Ecological Indicators 136 (2022) 108636

Contents lists available at ScienceDirect

Ecological Indicators
journal homepage: www.elsevier.com/locate/ecolind

A framework for a forest ecological base map – An example from Norway

Hans Ole Ørka *, Marie-Claude Jutras-Perreault , Erik Næsset , Terje Gobakken
Faculty of Environmental Sciences and Natural Resource Management, Norwegian University of Life Sciences, P.O. Box 5003, NO-1432 Ås, Norway

A R T I C L E I N F O A B S T R A C T

Keywords: In the management of forest ecosystems, spatial information about the extent, condition and pressures are
Ecosystem services essential. In the current study, we present a framework for a remote sensing-based forest ecological base map
Environmental management covering Norway. Combining remotely sensed imagery from optical satellite systems such as Sentinel-2 and
Naturalness
Landsat provides information about forest ecosystem extent and change over time. Utilizing a national dataset of
Remote sensing
Model-based inference
airborne laser scanning (ALS) data allowed predicting a range of attributes describing forest condition, including
naturalness. In total, seven definitions of naturalness were evaluated. Pressures on the forest ecosystems were
mapped using a change detection algorithm and satellite data from 1986 to 2020. Change detection is the
cornerstone in monitoring and for understanding the pressures on the ecosystems. The predicted forest extent
had an overall accuracy of 85 to 89% using Sentinel-2 imagery from 2020 and 71 to 81% using Landsat imagery
from 1986. For the forest condition attributes, the explained portion of the variances were >70% for biomass,
height and volume and from 21% to 64% for number of stems, crown coverage and a diversity index. Naturalness
was classified with accuracies of 77 to 98%, except for age-based definitions. Nevertheless, a large number of
false positives were present. Change detection was evaluated in terms of final harvest and was identified with an
overall accuracy of 84–92%. The land cover change classification had an overall accuracy of 70–92%. The
detailed maps of forest condition and forest pressures were aggregated to a local level using model-based
inference, providing estimates of mean values and uncertainty at a scale suitable for ecosystem indicator
development. The collection of map layers describing forest extent, condition and pressures form a forest
ecological base map important for environmental management.

1. Introduction ecosystems and their services. Providing a collection of map layers about
the extent, condition and pressures on the forests will together form a
Forest ecosystems provide a wide range of services, including forest ecological base map.
biomass, timber, carbon sequestration, recreational and cultural uses, as Remote sensing (RS) is often suggested as a tool to help mapping
well as habitat and biodiversity (Costanza et al. 1997). Indicators are ecosystem indicators, but the contribution of RS is most often indirectly
frequently used to evaluate and communicate the quality and quantity of linked rather than being predicted directly from the remotely sensed
ecosystem services and understand the ecosystems’ condition, trends data (Feld et al. 2010; Andrew et al. 2014). Nevertheless, RS can facil­
and rate of change (Layke 2009; Maes et al. 2016; Jakobsson et al. itate high spatial and temporal resolution mapping at multiple scales
2021). The ecosystem condition is defined as the quality or as the supporting ecosystem assessment (Vauhkonen 2018). The main contri­
physical, chemical and biological condition of an ecosystem at a bution of RS is to map the ecosystems’ state or condition and the drivers
particular point in time (Maes et al. 2018). In forest ecosystem man­ of change or pressures on the ecosystems and their extent (Burkhard and
agement, information on forest condition is a key attribute providing Maes 2017). These maps can later be used to produce indicators to assess
information on the current state. In addition to details of the forest a variety of ecosystem services. In addition to providing a foundation for
ecosystem condition, the drivers of change or pressures on the ecosys­ deriving indicators, such maps offer explicit spatial information appli­
tems are essential in managing the forests. Therefore, detailed maps of cable in many ecosystem management applications (Kerr and Ostrovsky
the extent, condition and pressures on the forest ecosystems over large 2003; Wulder et al. 2004; Andrew et al. 2014).
areas will provide substantial progress in managing the forest There is a long tradition of mapping forest ecosystem extent using

* Corresponding author.
E-mail addresses: [email protected] (H.O. Ørka), [email protected] (M.-C. Jutras-Perreault), [email protected] (E. Næsset),
[email protected] (T. Gobakken).

https://doi.org/10.1016/j.ecolind.2022.108636
Received 1 July 2021; Received in revised form 11 January 2022; Accepted 29 January 2022
Available online 13 February 2022
1470-160X/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

optical satellite images (Anderson et al. 1976; Nelson 1983; Wynne et al. 1. Mapping selected forest attributes describing ecosystem condition, i.
2000). Furthermore, traditional forest resource information describing e. biomass, volume, height, density (number of trees), crown cover.
the forest condition, such as volume, heights and density distribution, 2. Evaluating ALS as a source of providing information about forest
has been obtained using several RS techniques (Wulder 1998; Næsset naturalness.
2004a). Especially the use of three-dimensional (3D) point clouds from 3. Evaluating a methodology to detect and classify pressures on the
airborne laser scanning (ALS) has provided detailed and accurate maps forest ecosystems in terms of permanent (land-use changes) or tem­
(Maltamo and Packalen 2014; Næsset 2014). Today, ALS is the best- porary changes (harvests).
suited technology to map attributes related to forest structure (Hyyppä 4. Providing estimates with associated uncertainties on an aggregated
et al. 2000; Hyde et al. 2006). In many countries, national ALS acqui­ local level from the detailed maps, i.e., fine spatial resolution maps.
sition campaigns have been conducted. The main aim of collecting these
3D data is to provide accurate terrain elevation models, but they also 2. Material and methods
provide the possibilities to map forest condition for large areas (Nord-
Larsen and Schumacher 2012; Nilsson et al., 2017; Astrup et al. 2019). 2.1. Study area
However, the large area forest maps currently derived from these na­
tional campaigns mainly consider traditional provisional services such The study area was defined as Norway’s mainland extending from
as timber and biomass production (Nord-Larsen and Schumacher 2012; 58◦ N to 71◦ N and from 5◦ E to 31◦ E (Fig. 1). The total size of the
Nilsson et al., 2017; Astrup et al. 2019). Nevertheless, the same ALS data mainland is approximately 324 000 km2, of which 38% is covered with
have been proposed for deriving maps related to other forest condition forests (Bryn et al. 2018). The forested area is mainly part of the boreal
attributes, e.g. forest naturalness (NAT) (Valbuena et al. 2014; Sverdrup- zone, but also the boreonemoral zone is represented in the south. Nor­
Thygeson et al. 2016). Naturalness describes where the ecosystems are way spruce (Picea abies (L.) H. Karst.) and Scots Pine (Pinus sylvestris L.)
on a gradient mainly characterized by natural processes at one end and cover 10.2% and 9.4% of the area, respectively. Boral deciduous,
complete human influence at the other end (Angermeier 2000). In forest dominated by birch (Betula ssp.), is the most common forest type
ecosystems, plantations have the smallest degree of naturalness and (13.8%). In comparison, nemoral broadleaved deciduous forests cover
ecosystems characterized by terms such as “undisturbed”, “pristine” and less than 0.5% of the area (Bryn et al. 2018). The forests populate the
“virgin” at the other end of the gradient (McRoberts et al. 2012). area from sea level and up to approximately 1000 m above sea level. The
Naturalness is often used as an indicator of habitat quality and biodi­ country can be divided into five geographical regions (Fig. 1).
versity (Uotila et al. 2002).
The forest condition represents the quality of the ecosystems at a 3. Reference data
specific point in time. Repeating the mapping of the ecosystems’ state at
another point in time will provide means to monitor change. However, 3.1. National forest inventory data
mapping the pressures continuously over several years, even back in
time, is often desired. The main drivers of change in the forest ecosys­ In the current study, sample plots measured by the Norwegian NFI
tems in Europe are forest activities (Curtis et al. 2018). Change detection were used as reference data. The Norwegian NFI started in 1919, and
using a long time series of optical satellite imagery with relatively high since 1986 measurements on permanent sample plots have been carried
resolution enables annual detection of changes with high accuracy out regularly every five years (Breidenbach et al. 2020). The field pro­
(Kennedy et al. 2010; Hansen et al. 2013; Jutras-Perreault et al. 2021). tocol has changed several times since then, especially to consider
Several different algorithms for change detection have been developed, additional environmental information. Thus, some information is only
especially after 2008, when the Landsat archive was made freely available for the most recent years. However, on each sample plot, in­
available (Woodcock et al. 2008). formation about tree species, stem diameter, tree height and several
A key to utilize RS for mapping forest extent, condition and pressures other properties are collected. The positions of the sample plots have
is reliable reference data, typically precisely positioned field sample been measured with handheld Global Navigation Satellite Systems
plots (Valbuena et al. 2010). An important source of information for the (GNSS) receivers several times, but currently > 63% of the sample plots
forest ecosystems is the national forest inventories (NFIs), providing have been positioned with survey grade GNSS receivers and subsequent
detailed information for trees on sample plots (Kangas et al. 2018). The differential correction. The sample plots located in the lowland regions
NFIs measure many forest attributes when visiting the field plots in which coniferous tree species are economically important, are sys­
(Breidenbach et al. 2020), and additional maps could be developed to tematically distributed on a 3 km × 3 km sampling grid, while in the
describe, for example, forest “naturalness” and other ecosystem infor­ low-productive sub-alpine region, a 3 km × 9 km sampling grid is used.
mation (McRoberts et al. 2012; Knoke et al. 2021). Relating the NFI In Finnmark, the northernmost county, a 9 km × 9 km grid is used. These
sample plot information to the remotely sensed information using sta­ latter areas are less economically important (Breidenbach et al. 2020).
tistical models enables prediction for the entire area of interest and From the NFI data, volume (V), above-ground biomass (AGB), below-
provides detailed maps of the forest ecosystems. Such ecosystem maps, ground ground biomass (BGB), height (H), density (number of stems)
with a spatial resolution of, say, 10 to 20 m, are important for many (N) and crown coverage (CC) were calculated for all sample plots.
operational management applications, but they may also be used to Furthermore, the Gini diversity index (GINI) was computed based on the
estimate indicators or related condition and pressure attributes at any tree-level basal areas:
scale, from the size of individual map units of similar size as field plots ∑n
j=1 (2j − n − 1)baj
adopting model-assisted or model-based inference (McRoberts 2011; GINI = ∑n (1)
Ståhl et al. 2016). The characterization of the uncertainty is an essential j=1 baj (n − 1)

