual agents are represented at a fine

94
scale thereby producing emerging patterns at coarser scales based on the interactions
of system elements (Clifford, 2008). ABMs can encode many specific characteristics of
real complex systems such as evolution, emergence, self-organization, adaptation,
feedback loops, and bifurcation (Manson, 2001). The agents representing individuals of
a system are software routines following a set of transition rules that are programmed to
correspond to the real world interactions with each other and the environment (Messina
et al., 2008). At every iteration or step of the model all agents take actions based on their
rule sets and make decisions or move accordingly. These rule sets represent their
individual characteristics and local environment and can change given the actions of
other agents, moving to a new location, or modifying the environment to name a few.
Over many steps of the model, large-scale patterns emerge from the small-scale
changes made by the rule sets (Brown et al., 2005b). Agents of the same type operate
with identical rule sets where their differences in parameters and local environment
change the outcome of their actions. The environment is often based on geospatial
datasets typically represented through a two-dimensional (2D) regular grid of raster
cells, or three-dimensional (3D) voxels or sometimes spatial networks, where agents can
move over time in various ways and interacting through and with the environment thus
making the simulation models to be either 3D or 4D.

Modeling SRKW in 4D with the use of ABM will provide several advantages. The
first is the capability of the ABM to account for individuals of a system instead of a
generalization of their effects (Macatulad and Blanco, 2014). While some generalization
of individual behavior may occur due to computational processing limitations, individuals
can be represented with their own characteristics. The second is although the agents
representing SRKW follow the same rulesets, they have the flexibility to act
independently under the same external conditions based on their individual needs,
similar to role-playing games (Castella et al., 2005). The third is that the agents have
memory, make decisions, learn, and interact as the model progresses and thereby
adjust to changing conditions (Heppenstall et al., 2016). The use of ABM also comes
with challenges (Crooks et al., 2008; Lee et al., 2015) such as intricacy of model
development, demanding computational requirements (Richmond and Romano, 2008),
and the difficulty of model testing (Brown et al., 2005a). These challenges were
encountered and addressed in this research study.

95
 In the research literature, ABMs have been used for many applications including
modeling marine environments and oceanic mammals such as whales. While most
ABMs representing marine mammals are either non-spatial (Filatova and Miller, 2015;
Testa et al., 2012; Whitehead and Ford, 2018) or considered 2D over time (Chudzinska
et al., 2021; Mcloughlin et al., 2018), some are 4D (Anwar et al., 2007; Chion et al.,
2011; Daewel et al., 2008). The non-spatial models can use individuals variables,
processes, and probabilistic interactions to represent systems such as predator-prey
interactions (Testa et al., 2012), genetic mutation and evolution (Whitehead and Ford,
2018), and cultural mutation (Filatova and Miller, 2015) of killer whales. These models
do not include any spatial aspects and as such have no representation of movement or
spatial interactions. While these models can capture changes in populations over long
periods of time, they are not capable of simulating the locations of mammals in the
marine environment.

Spatial models in 2D are capable of representing the movement of agents

around a designated study area, such as the evolution of humpback whale songs within
a population in eastern Australia (Mcloughlin et al., 2018) or foraging locations of harbor
seals near Aberdeen and St. Andrews, Scotland (Chudzinska et al., 2021). By extending
models into 3D space, it allows for the influence of the bathymetry and visibility of
whales at the surface of the water (Anwar et al., 2007; Chion et al., 2011) or the
environmental variability both horizontally and vertically on larval fish (Daewel et al.,
2008). However, the identified 3D ABMs of marine environments are missing important
elements for a predator-prey model of SRKW. For example, there is absence of features
that represent the direct interaction between different agent types representing species,
or interactions with vessels or feedback from feeding. A 3D predator-prey ABM may
benefit from having both predator and prey agents in 3D. Movement is represented for
example, through a hydrodynamic model and the agents do not move using their own
decision-making processes therefore assessing whale movement is not possible. As
such, the main objectives of this research study are to: (a) design and implement a four-
dimensional agent-based model to simulate the spatial dynamics of the Southern
Resident Killer Whales, specifically the J pod, in the Salish Sea, Canada and USA, and
(b) investigate their movements considering bathymetry, vessel noise, and their primary
prey-predator relationship, specifically Chinook salmon. The following sections provide

96
specific information about the data and study area, the developed ABM, the simulation
modeling results and discussion, and conclusions about the implications of the research.

5.2. Methods

5.2.1. Study Area and Datasets

In this study, the Salish Sea (Canada and USA) is defined by the boundaries
identified by the Salish Sea Atlas hosted by the Western Washington University (2022)
and the location was chosen as it is the typical habitat for SRKW and Chinook salmon
(Figure 5.1). The area includes approximately 17,804km2 of marine waterways inland
from Sonora Island (Canada) to the north and Cape Flattery (USA) to the west. It also
includes the cities of Vancouver and Victoria (Canada) and Seattle (USA). Three
datasets were used to represent the marine environment in 3D based on the bathymetry
(Western Washington University, 2022) (Figure 5.1a), distance from shore (Figure 5.1b),
and vessel density (Office for Coastal Management, 2022) (Figure 5.1c). The Salish Sea
bathymetry data from the Salish Sea Atlas (Western Washington University, 2022) has a
maximum depth of 762.8m in Desolation Sound, and an average depth of 126.4m. This
dataset was used to determine the lower boundary of the study area and to derive
distance from shore to allow the navigation of all agent types to keep them in their
preferred regions and to avoid the shore. The vessel density dataset was obtained from
the National Oceanic and Atmospheric Administration’s Office for Coastal Management
(2022) containing the density of ships with Automatic Identification Systems (AIS). The
AIS transponders ping their location which is recorded in 100m grids and is then
converted into a relative density and the absolute vessel density cannot be calculated.
This dataset was used as a proxy for noise from vessels navigating the Salish Sea. The
model was developed in NAD 1983 UTM Zone 10N, the same as the input bathymetry
dataset.

97
Figure 5.1 Study area Salish Sea, Canada and USA, as study area with key
datasets: (A) Bathymetry, (B) Shore Distance, and (C) Vessel
Density.

5.2.2. Agent Based Model of Southern Resident Killer Whales

The developed 4D ABM incorporates the movements, behavior and interactions

of SRKWs as they swim, hunt, and live in the Salish Sea, as well as their main prey, the
Chinook salmon. The model simulates the predator-prey relationship between these two
species as well as the influence of the topography of the Salish Sea and the acoustic
disturbance caused by vessel traffic. A flowchart of the proposed 4D ABM is presented
in Figure 5.2. Information of the model Overview, Design, and Details (ODD) (Grimm et
al., 2010, 2006) are presented in Table 5.1 and details of the 4D ABM are explained in
the following subsections.

98
Figure 5.2 Flowchart of the 4D ABM approach for spatial simulations of
Southern Resident Killer Whales (SRKW).

99
Table 5.1 The Overview, Design, and Details (ODD) of the 4D ABM.

Model Overview
Purpose The purpose of the model is to represent in four dimensions (4D) the movement and
interactions of SRKWs along with Chinook Salmon. The model allows for various
scenarios to forecast the locations of SRKW in the Salish Sea, or any coastal area
where both species exist. The 4D ABM aims to present the representation of
biological complex geospatial systems with agent-based modeling.
Entities, state The model contains two agent types:
variables, and Whale agents (WAs): These agents represent the SRKW as they live in the Salish
scales Sea. They have many variables including but not limited to energy, various speeds,
hunger, stomach fullness, weight, age, sex, and others.
Salmon agents (SAs): These agents represent schools of chinook salmon as they
swim between various locations in the study area. They have several variables: a
swim speed, an evade speed, and school size.
3D movement process: as all agents move in the study and they avoid the ground at
the shore and the sea bottom represented in the 3D bathymetry dataset. As they
approach ground they will begin to either swim over it vertically or around it
horizontally. Vertically, they maintain their heading but follow the sea bottom.
Horizontally they search around them starting from their original heading in 45°
increments for a direction with sufficient depth.
Vessel noise: the WAs have a movement vector added to their movement that points
directly away from the highest vessel density that is above the minimum threshold and
is within range. This discourages the WAs from being in regions identified as having
high vessel density, and therefore high noise.
Shore distance: salmon agents have a preferred distance from shore and have
movement vectors applied to their movement directly away from or towards shore to
encourage them to be within this range.
Stranded: in the event that an agent gets stranded in a location with no visible exit to
open water, or when underground, they return to their last valid location.
Depth: WAs are limited by their breath, being able to dive for 15 minutes. At the end
of this time they begin to resurface. They dive to chase SAs that have a vertical vector
applied to their movement to bring them up when below 400m in the 3D model of the
marine environment. However, the large majority of the study area is shallower than
400m.
Control Modules with several functions:
Main control: defines the starting parameters of the model, including 3D spatial and
temporal scales, as well as the starting 3D locations of the agents
Data control: manages all the data inputs and storage in the model
Time control: manages the temporal extent of the model
The user is able to set variables such as: temporal resolution, study area extents,
spin-up duration, workspace location, agent starting locations, previously learned
locations, and agent parameters.

