Remote Sensing of Environment 139 (2013) 60–67

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Monitoring spruce volume and biomass with InSAR data from TanDEM-X
Svein Solberg a,⁎, Rasmus Astrup a, Johannes Breidenbach a, Barbi Nilsen a, Dan Weydahl b
a
Norwegian Forest and Landscape Institute, National Forest Inventory, P.O. Box 115, NO-1431 Ås, Norway
b
Norwegian Defence Research Establishment, Land and Airsystems Division, P.O. Box 25, NO-2027 Kjeller, Norway

a r t i c l e i n f o a b s t r a c t

Article history: There is a need for monitoring methods for forest volume, biomass and carbon based on satellite remote sensing.
Received 10 October 2012 In the present study we tested interferometric X-band SAR (InSAR) from the Tandem-X mission. The aim of
Received in revised form 23 July 2013 the study was to describe how accurate volume and biomass could be estimated from InSAR height and test
Accepted 29 July 2013
whether the relationships were curvilinear or not. The study area was a spruce dominated forest in southeast
Available online 27 August 2013
Norway. We selected 28 stands in which we established 192 circular sample plots of 250 m2, accurately
Keywords:
positioned by a Differential Global Positioning System (dGPS). Plot level data on stem volume and aboveground
InSAR biomass were derived from field inventory. Stem volume ranged from zero to 596 m3/ha, and aboveground bio-
Forest inventory mass up to 338 t/ha. We generated 2 Digital Surface Models (DSMs) from InSAR processing of two co-registered,
Volume HH-polarized TanDEM-X image pairs – one ascending and one descending pair. We used a Digital Terrain Model
Biomass (DTM) from airborne laser scanning (ALS) as a reference and derived a 10 m × 10 m Canopy Height Model
Nonlinear mixed models (CHM), or InSAR height model. We assigned each plot to the nearest 10 m × 10 m InSAR height pixel. We
applied a nonlinear, mixed model for the volume and biomass modeling, and from a full model we removed
effects with a backward stepwise approach. InSAR height was proportional to volume and aboveground biomass,
where a 1 m increase in InSAR height corresponded to a volume increase of 23 m3/ha and a biomass increase of
14 t/ha. Root Mean Square Error (RMSE) values were 43–44% at the plot level and 19–20% at the stand level.
© 2013 Elsevier Inc. All rights reserved.

1. Introduction 2008) and SAR data. The SAR methods include backscatter intensity
(Holopainen et al., 2010; Imhoff, 1995), polarimetry (Goncalves, Santos,
During the last two decades there has been a considerable increase & Treuhaft, 2007; Santos et al., 2003), InSAR coherence (Eriksson,
in the use of remote sensing for mapping of forest biomass and volume. Magnusson, Fransson, Sandberg, & Ulander, 2007; Hyyppä et al., 2008),
Monitoring of biomass, and carbon, is receiving increased attention InSAR height (Gama, dos Santos, & Mura, 2010; Neeff, Dutra,
because of the needs for reporting to international initiatives and Santos, Freitas, & Araújo, 2005; Solberg, Astrup, Gobakken,
conventions such as the REDD (Reduced Emissions from Deforestation Næsset, & Weydahl, 2010), PolinSAR (Le Toan et al., 2011) and
and forest Degradation in Developing Countries) and the UN-FCCC radargrammetry (Karjalainen, Kankare, Vastaranta, Holopainen, &
(United Nations Framework Convention on Climate Change). Satellite Hyyppä, 2012). Various limitations exist for satellite remote sensing
remote sensing is likely to be a major data provider in the future methods, i.e. saturation, low accuracy and effects of weather conditions
(Lynch, Maslin, Balzter, & Sweeting, 2013; Patenaude, Milne, & and clouds. There is a need for satellite remote sensing methods that
Dawson, 2005). Airborne Laser Scanning (ALS) enables wall-to-wall can overcome these limitations. BIOMASS is a P-band SAR satellite mis-
mapping with high accuracy (Hyyppä et al., 2008; Næsset, 2004a,b), sion dedicated to forest biomass monitoring (Le Toan et al., 2011). It will
and is increasingly used in forest management planning. Currently, provide biomass estimation based on a combination of backscatter, co-
photogrammetry based on aerial imagery is also receiving increased herence and PolInSAR (Polartimetric SAR Interferometry) techniques.
attention, due to commercial software that automatically generates 3D With its long wavelength (~69 cm) the microwaves will interact main-
point clouds, which have similar properties as ALS echoes (Bohlin, ly with coarser structures in the canopy, i.e. stems and coarse branches
Wallerman, & Fransson, 2012; Breidenbach & Astrup, 2012; St-Onge, which contain the majority of the biomass. However, it is planned for
Vega, Fournier, & Hua, 2008). However, data from airborne sensors launch in 2020.
may have too low aerial capacity and a lack of temporal consistency. In the present study we are testing 3D SAR with a short wavelength
Satellite data have been tested with variable results, both optical as an alternative that would be available already today, and which
(Fransson, Smith, Askne, & Olsson, 2001; Magnusson & Fransson, 2004; enjoys the advantages of high areal capacity from a satellite platform;
Mäkelä & Pekkarinen, 2004; Tomppo, Haakana, Katila, & Peräsaari, cloud-free imagery from the SAR sensor; and the correlation between
height and forest volume and biomass. There are four different 3D SAR
⁎ Corresponding author. Tel.: +47 64948996; fax: +47 64948001. methods, which can be used alone or in combination (Toutin & Gray,
E-mail address: [email protected] (S. Solberg). 2000). Relative or absolute elevations can be derived from phase-