feature in many management applications (Yousefpour et al. 2012).

The main objective of the current work was to develop a framework where j is the rank of the tree in ascending order, n is the total number of
for a forest ecological base map. The framework consists of a method­ trees on the sample plot and baj is the basal area of the tree with rank j.
ology to provide detailed maps of the entire country of Norway, of The Gini index was chosen because it has been found superior over many
condition and pressure of the forest ecosystems. Furthermore, the detail other indices in forest management for describing tree size diversity
maps are aggregated using model-based inference to local level (100 (Lexerød and Eid 2006). In addition, seven different definitions of
km2) more suitable for indicator assessment. Thus, these were the sub- naturalness were adopted and sample plots were classified as natural
objectives of the work: forest or not as a binary value for each definition (Tables 1, 2).

2
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 1. Location of Norway and the division into five geographical regions.

Table 1 Table 2
Overview of naturalness (NAT) definitions adopted for individual NFI plots. Minimum age (years) for a forest to be defined as biological old-growth forest
Label Description Reference
according to dominant species and site index (low, medium and high) (D6,
Table 1).
D1 “Old natural forest” in accordance with NFI (Viken 2018)
definition and recorded on each sample plot by Species Low Medium High
subjective assessment, i.e. forests without visible Spruce dominated 160 140 120
human disturbance, with native tree species, old Pine dominated 180 160 140
age, multilayer and dead wood in several stages of Broadleaved dominated 120 100 80
degradation and dead wood of large dimensions.
D2 Stand age of at least 160 years.
D3 Stand age of greater than 140 years.
has not been mapped. To classify land cover in non-mapped areas a
D4 GINI larger than the 25 percentile of GINI for the
region.
random sample of observations was drawn from the available map data
D5 GINI larger than the 50 percentile of GINI for the in each 10 km × 10 km (100 km2) block in a square grid covering the
region. entire country. A total of 41◦772 sample observations (250 m2) were
D6 Biological old-growth forests. Defined by age, site (Søgaard et al. collected across the country and they were used as reference observa­
index and dominant tree species. See Table 2. 2012)
tions for a land cover classification. The reference observations were
D7 Mature forests in the first visit of the permanent (Storaunet and
sample plot (1994–1998) which still is a mature Rolstad 2015) distributed on land cover categories with 33.2% located in forest, 8.1%
forest in the last visit of the plot. in cropland, 4.9% in grassland, 9.6% in wetland, 4.5% in settlements,
16.8% in other-land and 22.8% in water (freshwater and ocean).

3.1.1. Land cover information

The current land cover map in Norway (Ahlstrøm et al. 2019) was 3.2. Remotely sensed data and processing
reclassified to land cover categories; forest, cropland, grassland, peat­
land, settlements, other-land, freshwater and ocean (Jacobsen et al. 3.2.1. Satellite data
2020). The land cover map was originally created using manual photo In the current study, optical satellite imagery from Sentinel-2 from
interpretation and is being continuously updated in agricultural and 2020 and Landsat 4 to 5 from 1986 were used. From the available im­
populated areas (Ahlstrøm et al. 2019) . However, the land cover map ages, we constructed two national mosaics based on the 1986 and the
does not have complete wall-to-wall coverage and 38% of the land area 2020 data, respectively. We applied a generalized spatiotemporal
mosaic method implemented using the Geomosaic tool (Ørka et al. co-