100
Process The model begins with the main control setting up the 3D environment and agents.
overview and Each time step, the WAs perform the following:
scheduling Decides on and conducts one action:
finish eating: if the whale caught prey on the previous step;
breathe: moves towards the surface and takes a breath;
continue resting: if resting, it will stay at rest until fully rested;
chase prey: if it already has a target prey it will continue chasing it;
hunt prey: if it needs food but doesn’t have a target prey, it will find one and start
chasing it;
rest: if it needs rest and enough of the pod also needs rest, it will rest;
swim: either swim towards a memorized location, swim towards the matriarch, or
swim randomly to represent socializing.
The matriarch WA saves her location if it is suitable, and once per day adds one
saved location to her memory.
Hunger Management:
Reduce stomach content by one
If stomach is empty, lose weight equal to diet requirements per minute;
If stomach is more than half full, gain weight equal to diet requirements per minute;
Current location and time step (x,y,z,t) is output once per hour
Each time step, the SAs choose one action:
Swim away from a whale agent if nearby
Swims randomly if it recently arrived at its destination to represent feeding and
movement is modified for shore distance and depth.
Navigate
Selects a destination if it does not already have a destination
Swim towards selected destination, movement is modified for shore distance and
depth
Design concepts
Emergence The emergence of SRKW locations is represented by the density of WA locations in
3D over time. This is affected by the 3D location and abundance of SAs, and the
density of vessels
Adaptation The WAs adapt to the 3D environment by searching for more salmon, avoiding
running aground, and memorizing locations that have been historically suitable
Interaction WAs interact with each other through activities such as resting together and
swimming with the matriarch. The WAs interact with the SAs through chasing and
eating them. The SAs interact with the WAs by evading them.
Stochasticity Several random processes within the 4D ABM are set for the following situations:
The order in which the agents are processed;
WAs when they are too far or too close to the matriarch with no other actions to do;
SAs when they arrive at their destination;
When the matriarch adds a saved location to memory;
When the matriarch selects from nearby memories;
During the generation of previous memories;
When both directions are equally valid when an agent is avoiding the ground.
Collectives Collectives form in the WAs as they attempt to remain near the matriarch

101
 Calibration The agent parameters were calibrated through physiological traits of both species,
such as swim speeds under various scenarios, dietary requirements, average
weights, etc. as well as the estimated population sizes.
The spin-up length is calibrated to a long enough duration that the starting locations
cannot be identified while also not wasting processing time.
Outputs The whale agents’ locations are output at regular intervals, such as every 60 time
steps representing one hour of reality. The agents can be made to output any of their
variables at this time. The output locations in 3D are in the same coordinate system
as the input bathymetry data.
Details
Initialization Modeling Environment: Repast Simphony 2.8
Language: Java
Projection: NAD 1983 UTM Zone 10N
Origin lower left (x,y,z) [m]: 289484.9, 5160344.5, -500
Voxel size (x,y,z) [m]: 100, 100, 100
Extent (x,y,z) [voxels]: 3830, 4749, 5
Iteration length [minutes]: 1
Output interval [minutes]: 60
Time extent [minutes]: 525600 (365 days)
Number of whale agents: 25
Number of salmon agents: 1400
Input Data Bathymetry data source: Western Washington University Salish Sea Atlas
Link: https://salish-sea-atlas-data-
wwu.hub.arcgis.com/maps/f23310e286614728a823ae84ab865cc1/explore
Extent: clipped to study area
Horizontal resolution [m]: 90
Vertical precision [m]: 1

Distance from shore: calculated from bathymetry data

Vessel density data source: Office for Coastal Management, NOAA Fisheries
Link: https://www.fisheries.noaa.gov/inport/item/48932
Extent: clipped to study area
Horizontal resolution [decimal degrees]: 0.00001 [resampled to 90m]
Units: relative to entire dataset
Year: 2013

Agents

There are two agent types in the 4D ABM (Figure 5.2); Whale Agents (WAs) to
represent individual SRKWs and Salmon Agents (SAs) to represent schools of Chinook
salmon, both in 3D space. Each SA represents one school which can represent any
number of fish travelling together in 3D space depending on the scenario. These two
agent types navigate around the Salish Sea while interacting with their environment and

102
each other. WAs hunt SAs while they manage their hunger and their energy levels. SAs
attempt to evade the WAs and stay within specified distances of the shore.

Whale Agents (WAs)

The WAs possess several basic functions (Figure 5.2); hunt, rest, swim,
socialize, and breathe (Table 5.1), and they represent the functions as close to reality as
computationally reasonable. They hunt, also referred to as foraging, when they get
hungry, rest when they have expended their energy, and swim when they have a
destination to travel to, and when these needs are met, they will socialize (Ford, 1989).
WAs can dive for up to 15.9 minutes and must resurface to breathe (Reisinger et al.,
2015). Hunting consists of the two main actions of chasing and eating. When the WA
has decided it needs food it will find a nearby SA within its echolocation range of
approximately 700m (Simon et al., 2007) and begin to chase it. If the WA is located
within a region of high vessel density its echolocation range is reduced due to the noise
from the nearby ships. Based on the scientific literature describing the whales behavior,
the WA will use its travel speed while it moves around the study area, chase speed
when it finds prey, and will accelerate to a sprint in the final moments (Ford, 1989;
Noren, 2011; Williams and Noren, 2009). If the WA gets within eating distance of its
prey, it will eat one salmon from the school, which takes one additional step of the model
to complete. The salmon increases the fullness of the WA, which reduces its hunger and
desire to hunt. If the WA is full or another need becomes more urgent, the next step after
eating prey it will make a new decision. When not hunting, WAs movements have a
movement vector applied to encourage them to stay away from regions with high vessel
traffic. The amplitude of the vector is not strong enough to prevent them from passing
through, but enough to alter their course.

In the developed 4D ABM, each WA has been characterized with a set of

parameters that describes their movement behavior in the 3D space of the marine
environment that are informed from scientific literature. The WA maximum energy level
is set to 2000 units, because WA expends energy for every one-minute step, one unit for
swimming at a cruising speed, and five units during a sprint while chasing prey, at
speeds informed by the literature (Noren, 2011; Williams and Noren, 2009). As the WA
uses energy the desire to rest increases, however it will only rest if a majority of the pod
also would like to rest as the SRKW tend to rest as a pod (Heimlich-Boran, 1988). When

103
the conditions are met the entire pod will rest the WAs slowly swim (Noren, 2011; Wright
et al., 2017), gaining five units per step until all members are over 1800 units, or another
need is extremely high. These energy values were obtained from the model calibration
procedure starting with 1440 units, one for each minute of the day, and increased to
2000 to where they have enough energy to make it through the day. The need for rest is
calculated by the ratio of current energy to maximum energy and the desire to stop
resting begins at 200 units before the maximum, therefore the very low and very high
values will rarely be achieved. In order to maintain breathing, SRKWs only sleep with
one half of their brain at a time, and will continue to swim (Wright et al., 2017). While
resting, a WA swims at a slower pace and it will not hunt. If the WA decided that it does
not need to hunt or rest, it will swim, where it finds a destination and begins moving
towards it. The destination is dependent on several factors, with the most important
decision maker being the matriarch of the pod, who is the oldest female in the matriline
(Bigg et al., 1990).

