Pattern Based Compression of Multi Band Image Data for

Landscape Analysis

Visit the link below to download the full version of this book:

https://medipdf.com/product/pattern-based-compression-of-multi-band-image-data-f
or-landscape-analysis/

Click Download Now

PATTERN-BASED COMPRESSION OF
MULTI-BAND IMAGE DATA FOR
LANDSCAPE ANALYSIS

by

Wayne L. Myers
School of Forest Resources
The Pennsylvania State University, USA

and

Ganapati P. Patil
Center for Statistical Ecology and Environmental Statistics
The Pennsylvania State University, USA

Spri
ringer
Library of Congress Control Number: 2006932042

ISBN-10: 0-387-44434-3 e-ISBN-10: 0-387-44439-4

ISBN-13: 978-0-387-44434-5 e-ISBN-13: 978-0-387-44439-0

Printed on acid-free paper.

© 2006 Springer Science+Business Media, LLC

All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street,
New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly
analysis. Use in connection with any form of information storage and retrieval, electronic
adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed is forbidden.
The use in this publication of trade names, trademarks, service marks and similar terms, even if
they are not identified as such, is not to be taken as an expression of opinion as to whether or not
they are subject to proprietary rights.

9 8 7 6 5 4 3 2 1

springer.com
Dedication

This monograph is dedicated to Faith and Lalit, not as a work but as a la-
bor of love, in the same spirit with which our wives have so lovingly sup-
ported us over the years during which these analytical approaches and their
predecessors were in formative and convergent stages. They have pa-
tiently encouraged us, and have often assumed additional tasks while we
were occupied or preoccupied with these concepts and endeavors. Like-
wise, they have tolerated our travels and extended absences when circum-
stances made it difficult for them to accompany us. They have frequently
taken charge of domestic matters and sometimes serve as our surrogates
with family and friends while we concentrate attention on analysis, brain-
storming, or professional networking. So also for the numerous delays and
rescheduling of personal projects that they have gracefully accommodated.
Then they celebrate with us the auspicious outcomes of these ventures.
Our appreciation of them, and our gratitude to them, go way beyond
words. We salute them as best friends.
Preface

We offer here a non-conventional approach to muhivariate image-

structured data for which the basis is well tested but the analytical ramifi-
cations are still unfolding. Although we do not formally pursue them,
there are several parallels with the nature of neural networks. We employ
a systematic set of statistical heuristics for modeling multivariate image
data in a quasi-perceptual manner. When the human eye perceives a
scene, the elements of the scene are segregated heuristically into compo-
nents according to similarity and dissimilarity, and then the relationships
among the components are interpreted. Similarly, we segregate or seg-
ment the scene into hierarchically organized components that are subject to
subsequent statistical analysis in many modes for interpretive purposes.
We refer to the segregated scene segments as patterns, since they provide a
basis for perception of pattern. Since they are also hierarchically organ-
ized, we refer to them further as polypatterns. This leads us to our acro-
nym of Progressively Segmented Image Modeling As Poly-Patterns
(PSIMAPP). Likewise, we formalize our approach in terms of pattern
processes and segmentation sequences. In alignment with the terminology
of image analysis, we refer to our multivariate measures as being signal
bands.
There are several aspects of analytical advantage in our approach. We
model multi-band image data in terms of A-level aggregated and B-level
base patterns, with each level occupying an amount of computer media
equivalent to one signal band. Therefore, our models are parsimonious
relative to image data and can confer informational compression when
there are three or more signal bands. The models allow approximate resto-
ration of the image data, but since some residuals remain, the compression
is of a lossy nature which can obviate some of the issues regarding redis-
tribution and other proprietary provisions on the initial image information.
An additional advantage is that the A-level model is completely compati-
ble with facilities for handling raster maps in geographic information sys-
tems (GIS) that constitute the conventional contemporary computational
capability for conveying spatially specific geographical information. A
substantial suite of advantageous aspects are encompassed in our 4CS pat-
tern perspective of segmenting signal, spatial and sequential (temporal) in-
VIII Preface