3
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

submission). The mosaics were created by a desirability score calculated from the land cover classification to account for the minimum size of 0.5
for each scene. The score was the weighted sum of a relative distance ha used for the forest definition (Ahlstrøm et al. 2019). Thus, area
from the target date set to 15th of July, the aerosol optical thickness and patches smaller than 0.5 ha were removed and patches of non-forest <=
number of “good pixels”. The mosaics provided near cloud-free, anal­ 0.1 ha within the forest were reclassified to forest according to the
ysis-ready mosaics (Ørka et al. co-submission). minimum mapping sizes of four grid-cells of 250 m2 (Hengl 2006). This
operation was done using a sieve filter of four grid-cells, including eight
3.2.2. Airborne laser scanning data connectedness, i.e., the diagonal grid-cells were also considered neigh­
The ALS data were collected as part of the construction of a new bors to the central grid-cell. The resulting binary raster provided the
national elevation model for Norway. In total, 807 acquisitions covering forest ecosystem extent in 2020. The same approach with classification
nearly 150◦000 km2 were downloaded from https://hoydedata.no and and filtering was applied to the 1986 Landsat data. However, only the
processed (Table 3). First, height above ground was calculated by sub­ remotely sensed data were utilized to provide the forest ecosystem
tracting the terrain height calculated from echoes classified as ground extent in 1986. The Landsat and Sentinel-2 data were used because they
from the height of the current echo. This normalization was done by represent the best available spatial and spectral resolution data for the
Delaunay triangulation by which a linear interpolation was used for respective years.
each triangle. Second, for echoes along the edges of an acquisition
boundary, extrapolation based on nearest neighbor and inverse distance 3.3.2. Statistical modeling and prediction of forest condition
weighting was used. Furthermore, standard ALS metrics (cf. Næsset For modelling the forest attributes (AGB, BGB, CC, GINI, H, N and V),
2004b) were calculated for all first echoes above 2 m. The metrics the NFI sample plots were used as reference data. A challenge with ALS
included different statistical properties of the height distributions, such data is that sensors and flight parameters often vary among acquisitions.
as average or standard deviation and metrics calculated by the number This difference in flying altitude, point density and sensor type, com­
of echoes above a particular height relative to the total number of bined with a variation in phenology, influences the relationships be­
echoes. tween forest attributes and metrics derived from ALS data (Næsset
2004c; Ørka et al. 2010). To account for such acquisition-specific effects,
3.3. Methodological overview of framework mixed-effects modelling was applied (Pinheiro and Bates 2000). Models
were established independently for five geographical regions; eastern,
The applied framework for developing a forest ecological base map southern, western, central and northern. For the forest attributes, the
consisted of four steps (Fig. 2). First, the ecosystem extent was delin­ following models were used:
eated for two points in time, 1986 and 2020, respectively. The forest ( ) ( ) ( )
log AGBij = β0 + bi + β1 log Hmeanij + β2 log D0ij + ∊ij (2)
ecosystem extent for 1986 served as a reference forest extent to identify
changes over time. The land cover classification for 2020 served both as ( ) ( ) ( )
log BGBij = β0 + bi + β1 log Hmeanij + β2 log D0ij + ∊ij (3)
an up-to-date delineation of the forest area and to identify land cover
change classes. In the second step, we used ALS data for modelling
CCij = β0 + bi + β1 D0ij + ∊ij (4)
different attributes describing the forest condition. In the third step, we
identified changes from 1986 to 2020 and coupled this with the 2020
GINIij = β0 + bi + β1 Hcvij + β2 D0ij + ∊ij (5)
land cover classification to identify change classes. In the fourth and last
step, we used the modelled forest condition attributes and identified Hij = β0 + bi + β1 H90ij + ∊ij (6)
pressures and provided estimates on a local level, i.e., for squared blocks
of size of 100 km2 each. The forest extent was assumed to be without ( ) ( ) ( )
log Nij = β0 + bi + β1 log D0ij + β2 log Hcvij + ∊ij (7)
errors and was excluded in the uncertainty estimates related to forest
condition and change. ( ) ( ) ( )
log Vij = β0 + bi + β1 log Hmeanij + β2 log D0ij + ∊ij (8)
The predictions were carried out for individual grid-cells of 250 m2
(15.8114 m × 15.8114 m) tessellating the entire study area. This grid- where β0 , β1 and β2 are parameters to be estimated and ∈ij ∼ N(0, σ2 )
( )
cell size was selected because it is identical to the sample plot area and bi N 0, σ 2b . The ALS metrics used were the mean height of echoes >
used by the Norwegian NFI. Thus, avoiding potential problems with 2 m (Hmean), the 90th height percentile of echoes > 2 m (H90), the
respect to scale effects (cf. Packalen et al. 2019). The Sentinel-2 mosaics coefficient of variation of echoes > 2 m (Hcv) and the density of echoes
with spatial resolutions of 10 m and 20 m and the Landsat mosaics with a > 2 m to the total number of echoes (D0). The models were fitted using
spatial resolution of 30 m were resampled to grid-cells of 250 m2 by the function lme in the R package nlme (Pinheiro and Bates 2000). The
nearest-neighbor interpolation. most recent ALS acquisition covering any given sample plot was selected
and combined with the field plot data collected closest in time. Simi­
3.3.1. Ecosystem extent delineation larly, when the models subsequently were used for prediction, the most
The extent of the forest ecosystems was defined using the existing recent acquisitions were employed.
land cover map. The map was updated when information was missing The naturalness (NAT) definitions (D1-D7) were classified as bino­
based on a land cover classification with random forest (Breiman 2001). mial variables using generalized boosted regression modeling and the
The classification was based on the extracted land cover categories ALS metrics mean (Hmean), standard deviation (Hsd), kurtosis (Hkurt),
described in section 2.2.2. The forest class was extracted and filtered skewness (Hskew), coefficient of variation (Hcv) of echoes > 2 m and the
measures of canopy density D0 and D5 (density of echoes > 50% of the
Table 3 95th percentile height to the total number of echoes). Thus, no
Number of ALS acquisitions and area covered distributed on geographical acquisition-specific corrections (i.e., random effects) were included. To
region. account for the imbalanced sample, a simple oversampling of the mi­
Region Number of acquisitions Area covered (km2) nority class was applied (Ling and Li 1998). The models were used to
Eastern 229 47◦350
predict probability maps of naturalness as this is regarded more
Southern 97 22◦935 appropriate in ecological applications (Foody 1996).
Western 169 17◦769
Central 146 27◦682 3.3.3. Change detection and classification of ecosystem pressure
Northern 166 34◦143
In the current study, LandTrendr (Landsat-based detection of trends
Total 807 149◦881

4
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 2. Overview of the proposed framework.

in disturbance and recovery) was used to identify changes in the forest representing e.g., AGB, BGB, …, V, D7) for each sample plot j and ALS
landscape. LandTrendr is a time segmentation algorithm developed by acquisition i was computed as:
Kennedy et al. (2010), which has been widely used due to its ability to
y*ij = ̂ ε ij vij
y ij + ̂ (9)
detect both short-term disturbances and permanent changes (Griffiths
et al. 2012; Senf et al. 2015; Fragal et al. 2016; Vogeler et al. 2018).
LandTrendr models pixel’s spectral time series as a sequence of linear where ̂y ij is the fitted value of the original model and ̂ε ij is the residuals
segments bounded by breakpoints, or vertices, to detect the important of the model with the random effect that are multiplied with a random
*
features of a trajectory. A map of changes is created from the time po­ variable vij y*ij . The refitted model is used to predict a new value ̂
y k for
sitions and spectral values of these vertices. The changes are detected each grid-cell within the area of interest. Thus, each grid-cell has a
annually from satellite images acquired during the growing season. The predicted value for each bootstrap iteration.
recent implementation of LandTrendr on the Google Earth Engine For each bootstrap iteration, the mean estimate for the block m was
platform has substantially reduced data handling costs and processing calculated by averaging the individual grid-cells predictions for the k
time (Kennedy et al. 2018). grid-cell in the block:
The change detection was masked with a maximum forest ecosystem
∑Nm *
extent created by combining the forest extents produced from remotely ̂ *m = 1
Y y
̂ (10)
Nm k=1 k
sensed data in 1986 and 2020 and the existing land cover map (Ahlstrøm
et al. 2019). Identified changes overlapped with the forest extent in one
where Nm is the number of grid-cells in block m. The final estimates ( Y
̂ m)
of these maps were further utilized to mask out changes in the forest.
*
Finally, the identified changes were overlaid with the land cover clas­ were computed by averaging the individual bootstrap estimates( Y
̂ ):
m
sification carried out on Sentinel-2 images in 2020 to provide change ∑nboot *
class information. ̂m = 1
Y ̂
Y (11)
nboot b=1 m

3.3.4. Estimates and uncertainty where nboot is the number of bootstrap iterations. Then the uncer­
Indicators are most often connected to a larger area than the map­ tainty of Y
̂ m was computed by:
ping unit in the detailed maps. Thus, an aggregated spatial level is ∑nboot ( *
1 )
needed to support indirect or direct indicator development. In the cur­ ̂ m(Y
MSE ̂ m) = ̂ − Y
Y ̂m (12)
(nboot − 1) b=1 m
rent study, local-level estimates were provided by aggregating the
detailed maps into 100 km2 squared blocks. For each block, the mean For logarithmic models, we applied a back-transformation of the
values of the forest condition attributes and total pressures were esti­ predictions before estimates were produced. Due to the large distance
mated. Based on the size of the area, we used a model-based inference to between adjacent sample plots (3 km) and the large size of the area
provide the estimates. In addition to the mean values, we also estimated considered (100 km2), the residual variance and covariance were
the mean square errors of the respective estimates. ignored (McRoberts et al. 2018).
Wild bootstrap was applied to provide estimates of forest condition The proportion of area changed due to harvests in the period from
attributes (see 2.4.2.) (Wu 1986). The wild bootstrap was selected as it 1986 to 2020 was estimated for each block. As no model was used
retains the structure of the original residuals and can be applied to another approach was applied for estimation. The estimator used was a
different statistical models (Esteban et al. 2019). The residuals of the model-assisted post-stratified estimator (see Stehman 2013 for details).
mixed model excluding the random effect were used to maintain the Thus, the total harvested area for the period ( Y
̂ ma ) within the block was
data structure. In each bootstrap iteration, a new y value (y*ij estimated as:

