Upload
leicester
View
1
Download
0
Embed Size (px)
Citation preview
Forest woody biomass classification with satellite-based
radar coherence over 900 000 km2 in Central Siberia
David L.A. Gaveaua,*, Heiko Balztera, Stephen Plummerb
aCentre for Ecology and Hydrology, Monks Wood, Huntingdon, Cambridgeshire PE28 2LS, UKbESA/ESRIN, Via Galileo Galilei, Casella Postale 64, 00044 Frascati, Italy
Received 23 March 2001; received in revised form 20 September 2001; accepted 18 December 2001
Abstract
In the current context of global deforestation and global warming, a wide range of organisations, with local to international
remits, need estimates of forest biomass to assess the state of the World’s forests and their rate of change. The task would be
impossible without space-based Earth observation, which allows the rapid generation of extensive data sets describing land surface
properties. It is the task of remote sensing scientists to interpret these data into a meaningful source of forest information. Here, a fast
and easily automated method for classifying boreal forests in terms of growing stock volume is presented. The work was conducted
as part of the SIBERIA project, which has resulted in the recent publication of a map of forest growing stock volume covering
900 000 km2 in Central Siberia. The paper describes the use of satellite-based radar coherence to differentiate categories of forest
growing stock volume, the application of this method to classify and map Central Siberian forests, and the characterisation of the
forest classes to help in the interpretation. A list of acronyms and abbreviations used in the text is provided in Appendix A.
# 2002 Elsevier Science B.V. All rights reserved.
Keywords: Coherence; Forest growing stock volume; SAR interferometry; Forest classification
1. Introduction
The forests of Russia cover 7.5–7.7 million km2
and represent roughly 20% of the World’s forested
areas (Krankina and Dixon, 1992; FAO, 1999). Eco-
logically, Siberian forests are regarded as a critical
stabiliser of the planetary climate in maintaining
atmospheric gas balance (Bonan et al., 1992) and
are the habitat to a wide variety of plants and animals,
many of which are rare and/or endemic (WWF, 1994).
At the same time, Russia is going through major
political, social and economic transformations and
these forest resources could be used to help revitalise
the Russian economy. Nilsson and Shvidenko (1997)
comment that the world has seen too many forests used
as a short-term cash crop, without regard for the long-
term economic, environmental and social conse-
quences. The challenge is to avoid the same in Russia.
The SAR Imaging for Boreal Ecology and Radar
Interferometry Applications (SIBERIA) project was
initiated in this context (Schmullius and Rosenqvist,
1997).
Up-to-date information on forest status in the Cen-
tral Siberian region (Fig. 1) was not generally avail-
able or comprehensive until the SIBERIA project
published the first large-scale (900 000 km2) ‘forest
growing stock volume’ map of Central Siberia (Fig. 2).
The map was completed in October 2000 both as a
Forest Ecology and Management 174 (2003) 65–75
* Corresponding author. Tel.: þ44-1487-772-483;
fax: þ44-1487-773-467.
E-mail address: [email protected] (D.L.A. Gaveau).
0378-1127/02/$ – see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 1 1 2 7 ( 0 2 ) 0 0 0 2 8 - 2
digital data product, and also printed as 123 map
sheets on a scale of 1:200 000. These printed map
sheets will serve those Russian Forest Enterprises
without GIS capabilities, as background information
for sustainable forest management. The map also
provides a baseline for an accurate assessment of
future rates of change of growing stock volume (a
proxy to woody biomass) required to quantify the
amount of carbon either released into or absorbed
from the atmosphere (Nilsson and Shvidenko, 1997).
In SIBERIA, JERS-1 and ERS-1/2 SAR images
were used because microwaves offer the possibility to
‘see’ through clouds (Gerstl, 1990) and are more
directly responsive to differences in woody biomass
than from the optical spectrum (i.e. visible and near-
infrared). The satellites provided three types of ima-
gery: two ‘backscatter’ image types at two different
electromagnetic wavelengths, C-band (ERS-1/2) and
L-band (JERS-1), and a ‘coherence’ image, which is
derived by SAR interferometry (Wegmuller and Wer-
ner, 1995). The backscattering variables characterise
the dielectric properties of the imaged land surface.