0034-4257/$ – see front matter © 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.rse.2013.07.036
 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67 61

differences in an image pair (interferometry), from parallaxes in an set comprised 28 stands containing 192 sample plots. The size of these
image pair (radargrammetry), from shadow lengths and pixel brightness stands varied from 1.0 to 2.9 ha, with a mean value of 1.9 and a standard
(clinometry), or from the shift of the co-polarized polarimetric signature deviation of 0.5. Measurements according to the Norwegian National
maximum (polarimetry). We are here employing interferometric X-band Forest Inventory protocol were carried out on the sample plots (Anon,
SAR from the Tandem-X mission. This is a bistatic acquisition with two 2008). The location of the plot centers were accurately recorded with
satellites going in formation, where microwave pulses are sent from differential GPS measurements. Species and DBH were recorded for
one satellite at a time and the echo is received by both. The travel length each tree N 5 cm DBH. Tree height was measured for a subset of 10
of the echo differs between the two SAR sensors, and this generates phase trees per plot. The heights of the remaining trees were estimated
differences varying over the area. These phase differences can be based on height–diameter models for each plot. The volume including
processed into a DSM. The X-band has a short wavelength (3.1 cm) bark of the trees was calculated from standard Norwegian models
with a limited penetration in vegetation. The DSM is located close to the (Vestjordet, 1967), and this was aggregated per plot as per-hectare
top of the vegetation canopy, and contains both terrain topography and volume. Aboveground biomass of the trees was derived from the
vegetation heights. models of Marklund (1988) with height and diameter as predictor
We see two possible applications of an X-band InSAR DSM. First, if variables. It included stem wood, stem bark, stump, and live and dead
available, an external DTM can be subtracted from the DSM providing branches, and was aggregated to per-hectare biomass for each plot
InSAR heights, from which volume or biomass can be derived by (Table 1).
model inversion. Secondly, if a DTM is not available, temporal changes
in volume and biomass may still be monitored if the relationship 2.3. TanDEM-X data
between volume (biomass) and InSAR height is straight linear and
constant over time. In this case temporal changes in volume (biomass) We used two TanDEM-X strip map image pairs from the summer
could be derived directly from the temporal changes in the DSM. and autumn of 2011. One ascending acquisition was taken in the
There would be a fixed change in volume and biomass per meter change afternoon (16:54) on 23 July 2011. The weather in the study area was
in the InSAR DSM. rainy (30–50 mm rain during 24 h) and the mean temperature was
The aim of this study was to test whether bistatic X-band InSAR data around 15 degrees. A descending acquisition was taken in the early
from the TanDEM-X mission can be used to monitor stem volume and morning (05:40) on 1 September 2011. It was minor precipitation
aboveground forest biomass. The objectives were: during the 24 h before the acquisition (b10 mm) and a mean tempera-
ture of 5–10 degrees. The polarization was horizontal (H) both on send
1. To describe how accurate volume and biomass can be estimated, and
and receive, and the incidence angles were 36° and 42°, respectively.
2. To test whether the relationship is straight linear, whereby temporal
The data were received as co-registered Single Look Complex (SLC)
changes in biomass and volume could be monitored without having
data in CoSSC format. Each pair was processed to an InSAR DSM using
a DTM as input.
the Sarscape 5.0 module of the ENVI 5.0 software. An interferogram
was generated from each image pair, and this was further processed
2. Materials and methods
into a differential interferogram by removing the range dependent,
and the terrain topography dependent parts of the phase differences.
2.1. Study area
The ALS DTM was used as input in this step. These differential interfer-
ograms represented the phase differences caused by vegetation height,
The study was carried out in the Lardal municipality in southeast
in addition to phase noise, as well as possible inaccuracies in the orbit
Norway, and was centered at 59.4° north and 9.9° east. The study area
data and in the atmospheric corrections. The phase noise in the
contains a forest, with some intermixed peat lands and lakes, as well
TanDEM-X data could result from volume decorrelation and system
as a valley with agricultural lands. The forest is managed for timber
noise. The differential interferograms were filtered with the Boxcar
production, and is divided in stands of various age and development
adaptive filter, which reduces phase noise and enables an accurate
stages. For this study, the digitized stand map with stem volume data
phase unwrapping later. We removed phase offset and phase ramp
was made available, as was a DTM with 10 m × 10 m resolution
errors originating from possible orbit and atmospheric inaccuracies by
generated from an ALS data set with 10 pulses per m2.
fitting the following model to 30 Ground Control Points (GCPs):
2.2. Field data
Δφ ¼ k0 þ k1 RG þ k2 AZ; ð2Þ
Based on the stand map we initially selected 40 stands for field
sampling. First, we selected all compact stands with an area between
where Δφ was the phase difference at each GCP, k0, k1 and k2 were
one and three hectares. Compact stands were selected to avoid
correction factors, and RG and AZ were the range and azimuth coordi-
problems with defining stand borders in the field, and were defined as
nates (Table 2). The GCPs were subjectively laid out at locations without
trees, i.e. where the InSAR DSM should have the same elevation as the
√A=PN0:2 ð1Þ DTM. These locations were selected as having high coherence and
relatively flat terrain with low fringe density, i.e. in clear-cuts and
where A is the stand area and P is the perimeter. Secondly, in order agricultural fields. Hence, in these GCPs the final phase differences
to ensure a wide range of volume and biomass we stratified the should be minor.
stands in two groups, i.e. with high (N 150 m3/ha) and low volume
(b150 m3/ha) estimates in the forest management plan. Within each Table 1
of the volume strata, we randomly selected 20 stands for field sampling. Statistics for the 192 field plots.
In each stand, we made a 20 m × 20 m grid and randomly selected Variable Unit Mean Standard deviation Minimum Maximum
seven of the grid nodes as plot locations. At each location we established 3
Volume, total m /ha 216 132 0 596
a 250 m2, circular sample plot in which we carried out field inventory. Volume, spruce m3/ha 188 125 0 596
Three plot locations were discarded, because they were located in a Volume, pine m3/ha 10 38 0 267
part of the stand that had been clear-cut recently. For this study we Volume, broadleaves m3/ha 19 33 0 199
selected the spruce-dominated stands, by excluding 12 stands which Above ground t/ha 134 79 0 338
biomass
had less than 50% spruce volume. With these selections, the field data
62 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67

Table 2 model would be a direct proportionality between the forest variable

Correction factors for phase offset and phase ramp errors (radians), see Eq. (2), and the and InSAR height, i.e. where b = 1 and where c = 0 because we had
final accuracy (RMSE) for the 30 Ground Control Points (GCP).
removed phase offset and ramp errors, reducing (Eq. 4) to a linear
Acquisition k0 k1 k2 RMSE regression without intercept (6).
Ascending −2.749480 −0.000048 −0.000013 0.9 m
Descending 1.488414 −0.000093 0.000028 2.2 m y ¼ ax þ σ þ ε ð6Þ