5
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

∑2 nab
̂ ma =
Y Wma (13) where yi is the true value, ŷi is the predicted value and n is the number of
a=1 na+
observations. Furthermore, the squared Pearson’s correlation coefficient
where Wma is the area harvested in block m identified using the change (r2) between observed (yi ) and predicted (̂ y i ) was calculated and
detection procedure and n is a error matrix calculated based on the NFIs reported.
reference data of harvest or no-harvest, where nab is the area within the
reference class b for map class a and na+ is the total area classified in map 4. Results
class a in the error matrix. In our case, nab /na+ was assumed constant
within all blocks in a region. This ensures that the sum of the block es­ 4.1. Forest ecosystem extent
timates was consistent within the geographical regions. The standard
error of this estimate was computed as: The accuracy of the forest extent from 1986 varied from 71 to 81%,
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ with kappa values between 0.37 and 0.52, the precision was in the range
( )̅
√
√ nab nab 60 to 71, recall from 40 to 69 and F1 from 48 to 69 (Fig. 3). This indi­
√∑ 1 −
√ 2 na+ na+ cated a low to moderate accuracy of the classification.
̂ ma ) =
SE( Y W2 (14)
a=1 ma na+ − 1 The forest extent from 2020 was based on Sentinel-2 data and
showed greater classification accuracies compared to the extent based
on the 1986 data. The reasons for improved classification accuracy were
3.4. Evaluation and accuracy assessment finer spatial, spectral and radiometric resolutions. The accuracy values
varied between 85 and 89% and kappa between 0.67 and 0.77. The
Classification accuracy was assessed using an error matrix and precision values varied from 71 to 91, recall from 72 to 84 and F1 from
metrics such as overall accuracy, kappa and producer’s and user’s ac­ 76 to 85 (Fig. 3).
curacy (Congalton 1991). Overall accuracy was calculated as the num­
ber of observations that were correctly classified relative to the total 4.2. Forest ecosystem condition
number of observations. Kappa is a similar measure of accuracy as total
accuracy but adjusts the value with the probability of correct classifi­ The models for the forest attributes all explained over 70% of the
cation due to randomness (Cohen 1960). Kappa is therefore a more variation for the attributes AGB, BGB, H and V for all regions (Fig. 4). For
robust and comparable measure than accuracy. the attributes CC, N and GINI the explanatory power was smaller
In the case of binary classification of forest/non-forest and natural­ (22–64%). H was the attribute that showed the smallest RMSD and
ness we used the accuracy measures precision and recall (Goutte and largest correlation in all regions. For H, between 81 and 90% of the
Gaussier 2005). Precision was calculated by the number of true positives variation was explained using ALS data. The RRMSD value were be­
of the number classified as positive observations (true positives plus tween 12 and 17% for this attribute. The RRMSD values for V (37–62%)
false positives), i.e. similar to the user’s accuracy. Recall, on the other were slightly larger than for biomass (AGB and BGB).
hand, was computed as the number of true positives of the number of The classification of natural forests showed large to moderate accu­
positive observations in the reference data, i.e. like the producer’s ac­ racy for all definitions (77–98%) with the exception of the definition
curacy. The harmonic average of precision and recall is used to evaluate based on age (D2 and D3, 14–89%) where there was no correlation
these two measures together and is referred to as F1-score. (kappa values less than 0.05). Kappa values ranged from 0.33 to 0.93 for
Continuous attributes were evaluated using root mean square dif­ D1, D4, D5 and D7. The accuracy of natural forests is shown in Fig. 5.
ference (RMSD):
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑n 4.3. Change detection and classification
RMSD = (yi − ̂y i )2 (15)
n i=1
The accuracy of the change detection varied between the different
and the root mean square difference relative to the mean value parts of the country. Eastern and southern regions had the greatest ac­
(RRMSD): curacies. Western and central regions had intermediate accuracy values,
1∑n while the northern region had low accuracies (Fig. 6). The absolute
RRMSD = RMSD yi (16) deviations of the detected year of change and recorded year of change in
n i=1
NFI varied on average between two and three years in all regions except

Fig. 3. Accuracy for classification of forest/non-forest based on Landsat data from 1986 and Sentinel-2 data from 2020.

6
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 4. The correlation coefficient (r2) and RRMSD for basic forest attributes in different regions.

the northern region, in which the deviation was on average 4.7 years. estimates of the extent, condition and pressures on the forest ecosystems
The median absolute deviation was one year in the southern region, were outlined. The current version of the forest ecological base map
while it was three years in the northern region. The difference between builds on well-known methods widely used in operational forest man­
the detected year for change and the change registered by the NFI was agement inventories. Maps with country-wide coverage of attributes
mainly less than one year. such as H, V and AGB are already available in several countries (Nord-
Change classification was made according to the land cover classes. Larsen and Schumacher 2012; Nilsson et al., 2017; Astrup et al. 2019).
The accuracy varied from 70 to 92% and with a kappa from 0.57 to 0.89 The current study elaborates on this to support forest condition attri­
(Fig. 7). There was basically no geographical trend, but the collection of butes such as crown coverage, the GINI diversity index and naturalness.
reference data and the distribution of classes within the country played a Furthermore, the framework of mapping the forest extent and pressures
role in the final accuracy. The accuracy values for the individual classes for supporting monitoring. The objective was not to provide specific
were large for forests and water (Fig. 8). In contrast, cropland and other- indicators related to ecosystem services, ecosystem condition or specific
land were classified with medium accuracy, while classes that occur less national or international reporting systems (O’Connor et al., 2015; Maes
frequently, such as grassland, settlements and peatland, had smaller et al. 2016), but by mapping and estimating the essential forest
accuracy. ecosystem information, such indicators might be directly or indirectly
derived from the maps and the estimates. Specifically, the use of local-
4.4. Mapping and estimation level estimates opens up for calculating ecosystem indicators at a
spatial resolution not demonstrated earlier. This spatial level represents
The estimates of forest condition and pressure at the 100 km2 blocks an information level between detailed maps and national or regional
are presented in Fig. 9 and the respective standard errors appear in statistics provided by the NFI. Thus, providing a new tool for the
Fig. 10. These maps provide new insights in how the forest condition and comprehensive management of forest ecosystems, especially in conser­
pressures are distributed geographically. Areas with large AGB and V vation management. Many proposed condition attributes are included in
density appear in the same locations and are independent of, e.g. H. The inventories carried out for forest management planning purposes.
GINI was smaller in the most actively managed forest areas. However, such forest management plans often focus on timber produc­
The standard errors in Fig. 10 show a pattern that some local areas tion and, to a limited degree, other ecosystem conditions such as natu­
have larger errors than others. Likely, reasons for this are individual ralness. Furthermore, forest management plans have limitations in area
laser projects where the number of sample plots will influence the coverage and their accessibility to the public might be restricted. Finally,
estimated intercept of random effect. the inventory methods used may vary between districts or forest prop­
erties. Consequently, the possibility of aggregating the information
5. Discussion utilizing statistical sound estimators is limited.
To ensure proper mapping and monitoring of pressures on the forest
A methodology and framework for providing national maps and ecosystem, we established a reference level of the forest extent in 1986

7
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 5. Precision, recall, F1, accuracy and kappa for natural forests for different definitions and regions.

and an up-to-date forest extent. Combining these two ecosystem extents delineation of the forest ecosystems’ extent was that mountain areas
was necessary to map changes and changes classes. The 1986 Landsat were often classified as forests and seedling stands were classified as
map showed smaller accuracies but combined with the change detection cropland. Nevertheless, the accuracies obtained are in the order of what
and the 2020 Sentinel-2 land cover map, the influence of the lower ac­ have been presented in other studies (Wilkinson 2005).
curacy were minimized. The main challenge with the classification and The accuracy obtained for forest condition attributes was in general

8
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 6. Accuracy between detection in LandTrendr and registered final harvest in the NFI data.