Coherence measures the degree of correlation between
two radar images of the land surface. These images
were acquired during the ESA Tandem mission by
ERS-1 and ERS-2 from slightly different viewing
geometries and at a 1-day time lag. By comparing
the state of development in space and time of the
periodic electromagnetic fields captured by both
images through an interferometric average, the degree
of correlation (coherence) could be computed. The
apparent change in the separation between two objects
(e.g. two branches) when viewed from different posi-
tions and at different times introduces a random
parameter in the averaging process, which affects
the degree of correlation. As the random fluctuations
Fig. 1. Central Siberia region classified under the SIBERIA project. The grey areas represent 13 territories (comprising 50 test areas), for
which field-based forest inventory data were available.
66 D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75
increase, which correspond to an increasing number
and height of vegetation elements (leaves, branches,
etc.) mainly because of wind-induced movements in
the canopy, the coherence decreases within a range
from 1 to 0. Occasionally, an unusual event (e.g.
landslide, flood, ploughing activities) happening
between the two ERS-1/2 acquisition dates will make
both electromagnetic signals decorrelate strongly over
any land surface so that coherence loses the ability to
differentiate between forest and non-forest. These
sources of error are ignored in this study because they
are unlikely to occur within such a short time lag
(24 h).
Detailed field information on forest status from 50
test sites dispersed across Central Siberia (Fig. 1)
revealed that coherence images were far more sensi-
tive to changes in forest woody biomass (expressed
in growing stock volume) than the C- and L-band
Fig. 2. Overview of the published map of forest growing stock volume covering 900 000 km2 in Central Siberia. Copyright: EC ENV4-CT97-
0743-SIBERIA; A# ESA 97/98, NASDA GBFM, DLR.
D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75 67
backscatter images. Similar conclusions were reached
by Hyyppa et al. (2000) in a study comparing the
accuracies of various remote sensing techniques to
retrieve forest stand attributes. Therefore the pub-
lished map of forest growing stock volume presented
in Fig. 2 was successfully generated because of the
availability of coherence images over Central Siberia.
In contrast, L-band backscatter images were used
chiefly to map water bodies (lakes and rivers) although
they too contributed to a lesser extent to forest growing
stock volume mapping. An illustration of the coher-
ence image responsiveness to changes in woody bio-
mass is given in Fig. 3.
Here, a novel classification method based on coher-
ence images only, which maps Siberian forests into
classes of growing stock volume is presented. The
method is fast and easily automated because it avoids
co-registration issues between ERS-1/2 and JERS-1
images and reduces the data volume. It is applicable in
principle to the entire ESA ERS-1/2 Tandem data
archive. It only works in areas of low topography.
2. Classification of woody biomass in Siberia
2.1. Study area
The central-east Siberian economic region covers
an area of 4.1 million km2. It has a population of
10 million inhabitants of which 7 million are urban
and 3 million are rural citizens. Most of the land is
therefore sparsely populated (0.02 inhabitant ha�1).
The 900 000 km2 area classified under the SIBERIA
project is a mosaic of taiga forest (pine, spruce, fir,
Fig. 3. A coherence image example for a small area (33 � 34 km2) in the Bolshyamurta territory (578N, 928E). The white patches reveal the
presence of clear-cuts having rectangular shapes. The black patches indicate the presence of dense forest. The grey patches represent forested
areas of intermediate density.
68 D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75
larch, cedar, birch and aspen), wetland, open areas and
rivers. Agricultural land is predominantly seen around
the cities of Krasnoyarsk, Irkutsk and Bratsk. The
overall topography is relatively low except in the
southern part approaching the Mongolian border.
2.2. SAR interferometry data
The Tandem mission during which ERS-1 and ERS-
2 imaged the same area in a 1-day repeat mode
was operated over Central Siberia in autumn 1997.
The data were acquired with a mobile receiving
station of DLR in Ulaanbaatar, Mongolia. A total of
366 ERS 1-day coherence images were processed by
DLR covering an area of 90 426 512 ha. Image co-
registration was achieved within 0.05–0.3 pixel accu-
racy depending on topography. The common band
filtering algorithm (Gatelli et al., 1994) was applied
and an interferometric window size of 20 � 4 pixels
was used to maximise the visual contrast between
forest and non-forest. The final coherence images
were geo-corrected and re-sampled to a 50 � 50 m2
pixel size.