We carried out phase unwrapping using the ‘region growing’ method 3. Results
(Reigber & Moreira, 1997). The ascending acquisition had a normal base-
line of 122 m, and a corresponding 2π height of ambiguity of 23 m. This is The spatial variation in InSAR height reproduced well the spatial
close to, but slightly below the maximum tree height, and the distribution of forests (Fig. 1). Clear-cuts had InSAR heights around
unwrapping was required to avoid height ambiguity errors. The descend- zero, while old and dense stands had InSAR heights up to 20 m. The
ing acquisition had a considerably smaller normal baseline, 59 m. In this level of details was high, and even narrow forest roads in between old
case the 2π height of ambiguity was 122 m, i.e. more than the maximum stands could be seen as bands of low InSAR heights.
vegetation height, and we could have canceled the unwrapping and
converted phase values directly to InSAR heights. However, for conve- 3.1. Volume
nience we carried out one and the same processing chain, including
unwrapping, for both acquisitions. Finally, the unwrapped phases After backward, stepwise removal of non-significant effects, the final
were converted into elevation data and transformed from satellite model was a linear, no-intercept relationship between InSAR height and
slant-range geometry to geocoded DSMs. During the processing we stem volume (Table 3, Fig. 2). Neither the exponent, nor the intercept,
used a multi-looking of 7 azimuth × 4 range, which corresponded improved the model as seen from the likelihood-ratio tests. The log-
to about 10 m × 10 m, which was also the spatial resolution of the likelihood had minor and non-significant increases when these terms
geocoded DSMs obtained by 4th order cubic convolution resampling. were discarded. In addition to being non-significant the exponent
We subtracted the DTM from the DSM and obtained InSAR heights. estimate was fairly close to 1.0 (0.78), and the difference between the
We made a data set of weighted mean InSAR heights: straight and curvilinear models was minor (Fig. 2). The random effect
of stands, σ, was significantly different from zero. The log-likelihood
IHmean ¼ ððγ 1  IH1 Þ þ ðγ2  IH2 ÞÞ=ðγ1 þ γ2 Þ; ð3Þ increased by 16.2, from 2273.1 to 2289.3, when this effect was removed
and this was highly significant (p b 0.0001). This means that some
where IH1 and IH2 are the InSAR heights from the ascending and stands tended to have overestimated volumes and others to have
descending acquisitions, and γ1 and γ2 are the corresponding coher- underestimated volumes. Removing this random effect, however, pro-
ence values. Coherence was estimated in 3 × 3 pixel windows. In vided the estimated random error at the plot level, Var(ε) = 8836 t/ha.
each DSM the accuracy varied with coherence, generating inaccurate The random effects were recalculated to relative RMSE in % of mean
results in low coherence areas. This occurred mainly in some limited volume:
areas where the signal-to-noise ratio was low, i.e. over lakes and in
steep hillsides facing away from the satellites. We assigned one • RMSE at the plot or pixel level: 94/216 = 44%
10 m × 10 m InSAR height pixel to each plot, i.e. the pixel closest • RMSE at the stand level: 44/216 = 20%
to the center of the plot. This implies that in our study area, the volume could be estimated
for a random pixel with an accuracy of 94 m3/ha (44%), and at the
2.4. Statistical analyses stand level with an accuracy of 44 m3/ha (20%).

In order to determine the relationship between the two forest vari- 3.2. Aboveground biomass
ables and InSAR height we fitted the following nonlinear, mixed
model to the data: Aboveground biomass was strongly correlated to volume, with a
Pearson correlation coefficient r = 0.98. Hence, the results for biomass
b
y ¼ ax þ c þ σ þ ε; ð4Þ were similar to the results for volume – the parameters were only
differently scaled. As for the volume model, both the exponent and
where y is volume or biomass, a is a slope, x is InSAR height, b is an the intercept could be discarded because they were non-significant
exponent, c is an intercept, and finally σ and ε are zero-mean, Gaussian effects in the model (Table 4, Fig. 3). The exponent was now further
variables representing the random effect at the stand and plot levels, away from 1.0 (0.66) and nearly contributed significantly to the final
respectively. Of particular importance was the exponent: Any tendency model (p = 0.09). However, a visual inspection of Fig. 3 also revealed
of curve-linearity should result in b differing from the value 1. We that the straight and the curvilinear models appeared to fit equally
applied the NLMIXED procedure in the SAS software (Schabenberger well to the data. The exponent value being below 1.0 indicated a slight
& Pierce, 2002). This was an appropriate tool by enabling a testing of tendency of the opposite of a saturation effect. The random effect at
the curve-linearity, and by handling the hierarchical structure of the the stand level was highly significant (p = 0.0005). By removing this
random effects, i.e. plots within stands. The model fitting was carried we got the random effect at the plot level, Var(e) = 3249. The random
out with maximum likelihood estimation. Starting from the full model effects were recalculated to relative RMSE in % of mean biomass:
(4), we carried out a backward stepwise removal of non-significant
effects. The significance of an effect was assessed based on the • RMSE at the plot level: 57/134 = 43%
likelihood-ratio test, where an effect was retained only if its removal • RMSE at the stand level: 25/134 = 19%
generated a sufficiently large change in log-likelihood, i.e. if: This means that biomass could be estimated for a random pixel with
an accuracy of 57 t/ha (43%), and at the stand level with an accuracy of
2
−2lnðΔLikelihoodÞNχ n;0:05 ; ð5Þ 25 t/ha (19%).
The results for each acquisition separately were similar to what we
where n is the change in the number of parameters. For monitoring of obtained with the coherence-weighted, mean DSM model. The ascend-
temporal changes in volume or biomass without a DTM the preferred ing acquisition produced a slightly better result than the weighted
 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67 63

Fig. 1. Clip (850 m × 850 m) from aerial imagery and InSAR height showing the spatial correspondence between forests and InSAR height. The lines are forest stand borders.