Fig. 7. Accuracy and kappa for classification by land cover classes based on Sentinel-2 for different regions.

similar to already presented in the literature (Nilsson et al., 2017; Rahlf using ALS data (Næsset 2002).
et al. 2017; Astrup et al. 2019). Nevertheless, the models developed for Sverdrup-Thygeson et al. (2016) investigated the prediction of
the western region were not as good as those developed in other regions. naturalness, defined as the mature forest that had not previously been
The steep terrain in the western region might have influenced the re­ harvested by clear-cutting. In a study area in eastern Norway, they found
sults. The effect of steep terrain has been investigated in other studies that 86–94% of the localities were correctly classified using ALS and the
and an increasing error has been found with increasing terrain slope kappa values obtained were between 0.66 and 0.84. The definition used
(Hansen et al. 2017; Ørka et al. 2018). Hansen et al. (2017) presented by Sverdrup-Thygeson et al. (2016) is most similar to D7 in the current
methods for correcting for slope. However, these methods are relatively work where 77–91% of the sample plots were correctly classified and
computationally demanding and, therefore, time-consuming to imple­ with kappa values ranging from 0.45 to 0.75. Several factors suggest that
ment for large areas. CC was difficult to model for the southern region the accuracy in large-area applications will be smaller than in controlled
where mixed conifer and broadleaved species forests are common. studies covering a small geographical extent. First, a single homogenous
Studies have shown that mixtures of conifer and broadleaved species ALS acquisition at just a single point in time will tend to remove vari­
with a large portion of the minority species will tend to reduce overall ation related to systematic sensor differences and the phenological stage
accuracy (Næsset 2005). To our knowledge, no study has examined the of the vegetation (Næsset 2005; Ørka et al. 2010). Second, a small
accuracy of modelling of CC in mixed forests in Norway. However, the geographical extent also narrows down differences in how the forests
tendency of smaller accuracies obtained in the broadleaved forests have developed based on natural factors and human influence such as
compared to conifer forests are broadly recognized (Latifi et al. 2015). forest management. In local studies these factors are better captured in
Furthermore, several studies have shown that N is challenging to model the modelling by a higher field sampling intensity per area unit. Natural

9
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 8. Producer’s accuracy (PA) and user’s accuracy (UA) for different land cover classes.

factors will significantly impact predictions in larger areas because the calculating age from height and site index. Unfortunately, these methods
tree and forest structure will change across the landscape when would probably fail when used for areas with high naturalness due to the
approaching, for example, the alpine and tundra tree lines, or as wind general trend of decreasing height growth with increasing age.
and precipitation factors change along, for example, oceanic and con­ A methodological challenge in predicting the probability of natural
tinental climate gradients. Third, the study by Sverdrup-Thygeson et al. forests using data from ALS was to account for the random effect asso­
(2016) was also limited to older forests, while no such stratification was ciated with ALS acquisition. A small number of sample plots with natural
used at the national level in the current study. Finally, the size of the forests means that numerous individual acquisitions might not cover any
sample plots was eight times larger in the study by Sverdrup-Thygeson natural forest at all. On the other hand, other acquisitions, for example
et al. (2016) and the plots were located to minimize edge effects. in large conservation areas, might have many sample plots defined as
Ahlkrona et al. (2017) examined a survey of old natural forests in natural forests. Thus, a random effect will affect the results and the
Sweden based on data sources such as orthophotos, satellite images and predictions in a negative way. In the current work, the best solution was
ALS data. They followed a similar definition as Sverdrup-Thygeson et al. to ignore the effect of acquisition when modelling naturalness. However,
(2016) and similar change detection procedures as adopted in the cur­ when modelling the other forest condition attributes, a random effect of
rent study and included older orthophotos and satellite images. Vali­ the acquisition was included.
dation against stands in different stages of stand development in The long time series of the Landsat archive enables change detection
Värmland provided an accuracy of 80%, a kappa value of 0.28, a recall at 30 m spatial resolution back to 1986, which is a few years before the
of 0.40, a precision of 0.41 and an F1 score value of 0.40. These numbers establishment of the permanent sample plots of the Norwegian NFI.
were calculated based on stand-level data and cannot be compared Thus, although inaccuracies in the detection are present, Landsat pro­
directly with the numbers reported in the current work. They never­ vides the most accurate data describing a long time series of pressures on
theless indicate that the accuracy was similar to the accuracy obtained in the forest ecosystems in Norway. Inaccuracies may be improved by
the current study. Furthermore, they made a random interpretation of optimizing parameter settings, e.g. northern and mountain areas where
the classification of continuity forest against orthophotos from different alpine birch forests dominate. Moreover, clouds and possible insect in­
times. This validation showed an accuracy of 53–88% for areas >10 ha. festations in birch trees can indicate changes that do not correspond to
In the current study an slightly lower accuracy (46–69%) was found on changes related to logging that the NFI records. Because of the present
the sample plot level (250 m2). Both the current study and Ahlkrona accuracies, caution should be exercised when calculating indicators of
et al. (2017) pointed out a similar misclassification. In the Swedish forest change (Palahí et al. 2021). In the current study, we applied a
survey, a large proportion of managed forests was classified as high post-stratified estimator to ensure unbiasedness for the region but not on
naturalness. For example, in Värmland, areas mapped as continuity the local level. Systematic differences may be present at the same level
forests, but which were in fact managed forests, accounted for 60% of for the local level blocks as the correction factor will be identical.
the area. Similarly, in the current study all naturalness definitions The proposed framework in this study was applied in Norway, but
showed the same pattern with a high recall value, but a low precision applicable in other countries. The framework utilized freely available
value. Thus, forest with naturalness is classified correctly, but large satellite data for ecosystem delineation and to detect and classify
managed forest areas values appear with a high probability of natural­ ecosystem pressures. However, information on the forest condition re­
ness. Since the objective is to identify natural forests, this will be a better lies on available ALS and NFI data. The NFI data collection needs to be
starting point than the alternative misclassification where many forests carried out to allow the data to be used together with remotely sensed
with high naturalness are not identified. data. Thus, sample plots must be geolocated with very good accuracy
Forest age is an essential indicator of naturalness. Nevertheless, age and have a suitable plot size (Gobakken and Næsset 2009; Packalen et al.
is an attribute that is generally difficult to predict, especially for older 2019).
age groups, even when both spectral data and ALS data are used (Nie­
mann 1995; Racine et al., 2014; Kandare et al. 2017). This was also the 6. Conclusions
case for definitions based on age in the current study where results with
small accuracy and precision were obtained. Recently developed We have presented a framework for a RS-based forest ecological base
methods for calculating site index using repeated ALS data from map. Such a map can provide fine spatial resolution information and
different points in time (Noordermeer et al. 2020) open up for support calculation of statistically rigorous estimates about the extent,

10
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 9. Estimates at local level (100 km2) for forest condition attributes (AGB, BGB, CC, GINI, H, N, V, NAT (Naturalness definition D7)) and forest pres­
sures (Change).

condition and pressures on the forest ecosystems. The framework was probability-based maps enable user-defined prioritization, e.g. when
used to create the first version of a forest ecological base map for Nor­ using the probability maps as pre-information for more detailed field
way. The combined use of Sentinel-2, Landsat and ALS data enables surveys. The changes detection and change classification are the cor­
accurate forest extent and condition maps. Methods for naturalness nerstones in monitoring and understanding the pressures on the eco­
provided many false positives, but probability-based maps enable systems during that period. The presented framework establishes model-
communicating the uncertainty of such maps. Furthermore, the based estimates at a local level to support indicator development of

11
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

Fig. 10. Estimates of standard errors at local level (100 km2) for forest condition attributes (AGB, BGB, CC, GINI, H, N, V, NAT (Naturalness definition D7) and forest
pressures (Change).

forest condition and pressures on a detailed geographical level. Further adapted and used in other countries having high-quality forest infor­
research should focus on including additional map layers presenting mation on accurately geolocated sample plots and available remotely
attributes such as site productivity, species – both native and non-native sensed data.
and non-timber provisional services such as berries and mushrooms.
Many of these attributes can be predicted and estimated using the same
framework and data as described here. The framework can also be