2.3. Ground reference data
Field information on forest status in Central Siberia
was provided by IIASA. The forest inventory data set
comprised 50 test areas (each covering between
20 000 and 100 000 ha) grouped under 13 territories
scattered across the study area (Fig. 1). These GIS data
were polygon-based. Each polygon had a set of attri-
butes: land category, relative stocking, growing stock
volume, age, species composition, average height of
stand and average trunk diameter at breast height
(DBH) of stand. The attribute data were derived from
in situ surveys extrapolated to larger areas, using
1:10 000 or 1:20 000 aerial photography. The average
DBH of stand was calculated as the quadratic mean of
DBH of all living trees in a stand. The average height
of stand, being the average height of all trees of the
dominant species was calculated by a regression
curve, which is a function of DBH (Husch et al.,
1993). In the GIS, growing stock volume was recorded
as rounded integer values in steps of 5–20 m3 ha�1,
and in steps of 10 m3 ha�1 for greater values. The
data rarely exceeded 400 m3 ha�1. The forest cate-
gories (forest classes) required by the Russian Forest
Enterprises and by the scientific community were
expressed in growing stock volume and are given in
Table 1.
2.4. Classification strategy
To convert the coherence data into categories of
growing stock volume, the following strategy was
adopted:
1. Extraction of coherence statistics over areas where
ground reference measurements provided infor-
mation on growing stock volume.
2. Extrapolation of the coherence statistics extracted
in (1) providing a classification of the study area
into categories of growing stock volume (i.e. into
classes of growing stock volume).
3. Assessment of the quality of the classes derived in
(2) by comparing them with ground-based esti-
mates of growing stock volume measured over
different areas from those used in (1) to ensure the
independence of the assessment.
2.4.1. GIS pre-processing and image co-registration
To enable the extraction of the coherence statistics
over areas where ground-based estimates of growing
stock volume were available, the forest inventory data
were incorporated in the coherence images. In total 16
test areas of known forest status grouped under five
territories were registered to their respective coher-
ence image by using ground control points. The co-
registration accuracy was better than one pixel and the
boundary area of every polygon was eroded by one
pixel to further minimise misregistration errors.
The territories were selected for their low topogra-
phy to remove the dependence of coherence on slope
Table 1
Desired forest classes expressed in growing stock volume
Forest classes (m3 ha�1)
0 (bare soil)
0 (sparse shrub)
1–20
21–50
51–80
81–130
131–200
>200
D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75 69
(Lee and Liu, 2001). They were dispersed across the
study area in order to quantify the spatial variability of
the coherence statistics for a given class of growing
stock volume.
Since recent clear-cuts were easily identified in a
coherence image, all the polygons of ground reference
were removed through visual inspection where log-
ging had occurred between the inventory update and
the radar acquisition dates (Fig. 4).
To meet the requirements of the Russian Forest
Enterprises and the scientific community (Table 1), the
polygons in the GIS were aggregated into eight groups
of growing stock volume values (m3 ha�1): bare soil,
sparse shrub, 1–20, 21–50, 51–80, 81–130, 131–200
and >200 m3 ha�1.
2.4.2. Coherence signatures
The extracted coherence statistic results are pre-
sented in Table 2. Table 2 shows a decrease in the
coherence parameter (in a range from 1 to 0) with
increasing growing stock volume at the five territories
of known field-based forest status. The mean of the
Gaussian distributed coherence histogram, jgj ranges
from 0.83 to 0.86 over bare soil to 0.26 to 0.40 over the
highest growing stock volume class (>200 m3 ha�1). It
reaches a low of 0.15 over lakes and rivers because the
water surface is constantly moving. The standard
deviation, stdðjgjÞ is smaller over bare soil and rivers
(0.03–0.06) than over forests (0.11–0.18). This is
because forest stands are more heterogeneous (e.g.
greater spatial variation in height) than the bare soil
and river surfaces. The same explains the higher
variability in the mean coherence, jgj of forest stands
across the study area (e.g. 0.35–0.61 for 81–
130 m3 ha�1).