The matriarch WA is responsible for the large scale-movement of the pod as the
rest of the pod attempts to remain within a set distance of her while they hunt. The
matriarch beginning with a limited number of memories to represent knowledge learned
before the modeled period of time. Some memories are predefined to ensure coverage
of the study area and known locations SRKW are found, while others are randomly
generated to add flexibility to the model. As the matriarch finds suitable feeding
locations, she will add some of them to her memory. New memories cannot be created
in regions with high ship traffic. In the event of the death of the matriarch, the next oldest
female SA is selected to be the new matriarch and retains all the memories. Typically
other whales attempt to stay close to the matriarch (Miller, 2006), thus the WAs will
attempt to stay close to the matriarch if they are within auditory range for their calls. It is
possible for individuals to be separated from the pod and get stranded as seen in nature
(Foote et al., 2006), and the pod does not directly attempt to return to their last known
location. However, in most cases the pod naturally returns to the area and the stranded
individual can rejoin. During separation, the stranded whale will still hunt, swim in the
area, and rest, but will not find new distant destinations.

104
Salmon Agents (SAs)

The SAs (Figure 5.2) swim using a series of pre-set locations in the study area
similar to the WAs. These points represent regions identified as containing Chinook
salmon based off of spawning locations, migratory routes, fish tracking (Arostegui et al.,
2017a, 2017b; Sato et al., 2021), and fishing locations. Chinook salmon have an
estimated maximum daily travel distance of approximately 70km (Arostegui et al.,
2017b), and so the SAs select their next destination to be within this maximum distance.
They then wait for a set period of time after reaching their location before selecting a
new destination. During this wait time, SA are set to swim around the area according to
their preferences and evading WAs (Table 5.1). While swimming, SAs have a movement
vector applied to their movement that encourages being between 200m and 1700m from
shore as seen in Chinook salmon (Candy and Quinn, 1999), however the SAs are able
to surpass this vector and leave this range to travel longer distances. SAs also have a
depth limit of 400m where they are forced upwards if they try to go below this depth
(Candy and Quinn, 1999), however most of the study area is less than 400m deep.
When WAs come too close, the SA will evade being hunted by swimming directly away
from the nearest WA at high speed (Wightman and Taylor, 1976). When caught by a
WA, the SA subtracts one from the number of salmon it represents and swims away. If
the number of salmon a SA represents reaches zero, the SA is removed and is no
longer represented for the rest of the model simulation. The actions of the salmon are
generalized and therefore their feeding, health, and lifecycle functions are not
represented to simplify the processes due to computational challenges of model
simulations.

Model Computing Environment

The ABM was developed using Repast Simphony 2.8 (North et al., 2013), a
Java-based free and open source toolkit for developing ABMs (Table 5.1). While running
the model, it must first pass through a spin-up procedure that runs the simulations
without WAs for a set period of time to reach equilibrium of SA distribution before the
WAs are added and outputs are generated. The goal of the spin-up is to reduce the
impact of the input bias when selecting the initial start 3D locations for the SAs. Due to
the large number of SAs, it is not reasonable to select different starting location for each
agent, and the spin-up procedure will disperse the SAs around the study area. The
length of the spin-up procedure depends on the number and starting distribution of SAs.

105
For this research study, the length of a spin-up procedure corresponds to 30 days, or
43,200 one-minute model simulation steps. After the spin-up procedure, the WAs are
added and the model can start to generate outputs. The WAs output their locations in 3D
as X, Y, and Z coordinates in the same coordinate system as the input bathymetry
dataset, NAD 1983 UTM Zone 10N. In addition to the 3D coordinates, the WAs output
the current step count to allow for mapping of their 3D locations over time.

5.3. Simulation Results

The simulation outputs of the 4D ABM are series of points recorded as 3D

coordinates where each point represents a location of a WA every 60 time steps,
representing one hour in reality. Figure 5.3 depicts the simulation outputs in 3D for three
outputs in consecutive intervals of the model in the selected subsection of the study
area. Due to the large number of points generated, a total of 219000 for one iteration,
the kernel density function in ArcGIS Desktop (ESRI, 2018) was used to display the
density of locations of the WAs projected from the model outputs in 2D maps. The
density maps are presented in Figure 5.4 for a selected section of the study area
depicting differences among scenarios.

106
Figure 5.3 Locations of whale and salmon agents in 3D for three consecutive
output intervals at time step t=0, t=60 and t=120 and for a selected
section of the study area.

107
Figure 5.4 Simulation outputs for density of locations for whale agents for
Scenarios A, B, C, and D for all iterations of each scenario
representing 365 days.

108
 Four scenarios (A, B, C, D) were developed to represent different situations
about prey abundance. Each scenario consisted of ten iterations that are combined to
make one simulation outcome per scenario. Each iteration conducts a spin-up procedure
simulating 30 days of reality before the main run of 365 days is simulated, or 525600
model steps. SAs are spawned at seven starting locations selected to be spread
throughout the study area, and each SA represents 100 fish. Scenario A spawns 2800
SAs, 400 SAs at each starting location, to simulate high prey abundance (Figure 5.4A).
Scenario B spawns 2450 SAs, 350 at each starting location, to represent moderate prey
abundance (Figure 5.4A). Scenario C spawns 2100 SAs, 300 at each starting location,
to represent low prey abundance (Figure 5.4C). Scenario D introduces the ability to
spawn SAs during the model, instead of all initially at the start (Figure 5.4D). This
scenario begins the same as Scenario C, with 300 SAs spawned at each starting point
before spin-up, and spawns an additional SA every 720 steps, representing 12 hours of
reality, at the mouth of the Fraser River, approximately 20km southwest of Vancouver,
Canada. All scenarios have the WAs enter the study area from the Strait of Juan de
Fuca located between Vancouver Island (Canada) and the Olympic Peninsula in
Washington (USA).

Scenario C also had some WAs die in two of the iterations, almost all WAs died
after around 5 months of simulated time in one iteration, and around half the WAs died
around 7 months of simulated time in the other. Scenario B had a majority of the WAs
die approximate 7 months of simulated time into the model. All three iterations with WA
deaths occurred after they were stranded in Roche Cove and unable to find the exit
before starving. While this was a common location across all the iterations, only these
three resulted in deaths.

The density maps also indicate similarities among scenarios, especially in the
San Juan Islands between Vancouver Island and the mainland (Figure 5.5). This area
has a higher density of WA locations across all scenarios. Other areas with higher
density are around the City of Victoria, the entrance to Puget Sound in Washington, and
near the outer limits of the Gulf Islands in British Columbia. Some regions commonly
have a low density of WAs, including the Strait of Juan de Fuca, the Gulf Islands, and
the Strait of Georgia. Differences are present, with Scenario A having little to no WA
presence deeper in Puget Sound, but Scenario B having WA presence here, and
Scenario A having WA presence in the northern regions of the study area by Cortes

109
Island while Scenario B does not. The WAs in Scenario C are more centralized around
the Strait of Juan de Fuca than Scenarios A and B, with much less presence in the
northern San Juan Islands and in Puget Sound.

Figure 5.5 Simulation outputs for accumulated density of locations for whale
agents for the subsection around the San Juan Islands for
Scenarios A, B, C, and D.

110
5.3.1. Model Testing

There was no data available on the observed locations of the SRKWs to be

suitable for full model testing that allows for comparison between simulation outcomes
and real data. The Orca Network (Orca Network, 2022b) publishes their reported
sighting archive and is the best source available, but that data has several issues for
performing the model testing. First, the data is only the sightings reported to their
website and relies on someone both seeing a whale and reporting it accurately. This
biases the data towards high traffic areas and for whales that are at the surface. Second,
the sightings are sorted by whale type, but if it is not known if an orca is transient or
resident it is included in the resident category. Third, the sighting locations are
generalized and are not precise. It is for these reasons the 3D simulation outputs of the
four scenarios are compared against each other to discover the relative differences
between the scenarios produced by various initial conditions and the constant input of
SAs into the marine environment. The 3D Kappa and 3D Fuzzy Kappa measures (Smith
and Dragićević, 2022) were used for comparisons.

The subsection of the study area (Figure 5.5) was used for the scenario
comparisons due to the computational processing requirements of such large numbers
of 3D locations of WAs. This subsection was selected due to it being a common area in
all scenarios with WA presence identifying it as an important location for this marine
mammal. The values for the 3D Kappa and 3D Fuzzy Kappa measures for comparisons
of four scenarios are presented in Table 5.2. Both metrics were measured using a 100m
voxel resolution and classified into five classes depending on the number of WAs
present in each voxel: class 0 with no WAs present, class 1 with 1-3 WAs present, class
2 with 4-6, class 3 with 7-10, and class 4 with over 10 WAs present.