12
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

CRediT authorship contribution statement Fragal, E.H., Silva, T.S.F., Novo, E.M.L. de M., 2016. Reconstructing historical forest
cover change in the Lower Amazon floodplains using the LandTrendr algorithm. Acta
Amazonica 46, 13–24. https://doi.org/10.1590/1809-4392201500835.
Hans Ole Ørka: Conceptualization, Methodology, Formal analysis, Gobakken, T., Næsset, E., 2009. Assessing effects of positioning errors and sample plot
Investigation, Writing – original draft, Visualization, Project adminis­ size on biophysical stand properties derived from airborne laser scanner data.
tration, Funding acquisition. Marie-Claude Jutras-Perreault: Formal Canadian journal of forest research. Journal canadien de la recherche forestiere 39
(5), 1036–1052. https://doi.org/10.1139/X09-025.
analysis, Investigation, Writing – review & editing. Erik Næsset: Goutte, C., Gaussier, E., 2005. A probabilistic interpretation of precision, recall and F-
Writing – review & editing, Supervision, Funding acquisition. Terje Score, with implication for evaluation. Adv. Information Retrieval 345–359. https://
Gobakken: Methodology, Writing – review & editing, Supervision, doi.org/10.1007/978-3-540-31865-1_25.
Griffiths, P., Kuemmerle, T., Kennedy, R.E., Abrudan, I.V., Knorn, J., Hostert, P., 2012.
Project administration, Funding acquisition. Using annual time-series of Landsat images to assess the effects of forest restitution
in post-socialist Romania. Remote Sens. Environ. 118, 199–214. https://doi.org/
Declaration of Competing Interest 10.1016/j.rse.2011.11.006.
Hansen, E.H., Ene, L.T., Gobakken, T., Ørka, H.O., Bollandsås, O.M., Næsset, E., 2017.
Countering negative effects of terrain slope on airborne laser scanner data using
The authors declare that they have no known competing financial procrustean transformation and histogram matching. Forests 8, 401. https://doi.org/
interests or personal relationships that could have appeared to influence 10.3390/f8100401.
Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.A., Tyukavina, A.,
the work reported in this paper. Thau, D., Stehman, S.V., Goetz, S.J., Loveland, T.R., Kommareddy, A., Egorov, A.,
Chini, L., Justice, C.O., Townshend, J.R.G., 2013. High-resolution global maps of
Acknowledgment 21st-century forest cover change. Science 342 (6160), 850–853.
Hengl, T., 2006. Finding the right pixel size. Comput. Geosci. 32 (9), 1283–1298. https://
doi.org/10.1016/j.cageo.2005.11.008.
The Norwegian Environment Agency funded this study as a part of Hyde, P., Dubayah, R., Walker, W., Blair, J.B., Hofton, M., Hunsaker, C., 2006. Mapping
the project “Remote sensing based mapping and monitoring of the forest forest structure for wildlife habitat analysis using multi-sensor (LiDAR, SAR/InSAR,
ETM+, Quickbird) synergy. Remote Sens. Environ. 102, 63–73.
ecosystem” [grant number N◦18087221]. Furthermore, the study
Hyyppä, J., Hyyppä, H., Inkinen, M., Engdahl, M., Linko, S., Zhu, Y.H., 2000. Accuracy
received funding under the umbrella of ERA-NET Cofund ForestValue comparison of various remote sensing data sources in the retrieval of forest stand
project NOBEL, “Novel business models and mechanisms for the sus­ attributes. For. Ecol. Manage. 128, 109–120.
tainable supply of and payment for forest ecosystem services”. Forest­ Jacobsen A.Z., Jabot J., Holmengen N., Ekre T.H., Rasch M.K., Lillesund V.F. et al.
(2020). Klimagassregnskap for kommuner og fylker - Dokumentasjon av metode –
Value was funded by the European Union’s Horizon 2020 research and versjon 3. Miljødirektoratet.
innovation program [grant number N◦773324]. We also want to thank Jakobsson, S., Evju, M., Framstad, E., Imbert, A., Lyngstad, A., Sickel, H., Sverdrup-
two anonymous reviewers for the valuable comments on an earlier Thygeson, A., Töpper, J.P., Vandvik, V., Velle, L.G., Aarrestad, P.A., Nybø, S., 2021.
Introducing the index-based ecological condition assessment framework (IBECA).
manuscript version. Ecol. Ind. 124, 107252. https://doi.org/10.1016/j.ecolind.2020.107252.
Jutras-Perreault, M.-C., Gobakken, T., Ørka, H.O., 2021. Comparison of two algorithms
References for estimating stand-level changes and change indicators in a boreal forest in
Norway. Int. J. Appl. Earth Obs. Geoinf. 98, 102316. https://doi.org/10.1016/j.
jag.2021.102316.
Ahlkrona, E., Giljam, C., Wennberg, S., Jönsson, C., Alvarez, M., Bovin, M., et al., 2017.
Kandare, K., Ørka, H.O., Dalponte, M., Næsset, E., Gobakken, T., 2017. Individual tree
Kartering av kontinuitetsskog i boreal region. Metria AB på uppdrag av
crown approach for predicting site index in boreal forests using airborne laser
Naturvårdsverket.
scanning and hyperspectral data. Int. J. Appl. Earth Obs. Geoinf. 60, 72–82. https://
Ahlstrøm, A., Bjørkelo, K., Fadnes, K.D., 2019. AR5 Klassifikasjonssystem. NIBIO.
doi.org/10.1016/j.jag.2017.04.008.
Anderson, J.R., Hardy, E.E., Roach, J.T., Witmer, R.E., 1976. A land use and land cover
Kangas, A., Astrup, R., Breidenbach, J., Fridman, J., Gobakken, T., Korhonen, K.T., et al.,
classification system for use with remote sensor data. Geological Survey Professional
2018. Remote sensing and forest inventories in Nordic countries–roadmap for the
Paper 964, 28. https://doi.org/10.3133/pp964.
future. Scandinavian J. Forest Res. / issued bimonthly by the Nordic Forest Research