mean, in the sense that the random effect at the stand level was slightly The fitted models indicate that in a spruce forest, an increase of 1 m
lower (Table 5). The descending pair produced slightly larger random in the InSAR DSM corresponds to a volume increase of 23 m3/ha and a
errors. This was not due to differences in coherence, as the descending biomass increase of 14 t/ha. The magnitude of this is fairly similar to
acquisition had clearly higher coherence values (Table 6). InSAR heights the results of other studies. Solberg, Astrup, Gobakken et al. (2010),
from the descending and ascending acquisitions were fairly similar Solberg, Astrup, Bollandsås, Næsset, and Weydahl (2010) found similar
(Fig. 4). However, there were some random errors creating scattering slope parameters, i.e. 15 and 12 t/ha/m, for spruce biomass, even if
along the 1:1 line. Based on the results in Table 6, it is likely that these those studies were based on winter acquisitions, i.e. the February
random errors were dominated by errors from the descending 2000 X-band SRTM acquisitions, while we in the present study used
acquisition. summer acquisitions. In the Brazilian X-band InSAR studies on biomass,
Neeff et al. (2005) had a slope parameter of 13 t/ha/m, while Gama
et al.'s (2010) curvilinear relationships correspond to 2–6 t/ha/m.
4. Discussion Altogether, although the slope may vary between forest types, the
order of magnitude for any forest type seems to be 20 m3/ha and
We found linear, no-intercept relationships, which mean a direct 10 t/ha per m change in X-band InSAR height.
proportionality between InSAR height and the forest variables volume The straightness of the relationship was clear in the present data
and biomass. The height aboveground of the X-band RADAR echo is set. When we included an exponent in the model it was estimated to
largely determined by vegetation height and vegetation density, and be somewhat below one, i.e. 0.78 for volume and 0.66 for biomass.
these two properties are also largely determining volume and biomass These effects were non-significant and were removed from the model.
(Treuhaft & Siqueira, 2004). The 3.1 cm wavelength X-band SAR has If anything, these estimates, being lower than 1.0, indicated a tendency
the majority of scatterers in the foliage, i.e. in leaves, needles and towards the opposite of saturation. It has been demonstrated both from
smaller branches. The height of the SAR backscattering center is theoretical approaches and empirical studies that the relationship
determined by the vertical distribution of such smaller canopy objects. between stem volume (V) and tree height (H) is curvilinear and
The volume and biomass are determined largely by the size and density follows a power law function:
of tree stems. Hence, the relationship we have found implies that the
α
height and density of foliage are related to the height and density of V∝H ; ð7Þ
tree stems. Due to this, the relationship might be unstable over time if
there were seasonal variations in foliage density, but this should not where α normally takes a value in the range 1.5–2.0, which implies
be a problem because Norway spruce is an evergreen tree with many a curvilinear relationship with a tendency of saturation (Askne,
year classes of needles. The relationship might be violated in cases of Dammert, Ulander, & Smith, 1997; Mette, Papathanassiou, & Hajnsek,
defoliation from stress or forest damage, which would reduce the height 2004; Woodhous, 2006). However, this should not be seen as a
of the backscattering center while volume and biomass would remain contradiction to the relationship found in the present study. In model
almost unaffected. Moisture and frost in the vegetation might also affect (7) the height variable used was tree height, either mean tree height
the penetration of the microwaves into the vegetation, and affect the or top height (H100). The relationship may well be curvilinear for such
relationships. This deserves further studies. Dryness and frost in the tree height variables, while it is linear (α = 1) for a canopy height
foliage reduces the dielectric constant and the SAR microwaves will variable derived from remote sensing, such as InSAR height. The crucial
penetrate deeper into the vegetation (Way et al., 1994). distinction between a tree height variable and a canopy height variable

Table 3
Fitting the volume model: Backward stepwise selection of effects based on nonlinear, mixed modeling with the NLMIXED procedure. N = number of parameters, y = volume, x = InSAR
height, a, b and c were fixed effect parameters, σ and ε were zero-mean, Gaussian variables, where σ represented the random effect of stand, and ε was the residual error at the plot level
given as standard deviation.

Step Model N ΔN -2LL -2ΔLL p N -2ΔLL a b c σ ε

0 y = axb + c + σ + ε 5 2271.4 – 45 .78 −31 47 83

1 y = ax + c + σ + ε 4 1 2272.6 1.2 0.27 22 – 18 47 83
2 y = ax + σ + ε 3 1 2273.1 0.5 0.48 23 – – 44 83
3 y = ax + ε 2 1 2289.3 16.2 b0.0001 26 – – – 94
64 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67

Fig. 2. Stem volume against InSAR height with the fitted linear model Volume = 23 IH Fig. 3. Aboveground biomass against InSAR height with the fitted linear model Bio-
(solid line) and the curvilinear model Volume = 45 IH0.78 – 31 (hatched line). Small mass = 14 IH (solid line) and the curvilinear model Biomass = 41 IH0.66 – 40 (hatched
dots: 250 m2 plots and 10 m × 10 m pixels, N = 192. Large dots: stand means, N = 28. line). Small dots: 250 m2 plots and 10 m × 10 m pixels, N = 192. Large dots: stand
means, N = 28.