111
Table 5.2 Results of simulation outcome comparisons for Scenarios A, B, C,
and D.

3D Kappa Scenario A Scenario B Scenario C Scenario D

Scenario A 1.00 0.472 0.409 0.482

Scenario B 1.00 0.416 0.465

Scenario C 1.00 0.418

Scenario D 1.00

3D Fuzzy Kappa Scenario A Scenario B Scenario C Scenario D

Scenario A 1.00 0.491 0.365 0.500

Scenario B 1.00 0.400 0.487

Scenario C 1.00 0.402

Scenario D 1.00

The difference between the four scenarios can be assessed through the values
obtained for 3D Kappa and 3D Fuzzy Kappa. A higher value represents a higher degree
of similarity between two observed simulation outputs. The most similar scenarios are
Scenarios A and D, showing that although the model can start with fewer SAs, the
addition of SAs over time can account for this initial difference. Scenarios A and C are
the most different, which was expected as they have the most different initial conditions.
The loss of WAs further increased the differences between the scenarios in several
ways. With fewer WAs present to generate outputs, there is a lower density of WAs
presented as fewer SAs are consumed, leaving more SAs available for the surviving
WAs, changing their behavior. Additionally, when the matriarch dies, a new matriarch
takes leadership.

5.4. Discussion

The spatial locations of SRKW movement as forecasted with the developed 4D

ABM methodology can be explained through several attributes and properties of the
model when compared to the observed SRKW and Chinook salmon behaviors. The
presence of WAs residing near the coast has several explanations. The first is the SAs
have a distance from shore preference. With the SAs preferring to be 200m to 1700m

112
from shore, the WAs will also be located in this range to feed on the SAs. The second is
most of the marine vessels for shipping and ferries based on the available dataset are in
deeper water further from shore. As the WAs avoid these areas due to the noise, they
generally move towards shore. This can be seen in the Haro Strait, especially around
Saturna Island, where the WAs avoid the center of the strait where high vessel density is
located. The third is since the agents’ navigation does not avoid land until it is close, this
causes them to follow the shoreline and occasionally getting caught in coves and inlets
where they backtrack out to more open water.

There are several overlapping regions between the sightings (Orca Network,
2022b) and the locations forecasted by 4D ABM simulations. These were mainly around
the San Juan Islands, the entrance to Puget Sound, and on the outer edges of the Gulf
Islands. The preference by the WAs for the San Juan Islands in all scenarios can be
explained by the closeness of the islands. This region contains a large area within the
SA’s preferred distance from shore as well as having low vessel density, with the
exception of the ferry route between Friday Harbor on San Juan Island and the
Anacortes Ferry Terminal on mainland Washington. The many islands also create
obstructions, slowing down the SAs and WAs as they navigate between locations and
find ways around them.

The Strait of Juan de Fuca has a straight section of higher WA density in all
scenarios. This is due to all the WAs starting at this location and moving along this path
to enter the study area and is not representative of SRKW movement. Outside of this
section, the strait has lower density of WAs, with higher density along the shore of
Vancouver Island matching a report on SRKW sighting by Fisheries and Oceans
Canada (Ford et al., 2017). The Gulf Islands and the deeper regions of Puget Sound
have almost no WA presence, possibly due to WAs and SAs having difficulty navigating
the channels and neither being a passage between different locations in the study area.

The death of WAs is an important event that also threatens the SRKW
population. In the ABM, the death of WAs is only represented by starvation and the
model does not consider death by vessel collisions, sickness, or other factors. Scenario
A and D had no deaths of WAs, while Scenario B had one iteration in which the WAs
died. Scenario C had two iterations where WAs died, showing that the abundance of
SAs has a major impact on the health of the WAs. While the WAs do not eat all of the

113
SAs, the SAs become less dense, and it takes the WAs longer to find prey. The similarity
between Scenarios A and D, and with both containing no deaths, shows that a constant
replenishment of the SA population is very important to the WA survival, and
connections can be made with reality. The spawning ground of Chinook salmon and
other fish in major rivers such as the Fraser River in Canada, Yukon River in Alaska,
USA, and the Columbia River in Washington, USA, must be protected to ensure the
survival of marine mammals such as the SRKW.

The developed 4D ABM model as a tool to represent the SRKW is beneficial in

the case of forecasting SRKW locations in 3D that could be further translated into marine
protection and conservation actions. Through the representation of individual whales and
their movement, the model uses geographic features to simulate their locations. For
example, while the San Juan Islands and Gulf Islands both seem to be desirable
locations, the WAs shows a clear preference for the San Juan Islands. The main
difference being the amount of time spent travelling through these regions on the way to
other locations. Commonly used top-down modelling methods such as equation models
are unable to represent the movement of individuals and thereby making this type of
spatio-temporal analysis not possible. Another benefit of using 4D ABM is the ability for
agents to learn and change during the simulations and so allows the WAs to learn and
adapt to their environment using a memory instead of their locations and decisions being
predetermined.

There are some challenges associated with the developed 4D ABM approach
and they are typical of ABMs. The SRKW are very intelligent animals with complex
social structures living in an environment inhospitable to humans. This makes them
difficult to study, observe and represent due to the lack of real data and thus their
behavior must be generalized. The SRWK and Orcas in general have complex hunting
strategies. At the time scale in this model, hunting is reduced to being sufficiently near
Chinook salmon. However, Orcas are known to hunt in packs, trapping their prey
(Heimlich-Boran, 1988), and in some locations in the world, Orcas will partially beach
themselves to hunt seals and sea lions in the surf (Guinet and Bouvier, 1995; Lopez and
Lopez, 1985). Consequently, generalization means some specific behavior may not be
adequately represented by the 4D ABM, especially when simulating at much coarser
scales.

114
 Another challenge of using 4D ABMs is the difficulty of performing full model
testing, including model verification, calibration, and validation. This is mainly due to lack
of data on whales in both 2D and 3D, but also being 3D and 4D adds an additional level
of difficulty in terms of available tools and procedures that can evaluate multiple
dimensional model outputs. Development of 3D model testing or 3D Kernel Density
approaches would be useful for the display and spatio-temporal analysis of 4D ABM
outputs.

While some refinement could be made to the developed 4D ABM, they would
likely require additional information on the SRKWs, Chinook salmon, and other species
in the Salish Sea, as well as more real-world data at higher resolutions, including vessel
traffic without AIS and impacts of commercial fishing (Adams et al., 2016). The first
refinement would be to increase the social complexity of the SRKW. This could allow
both more independence of individuals, sub-pod formation, and intra-pod interactions.
The second refinement would be to use the physiological characteristics of the SRKW
and the Chinook salmon to improve biological representations, such as dietary
requirements, energy expended, and life cycle characteristics. While the SRKW have
proxies for the various physiological processes in the WAs, the SA do not and so follow
basic processes. Improving these aspects would improve the representation of
swimming patterns in 3D and 4D and preferred locations of both the SRKW and Chinook
salmon, which may create different patterns of WAs and SAs. The third refinement
would be to incorporate more environmental variables such as tides, currents, and
additional marine species. However, useful datasets were not available on both sides of
the international border, or available datasets were not compatible to each other. These
improvements could also allow for circadian, seasonal, and annual cycles within the
model, to allow for difference in time of day, season, and year. Nevertheless, there is
always a trade-off between the complexity and validity of the model, data availability,
overall tractability of the data and processes, and ultimately the goal of the modeling
process.

5.5. Conclusions

This study designed and implemented a four-dimensional (4D) agent-based

model to simulate the spatial dynamics of the Southern Resident Killer Whales,
specifically the J pod, in the Salish Sea, Canada and USA. The developed 4D ABM

115
approach was capable of producing outputs that represent 3D locations in time of the
SRKW and these correlate with reported real sightings of the whales. The presented
work advances previous research that represent 4D complex systems and ABMs by
incorporating additional capabilities such as agents representing organisms that can
change behavior based on their previous experiences. The model used generalized
behavior of multiple whale individuals at a fine scale to represent patterns over a large
geographic area. Further improvements of the model could include representing a larger
number of different species in various geographic locations in the ocean, and more data
to support these species and additional detailed physiological and behavioral traits. This
could include migratory whales such as humpback whales and their natural preys. This
model’s prime geographic locations are along coastlines where the land provides a
heterogeneous landscape that influences the behavior of the various species present.