Andrew, M.E., Wulder, M.A., Nelson, T.A., 2014. Potential contributions of remote
Cooperation Committee 33, 397–412.
sensing to ecosystem service assessments. Prog. Phys. Geography: Earth Environ. 38
Kennedy, R.E., Yang, Z., Cohen, W.B., 2010. Detecting trends in forest disturbance and
(3), 328–353. https://doi.org/10.1177/0309133314528942.
recovery using yearly Landsat time series: 1. LandTrendr—Temporal segmentation
Angermeier, P.L., 2000. The natural imperative for biological conservation. Conserv.
algorithms. Remote Sens. Environ. 114, 2897–2910.
Biol. 14 (2), 373–381. https://doi.org/10.1046/j.1523-1739.2000.98362.x.
Kennedy R.E., Yang Z., Gorelick N., Braaten J., Cavalcante L., Cohen W.B. et al. (2018).
Astrup, R., Rahlf, J., Bjørkelo, K., Debella-Gilo, M., Gjertsen, A.-K., Breidenbach, J., 2019.
Implementation of the LandTrendr Algorithm on Google Earth Engine. Remote
Forest information at multiple scales: development, evaluation and application of
Sensing 10.
the Norwegian forest resources map SR16. Scandinavian J. Forest Res. / issued
Kerr, J.T., Ostrovsky, M., 2003. From space to species: ecological applications for remote
bimonthly by the Nordic Forest Research Cooperation Committee 34 (6), 484–496.
sensing. Trends Ecol. Evol. 18 (6), 299–305. https://doi.org/10.1016/S0169-5347
https://doi.org/10.1080/02827581.2019.1588989.
(03)00071-5.
Breidenbach, J., Granhus, A., Hylen, G., Eriksen, R., Astrup, R., 2020. A century of
Knoke, T., Kindu, M., Schneider, T., Gobakken, T., 2021. Inventory of forest attributes to
National Forest Inventory in Norway – informing past, present, and future decisions.
support the integration of non-provisioning ecosystem services and biodiversity into
Forest Ecosystems 7, 46. https://doi.org/10.1186/s40663-020-00261-0.
forest planning—from collecting data to providing information. Curr. Forestry Rep. 7
Breiman, L., 2001. Random forests. Machine Learning 45, 5–32.
(1), 38–58. https://doi.org/10.1007/s40725-021-00138-7.
Bryn, A., Strand, G.-H., Angeloff, M., Rekdal, Y., 2018. Land cover in Norway based on an
Latifi, H., Fassnacht, F.E., Hartig, F., Berger, C., Hernández, J., Corvalán, P., Koch, B.,
area frame survey of vegetation types. Norsk Geografisk Tidsskrift – Norwegian J.
2015. Stratified aboveground forest biomass estimation by remote sensing data. Int.
Geography 72 (3), 131–145. https://doi.org/10.1080/00291951.2018.1468356.
J. Appl. Earth Obs. Geoinf. 38, 229–241. https://doi.org/10.1016/j.
Burkhard B., Maes J., 2017. Mapping ecosystem services. Advanced books 1: e12837.
jag.2015.01.016.
Cohen, J., 1960. A coefficient of agreement for nominal scales. Educ. Psychol. Measur.
Layke, C., 2009. Measuring Nature’s Benefits: A Preliminary Roadmap for Improving
20, 37–46.
Ecosystem Service Indicators. World Resources Institute, Washington, pp. 1–36.
Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely
Lexerød, N.L., Eid, T., 2006. An evaluation of different diameter diversity indices based
sensed data. Remote Sens. Environ. 37 (1), 35–46. https://doi.org/10.1016/0034-
on criteria related to forest management planning. For. Ecol. Manage. 222 (1-3),
4257(91)90048-B.
17–28. https://doi.org/10.1016/j.foreco.2005.10.046.
Costanza, R., d’Arge, R., de Groot, R., Farber, S., Grasso, M., Hannon, B., Limburg, K.,
Ling, C.X., Li, C., 1998. Data mining for direct marketing: problems and solutions. Kdd
Naeem, S., O’Neill, R.V., Paruelo, J., Raskin, R.G., Sutton, P., van den Belt, M., 1997.
98, 73–79.
The value of the world’s ecosystem services and natural capital. Nature 387 (6630),
Maes, J., Liquete, C., Teller, A., Erhard, M., Paracchini, M.L., Barredo, J.I., Grizzetti, B.,
253–260. https://doi.org/10.1038/387253a0.
Cardoso, A., Somma, F., Petersen, J.-E., Meiner, A., Gelabert, E.R., Zal, N.,
Curtis, P.G., Slay, C.M., Harris, N.L., Tyukavina, A., Hansen, M.C., 2018. Classifying
Kristensen, P., Bastrup-Birk, A., Biala, K., Piroddi, C., Egoh, B., Degeorges, P.,
drivers of global forest loss. Science 361 (6407), 1108–1111.
Fiorina, C., Santos-Martín, F., Naruševičius, V., Verboven, J., Pereira, H.M.,
Esteban, J., McRoberts, R.E., Fernández-Landa, A., Tomé, J.L., Nӕsset, E., 2019.
Bengtsson, J., Gocheva, K., Marta-Pedroso, C., Snäll, T., Estreguil, C., San-Miguel-
Estimating forest volume and biomass and their changes using random forests and
Ayanz, J., Pérez-Soba, M., Grêt-Regamey, A., Lillebø, A.I., Malak, D.A., Condé, S.,
remotely sensed data. Remote Sensing 11, 1944. https://doi.org/10.3390/
Moen, J., Czúcz, B., Drakou, E.G., Zulian, G., Lavalle, C., 2016. An indicator
rs11161944.
framework for assessing ecosystem services in support of the EU Biodiversity
Feld, C.K., Sousa, J.P., da Silva, P.M., Dawson, T.P., 2010. Indicators for biodiversity and
Strategy to 2020. Ecosyst. Serv. 17, 14–23. https://doi.org/10.1016/j.
ecosystem services: towards an improved framework for ecosystems assessment.
ecoser.2015.10.023.
Biodivers. Conserv. 19 (10), 2895–2919. https://doi.org/10.1007/s10531-010-
Maes, J., Teller, A., Erhard, M., Grizzetti, B., Barredo, J.I., Paracchini, M.L., et al., 2018.
9875-0.
Mapping and assessment of ecosystems and their services : an analytical framework
Foody G.M., 1996. Fuzzy modelling of vegetation from remotely sensed imagery.
Ecological modelling 85: 3–12. doi: 10.1016/0304-3800(95)00012-7.