is that the latter is influenced by the amount of gaps in the canopy, i.e. by
vegetation or stand density, while the former is not. If this distinction is et al., 2005; Rombach & Moreira, 2003), difference between X- and L-
obscured it may generate the perception that the relationship between band heights from airborne sensors (Balzter, Rowland, & Saich, 2007)
biomass and a canopy height variable, such as InSAR height, must be and C-band SRTM data (Kellndorfer et al., 2004).
curvilinear. InSAR height relationships that are close to linear are also An additional error component originating from the allometric
found by Solberg, Astrup, Gobakken, Næsset, and Weydahl (2010) for models may be present. It can be estimated to about 14% at the stand
a similar forest type. They used another bistatic acquisition – the X- level based on analyses of variance provided by Marklund (1987),
band SRTM. Neeff et al. (2005) also found a similar, straight linear however, this source of error is normally not included in other studies
relationship in a native rainforest in Brazil, based on bistatic X-band and is excluded here as well for the sake of comparison.
InSAR data, where repeat-pass P-band heights representing the terrain There is a wide range of remote sensing methods for volume and
heights were subtracted. However, whether the relationship is straight biomass mapping; however, there is also a wide range of limitations.
linear or curvilinear may vary between forest types. Gama et al. (2010) Many methods suffer from low accuracy, saturation at low levels and
used the same method as Neeff et al. (2005) and found a curvilinear unstable relationships (Eriksson et al., 2007; Imhoff, 1995; Luckman,
relationship for a Eucalyptus plantation. They squared InSAR heights, Baker, Honzak, & Lucas, 1998; Santos et al., 2003; Tomppo et al.,
i.e. used a predetermined exponent of 2. 2008). With optical imagery the relationships vary from acquisition to
The results demonstrate that interferometric X-band SAR, as one acquisition due to variations in atmospheric conditions, sun angle and
type of 3D SAR, has a potential for forest volume, biomass and carbon season. For example, Mäkelä and Pekkarinen (2004) obtained RMSE
monitoring. The results show a fairly high accuracy and no saturation values at the stand level of 48% for total volume, 81% for spruce and
problem. This is in line with other studies based on InSAR. For more than 100% for pine and broadleaf volumes, and concluded that
10 m × 10 m pixels the accuracy in terms of RMSE was about 43%, volume estimation based on Landsat TM is not accurate enough for
while at the stand level it was about 19%. This is close to the accuracy forest management. To some extent these problems might be overcome
obtained with X-band InSAR data from the Shuttle Radar Topographic by more advanced modeling. Using multiplicative models on optical
Mission (SRTM). In Solberg, Astrup, Gobakken, Næsset, and Weydahl data from SPOT-4 Fransson et al. (2001) and Magnusson and Fransson
(2010), Solberg, Astrup, Bollandsås, Næsset, and Weydahl (2010) the (2004) obtained stand level volume estimates having RMSE values of
RMSE with X-band SRTM was 18–19% at the stand level and 36–40% 23–38%, and with a less pronounced saturation effect. With SAR back-
at the plot level. In general, biomass and vegetation height have scatter data, the relationship with biomass is curvilinear with a satura-
been strongly correlated to InSAR height, being the difference between tion effect. The level of biomass where this effect occurs increases with
X- and P-band heights from airborne sensors (Gama et al., 2010; Neeff the SAR wavelength, as demonstrated by Imhoff (1995), who compared

Table 4
Fitting the biomass model: Backward stepwise selection of effects based on nonlinear, mixed modeling with the NLMIXED procedure. N = number of parameters, Y = biomass, x = InSAR
height, a, b and c were fixed effect parameters, σ and ε were zero-mean, Gaussian variables, where σ represented the random effect of stand, and ε was the residual error at the plot level
given as standard deviation.

Step Model N ΔN -2LL -2ΔLL p N -2ΔLL a b c σ ε

0 y = axb + c + σ + ε 5 2082.2 – 41 0.66 −40 26 51

1 y = ax + c + σ + ε 4 1 2085.1 2.9 0.09 13 – 15 26 51
2 y = ax + σ + ε 3 1 2086.3 1.2 0.27 14 – – 25 52
3 y = ax + ε 2 1 2098.4 12.1 0.0005 14 – – – 57
 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67 65

Table 5
Comparison of the results for volume estimation based on InSAR height derived from each
of the two acquisitions and the weighted mean: y = ax + σ + ε, where y was volume, x
was InSAR height, and σ and ε were zero-mean, Gaussian variables, where σ represented
the random effect of stand, and ε was the residual error at the plot level given as standard
deviation.

Data a σ ε

Ascending 24.1 42 83
Descending 22.0 50 86
Weighted mean 23.2 44 83

C, L and P-band and obtained saturation at 20, 40 and 100 t/ha, respec-
tively. Saturation levels vary considerably between studies, partly be-
cause they are defined and derived in different ways, and partly
because they are influenced by biophysical properties such as stem
number density and moisture. The saturation level is in the range 0–
100 t/ha for C-band (Dobson et al., 1992; Imhoff, 1995; Luckman,
Baker, Kuplich, Yanasse, & Frery, 1997; Ranson & Sun, 1994), 40–
270 t/ha for L-band (Dobson et al., 1992; Englhart, Keuck, & Siegert,
2011; Imhoff, 1995; Lucas et al., 2010; Luckman et al., 1997, 1998;
Ranson, Saatchi, & Sun, 1995; Ranson & Sun, 1994), and 100–200 t/ha
for P-band (Dobson et al., 1992; Imhoff, 1995; Le Toan et al., 2011;
Neeff et al., 2005; Ranson & Sun, 1994). These are all well below the
Fig. 4. InSAR heights from the descending acquisition plotted against InSAR heights from
maximum forest biomass values. Interferometric coherence may also the ascending acquisition, and the 1:1 line.
be used for volume and biomass estimation; however, the relationships
may be temporally unstable (Askne et al., 1997; Balzter, 2001; Hagberg,
Ulander, & Askne, 1995; Santoro, Askne, & Dammert, 2005; Santoro,
Shvidenko, McCallum, Askne, & Schmullius, 2007). To some extent the costs, lower aerial capacity, and it is less feasible to obtain data sets
limitations with SAR data can be overcome by using time series or by that are consistent in time and space. Altogether, a satellite based 3D
combining bands. Using hyper-temporal series of C-band backscatter SAR approach appears to be among the remote sensing technologies
intensity from ASAR in ScanSAR mode, Santoro et al. (2011) got no sat- for forest monitoring that deserves further studies and development.
uration problems for volumes estimated up to 300 m3/ha, and RMSE Our processing of the TanDEM-X data might be further improved,
values of 34–48% for 1 km × 1 km pixels, with their BIOMASAR algo- although it is likely that the potential for this is minor. Firstly, there
rithm. With multi-temporal data, i.e. four ERS-1/2 1-day repeat-pass co- might be remaining errors from inaccuracies in the orbital data and in
herence images, Santoro, Askne, Smith, and Fransson (2002) estimated the corrections for atmospheric effects. In principle these errors have
volume at the stand level with an RMSE value of only 10 m3/ha, which been removed by the correction for residual phase offset and ramp.
corresponded to a relative RMSE of only 7%. By combining a short (X or However, the performance of this step relies on the correct selection
C) and a long (L or P) SAR wavelength the saturation problem can be re- of ground control points (GCPs). The GCPs should have been placed in
duced (Englhart, Keuck, & Siegert, 2012; Englhart et al., 2011; Ranson & locations without any trees, so the InSAR DSM should be equal to the
Sun, 1994; Ranson et al., 1995). Various indices from polarimetric, L- DTM. However, some residual height errors remained after that
band SAR has been correlated to stem volume (Goncalves et al., correction. Secondly, the degree of smoothing of the differential inter-
2007). A satellite mission that is tailored for forest biomass monitoring ferogram is a trade-off between removing phase noise and retaining
is BIOMASS. Working with P-band SAR a combination of backscatter true details, and a further optimization might be possible. Today's forest
intensity, coherence and PolInSAR will provide aboveground forest bio- management in Norway has distinct stands of different ages, and also
mass with an accuracy of at least 20% (10 t/ha) from 70° N to 56° S, at a leaves a number of solitary, large trees in the clear-cuts (seed trees
spatial scale of 100–200 m, and with a 25–45 day repeat cycle (Le Toan and leftover trees for biodiversity and landscape issues). The stand
et al., 2011). BIOMASS was recently selected to become ESA's 7th Earth edges and the solitary trees may generate discontinuities in the
Explorer mission. However, it is planned for launch in 2020, and in the interferograms. On the one hand, a removal of discontinuities is re-
meantime there is a need for alternatives. Finally, airborne remote quired for the phase unwrapping to work, while on the other hand
sensing methods can provide very accurate results. Height metrics of it removes true details of the forest. The potential for further improve-
point clouds from laser scanning or photogrammetry contains detailed ments is also indicated by the difference in the results obtained for the
information. The accuracy of volume estimation with ALS is in the range ascending and descending acquisitions. The 122 m normal baseline of
8–14% for plot sizes ranging from 200 to 4000 m2 (Holmgren, 2004, the ascending acquisition appeared to be more appropriate than the
Holmgren & Jonsson, 2004; Maltamo, Eerikäinen, Packalén, & Hyyppaa, 59 m of the descending acquisition. Possibly, this difference could be
2006; Næsset 2002, 2004a,b, 2007). With photogrammetry on aerial reduced by a careful fine-tuning of the processing.
imagery Bohlin et al. (2012) obtained a similar accuracy (13–14% Radargrammetry is another 3D SAR method that might be more
RMSE) for 314 m2 plots, while Breidenbach and Astrup (2012) obtained useful than InSAR from a mission like Tandem-X. The advantage is
43% for 250 m2 plots. However, airborne remote sensing has some dis- that a bistatic acquisition is not required. Repeat-pass acquisitions can
advantages in comparison with satellites, by normally having higher be used, and data can be obtained from a higher number of ongoing
SAR missions, such as Cosmo Skymed, TerraSAR-X and Radarsat-2.
Table 6
Radargrammetry has provided promising results for forest inventory
Coherence statistics for field inventory plots and stands. (Chen, Shi, Deng, & Li, 2007; Karjalainen et al., 2012; Perko et al., 2011).
In conclusion, the 3D SAR data from Tandem-X is a promising
Variable Plot mean (min–max) Stand mean (min–max)
method for mapping and monitoring of spruce forests. Volume and
Ascending acquisition 0.68 (0.26–0.91) 0.68 (0.41–0.85) biomass could be estimated fairly accurately in comparison with other
Descending acquisition 0.86 (0.41–0.96) 0.86 (0.71–0.92)
satellite remote sensing methods, and the relationship was straight
66 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67