formation for consideration of contrast, content, context and change at a

landscape level spanning several spatial scales.
Convenient contrast control in portraying prominent patterns is provided
by preparing pictorial presentations through tonal transfer tables. Depic-
tions are done by developing columns of transfer tables using pattern prop-
erties to produce palettes for patterns instead of the slower scanning of
scenes. This supports special selection of scene segments for showcasing
or suppression. Pictorial presentations are managed mosdy as mosaic
maps of patterns. Investigation of innovative integration of several signals
as image indicators can be done in spreadsheet software instead of com-
plex computational configurations of pixel processing packages. Integra-
tive indexing culminates in computing composite intensity as average
across signal bands in each of the A-level patterns and using the rank order
of that intensity as an identifying index for the particular pattern. The map
of these index identifiers can be displayed directiy as a single-band image
of what we call ordered overtones for the A-level patterns of our model.
Image information processing packages are publicly available for
downloading without charge that can be used to portray pattern pictures.
Parsing of patterns permits auto-adaptive contrast control, which auto-
matically adjusts contrast relations among patterns by an iterative algo-
rithm using statistical criteria that would be unworkable for initial image
information. This is in the nature of a statistically guided sequential linear
'stretch' with 'saturation'. Stretching involves proportional extension of
dynamic range of signals, whereas saturation entails truncation of signal
ranges. Contrast is perceived as differences in an image, with differences
due to variation which is expressed statistically as variance. Increasing
contrast is associated with increasing variance of intensities in the image.
Truncation by saturation induces areas of uniformity in high intensity
while increasing differences for lower intensities. There is also an issue of
tonal balance, whereby the average intensity should have intermediate
brightness. The adaptive algorithm proceeds with modification by satura-
tion as long as variance increases and average intensity remains at or be-
low a preset level.
Polypattems provide for interactive interpretive classification of scene
content such as land cover for thematic mapping in a manner that would
not be accommodated by conventional image analysis software systems.
As with pictorial presentations, this can be accomplished with software for
personal computers that is publicly available for downloading without
charge. Polypatterns also accommodate algorithmic approaches that are
usually called supervised and unsupervised classification, as well as adap-
tive strategies that hybridize the two approaches. Transfer tables for picto-
rial presentation can also be integrated into the mapping methods so that
 IX

particular patterns can be shown as categorical color imbedded directly

into grayscale renderings of the surrounding scenery.
Contextual considerations concern the fragmentation, juxtaposition and
interspersion of the pattern components comprising the image model.
Much of the visual perception of context in a picture is geometrically gen-
erated by mental models from information that is implicit rather than ex-
plicit in the image. Crucial contextual capability comes in segregating
segments to enable explicit analysis of their arrangements. With the mod-
eled mapping of pattern segments it becomes possible to investigate their
joint Distribution of POSITION (or DISPOSITION) in multiple modes.
One line of inquiry that would not be available for ordinary images is to
examine edge affinities, or preferential juxtaposition of particular patterns
relative to expectations in a RANdom landSCAPE (or RANDSCAPE). By
comparing edge affinities and signal similarities as dual domains, it be-
comes possible to explore several generations of generalization to produce
primary patterns of the landscape.
Innovative investigation of multi-scale landscape characteristics can be
conducted by compiling frequency profiles of patterns in nested blocks of
cellular image elements (pixels). Such profiles capture composite charac-
teristics of vicinities in the various settings of a landscape. A newly de-
veloped measure of profile differences as accordion distance allows
guided grouping of similar settings across landscapes, and thus systematic
study of landscape linkages. Equally innovative is investigation of compo-
sitional components as families of patterns through partial ordering and
rank range runs.
The topological spatial structure of the step surface formed by the
model's multivariate ordered overtones can be expressed in terms of eche-
lon hierarchies. Topological territories and transitions can thereby be iso-
lated for intensive investigation. Restored regions of the models can be
subtracted from the initial image information for purposes of multivariate
detrending so that assumptions of homogeneity and isotropy are more ap-
plicable for analysis of residuals by the methods of spatial statistics.
A principal components approach can be adapted to explore the signal
structure of the models in relation to that of the original image data. This
entails compiling covariance and correlation matrices for restorations of
the signal bands from the models. Extracting eigenvalues and eigenvectors
from these matrices give what we call principal properties in parallel to
and for comparison with principal components of the original image in-
formation.
Much of the impetus for our investigations has come in the interest of
detecting differences between instances of imaging in order to capture
change. The classic case is for sequences of scenes taken over time with
X Preface