2.4.3. Generalised coherence signatures and
classification
The variability of the coherence statistics across the
study area was removed by weighted averaging the
coherence signatures of Table 2, thus generating a
generalised set of coherence signatures (Table 3). This
generalised set was used in a Gaussian maximum
likelihood classification (MLC) algorithm (Schowen-
gerdt, 1997), thus providing a classification of the
study area into categories of growing stock volume
(i.e. into classes of growing stock volume). Plots
comparing the GIS growing stock volume polygons
versus the identified eight forest classes were obtained
at three independent territories (eight new test areas)
dispersed across the region (Fig. 5). The clusters
derived from the classification correspond to forest
growing stock volume (Fig. 5) but the separation
between the classes 21–50 and 51–80 m3 ha�1 and
between those above 80 m3 ha�1 is poor, which means
that in practice only four classes of growing stock
volume can be identified: bare soil/sparse shrub, 1–20,
21–80 and >80 m3 ha�1. A classification example for
Fig. 4. A coherence image example for a small area in the Ust-Ilimsk territory (588N, 1038E) and corresponding field-based growing stock
volume information. Note the presence of new clear-cuts in the coherence image (bright rectangular patches) not yet identified in the field-
based inventory data (black rectangular patches) showing evidence of logging activities between 1991 and 1997.
70 D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75
a small area in the Chunsky territory based on the
generalised coherence signatures (Table 3) is given in
Fig. 6.
2.4.4. Quality assessment
A commonly used tool for assessing the quality of a
map produced from remotely sensed images is the
confusion matrix (Aronoff, 1982; Foody, 1992). This
matrix is obtained by calculating the area of correspon-
dence between the growing stock volume polygons
of forest inventory data and the classified categories
of growing stock volume. The most intuitive statistic
from the confusion matrix is the overall accuracy, p0,
which is the total area (expressed in percentage pixels)
for which the growing stock volume polygons of forest
inventory data match the classified categories of grow-
ing stock volume. However, p0 depends on the number
of classes and the expected chance agreement, which
makes it impossible to compare p0 values from dif-
ferent classifications. The kappa coefficient, k was
developed by Cohen (1960) to correct for the expected
chance agreement. A k ¼ 0 indicates a pure chance
agreement and a k ¼ 1 indicates a perfect agreement.
For ranked classes like the growing stock volume
classes, a modified k coefficient, k0 was used. k0 is
weighted by the seriousness of the classification error
so that classifying a pixel of class bare soil/sparse
shrub in the forest inventory data as 1–20 m3 ha�1 is
less serious than classifying it as 21–50 m3 ha�1. This
choice was justified by the important overlap between
two adjacent classes (Fig. 5).
Three new territories of known field-based forest
inventory data (Ust-Ilimsk, Ulkansky and Hrebtovsky)
with low topography and dispersed across the study
area (Fig. 1) were selected for an independent assess-
ment of the coherence-based classification method.
Table 2
Coherence signatures (mean and standard deviation) extracted at five territories of known field-based forest status
Territories
Bolshyamurta Primorsky Chunsky A Chunsky B Nishne
gj j � stdðjgjÞForest classes (m3 ha�1)
0 (bare soil) 0.83 � 0.06 0.86 � 0.03 No data 0.85 � 0.05 0.84 � 0.04
0 (sparse shrub) 0.80 � 0.05 0.76 � 0.05 0.80 � 0.04 0.82 � 0.04 No data
1–20 0.54 � 0.14 0.68 � 0.11 0.61 � 0.15 No data 0.80 � 0.09
21–50 0.49 � 0.14 0.61 � 0.11 0.58 � 0.15 0.60 � 0.11 0.63 � 0.14
51–80 0.44 � 0.14 0.54 � 0.12 0.45 � 0.18 0.54 � 0.11 0.56 � 0.12
81–130 0.35 � 0.14 0.50 � 0.13 0.41 � 0.18 0.49 � 0.11 0.61 � 0.14
131–200 0.31 � 0.12 0.39 � 0.13 0.29 � 0.14 0.43 � 0.13 0.40 � 0.14
> 200 0.30 � 0.13 0.35 � 0.12 0.26 � 0.12 0.40 � 0.13 0.32 � 0.11
Lake/river 0.15 � 0.06 0.15 � 0.06 No data 0.15 � 0.06 No data
Acquisition dates (1997) 25 September 9 October 5 October 8 October 26 September
26 September 10 October 6 October 9 October 27 September
Baseline, B? (m) 219.9 180.3 230 187.4 227.2
Tmin � Tmax (8C) 4–15 5–10 0–8 0–8 3–16
Rainfall (mm) up to 4 days before 6 0 No data No data 6
Forest inventory update 1998 1996 1998 1998 1997
Ground data location 578N, 928E 568N, 1028E 588N, 958E 588N, 968E 558N, 1008E
Table 3
Generalised coherence signatures
Forest classes (m3 ha�1) jgj � stdðjgjÞ
0 (bare soil) 0.85 � 0.04
0 (sparse shrub) 0.79 � 0.05
1–20 0.68 � 0.13
21–50 0.53 � 0.13
51–80 0.45 � 0.13
81–130 0.40 � 0.13
131–200 0.33 � 0.13
>200 0.29 � 0.12
D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75 71
Fig. 5. Plots comparing the GIS growing stock volume polygons (at three independent territories) versus eight classes of growing stock
volume derived from the generalised coherence statistics for (a) Ust-Ilimsk, (b) Ulkansky and (c) Hrebtovsky. The box boundaries are the 25-
and 75-percentiles, the line in the box is the median and the whiskers are the extreme values or 1.5 times the interquartile range whichever is
the lowest. The lines outside the whiskers are outliers.