linear whereby temporal changes in biomass and volume might be parameters from SAR data for land applications, Bari, Italy. : ESA Istituto di Studi
sui Sistemi Intelligenti per l'Automazione. CD.
monitored without having a DTM as input. This would require that Fransson, J. E. S., Smith, G., Askne, J., & Olsson, H. (2001). Stem volume estimation in
the relationship between these forest variables and InSAR height is boreal forests using ERS-1/2 coherence and SPOT XS optical data. International Journal
stable over time. This requires further studies. In the present study the of Remote Sensing, 22(14), 2777–2791.
Gama, F. F., dos Santos, J. R., & Mura, J. C. (2010). Eucalyptus biomass and volume
relationship was fairly similar in the two acquisitions; however, some estimation using interferometric and polarimetric SAR data. Remote Sensing, 2(4),
random errors were still present. The differences might be due to the 939–956.
different baselines. The shorter baseline of the descending acquisition Goncalves, F. G., Santos, J. R., & Treuhaft, R. N. (2007). Stem volume of tropical forests from
polarimetric radar. International Journal of Remote Sensing, 32(2), 503–522.
(59 m) appeared to be less appropriate. Apparently, baselines should Hagberg, J. O., Ulander, L. M. H., & Askne, J. (1995). Repeat-pass SAR interferometry
be kept fairly stable and within certain ranges in multi-temporal data over forested terrain. IEEE Transactions on Geoscience and Remote Sensing, 33(2),
sets for monitoring. 331–340.
Holmgren, J. (2004). Prediction of tree height, basal area and stem volume in forest stands
One application of this is to derive volume and biomass estimates
using airborne laser scanning. Scandinavian Journal of Forest Research, 19, 543–553.
wall-to-wall at the regional or national scale as a supplement to the Holmgren, J., & Jonsson, T. (3–6 October). Large scale airborne laser scanning of
systematic field surveys in a National Forest Inventory. At a national forest resources in Sweden. In M. Thies, B. Koch, H. Spiecker, & H. Weinacker
level this provides cost-efficient combinations of two complementary (Eds.), Proceedings of the ISPRS working group VIII/2. Laser-scanners for forest
and landscape assessment, Freiburg, Germany. International archives of photogram-
data sets (Næsset et al. 2011), and at the same time provides accurate metry, remote sensing and spatial information sciences, Vol. XXXVI. (pp. 157–160)
data at the local level for municipalities and forest owners. Another (Part 8/W2).
application which does not require a DTM is monitoring of temporal Holopainen, M., Haapanen, R., Karjalainen, M., Vastaranta, M., Hyyppä, J., Yu, X., et al.
(2010). Comparing accuracy of airborne laser scanning and TerraSAR-X radar
changes in volume or biomass, derived from changes in the DSM. This images in the estimation of plot-level forest variables. Remote Sensing, 2(2),
requires that the relationship between biomass and InSAR height is 432–445 (2010).
straight linear, and the present study has demonstrated that this is the Hyyppä, J., Hyyppä, H., Leckie, D., Gougeon, F., Yu, X., & Maltamo, M. (2008). Review of
methods of small-footprint airborne laser scanning for extracting forest inventory
case in a spruce dominated forest. Further studies may reveal whether data in boreal forests. International Journal of Remote Sensing, 29, 1339–1366.
this is a suitable method for biomass and carbon monitoring under Imhoff, M. L. (1995). Radar backscatter and biomass saturation: ramifications for global
international initiatives and conventions. biomass inventory. IEEE Transactions on Geoscience and Remote Sensing, 33, 511–518.
Karjalainen, M., Kankare, V., Vastaranta, M., Holopainen, M., & Hyyppä, J. (2012). Predic-
tion of plot-level forest variables using TerraSAR-X stereo SAR data. Remote Sensing
of Environment, 117, 338–347.
Acknowledgments
Kellndorfer, J. M., Walker, W. S., Pierce, L. E., Dobson, M. C., Fites, J., Hunsaker, C., et al.
(2004). Vegetation height derivation from Shuttle Radar Topography Mission and
We acknowledge Deutsches Zentrum für Luft und Raumfahrt (DLR) National Elevation data sets. Remote Sensing of Environment, 2004(93), 339–358.
for TanDEM-X science data free of charge, the European Space Agency Le Toan, T., Quegan, S., Davidson, M., Balzter, H., Paillou, P., Papathanassiou, K., et al.
(2011). The BIOMASS mission: Mapping global forest biomass to better understand
for funding through PRODEX, and the Norwegian Research Council for the terrestrial carbon cycle. Remote Sensing of Environment, 115, 2850–2860.
funding through the Development fund for forestry. Lucas, R., Armston, J., Fairfax, R., Fensham, R., Accad, A., Carreiras, J., et al. (2010). An eval-
uation of the ALOS PALSAR L-Band Backscatter-Above Ground Biomass Relationship
Queensland, Australia: Impacts of surface moisture condition and vegetation struc-
Appendix A. Supplementary data ture. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,
3(4), 576–593.
Luckman, A., Baker, J., Kuplich, T. M., Yanasse, C. D. F., & Frery, A. C. (1997). A study of the
Supplementary data associated with this article can be found in the relationship between radar backscatter and regenerating tropical forest biomass for