Overall, the developed 4D ABM captured the complexity of the 4D representation

of SRKW and Chinook salmon within the Salish Sea study area and generated results
that correlate with real data and behaviors. Further, the model can be extended to other
species that live in marine environments and can be used as an exploratory laboratory to
represent various species’ behavior and test scenarios in data rich or data scarce
situations. The ultimate goal is to use these models as decision support tools that inform
and improve marine conservation and management practices.

5.6. References

Adams, J., Kaplan, I.C., Chasco, B., Marshall, K.N., Acevedo-Gutiérrez, A., Ward, E.J.,
2016. A century of Chinook salmon consumption by marine mammal predators in
the Northeast Pacific Ocean. Ecological Informatics 34, 44–51.
https://doi.org/10.1016/j.ecoinf.2016.04.010

Anwar, S.M., Jeanneret, C.A., Parrott, L., Marceau, D.J., 2007. Conceptualization and
implementation of a multi-agent model to simulate whale-watching tours in the St.
Lawrence Estuary in Quebec, Canada. Environmental Modelling and Software
22, 1775–1787. https://doi.org/10.1016/j.envsoft.2007.02.007

Arostegui, M.C., Essington, T.E., Quinn, T.P., 2017a. Interpreting vertical movement
behavior with holistic examination of depth distribution: A novel method reveals
cryptic diel activity patterns of Chinook salmon in the Salish Sea. Animal
Biotelemetry 5, 1–11. https://doi.org/10.1186/s40317-016-0116-5

116
Arostegui, M.C., Smith, J.M., Kagley, A.N., Spilsbury-Pucci, D., Fresh, K.L., Quinn, T.P.,
2017b. Spatially clustered movement patterns and segregation of subadult
chinook salmon within the salish sea. Marine and Coastal Fisheries 9, 1–12.
https://doi.org/10.1080/19425120.2016.1249580

Bigg, M.A., Olesiuk, P.F., Ellis, G.M., Ford, J.K.B., Balcomb, K.C., 1990. Social
organization and genealogy of resident killer whales (Orcinus orca) in the coastal
waters of British Columbia and Washington State. Report - International Whaling
Commission 383–405.

Brown, D.G., Page, S., Riolo, R., Zellner, M., Rand, W., 2005a. Path dependence and
the validation of agent-based spatial models of land use. International Journal of
Geographical Information Science 19, 153–174.
https://doi.org/10.1080/13658810410001713399

Brown, D.G., Riolo, R., Robinson, D.T., North, M., Rand, W., 2005b. Spatial process and
data models: Toward integration of agent-based models and GIS. Journal of
Geographical Systems 7, 25–47. https://doi.org/10.1007/s10109-005-0148-5

Candy, J.R., Quinn, T.P., 1999. Behavior of adult chinook salmon (Oncorhynchus
tshawytscha) in British Columbia coastal waters determined from ultrasonic
telemetry. Canadian Journal of Zoology 77, 1161–1169.
https://doi.org/10.1139/cjz-77-7-1161

Castella, J.-C., Trung, T.N., Boissau, S., 2005. Participatory simulation of land-use
changes in the northern mountains of vietnam: the combined use of an agent-
based model, a role-playing game, and a geographic information system.
Ecology and Society 10, 27.

Chen, K., Chan, K.S., Bailey, K.M., Aydin, K., Ciannelli, L., 2012. A probabilistic cellular
automata approach for predator-prey interactions of arrowtooth flounder
(Atheresthes stomias) and walleye pollock (Theragra chalcogramma) in the
eastern bering sea. Canadian Journal of Fisheries and Aquatic Sciences 69,
259–272. https://doi.org/10.1139/F2011-160

Chion, C., Lamontagne, P., Turgeon, S., Parrott, L., Landry, J.A., Marceau, D.J.,
Martins, C.C.A., Michaud, R., Ménard, N., Cantin, G., Dionne, S., 2011. Eliciting
cognitive processes underlying patterns of human-wildlife interactions for agent-
based modelling. Ecological Modelling 222, 2213–2226.
https://doi.org/10.1016/j.ecolmodel.2011.02.014

Chudzinska, M., Nabe-Nielsen, J., Smout, S., Aarts, G., Brasseur, S., Graham, I.,
Thompson, P., McConnell, B., 2021. AgentSeal: Agent-based model describing
movement of marine central-place foragers. Ecological Modelling 440, 109397.
https://doi.org/10.1016/j.ecolmodel.2020.109397

Clifford, N.J., 2008. Models in geography revisited. Geoforum 39, 675–686.

https://doi.org/10.1016/j.geoforum.2007.01.016

117
Coro, G., Ellenbroek, A., Pagano, P., 2021. An Open Science approach to infer fishing
activity pressure on stocks and biodiversity from vessel tracking data. Ecological
Informatics 64, 101384. https://doi.org/10.1016/j.ecoinf.2021.101384

Crooks, A., Castle, C., Batty, M., 2008. Key challenges in agent-based modelling for
geo-spatial simulation. Computers, Environment and Urban Systems 32, 417–
430. https://doi.org/10.1016/j.compenvurbsys.2008.09.004

Daewel, U., Peck, M.A., Kuhn, W., St. John, M.A., Alekseeva, I., Schrum, C., 2008.
Coupling ecosystem and individual-based models to simulate the influence of
environmental variability on potential growth and survival of larval sprat ( Sprattus
sprattus L.) in the North Sea. Fisheries Oceanography 17, 333–351.
https://doi.org/10.1111/j.1365-2419.2008.00482.x

Eriksen, M., Lebreton, L.C.M., Carson, H.S., Thiel, M., Moore, C.J., Borerro, J.C.,
Galgani, F., Ryan, P.G., Reisser, J., 2014. Plastic pollution in the world’s oceans:
More than 5 trillion plastic pieces weighing over 250,000 tons afloat at sea. PLoS
ONE 9, 1–15. https://doi.org/10.1371/journal.pone.0111913

ESRI, 2018. ArcGIS Desktop: Release 10.5.

Filatova, O.A., Miller, P.J.O., 2015. An agent-based model of dialect evolution in killer
whales. Journal of Theoretical Biology 373, 82–91.
https://doi.org/10.1016/j.jtbi.2015.03.020

Foote, A.D., Griffin, R.M., Howitt, D., Larsson, L., Miller, P.J.O., Hoelzel, A.R., 2006.
Killer whales are capable of vocal learning. Biology Letters 2, 509–512.
https://doi.org/10.1098/rsbl.2006.0525

Ford, J.K.B., 1989. Acoustic behaviour of resident killer whales (Orcinus orca) off
Vancouver Island, British Columbia. Canadian Journal of Zoology 67, 727–745.
https://doi.org/10.1139/z89-105

Ford, J.K.B., Pilkington, J.F., Reira, A., Otsuki, M., Gisborne, B., Abernethy, R.M.,
Stredulinsky, E.H., Towers, J.R., Ellis, G.M., 2017. Habitats of special importance
to resident killer whales (Orcinus orca) off the west coast of Canada. DFO Can.
Sci. Advis. Sec. Res. Doc. vii + 57 p.