Fig. 6. A classification example based on the generalised coherence statistics for a small area (33 � 33 km2) in the Chunsky territory (588N,
958E). The presence of clear-cuts having rectangular shapes and road access reveal a characteristic logging pattern.
72 D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75
The overall accuracy, p0 and the weighted k coeffi-
cient, k0 were calculated for four classes of growing
stock volume (bare soil/sparse shrub, 1–20, 21–80 and
>80 m3 ha�1). Results are given in Table 4.
2.5. Forest class characterisation
In the literature on forest mapping by remote sen-
sing, there is usually confusion regarding what con-
stitutes ‘forest’ (Singh, 1988). Lack of an accurate
definition of the forest classes of a given map can
result in selective interpretation of the map and can
therefore lead to erroneous conclusions. To ensure the
quality and usefulness of the forest growing stock
volume map published within the context of the
SIBERIA project (Fig. 2), the complementary forest
attributes such as average height of stand, average
DBH of stand and species composition were added to
the eight categories of forest growing stock volume
initially desired by the Russian Forest Enterprises and
the scientific community. These boreal forest statistics
were extracted from the field-based forest inventory
data set at the five territories of Table 2. The results are
shown in Table 5.
As expected, the results in Table 5 show an increase
in average height of stand and average DBH of stand
with increasing growing stock volume. The results
also reveal that the deciduous birch and aspen species
predominate in the low growing stock volume classes,
to gradually give way to the four main conifer species
(pine, spruce, larch, cedar) in the high volume classes.
This trend is due to the fact that birch and aspen are
pioneer species. The rarity of the cedar specie is
noticed.
3. Discussion
The coherence data were classified into eight
classes of growing stock volume over 16 test areas
scattered across 900 000 km2 in Central Siberia
(Table 2). In all cases, the coherence responsiveness
was to decrease within a range from 1 to 0 with
increasing growing stock volume. This demonstrates
the boreal forest woody biomass mapping capabilities
of coherence images over large areas. A generalised
Table 4
Quality assessment of the coherence-based classification method at
three independent territories dispersed across the study area (Ust-
Ilimsk, Ulkansky and Hrebtovsky)
GIS classes
(m3 ha�1)
Derived classes (m3 ha�1)
Bare soil/
sparse shrub
1–20 21–80 >80
Bare soil/
sparse shrub
9318 22716 251 1046
1–20 451 61988 23922 23738
21–80 4 25721 29702 78529
>80 1 4201 17869 248989
Class accuracy (%) 95 54 41 71
Overall accuracy,
p0 (%)
64
Weighted kappa, k0 0.69
Table 5
Complementary forest attributes of eight growing stock volume classes extracted from the field-based forest inventory database at the five
territories of Table 2
Mensurations Species composition
Growing stock
volume (m3 ha�1)
Average height
of stand (m)
Average DBH
of stand (cm)
Pine (%) Spruce (%) Larch (%) Cedar (%) Birch and
aspen (%)
0 (bare soil) 0 0 No data No data No data No data No data
0 (sparse shrub) <2.21 <1.24 No data No data No data No data No data
1–20 2.21 � 1.29 1.24 � 1.48 20 10 2 2 66
21–50 6.00 � 3.11 5.70 � 3.34 14 14 6 4 62
51–80 9.69 � 4.14 9.64 � 4.96 14 16 8 4 58
81–130 14.45 � 5.50 15.84 � 7.82 14 22 12 4 48
131–200 18.72 � 4.47 21.78 � 7.22 18 24 10 6 42
>200 21.18 � 3.72 25.84 � 6.26 28 24 12 10 26
D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75 73
set of coherence signatures (Table 3) was obtained
from those in Table 2 and was tested in an MLC
algorithm at three independent territories (eight test
areas) scattered across the region. Plots comparing the
GIS forest growing stock volume polygons versus
eight forest classes confirmed the validity of using
the generalised set of coherence statistics (Table 3) to
estimate woody biomass (Fig. 5). Nonetheless, the
poor separation between the classes 21–50 and 51–
80 m3 ha�1 and between those above 80 m3 ha�1
reduced the number of extracted classes to four: bare