online version, at http://dx.doi.org/10.1016/j.rse.2013.07.036. These spaceborne SAR instruments. Remote Sensing of Environment, 60(1), 1–13.
Luckman, A., Baker, J., Honzak, M., & Lucas, R. (1998). Tropical forest biomass density
data include Google map of the most important areas described in this
estimation using JERS-1 SAR: Seasonal variation, confidence limits, and application
article. to image mosaics. Remote Sensing of Environment, 63, 126–139.
Lynch, J., Maslin, K., Balzter, H., & Sweeting, M. (2013). Sustainability: Choose satellites to
monitor deforestation. Nature, 496, 293–294.
References Magnusson, M., & Fransson, J. E. S. (2004). Combining airborne CARABAS-II VHF SAR data
and optical SPOT-4 satellite data for estimation of forest stem volume. Canadian
Anon (2008). Landsskogtakseringens feltinstruks 2008, Håndbok fra Skog og landskap 05/08. Journal of Remote Sensing, 30(4), 661–670.
Ås, Norway: Landsskogtakseringen, Skog og landskap (In Norwegian). Mäkelä, H., & Pekkarinen, A. (2004). Estimation of forest stand volumes by Landsat
Askne, J. I. H., Dammert, P. B. G., Ulander, L. M. H., & Smith, G. (1997). C-band repeat-pass TM imagery and stand-level field-inventory data. Forest Ecology and Management,
interferometric SAR observations of the forest. IEEE Transactions on Geoscience and 196(2–3), 245–255.
Remote Sensing, 35(1), 25–35. Maltamo, M., Eerikäinen, K., Packalén, P., & Hyyppaa, J. (2006). Estimation of stem volume
Balzter, H. (2001). Forest mapping and monitoring with interferometric synthetic using laser scanning-based canopy height metrics. Forestry, 79, 217–229.
aperture radar (InSAR). Progress in Physical Geography, 25(2), 159–177. Marklund, L. G. (1987). Biomass functions for Norway spruce in Sweden. Swedish Univer-
Balzter, H., Rowland, C. S., & Saich, P. (2007). Forest canopy height and carbon estimation sity of Agricultural sciences. Department of Forest Survey, Rapport, 43. (pp. 1–127).
at Monks Wood National Nature Reserve, UK, using dual-wavelength SAR interfer- Marklund, L. G. (1988). Biomassafunktioner för tall, gran och björk i Sverige. Sveriges
ometry. Remote Sensing of Environment, 108, 224–239. Landbruksuniversitet. Rapporter Skog, 45, 1–73 (In Swedish with English summary).
Bohlin, J., Wallerman, J., & Fransson, J. E. S. (2012). Forest variable estimation using Mette, T., Papathanassiou, K. P., & Hajnsek, I. (2004). Biomass estimation from
photogrammetric matching of digital aerial images in combination with a high- polarimetric SAR interferometry over heterogeneous forest terrain. Proc. IEEE Intl.
resolution DEM. Scandinavian Journal of Forest Research. http://dx.doi.org/10.1080/ Geosci. Remote Sensing Symp. (IGARSS), Anchorage, USA, Vol. 1. (pp. 511–514)
02827581.2012.686625. 0-7803-8742-2 (Sept. 2004).
Breidenbach, J., & Astrup, R. (2012). Small area estimation of forest attributes in the Næsset, E. (2002). Predicting forest stand characteristics with airborne scanning laser
Norwegian National Forest Inventory. European Journal of Forest Research, 131, using a practical two-stage procedure and field data. Remote Sensing of Environment,
1255–1267. 80, 88–99.
Chen, Y., Shi, P., Deng, L., & Li, J. (2007). Generation of a top-of-canopy Digital Elevation Næsset, E. (2004a). Practical large-scale forest stand inventory using a small-footprint
Model (DEM) in tropical rain forest regions using radargrammetry. International airborne scanning laser. Scandinavian Journal of Forest Research, 19, 164–179.
Journal of Remote Sensing, 28(19), 4345–4349. Næsset, E. (2004b). Accuracy of forest inventory using airborne laser-scanning:
Dobson, M. C., Ulaby, F. T., Le Toan, T., Beaudoin, A., Kasischke, E. S., & Christensen, N. (1992). Evaluating the first Nordic full-scale operational project. Scandinavian Journal of Forest
Dependence of radar backscatter on coniferous forest biomass. IEEE Transactions on Research, 19, 554–557.
Geoscience and Remote Sensing, 30, 412–415. Næsset, E. (2007). Airborne laser scanning as a method in operational forest inventory:
Englhart, S., Keuck, V., & Siegert, F. (2011). Aboveground biomass retrieval in tropical Status of accuracy assessments accomplished in Scandinavia. Scandinavian Journal
forests — The potential of combined X- and L-band SAR data use. Remote Sensing of of Forest Research, 22, 433–442.
Environment, 115(5), 1260–1271. Næsset, E., Gobakken, T., Solberg, S., Gregoire, T. G., Nelson, R., Ståhl, G., et al. (2011).
Englhart, S., Keuck, V., & Siegert, F. (2012). Modeling aboveground biomass in tropical Model-assisted regional forest biomass estimation using LiDAR and InSAR as auxiliary
forests using multi-frequency SAR data—a comparison of methods. IEEE Journal of data: A case study from a boreal forest area. Remote Sensing of Environment, 115,
Selected Topics in Applied Earth Observations and Remote Sensing, 5(1), 298–306. 3599–3614.
Eriksson, L. E. B., Magnusson, M., Fransson, J. E. S., Sandberg, G., & Ulander, L. M. H. Neeff, T., Dutra, L. V., Santos, J. R., Freitas, C. C., & Araújo, L. S. (2005). Tropical forest
(September 25–28). Stem volume estimation for boreal forest using ALOS measurement by interferometric height modeling and P-band backscatter. Forest
PALSAR. 5th international symposium on retrieval of bio- and geophysical Science, 51(6), 585–594.
 S. Solberg et al. / Remote Sensing of Environment 139 (2013) 60–67 67