Ford, M.J., Hempelmann, J., Hanson, M.B., Ayres, K.L., Baird, R.W., Emmons, C.K.,
Lundin, J.I., Schorr, G.S., Wasser, S.K., Park, L.K., 2016. Estimation of a killer
whale (Orcinus orca) population’s diet using sequencing analysis of DNA from
feces. PLoS ONE 11, 1–14. https://doi.org/10.1371/journal.pone.0144956

Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., Goss-Custard,
J., Grand, T., Heinz, S.K., Huse, G., Huth, A., Jepsen, J.U., Jørgensen, C., Mooij,
W.M., Müller, B., Pe’er, G., Piou, C., Railsback, S.F., Robbins, A.M., Robbins,
M.M., Rossmanith, E., Rüger, N., Strand, E., Souissi, S., Stillman, R.A., Vabø,
R., Visser, U., DeAngelis, D.L., 2006. A standard protocol for describing
individual-based and agent-based models. Ecological Modelling 198, 115–126.
https://doi.org/10.1016/j.ecolmodel.2006.04.023

118
Grimm, V., Berger, U., DeAngelis, D.L., Polhill, J.G., Giske, J., Railsback, S.F., 2010.
The ODD protocol: A review and first update. Ecological Modelling 221, 2760–
2768. https://doi.org/10.1016/j.ecolmodel.2010.08.019

Guinet, C., Bouvier, J., 1995. Development of intentional stranding hunting techniques in
killer whale (Orcinus orca) calves at Crozet Archipelago. Canadian Journal of
Zoology 73, 27–33. https://doi.org/10.1139/z95-004

He, C., Huang, Z., Wang, R., 2013. Land use change and economic growth in urban
China: A structural equation analysis. Urban Studies 51, 2880–2898.
https://doi.org/10.1177/0042098013513649

Heimlich-Boran, J.R., 1988. Behavioral ecology of killer whales (Orcinus orca) in the
Pacific Northwest. Canadian Journal of Zoology 66, 565–578.
https://doi.org/10.1139/z88-084

Heppenstall, A., Malleson, N., Crooks, A., 2016. “Space, the Final Frontier”: How good
are agent-based models at simulating individuals and space in cities? Systems 4,
9. https://doi.org/10.3390/systems4010009

Jansson, T., Waldo, S., 2022. Managing marine mammals and fisheries: A calibrated
programming model for the seal-fishery interaction in Sweden. Environmental
and Resource Economics 81, 501–530. https://doi.org/10.1007/s10640-021-
00637-y

Joy, R., Tollit, D., Wood, J., MacGillivray, A., Li, Z., Trounce, K., Robinson, O., 2019.
Potential benefits of vessel slowdowns on endangered southern resident killer
whales. Frontiers in Marine Science 6, 1–20.
https://doi.org/10.3389/fmars.2019.00344

Lee, J.S., Filatova, T., Ligmann-Zielinska, A., Hassani-Mahmooei, B., Stonedahl, F.,
Lorscheid, I., Voinov, A., Polhill, G., Sun, Z., Parker, D.C., 2015. The
complexities of agent-based modeling output analysis. Journal of Artificial
Societies and Social Simulation 18, 1–26. https://doi.org/10.18564/jasss.2897

Lopez, J.C., Lopez, D., 1985. Killer whales (Orcinus orca) of Patagonia, and their
behavior of intentional stranding while hunting nearshore. Journal of Mammalogy
66, 181–183. https://doi.org/10.2307/1380981

Macatulad, E.G., Blanco, A.C., 2014. 3DGIS-based multi-agent geosimulation and

visualization of building evacuation using gama platform. International Archives
of the Photogrammetry, Remote Sensing and Spatial Information Sciences -
ISPRS Archives 40, 87–91. https://doi.org/10.5194/isprsarchives-XL-2-87-2014

Manson, S.M., 2001. Simplifying complexity: a review of complexity theory. Geoforum

32, 405–414. https://doi.org/10.1016/S0016-7185(00)00035-X

Mcloughlin, M., Lamoni, L., Garland, E.C., Ingram, S., Kirke, A., Noad, M.J., Rendell, L.,
Miranda, E., 2018. Using agent-based models to understand the role of
individuals in the song evolution of humpback whales (Megaptera novaeangliae).
Music and Science 1, 1–17. https://doi.org/10.1177/2059204318757021

119
Messina, J.P., Evans, T.P., Manson, S.M., Shortridge, A.M., Deadman, P.J., Verburg,
P.H., 2008. Complex systems models and the management of error and
uncertainty. Journal of Land Use Science 3, 11–25.
https://doi.org/10.1080/17474230802047989

Miller, P.J.O., 2006. Diversity in sound pressure levels and estimated active space of
resident killer whale vocalizations. Journal of Comparative Physiology A:
Neuroethology, Sensory, Neural, and Behavioral Physiology 192, 449–459.
https://doi.org/10.1007/s00359-005-0085-2

Noren, D.P., 2011. Estimated field metabolic rates and prey requirements of resident
killer whales. Marine Mammal Science 27, 60–77. https://doi.org/10.1111/j.1748-
7692.2010.00386.x

North, M.J., Collier, N.T., Ozik, J., Tatara, E.R., Macal, C.M., Bragen, M., Sydelko, P.,
2013. Complex adaptive systems modeling with Repast Simphony. Complex
Adaptive Systems Modeling 1, 3. https://doi.org/10.1186/2194-3206-1-3

Office for Coastal Management, 2022. 2013 Vessel Density [WWW Document]. NOAA
National Centers for Environmental Information. URL
https://www.fisheries.noaa.gov/inport/item/48932

Olson, J., Wood, J., Osborne, R., Barrett-Lennard, L., Larson, S., 2018. Sightings of
southern resident killer whales in the Salish Sea 1976-2014: the importance of a
long-term opportunistic dataset. Endangered Species Research 37, 105–118.
https://doi.org/10.3354/esr00918

Orca Network, 2022a. Southern resident orca community demographics, composition of

pods, births and deaths since 1998 [WWW Document]. URL
https://www.orcanetwork.org/Main/index.php?categories_file=Births and Deaths

Orca Network, 2022b. Sightings report summary archives [WWW Document]. URL
https://indigo-ukulele-jm29.squarespace.com/sightings-report-archive (accessed
5.12.22).

Parker, D.C., Manson, S.M., Janssen, M.A., Hoffmann, M.J., Deadman, P., 2003. Multi-
agent systems for the simulation of land-use and land-cover change: A review.
Annals of the Association of American Geographers 93, 314–337.
https://doi.org/10.1111/1467-8306.9302004

Reisinger, R.R., Keith, M., Andrews, R.D., de Bruyn, P.J.N., 2015. Movement and diving
of killer whales (Orcinus orca) at a Southern Ocean archipelago. Journal of
Experimental Marine Biology and Ecology 473, 90–102.
https://doi.org/10.1016/j.jembe.2015.08.008

Richmond, P., Romano, D., 2008. Agent based GPU , a real-time 3D simulation and
interactive visualisation framework for massive agent based modelling on the
GPU. Proceedings International Workshop on Super Visualization.
https://doi.org/10.1.1.144.734

120
Rios, L.M., Moore, C., Jones, P.R., 2007. Persistent organic pollutants carried by
synthetic polymers in the ocean environment. Marine Pollution Bulletin 54, 1230–
1237. https://doi.org/10.1016/j.marpolbul.2007.03.022

Sato, M., Trites, A.W., Gauthier, S., 2021. Southern resident killer whales encounter
higher prey densities than northern resident killer whales during summer.
Canadian Journal of Fisheries and Aquatic Sciences 78, 1732–1743.
https://doi.org/10.1139/cjfas-2020-0445

Simon, M., Wahlberg, M., Miller, L.A., 2007. Echolocation clicks from killer whales (
Orcinus orca ) feeding on herring ( Clupea harengus ) . The Journal of the
Acoustical Society of America 121, 749–752. https://doi.org/10.1121/1.2404922

Smith, A.K., Dragićević, S., 2022. Map comparison methods for three-dimensional space
and time voxel data. Geographical Analysis 54, 149–172.
https://doi.org/10.1111/gean.12279

Testa, J.W., Mock, K.J., Taylor, C., Koyuk, H., Coyle, J.R., Waggoner, R., 2012. Agent-
based modeling of the dynamics of mammal-eating killer whales and their prey.
Marine Ecology Progress Series 466, 275–291.
https://doi.org/10.3354/meps09845

Torrens, P.M., Benenson, I., 2005. Geographic Automata Systems. International Journal
of Geographical Information Science 19, 385–412.
https://doi.org/10.1080/13658810512331325139

Veirs, S., Veirs, V., Wood, J.D., 2016. Ship noise extends to frequencies used for
echolocation by endangered killer whales. PeerJ 4, e1657.
https://doi.org/10.7717/peerj.1657

Western Washington University, 2022. Salish Sea Atlas [WWW Document]. URL
https://wp.wwu.edu/salishseaatlas/ (accessed 3.22.22).

Whitehead, H., Ford, J.K.B., 2018. Consequences of culturally-driven ecological

specialization: Killer whales and beyond. Journal of Theoretical Biology 456,
279–294. https://doi.org/10.1016/j.jtbi.2018.08.015

Wightman, J.C., Taylor, G.D., 1976. Salmonid swimming performance in relation to

passage through culverts. Ministry of Recreation and Conservation.