soil/sparse shrub, 1–20, 21–80 and >80 m3 ha�1. The
important overlap between two adjacent forest classes
was caused by the high spatial heterogeneity of forest
stands giving them large standard deviation values
(0.11–0.18) and by the high variability in the mean
coherence of forest stands across the study area. This
partly explains why the published map of forest
growing stock volume (Fig. 2) failed to extract classes
beyond 80 m3 ha�1 (see legend in Fig. 2). Still, the
classes 21–50 and 51–80 m3 ha�1 were not merged in
the published map and their validity is questioned
here. The overall accuracy (i.e. the percentage of
correctly classified pixels) of the coherence-based
forest map reached a 64% agreement and the weighted
k0 ranked 0.69. Accuracy results per class gave the
best correspondence values for bare soil/sparse shrub
and >80 m3 ha�1 (95 and 71%, respectively).
Classification results obtained here only apply to
‘ERS repeat-pass 1-day’ coherence images and over
low topography areas. For the published ‘forest grow-
ing stock volume’ map, for which JERS-1 backscatter
images were also used and a topographic mask was
applied to mask out high relief areas, the overall
accuracy reached 80% and the weighted k0 ranked
0.72.
By assigning additional forest attributes to eight
growing stock volume classes initially desired
(Table 5), we have attempted to provide a more
intelligent explanation of what the growing stock
volume classes mean so that anyone using the pub-
lished map (Fig. 2) can relate those classes to other
forest map definitions. However, it is clear that this
needs a more comprehensive examination of inter-
comparability to reconcile classifications, e.g. that of
FAO (1996) that are a function of tree height and
fractional cover. This will be the focus of a further
study.
4. Conclusion
A simple and quick forest growing stock volume
classification method applicable across the whole
Central Siberian region (Fig. 1) based on 1-day coher-
ence images derived from the ESA Tandem mission is
presented. The purpose was to show the potential of
coherence with regard to forest woody biomass clas-
sification and hence explain how growing stock
volume was successfully estimated in the SIBERIA
project. The proposed method is quicker and easier to
use than the classification method used to publish the
forest growing stock volume map shown in Fig. 2 but
still produces very similar results. It is applicable in
principle to the entire ESA ERS Tandem data archive.
The only drawback is that it does not map water bodies
and, like the classification method which generated
Fig. 2, it only works in areas of low topography.
To increase the value of the published map, the
forest classes were further characterised by addition of
other forest attributes. About 110 000 ERS-1/2 pair
images covering nearly the total global land surface
were acquired during the ESA Tandem mission.
Whilst over South America and parts of Southeast
Asia, just one data pair was acquired: as many as five
or six interferometric pairs are available for Europe
and North America. Therefore, there is strong poten-
tial to develop this technique further for forest woody
biomass estimation in other parts of the world.
Acknowledgements
SIBERIA was partly funded by the Framework 4
programme of the European Commission (Contract
No. ENV4-CT97-0743-SIBERIA). ESA and NASDA
provided the ERS and JERS SAR data through data
initiatives. We wish to thank the other members of the
SIBERIA team for their contribution to the project,
namely Kevin Tansey, Adrian Luckman (University of
Wales), Christiane Schmullius, Andrea Holz, Ursula
Marschalk, Wolfgang Wagner, Jan Vietmeier (DLR),
Malcolm Davidson, Thuy Le Toan (CESBIO), Yrjo
Rauste (VTT), Jiong Jiong Yu, Shaun Quegan (Shef-
field University), Michael Gluck, Anatoly Shvidenko
(IIASA), Tazio Strozzi, Urs Wegmuller, Andreas
Wiesmann (Gamma Remote Sensing), Hans Jonsson,
Marianne Orrmalm, Roland Utsi and Torbjørn Westin
74 D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75
(Satellus). In addition, we wish to thank Barry Wyatt
and Robin Fuller for their useful comments and sug-
gestions while drafting this paper.