the same sensor. In this relatively simple scenario, the signal smoothing
from pattern processes helps to make perturbed patches prominent. Addi-
tional advantage and innovative improvement comes from matching mosa-
ics of patterns in paired instances of imaging. Perturbation of patterns in
matched mosaics provides difference detection even with a shift in sensing
systems. Segregating sets of signals from a sensor also allows assessment
of differences in detectors. Still more substantial is the second-order ad-
vantage available by polypattem processing after compositing indicators of
change across multiple instances of imaging to track temporal trajectories
in landscape dynamics.
We have provided a substantial sampling of pattern pictures to illustrate
applications of our approaches, but have been considerably constrained in
do so by need to forego color in order to control costs. Color can make an
amazing difference in preparation of pattern pictures. Likewise, the sim-
pler printing processes do not provide the richness of resolution that is
available directly on a computer display screen. We ask that you do a web
search for PSIMAPP to access available software for applications. The
web is also a medium through which we can also offer case studies in
color, although still not with the resolution capabilities of a computer con-
sole.
Contents

1 Innovative Imaging, Parsing Patterns and Motivating Models 1

1.1 Image Introductory 2
1.2 Satellite Sensing Scenario 9
1.3 Innovative Imaging of Ecological and Environmental Indicators 11
1.4 Georeferencing and Formatting Image Data 16
1.5 The 4CS Pattern Perspective On Image Modeling 18
References 21

2 Pattern Progressions and Segmentation Sequences for IMAGE

Intensity Modeling and Grouped Enhancement 23
2.1 Pattern Process, Progression, Prominence and Potentials 23
2.2 Polypatterns 25
2.3 Pattern Pictures, Ordered Overtones and Mosaic Models of Images 26
2.4 Pattern Processes for Image Compression by Mosaic Modeling 29
2.5 a-Scenario Starting Stages 31
2.6 a-Scenario Splitting Stage 32
2.7 a-Scenario Shifting Stage 33
2.8 p-Scenario Starting Stages 36
2.9 P-Scenario Splitting Stage 37
2.10 Tree Topology and Level Loss 39
2.11 y-Scenario for Parallel Processing 40
2.12 Regional Restoration 42
2.13 Relative Residuals 42
2.14 Pictorial Presentation and Grouped Versus Global Enhancement 47
2.15 Practicalities of Pattern Packages 47
References 48

3 Collective and Composite Contrast for Pattern Pictures 51

3.1 Indirect Imaging by Tabular Transfer 51
3.2 Characteristics of Colors 53
3.3 Collective Contrast 54
3.4 Integrative Image Indicators 55
3.5 Composite Contrast for Pattern Pictures 60
3.6 Tailored Transfer Tables 61
References 62
XII

4 Content Classification and Thematic Transforms 63

4.1 Interpretive Identification 64
4.2 Thematic Transforms 67
4.3 Algorithmic Assignments 69
4.4 Adaptive Assignment Advisor 70
4.5 Mixed Mapping Methods 75
References 78

5 Comparative Change and Pattern Perturbation 79

5.1 Method of Multiple Mappings 80
5.2 Compositing Companion Images 81
5.3 Direct Difference Detection 82
5.4 Pattern Perturbation 87
5.5 Integrating Indicators 90
5.6 Spanning Three or More Dates 92
References 95

6 Conjunctive Context 99
6.1 Direct Detrending 99
6.2 Echelons of Explicit Spatial Structure 103
6.3 Disposition and Situation 106
6.4 Joint Disposition 106
6.5 Edge Affinities 109
6.6 Patch Patterns and Generations of Generalization 114
6.7 Parquet Polypattern Profiles 115
6.8 Conformant/Comparative Contexts and Segment Signal Sequences 117
6.9 Principal Properties of Patterns 125
References 128