Williams, R., Noren, D.P., 2009. Swimming speed, respiration rate, and estimated cost
of transport in adult killer whales. Marine Mammal Science 25, 327–350.
https://doi.org/10.1111/j.1748-7692.2008.00255.x

Wright, A., Scadeng, M., Stec, D., Dubowitz, R., Ridgway, S., Leger, J.S., 2017.
Neuroanatomy of the killer whale (Orcinus orca): a magnetic resonance imaging
investigation of structure with insights on function and evolution. Brain Structure
and Function 222, 417–436. https://doi.org/10.1007/s00429-016-1225-x

121
Chapter 6.

Conclusions

6.1. General Conclusions

In order to represent our world closer to reality, advancing the representation of

spatial complex systems (CS) into three spatial dimensions (3D) and time (4D) is
important for theoretical and applied aspects of geographic information science
(GIScience). The 4D agent-based modeling (ABM) framework presented in this
dissertation provides a useful option for the representation of many different spatial CS
such as forest fire smoke propagation or marine species interactions, and also could be
expanded to other 3D spatial dynamic phenomena. To facilitate multidimensional
geographic automata model evaluations, the comparison of 3D data, maps and
simulation outcomes have been achieved with novel 3D Kappa and 4D Fuzzy Kappa
techniques. The outputs from this dissertation can be summarized as: 1) the theoretical
framework and practical implementation of basic 4D ABM; 2) the 3D map comparison
methods 3D and 4D Fuzzy Kappa; and 3) the practical implementation of a 4D ABM of
higher complexity.

6.2. Summary of Findings

Chapter 2 proposes a theoretical framework that defines 4D ABMs for

representing spatial dynamic CS. This framework advances ABMs from two spatial
dimensions (2D) and adds a third spatial dimension, allowing the representation of 3D
movement, interactions and change in time. Traditionally, ABMs use 2D space, where
agents and the environment were restricted to a plane representing the local surface of
the Earth, such as in land-use change models. While this restriction may be acceptable
for representing some spatial CS, terrestrial and marine environments exist in 3D and
have significant processes that require multi-dimensional representation in order to be
accurately characterized. This chapter implements the proposed 4D modeling framework
to present a novel method to represent forest fire smoke propagation in 4D that
simulates smoke and other airborne particulate matter moving around a study area
considering atmospheric conditions and 3D terrain. The Smoke Agents move based off

122
the speed and direction of the wind, diffusing away from other Smoke Agents, and
following the terrain as defined by a digital elevation model. The agents are guided with
physics-based rules and do not incorporate any complex decision making processes.
Two scenarios were conducted, one where the Fire Agents emitting the Smoke Agents
were stationary, and one where the Fire Agents had movement to represent a fire being
spread by the wind. Both scenarios used atmospheric data from August 10th to 25th,
2017, and smoke locations and times coincided approximately with the smoke in
Calgary, Alberta, Canada on August 14th that originated from real fires. The results of the
developed model indicate that 4D ABMs are capable of representing realistically
propagation of smoke and its physical process effectively in 4D. The proposed
methodology can be used for air quality alerts and planning the need for medical
resources for spatial decision-making by policy makers. While the model generated
realistic spatial patterns and locations of smoke propagation, there was a lack of ability
in the comparison of obtained simulation results to themselves and reality. This
demonstrated the necessity to develop 3D voxel-based map comparison methods.

Chapter 3 extends the application of accuracy and Cohen’s Kappa into 3D and
presents their use in the quantitative comparison of simulation outcomes. The results
indicate that compared scenarios of the smoke propagation model were influenced by
the amount of voxels of a single class, specifically voxels without smoke present. While
3D accuracy showed different scenarios as highly similar, 3D Kappa was able to identify
the scenarios with more similar conditions. This shows that Kappa is a good candidate to
advance into 3D and 4D for map comparison while using fuzzy logic for space, time, and
class membership. The use of fuzzy logic allows for imperfect and imprecise
relationships (Openshaw, 1998).

The more advanced comparison methods particularly 3D and 4D Fuzzy Kappa

techniques that can compare in general 3D data, 3D maps, and 3D simulation outcomes
were presented in Chapter 4. 3D Fuzzy Kappa and 4D Fuzzy Kappa techniques aim to
address this need through providing a method in 3D and 4D that can account for chance
agreement between two maps and use fuzzy logic for small differences in space, time,
and classification. The developed techniques were implemented on multiple 3D and 4D
datasets, including hypothetical datasets, voxelized LiDAR data, soil horizon data, and
the simulation outputs from the 4D ABM of smoke propagation, as different case studies.
The results from the implementations met expected values to depict similarities and

123
differences among 3D voxel maps and provided an improvement over simpler 3D
Accuracy and 3D Kappa techniques. 3D and 4D Fuzzy Kappa provide new tools in the
comparison of 3D and 4D maps and can assist in the model testing procedures of 4D
ABMs or other multidimensional geospatial models, detecting spatial change in 3D and
3D spatial patterns. It also demonstrates the strength of using existing 2D methods that
have known knowledge of their use and their strengths and weaknesses. Advancing
existing 2D map comparison methods to 3D and 4D will help ease the transition and
acceptance of 3D and 4D metrics compared to unknown 3D and 4D metrics.

With an established 4D ABM framework and 3D model output comparison

methods, Chapter 5 presented a 4D ABM where agents use more complex decision-
making processes. This model shows the potential of 4D ABMs to represent dynamic
prey-predator systems involving living and thinking individuals, such as the Southern
Resident Killer Whales (SRKW) and Chinook Salmon. The model used several datasets
including bathymetry, distance from shore, and vessel density to inform the agents’
movements, decision making and interactions. The Whale Agents selected from various
actions, including hunting and eating the Salmon Agents, while the Salmon Agents
moved around the study area evading the Whale Agents. Whale Agents regularly
outputs their 3D coordinates that are captured as 3D maps of visited locations. These
locations were used to compare four different scenarios of salmon abundance. The
simulation outputs in 3D indicate important locations and behaviors to the SRKW and
the importance of the supply of Chinook salmon. The proposed methodology can be
used in spatial decision-making by policy makers for marine conservation. Software
routines have been created in each of these research studies.

The models and map comparisons methods developed in this thesis also had to
undergo testing. While aspects such as calibration and validation are presented in each
chapter, verification is not presented although it was conducted. Verification was done in
two main steps continually during development. The first step was in the verification of
the code, where variables were tracked through functions to ensure the intended and
expected changes were completed. The second step was in the verification of the
system processes being represented, where the range of potential inputs were tested to
ensure results are as expected. Part of this processes is shown for Chapter 4 in Table
4.2, where inputs such as random datasets or perfectly matching datasets are tested to
ensure the final output is correct.

124
6.3. Future Research

Data issues were a constant challenge that can be presented in several ways.
The first was the lack of available real 3D or 4D data of sufficient quality to use as inputs
to the 4D ABMs, which limited the number of case studies that were possible to select
for the implementation of proposed methodologies. The 4D ABM of smoke propagation
would benefit from increased spatial resolution in atmospheric data, both horizontal and
vertical axes, in order to improve the movement of the Smoke Agents. The relatively
large voxel size of the atmospheric data caused issues around physical features that
influence the movement of the wind around them, making Smoke Agents attempt to
move directly into mountains instead of the wind carrying them up and around. The 4D
ABM of SRKW would benefit from available data both 2D and 3D on location of whales
and salmon, for model input and building but also for model testing and validation. More
complete data could also include ocean currents, temperature patterns, and the
availability of species that could be prey for the Chinook salmon. In addition to improving
the case studies, data of higher quality and resolution would also allow for more and
different case studies to be implemented as ABMs. The second issue was the lack of
data further implied the limitations to perform full model testing. Therefore, the 4D ABM
had not passed rigorous model calibration or validation. The smoke propagation 4D
ABM had ground level recordings of airborne particulate matter that were used to
compare real smoke events to simulated smoke events. The SRKW 4D ABM had less
data, as the available data was not trustworthy enough for the purposes of model
testing. With increased data availability, both models could be better calibrated and
validated.

As future research work, the proposed 4D ABM can be coupled with voxel
automata (VA). The ABM would represent agents as the individuals of a system, such as
the SRKW and the Chinook salmon or other species, or fishermen’s activities, while the
VA could represent the changing state of various other systems, such as water
conditions, algae and other microorganisms, contaminants, marine vessel noise
propagation, and similar. This could improve the interactions between the system
elements that are represented in different ways and change the behavior or the system.
The agents could incorporate the voxel states into their decision-making process while
also being able to alter the voxel states directly. The VA transition rules could also

125
account for the presence or density of agents. This could also reduce some of the
computational intensity of representing large number of agents that would otherwise be
needed to represent all of these components.