Patenaude, G., Milne, R., & Dawson, T. P. (2005). Synthesis of remote sensing approaches Santos, J. R., dos Freitas, C., da Costa Araújo, L., Dutra, L. V., Mura, J. C., Gama, F., et al.
for forest carbon estimation: Reporting to the Kyoto Protocol. Environmental Science (2003). Airborne P-band SAR applied to the above ground biomass studies in the
& Policy, 8(2), 161–178. Brazilian tropical rainforest. Remote Sensing of Environment, 87(4), 482–493.
Perko, R., Raggam, H., Deutscher, J., Deutscher, J., Gutjahr, K., & Schardt, M. (2011). Forest Schabenberger, O., & Pierce, F. J. (2002). Contemporary statistical models for the plant and
assessment using high resolution SAR data in X-band. Remote Sensing, 3(4), 792–815. soil sciences. Boca Raton: CRC-Press (738 pp.).
http://dx.doi.org/10.3390/rs3040792. Solberg, S., Astrup, R., Bollandsås, O. M., Næsset, E., & Weydahl, D. J. (2010). Deriving
Ranson, K. J., Saatchi, S., & Sun, G. Q. (1995). Boreal forest ecosystem characterization with forest monitoring variables from X-band InSAR SRTM height. Canadian Journal of
SIR-C/X-SAR. IEEE Transactions on Geoscience and Remote Sensing, 33, 867–876. Remote Sensing, 36(1), 68–79.
Ranson, K. J., & Sun, G. Q. (1994). Mapping biomass of a northern forest using Solberg, S., Astrup, R., Gobakken, T., Næsset, E., & Weydahl, D. J. (2010). Estimating spruce
multifrequency SAR data. IEEE Transactions on Geoscience and Remote Sensing, and pine biomass with interferometric X-band SAR. Remote Sensing of Environment,
32, 388–396. 114, 2353–2360.
Reigber, A., & Moreira, J. (1997). Phase unwrapping by fusion of local and global methods. St-Onge, B., Vega, C., Fournier, R. A., & Hua, Y. (2008). Mapping canopy height using a
Proceedings of IGARSS'97 Symposium 3–8 Aug 1997. Geoscience and Remote Sensing combination of digital stereo-photogrammetry and lidar. International Journal of
(1997). Remote Sensing, 29(11), 3343–3364.
Rombach, M., & Moreira, J. (2003). Description and applications of the multipolarized dual Tomppo, E., Haakana, M., Katila, M., & Peräsaari, J. (2008). Multi-source National Forest
band OrbiSAR-1 InSAR sensor. Proceedings of the International Radar Conference. IEEE. Inventory. Methods and applications: Springer (373 pp.).
(pp. 245–250) (3–5 Sept. 2003). Toutin, T., & Gray, L. (2000). State-of-the-art of elevation extraction from satellite SAR
Santoro, M., Askne, J., & Dammert, P. B. G. (2005). Tree height influence on ERS interfer- data. ISPRS Journal of Photogrammetry and Remote Sensing, 55(1), 13–33.
ometric phase in boreal forest. IEEE Transactions on Geoscience and Remote Sensing, Treuhaft, R. N., & Siqueira, P. R. (2004). The calculated performance of forest structure and
43(2), 207–217. biomass estimates from interferometric radar. Waves in Random Media, 14(2),
Santoro, M., Askne, J., Smith, G., & Fransson, J. E. S. (2002). Stem volume retrieval in boreal 345–358.
forests from ERS-1/2 interferometry. Remote Sensing Of Environment, 81(1), 19–35. Vestjordet, E. (1967). Functions and tables for volume of standing trees. Norway spruce.
Santoro, M., Beer, C., Cartus, O., Schmullius, C., Shvidenko, A., McCallum, I., et al. (2011). Meddr. Norske SkogforsVes. (pp. 543–574) (In Norwegian with English summary).
Retrieval of growing stock volume in boreal forest using hyper-temporal series of Way, J., Rignot, E. J. M., Mcdonald, K. C., Oren, R., Kwok, R., Bonan, G., et al. (1994).
Envisat ASAR ScanSAR backscatter measurements. Remote Sensing of Environment, Evaluating the type and state of Alaska Taiga forests with imaging radar for use in
115, 490–507. ecosystem models. IEEE Transactions on Geoscience and Remote Sensing, 32, 353–370.
Santoro, M., Shvidenko, A., McCallum, I., Askne, J., & Schmullius, C. (2007). Properties of Woodhous, I. H. (2006). Predicting backscatter-biomass and height-biomass trends using
ERS-1/2 coherence in the Siberian boreal forest and implications for stem volume a macroecology model. IEEE Transactions on Geoscience and Remote Sensing, 44(4),
retrieval. Remote Sensing of Environment, 106(2), 154–172. 871–877.