7 Advanced Aspects and Anticipated Applications 129

7.1 Advantageous Alternative Approaches 129
7.2 Structural Sectors of Signal Step Surfaces 131
7.3 Thematic Tracking 133
7.4 Compositional Components 134
7.5 Scale and Scope 136
References 136

Appendix A. Public Packages for Portraying Polypatterns 139

A.l MultiSpec for Multiband Images and Ordered Overtones 139
A.2 ArcExplorer 147

Appendix B. a-Scenario with PSIMAPP Software 149

B.l Polypatterns from Pixels 151
B.2 Supplementary Statistics 153
B.3 Collective Contrast 153
B.4 Tonal Transfer Tables 156
 XIII

B.5 Combinatorial Contrast 159

B.6 Regional Restoration 160
B.7 Relative Residuals 161
B.8 Direct Differences 163
B.9. Detecting Changes from Perturbed Patterns 165
B.IO Edge Expression 167
B.l 1 Covariance Characteristics 168

Appendix C. Details of Directives for PSIMAPP Modules 171

Glossary 175

Index 177
Contributing Authors:

Wayne L. Myers, Ph.D.

Professor of Forest Biometrics, School of Forest Resources
Director, Office for Remote Sensing and Spatial Information Resources
Penn State Institutes of Environment
The Pennsylvania State University

Ganapati P. Patil, Ph.D., D.Sc.

Distinguished Professor of Mathematical Statistics, Department of Statistics
Director, Center for Statistical Ecology and Environmental Statistics
Department of Statistics
The Pennsylvania State University
Acknowledgments

We appreciatively acknowledge an array of assistance spanning several

sorts. First on this list is the late Charles Taillie whose intellectual inspira-
tion and insight permeated our cooperative working arrangements for so
many years. Genuine genius seems to be the best way of phrasing his phe-
nomenal mathematical mastery. Close behind are a number of former
graduate students, including but not limited to Dr. Glen Johnson, Joe
Bishop, Dr. Ningning Kong, Dr. Shih-Tsen Liu, Dr. Brian Lee, Dr. Som-
boon Kiratiprayoon, Francis Beck, Apinya Ramakomud and Dr. Salleh
Mohd. Nor. Emily Hill receives recognition for her capable conversion of
diverse document files to camera ready final form in meeting our time-
lines.
Agency, organizational and institutional acknowledgments are also in
order. The Penn State Institutes of Environment, and the School of Forest
Resources through the Pennsylvania Agricultural Experiment Station have
provided continuing support. The national GAP Analysis Program of bio-
diversity assessment furnished foundational support as it migrated through
agency settings from the U.S. Fish and Wildlife Service to the U.S. Geo-
logical Survey. This was augmented by funding for change detection work
through NASA Goddard Space Flight Center and the encouragement of
Dr. Darrel Williams. The National Science Foundation and Environmental
Protection Agency partnering programs in Water and Watersheds initia-
tives enabled emphasis on landscapes and linkages across scales. Most re-
cently, the National Geospatial Intelligence Agency contributed to contex-
tual concepts and computations. Finally, KUSTEM University in
Malaysia was a gracious host for Fellowship work by which to make the
manuscript manifest.
Also, all along, the Center for Statistical Ecology and Environmental
Statistics in the Department of Statistics at Penn State has provided condu-
cive work facilities. And at the turn of this century, the National Science
Foundation Digital Government Research Program and the Environmental
Agency STAR Grants Program have been supporting related research,
training, and outreach initiatives.
1 Innovative Imaging, Parsing Patterns and
Motivating Models