13
H.O. Ørka et al. Ecological Indicators 136 (2022) 108636

for mapping and assessment of ecosystem condition in EU : discussion paper. Palahí, M., Valbuena, R., Senf, C., Acil, N., Pugh, T.A.M., Sadler, J., Seidl, R., Potapov, P.,
Publications Office of the European Union. Gardiner, B., Hetemäki, L., Chirici, G., Francini, S., Hlásny, T., Lerink, B.J.W.,
Maltamo, M., Packalen, P., 2014. Species-specific management inventory in Finland. In: Olsson, H., González Olabarria, J.R., Ascoli, D., Asikainen, A., Bauhus, J.,
Maltamo, M., Næsset, E., Vauhkonen, J. (Eds.), Forestry Applications of Airborne Berndes, G., Donis, J., Fridman, J., Hanewinkel, M., Jactel, H., Lindner, M.,
Laser Scanning – Concepts and Case Studies. Springer, Dordrecht-Heidelberg-New Marchetti, M., Marušák, R., Sheil, D., Tomé, M., Trasobares, A., Verkerk, P.J.,
York-London, pp. 241–252. Korhonen, M., Nabuurs, G.-J., 2021. Concerns about reported harvests in European
McRoberts, R.E., 2011. Satellite image-based maps: scientific inference or pretty forests. Nature 592 (7856), E15–E17. https://doi.org/10.1038/s41586-021-03292-
pictures? Remote Sens. Environ. 115 (2), 715–724. https://doi.org/10.1016/j. x.
rse.2010.10.013. Pinheiro, J.C., Bates, D.M., 2000. Mixed-effects Models in S and S-PLUS. Springer, New
McRoberts, R.E., Winter, S., Chirici, G., LaPoint, E., 2012. Assessing forest naturalness. York, p. 528p.
Forest Sci. 58 (3), 294–309. https://doi.org/10.5849/forsci.10-075. Racine, E.B., Coops, N.C., St-Onge, B., Bégin, J., 2014. Estimating forest stand age from
McRoberts, R.E., Næsset, E., Gobakken, T., Chirici, G., Condés, S., Hou, Z., Saarela, S., LiDAR-derived predictors and nearest neighbor imputation. Forest Sci. 60 (1),
Chen, Q.i., Ståhl, G., Walters, B.F., 2018. Assessing components of the model-based 128–136. https://doi.org/10.5849/forsci.12-088.
mean square error estimator for remote sensing assisted forest applications. Can. J. Rahlf, J., Breidenbach, J., Solberg, S., Næsset, E., Astrup, R., 2017. Digital aerial
Forest Res. Journal canadien de la recherche forestiere 48 (6), 642–649. https://doi. photogrammetry can efficiently support large-area forest inventories in Norway.
org/10.1139/cjfr-2017-0396. Forestry 90 (5), 710–718. https://doi.org/10.1093/forestry/cpx027.
Næsset, E., 2002. Predicting forest stand characteristics with airborne scanning laser Senf, C., Pflugmacher, D., Wulder, M.A., Hostert, P., 2015. Characterizing
using a practical two-stage procedure and field data. Remote Sens. Environ. 80 (1), spectral–temporal patterns of defoliator and bark beetle disturbances using Landsat
88–99. https://doi.org/10.1016/S0034-4257(01)00290-5. time series. Remote Sens. Environ. 170, 166–177. https://doi.org/10.1016/j.
Næsset, E., 2004a. Accuracy of forest inventory using airborne laser scanning: evaluating rse.2015.09.019.
the first nordic full-scale operational project. Scandinavian J. Forest Res. / issued Søgaard G., Eriksen R., Astrup R.A., Øyen B.-H., 2012. Effekter av ulike miljøhensyn på
bimonthly by the Nordic Forest Research Cooperation Committee 19 (6), 554–557. tilgjengelig skogareal og volum i norske skoger. Norsk institutt for skog og landskap.
https://doi.org/10.1080/02827580410019544. Ståhl, G., Saarela, S., Schnell, S., Holm, S., Breidenbach, J., Healey, S.P., Patterson, P.L.,
Næsset, E., 2004b. Practical large-scale forest stand inventory using a small-footprint Magnussen, S., Næsset, E., McRoberts, R.E., Gregoire, T.G., 2016. Use of models in
airborne scanning laser. Scandinavian J. Forest Res. / issued bimonthly by the large-area forest surveys: comparing model-assisted, model-based and hybrid
Nordic Forest Research Cooperation Committee 19 (2), 164–179. https://doi.org/ estimation. Forest Ecosystems 3 (1). https://doi.org/10.1186/s40663-016-0064-9.
10.1080/02827580310019257. Stehman, S.V., 2013. Estimating area from an accuracy assessment error matrix. Remote
Nasset, E., 2004. Effects of different flying altitudes on biophysical stand properties Sens. Environ. 132, 202–211. https://doi.org/10.1016/j.rse.2013.01.016.
estimated from canopy height and density measured with a small-footprint airborne Storaunet, K.O., Rolstad, J., 2015. Mengde og utvikling av død ved i produktiv skog i
scanning laser. Remote Sens. Environ. 91 (2), 243–255. Norge. Med basis i data fra Landsskogtakseringens 7.(1994-1998) og 10. takst (2010-
Næsset, E., 2005. Assessing sensor effects and effects of leaf-off and leaf-on canopy 2013). Oppdragsrapport fra Skog og Landskap, Norsk institutt for skog og landskap.
conditions on biophysical stand properties derived from small-footprint airborne Sverdrup-Thygeson, A., Ørka, H.O., Gobakken, T., Næsset, E., 2016. Can airborne laser
laser data. Remote Sens. Environ. 98, 356–370. scanning assist in mapping and monitoring natural forests? For. Ecol. Manage. 369,
Næsset, E., 2014. Area-based inventory in Norway–from innovation to an operational 116–125. https://doi.org/10.1016/j.foreco.2016.03.035.
reality. In: Næsset, E., Vauhkonen, J. (Eds.), Maltamo, M. Forestry Applications of Uotila, A., Kouki, J., Kontkanen, H., Pulkkinen, P., 2002. Assessing the naturalness of
Airborne Laser Scanning, Springer, pp. 215–240. boreal forests in eastern Fennoscandia. For. Ecol. Manage. 161 (1-3), 257–277.
Nelson, R.F., 1983. Detecting forest canopy change due to insect activity using Landsat https://doi.org/10.1016/S0378-1127(01)00496-0.
MSS. Photogramm. Eng. Remote Sens. 49, 1303–1314. Valbuena, R., Mauro, F., Rodriguez-Solano, R., Manzanera, J.A., 2010. Accuracy and
Niemann, K.O., 1995. Remote sensing of fores stand age using airborne spectrometer precision of GPS receivers under forest canopies in a mountainous environment.
data. Photogramm. Eng. Remote Sens. Spanish journal of agricultural research = Revista de investigacion agraria 8:
Nilsson, M., Nordkvist, K., Jonzén, J., Lindgren, N., Axensten, P., Wallerman, J., 1047–1057. doi: 10.5424/sjar/2010084-1242.
Egberth, M., Larsson, S., Nilsson, L., Eriksson, J., Olsson, H., 2017. A nationwide Valbuena, R., Vauhkonen, J., Packalen, P., Pitkänen, J., Maltamo, M., 2014. Comparison
forest attribute map of Sweden predicted using airborne laser scanning data and field of airborne laser scanning methods for estimating forest structure indicators based
data from the National Forest Inventory. Remote Sens. Environ. 194, 447–454. on Lorenz curves. ISPRS J. Photogrammetry Remote Sens. 95, 23–33. https://doi.
https://doi.org/10.1016/j.rse.2016.10.022. org/10.1016/j.isprsjprs.2014.06.002.
Noordermeer, L., Gobakken, T., Næsset, E., Bollandsås, O.M., 2020. Predicting and Vauhkonen, J., 2018. Predicting the provisioning potential of forest ecosystem services
mapping site index in operational forest inventories using bitemporal airborne laser using airborne laser scanning data and forest resource maps. Forest Ecosystems 5,
scanner data. For. Ecol. Manage. 457, 117768. https://doi.org/10.1016/j. 24. https://doi.org/10.1186/s40663-018-0143-1.
foreco.2019.117768. Viken, K.O., 2018. Landsskogtakseringens feltinstruks 2018. NIBIO,.
Nord-Larsen, T., Schumacher, J., 2012. Estimation of forest resources from a country Vogeler, J.C., Braaten, J.D., Slesak, R.A., Falkowski, M.J., 2018. Extracting the full value
wide laser scanning survey and national forest inventory data. Remote Sens. of the Landsat archive: Inter-sensor harmonization for the mapping of Minnesota
Environ. 119, 148–157. https://doi.org/10.1016/j.rse.2011.12.022. forest canopy cover (1973–2015). Remote Sens. Environ. 209, 363–374. https://doi.
O’Connor, B., Secades, C., Penner, J., Sonnenschein, R., Skidmore, A., Burgess, N.D., org/10.1016/j.rse.2018.02.046.
Hutton, J.M., 2015. Earth observation as a tool for tracking progress towards the Wilkinson, G.G., 2005. Results and implications of a study of fifteen years of satellite
Aichi Biodiversity Targets. Remote Sens. Ecol. Conserv. 1 (1), 19–28. https://doi. image classification experiments. IEEE Trans. Geosci. Remote Sens. 43, 433–440.
org/10.1002/rse2.4. Woodcock, C.E., Allen, R., Anderson, M., Belward, A., Bindschadler, R., Cohen, W., et al.,
Ørka, H.O., Næsset, E., Bollandsås, O.M., 2010. Effects of different sensors and leaf-on 2008. Free access to Landsat imagery. Science 320, 1011.
and leaf-off canopy conditions on echo distributions and individual tree properties Wu, C.F.J., 1986. Jackknife, Bootstrap and other resampling methods in regression
derived from airborne laser scanning. Remote Sens. Environ. 114 (7), 1445–1461. analysis. Ann. Stat. 14, 1261–1295.
https://doi.org/10.1016/j.rse.2010.01.024. Wulder, M., 1998. Optical remote-sensing techniques for the assessment of forest
Ørka, H.O., Bollandsås, O.M., Hansen, E.H., Næsset, E., Gobakken, T., 2018. Effects of inventory and biophysical parameters. Prog. Phys. Geogr. 22, 449–476.
terrain slope and aspect on the error of ALS-based predictions of forest attributes. Wulder, M.A., Hall, R.J., Coops, N.C., Franklin, S.E., 2004. High spatial resolution
Forestry 91, 225–237. https://doi.org/10.1093/forestry/cpx058. remotely sensed data for ecosystem characterization. Bioscience 54, 511–521.
Ørka H.O., Gailis J., Vege M., Gobakken T., Hauglund K. (co-submission). Analysis-ready https://doi.org/10.1641/0006-3568(2004)054[0511:HSRRSD]2.0.CO;2.
satellite data mosaics from Landsat and Sentinel-2 imagery. MethodsX. Wynne, R.H., Oderwald, R.G., Reams, G.A., Scrivani, J.A., 2000. Optical remote sensing
Packalen, P., Strunk, J., Packalen, T., Maltamo, M., Mehtätalo, L., 2019. Resolution for forest area estimation. J. Forest. 98, 31–36.
dependence in an area-based approach to forest inventory with airborne laser Yousefpour, R., Jacobsen, J.B., Thorsen, B.J., Meilby, H., Hanewinkel, M., Oehler, K.,
scanning. Remote Sens. Environ. 224, 192–201. https://doi.org/10.1016/j. 2012. A review of decision-making approaches to handle uncertainty and risk in
rse.2019.01.022. adaptive forest management under climate change. Ann. Forest Sci. 69 (1), 1–15.
https://doi.org/10.1007/s13595-011-0153-4.

14