Appendix A. Glossary of acronyms and
abbreviations
backscatter electromagnetic radiation reflected
back the SAR sensor
C-band a band of microwave frequency
around 5.3 GHz (5.7 cm wavelength)
DLR German Space Agency
ERS-1 European Research Satellite 1
ERS-2 European Research Satellite 2
ESA European Space Agency
GHz gigahertz (measure of the frequency
of electromagnetic radiation)
GIS geographical information system
IIASA International Institute for Applied
Systems Analysis
JERS-1 Japanese Earth Resources Satellite 1
L-band a band of microwave frequency
around 1.25 GHz (24 cm wavelength)
MLC maximum likelihood classification
SAR synthetic aperture radar
SIBERIA SAR imaging for boreal ecology and
radar interferometry applications
References
Aronoff, S., 1982. The map accuracy report: a user’s view.
Photogrammetric Eng. Remote Sens. 48, 1309–1312.
Bonan, G.B., Pollard, D., Thompson, S.L., 1992. Effects of boreal
forest vegetation on global climate. Nature 359, 716–718.
Cohen, J., 1960. A coefficient of agreement for nominal scales. Ed.
Psychol. Measure. 20, 37–46.
FAO, 1996. Forest resources assessment 1990. FAO Forestry Paper
No. 130.
FAO, 1999. State of the world’s forests 1999. FAO Report.
Foody, G.M., 1992. On the compensation for chance agreement in
image classification accuracy assessment. Photogrammetric
Eng. Remote Sens. 58, 1459–1460.
Gatelli, F., Guarnieri, A.M., Parizzi, F., Pasquali, P., Prati, C.,
Rocca, F., 1994. The wavenumber shift in SAR interferometry.
IEEE Trans. Geosci. Remote Sens. 32, 855–865.
Gerstl, S., 1990. Physics concepts of optical and radar reflectance
signatures: a summary review. Int. J. Remote Sens. 11, 1109–
1117.
Husch, B., Miller, Ch.I., Beers, T.W., 1993. Forest Mensuration,
3rd Edition. Krieger Publishing Company, Malabar, FL, 402 pp.
Hyyppa, J., Hyyppa, H., Inkinen, M., Engdahl, M., Linko, S., Zhu,
Y., 2000. Accuracy comparison of various remote sensing data
sources in the retrieval of forest stand attributes. For. Ecol.
Mgmt. 128, 109–120.
Krankina, O.N., Dixon, R.K., 1992. Forest management in
Russia—challenges and opportunities in the era of Perestroika.
J. For. 90, 29–34.
Lee, H., Liu, J.G., 2001. Analysis of topographic decorrelation in
SAR interferometry using ratio coherence imagery. IEEE
Trans. Geosci. Remote Sens. 39, 223–232.
Nilsson, S., Shvidenko, A., 1997. The Russian forest sector. A
Position Paper for the World Commission on Forests and
Sustainable Development. WCFSD, St Petersburg, Russia, p. 66.
Schmullius, C.C., Rosenqvist, A., 1997. Closing the gap—a
Siberian boreal forest map with ERS-1/2 and JERS-1. In:
Proceedings of the Third ERS Symposium on Space at the
Service of our Environment, Florence, Italy, ESA, pp. 1885–
1890.
Schowengerdt, R.A., 1997. Models and Methods for Image
Processing. Academic Press, London.
Singh, K.D., 1988. Comments on: tropical deforestation and remote
sensing, by Norman Myers (1988). For. Ecol. Mgmt. 24, 312–
313.
Wegmuller, U., Werner, C.L., 1995. SAR interferometric signatures
of forests. IEEE Trans. Geosci. Remote Sens. 33, 1153–1161.
WWF, 1994. Conserving Russia’s biological diversity. World
Wildlife Fund Report.
D.L.A. Gaveau et al. / Forest Ecology and Management 174 (2003) 65–75 75