It seems that much of science has a propensity to concern itself with mi-
croscopic scales by focusing on genomic, molecular, atomic and sub-
atomic phenomena from a perspective of reductionism. The sciences of
ecology and sustainable environment, however, must counterbalance this
by investigating patterns of interaction in space and time at larger scales
covering landscapes and regions. This is particularly true for the ecologi-
cal discipline of landscape ecology (Forman & Godron, 1986; Forman,
1995; McGarigal & Marks, 1995; Turner, Gardner & O'Neill, 2001; Myers
et al., 2006). Image data and innovative imaging play important roles in
such investigations by serving as macroscopes to reveal patterns of ar-
rangement and change over substantial spatial extents at several scales. It
is to these macroscopic scales that we turn our attention in this work, with
particular emphasis on patterns in images.
The data that drive our imaging must be synoptic, by which we mean
that it is possible and sensible to ascribe a value to any location within the
geographic extent of interest - whether by direct determination or by inter-
polation. The data need not necessarily arise, however, from conventional
remote sensing based on signal scenarios involving radiant energy of the
electromagnetic spectrum. One of our thrusts is innovative imaging
whereby a variety of environmental indicators are cast in the manner of
multiple 'bands' of images so that spatial patterns and interactions having
ecological implications can be more readily detected, analyzed, and
tracked over time.
With this pattern purview, we focus on advanced approaches to parsing
patterns into components so that the spatial arrangements of the compo-
nents can be subjected to systematic study both statistically and structur-
ally. Patterns are often found to have dominant and subordinate contribu-
tions that express at different scales, in what might be considered as more
evident overtones and more subtie undertones. We seek to facilitate cog-
nizance and characterization of these multi-scale manifestations of spatial
relations in landscapes.
2 1 Innovative Imaging, Parsing Patterns and Motivating Models

Principal parts of patterns and their area arrangements present possibili-

ties of parsimonious packaging as multi-models of images that confer
compression as well as enabling enhancements of pattern groupings for
pictorial presentations in a more flexible fashion than regular renderings.
If pattern positions are empirically expressed in the mode of modeling,
then alternative analysis of spatial structure can be done directly by means
of the model without recourse to restoration. Analytical approaches to
crucial concerns of contrast, content, context and change are also aug-
mented.
Introduction of image innovations will be facilitated by an opening
overview of certain conventional concepts regarding signals, scenes and
sensors. This provides a path for pursuit of Progressively Segmented Im-
age Models As Poly-Patterns (PSIMAPP), with poly-pattern henceforth
being written as ]ust polypattern.

1.1 Image Introductory

Images are composed of pixels (picture elements) forming a regular lattice

L of area units that limit the spatial resolution of the image. Lattice ele-
ments of essentially equivalent nature are called cells in the terminology of
geographic information systems (GIS) for mapping landscapes (Bolstad,
2005; Chrisman, 2002; DeMers, 2000). The terms cell and pixel will be
used interchangeably in the course of this presentation, although the latter
term usually implies a rectangular lattice (raster arrangement) whereas the
former might be non-rectangular as for example hexagonal. In the termi-
nology of spatial statistics, the area comprising a lattice element is its spa-
tial support (Schabenberger & Gotway, 2005). Associated with each lat-
tice element (pixel or cell) is an ordered list (vector) of numeric values
with regard to particular properties. We will refer to values comprising
such vectors as signals in a generic manner. As used here, a signal band is
a particular component of the signal vector (i.e., specific position in the list
of signal values). Accordingly, we denote the set of all distinct signal vec-
tors by V without reference to particular pixel positions.
An image is formed electronically by translating the values of one or
more bands into intensities of gray-tone or color for displaying in the re-
spective lattice positions of the pixels. Fig. 1.1 is an image formed by ren-
dering a band of red light recorded by a sensor aboard a satellite (MSS
sensor of Landsat satellite) as gray-tone intensities. An image of this sort
is based on reception of a physical signal by a remotely positioned sensing
device, and thus falls under the rubric of remote sensing. Since remote
 1.1 Image Introductory 3

Fig. 1.1 Gray-tone image of band 2 (red light) from MSS sensor of Landsat satel-
lite showing wooded ridges (darker) and agricultural valleys (lighter) characteris-
tic of the Ridge and Valley Physiographic Province of central PA. Note clouds in
upper left (northwest) corner, and Raystown Lake reservoir in the southern por-
tion.