Challenges with data also extended to the computational abilities of the software
developed that impede the processing time of developed modeling methods. The
computational intensity of working with 3D and 4D data is much greater than 2D data
due to the exponential increase in the amount of 3D data. The increased computation
time affects the models by reducing the potential complexity and number of agents that
can be represented. It also reduces the number of iterations of scenarios due to the time
each iteration takes to complete, impacting the model testing process. As computational
power increases over time this challenge will become smaller, and larger and more
complex studies will be more practical. The benefit to the challenge is that is encourages
more efficient code as the increased time in development is matched with a reduction in
computation time.

It is well known that agent-based models have been used in support of spatial
decision support systems (Sengupta and Bennett, 2003) or in participatory planning
(Zellner et al., 2014, 2012). The proposed 4D ABM methodologies can therefore be used
in spatial decision-making, evacuation or spatial planning, environmental management,
or conservation policy evaluation. However, the use of the 4D ABM by various
stakeholders and interested parties demands a friendly user interface that should be
created as an addition to the modeling software tools that have been already developed.
Displaying the 3D and 4D model outputs was a very difficult process, with a lot of effort
put into creating 2D figures that could depict 3D information of model outputs. The
enhanced visualization capabilities (Kornhauser et al., 2009) for the model outputs and
better user interface (Rydvanskiy and Hedley, 2020) would be highly beneficial
improvements that could facilitate other users than modeling experts to handle the 4D
ABM procedures in the decision making process.

Another challenge is the availability of 3D and 4D data analysis and comparison

methods. 3D Fuzzy Kappa and 4D Fuzzy Kappa work towards addressing this issue,
however there is still a lot of progress needed. There is a need especially in 3D data
analysis, as it can be used within model logic and to use as input to map comparison
methods for model testing. Similar to 3D and 4D Fuzzy Kappa, existing aspatial and 2D

126
techniques can be extended into 3D and 4D. A promising candidate for this is to
advance kernel density method into 3D. Current studies must use a 2D kernel density
where the data is projected into 2D, such as the distribution of elevation of points for
LiDAR analysis of trees (Wang et al., 2022). While other fields such as computer vision
have developed relevant techniques (Zhang et al., 2021), they need to be further
adapted to GIScience and incorporated in geospatial software packages.

The proposed methodologies in this thesis can be applied to other spatial

phenomena that operate in 4D. The use of 4D ABM also unlocks the increased
capability to represent 4D systems that have either been too difficult to represent with
traditional 2D ABMs, or can be revisited in 4D to more completely represent the
systems. One example is urban land use change modeling which has typically been
represented in 2D space and time and with use of 2D ABMs (Crooks et al., 2021).
However, urban areas are expanding upwards, with cities building high-rise buildings for
residential or commercial use. A 2D ABM cannot represent the vertical variation of
different users and uses within these buildings. A 4D ABM would be able to represent
the city in 4D, and agents that are occupying different units and interacting based on
agents’ rules that can represent their decisions while accounting for variables such as
views from buildings, height of their condo, and more. This could be done through the
coupling of a 3D geosimulation method such as iCity 3D (Koziatek and Dragicevic, 2017)
with a 4D ABM, and with a 3D Logic Scoring Preference method (Dragićević and Hatch,
2018; Munn et al., 2022) informing the decision making of agents.

By defining a 4D ABM framework, future work in 4D ABMs will have an initial

framework to develop further and compare against for various cases. The developed 4D
ABMs begin to incorporate 3D and 4D processes into ABMs. As more use and
advancement occurs, these processes can be incorporated into the ABM development
platforms such as Repast Simphony (North et al., 2013, 2007). Repast Simphony was
able to handle 3D positions and movement of agents, however it had few 3D geospatial
processes. With more additions such as geographic projections and coordinate systems
in 3D, the platform will be much more user friendly and encourage more development of
4D ABMs.

127
6.4. Research Contributions

The research in this dissertation contributes to the fields of GIScience,

particularly in advancement of multidimensional geographic automata systems (GAS)
and geosimulation approaches. The contributions to GIScience were made through the
representation of spatial dynamic Complex System (CS) phenomena with development
of novel 4D ABM with 3D and 4D map comparison methods. Further the thesis
contributes to applications in representing forest-fire propagation dynamics and species
interactions in marine environments that can be used for spatial planning, decision-
making and environmental management or protection purposes.

6.5. References

Crooks, A., Heppenstall, A., Malleson, N., & Manley, E. (2021). Agent-Based Modeling
and the City: A Gallery of Applications. Urban Book Series. Springer Singapore.
https://doi.org/10.1007/978-981-15-8983-6_46

Dragićević, S., & Hatch, K. (2018). Urban geosimulations with the Logic Scoring of
Preference method for agent-based decision-making. Habitat International, 72,
3–17. https://doi.org/10.1016/j.habitatint.2017.09.006

Kornhauser, D., Wilensky, U., & Rand, W. (2009). Design guidelines for agent based
model visualization. Journal of Artificial Societies and Social Simulation, 12(2), 1.

Koziatek, O., & Dragicevic, S. (2017). iCity 3D : A geosimualtion method and tool for
three-dimensional modeling of vertical urban development. Landscape and
Urban Planning, 167, 356–367. https://doi.org/10.1016/j.landurbplan.2017.06.021

Munn, K., Dragićević, S., & Feick, R. (2022). Spatial Decision-Making for Dense Built
Environments: The Logic Scoring of Preference Method for 3D Suitability
Analysis. Land, 11(3). https://doi.org/10.3390/land11030443

North, M. J., Collier, N. T., Ozik, J., Tatara, E. R., Macal, C. M., Bragen, M., & Sydelko,
P. (2013). Complex adaptive systems modeling with Repast Simphony. Complex
Adaptive Systems Modeling, 1(1), 3. https://doi.org/10.1186/2194-3206-1-3

North, M. J., Tatara, E., Collier, N. T., & Ozik, J. (2007). Visual Agent-based Model
Development with Repast Simphony. Proceedings of the Agent 2007 Conference
on Complex Interaction and Social Emergence, (1), 1–20. Retrieved from
http://www.osti.gov/energycitations/product.biblio.jsp?osti_id=973007

Openshaw, S. (1998). Some trends and future perspectives for spatial analysis in gis.
Geographic Information Sciences, 4(1–2), 5–13.
https://doi.org/10.1080/10824009809480497

128
Rydvanskiy, R., & Hedley, N. (2020). 3D Geovisualization Interfaces as Flood Risk
Management Platforms: Capability, Potential, and Implications for Practice.
Cartographica: The International Journal for Geographic Information and
Geovisualization, 55(4), 281–290. https://doi.org/10.3138/cart-2020-0003

Sengupta, R. R., & Bennett, D. A. (2003). Agent-based modelling environment for spatial
decision support. International Journal of Geographical Information, 17(2), 157–
180. https://doi.org/10.1080/1365881022000015990

Wang, Y., Xi, X., Wang, C., Yang, X., Wang, P., Nie, S., & Du, M. (2022). A Novel
Method Based on Kernel Density for Estimating Crown Base Height Using UAV-
Borne LiDAR Data. IEEE Geoscience and Remote Sensing Letters, 19.
https://doi.org/10.1109/LGRS.2022.3171316

Zellner, M. L., Lyons, L. B., Hoch, C. J., Weizeorick, J., Kunda, C., & Milz, D. C. (2012).
Modeling, learning, and planning together: An application of participatory agent-
based modeling to environmental planning. URISA Journal, 24(1), 77–92.

Zellner, M., Watkins, C., Massey, D., Westphal, L., Brooks, J., & Ross, K. (2014).
Advancing Collective Decision-Making Theory with Integrated Agent-Based
Modeling and Ethnographic Data Analysis : An Example in Ecological
Restoration. Journal of Artificial Societies and Social Simulation, 17(4), 11.

Zhang, Y., Li, C., Guo, B., Guo, C., & Zhang, S. (2021). KDD: A kernel density based
descriptor for 3D point clouds. Pattern Recognition, 111, 107691.
https://doi.org/10.1016/j.patcog.2020.107691

129