Vegetation height estimation from Shuttle Radar Topography Mission and National Elevation Datasets Josef Kellndorfer a, * , Wayne Walker a , Leland Pierce a , Craig Dobson a , Jo Ann Fites b , Carolyn Hunsaker c , John Vona d , Michael Clutter e a Radiation Laboratory, EECS Department, The University of Michigan, 1301 Beal Avenue, Ann Arbor, MI 48109, United States b USDA Forest Service, Adaptive Management Services, Tahoe National Forest, Nevada City, CA, United States c USDA Forest Service, Forestry Sciences Laboratory, Pacific South West Research Station, Fresno, CA, United States d Plum Creek Timber Company, Watkinsville, GA, United States e Warnell School of Forest Resources, The University of Georgia, Athens, GA, United States Received 16 April 2004; received in revised form 26 July 2004; accepted 27 July 2004 Abstract A study was conducted to determine the feasibility of obtaining estimates of vegetation canopy height from digital elevation data collected during the 2000 Shuttle Radar Topography Mission (SRTM). The SRTM sensor mapped 80% of the Earth’s land mass with a C-band Interferometric Synthetic Aperture Radar (InSAR) instrument, producing the most complete digital surface map of Earth. Due to the relatively short wavelength (5.6 cm) of the SRTM instrument, the majority of incoming electromagnetic energy is reflected by scatterers located within the vegetation canopy at heights well above the bbald-EarthQ surface. Interferometric SAR theory provides a basis for properly identifying and accounting for the dependence of this scattering phase center height on both instrument and target characteristics, including relative and absolute vertical error and vegetation structural attributes. An investigation to quantify the magnitude of the vertical error component was conducted using SRTM data from two vegetation-free areas in Iowa and North Dakota, revealing absolute errors of 4.0 and 1.1 m, respectively. It was also shown that the relative vertical error due to phase noise can be reduced significantly through sample averaging. The relative error range for the Iowa site was reduced from 13 to 4 m and for the North Dakota site from 7 to 3 m after averaging of 50 samples. Following error reduction, it was demonstrated that the SRTM elevation data can be successfully correlated via linear regression models with ground-measured canopy heights acquired during the general mission timeframe from test sites located in Georgia and California. Prior to outlier removal and phase noise reduction, initial adjusted r 2 values for the Georgia and California sites were 0.15 and 0.20, respectively. Following outlier analysis and averaging of at least 20 SRTM pixels per observation, adjusted r 2 values for the Georgia and California sites improved to 0.79 (rmse=1.1 m) and 0.75 (rmse=4.5 m), respectively. An independent validation of a novel bin-based modeling strategy designed for reducing phase noise in sample plot data confirmed both the robustness of the California model (adjusted r 2 =0.74) as well as the capacity of the binning strategy to produce stable models suitable for inversion (validated rmse=4.1 m). The results suggest that a minimum mapping unit of approximately 1.8 ha is appropriate for SRTM-based vegetation canopy height mapping. D 2004 Elsevier Inc. All rights reserved. Keywords: SRTM; InSAR; NED; Vegetation canopy height; Biomass; Carbon; Noise reduction 1. Introduction Accurate estimates of aboveground terrestrial biomass and carbon stocks are dependent on the availability of biophysical measures that capture both the horizontal and vertical structural character of vegetation. In general, 0034-4257/$ - see front matter D 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2004.07.017 * Corresponding author. Tel.: +1 734 763 9442; fax: +1 734 647 2106. E-mail address: [email protected] (J. Kellndorfer). Remote Sensing of Environment 93 (2004) 339 – 358 www.elsevier.com/locate/rse
20
Embed
Vegetation height estimation from Shuttle Radar Topography Mission and National Elevation Datasets
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
www.elsevier.com/locate/rse
Remote Sensing of Environm
Vegetation height estimation from Shuttle Radar Topography Mission and
National Elevation Datasets
Josef Kellndorfera,*, Wayne Walkera, Leland Piercea, Craig Dobsona, Jo Ann Fitesb,
Carolyn Hunsakerc, John Vonad, Michael Cluttere
aRadiation Laboratory, EECS Department, The University of Michigan, 1301 Beal Avenue, Ann Arbor, MI 48109, United StatesbUSDA Forest Service, Adaptive Management Services, Tahoe National Forest, Nevada City, CA, United States
cUSDA Forest Service, Forestry Sciences Laboratory, Pacific South West Research Station, Fresno, CA, United StatesdPlum Creek Timber Company, Watkinsville, GA, United States
eWarnell School of Forest Resources, The University of Georgia, Athens, GA, United States
Received 16 April 2004; received in revised form 26 July 2004; accepted 27 July 2004
Abstract
A study was conducted to determine the feasibility of obtaining estimates of vegetation canopy height from digital elevation data collected
during the 2000 Shuttle Radar Topography Mission (SRTM). The SRTM sensor mapped 80% of the Earth’s land mass with a C-band
Interferometric Synthetic Aperture Radar (InSAR) instrument, producing the most complete digital surface map of Earth. Due to the
relatively short wavelength (5.6 cm) of the SRTM instrument, the majority of incoming electromagnetic energy is reflected by scatterers
located within the vegetation canopy at heights well above the bbald-EarthQ surface. Interferometric SAR theory provides a basis for properly
identifying and accounting for the dependence of this scattering phase center height on both instrument and target characteristics, including
relative and absolute vertical error and vegetation structural attributes.
An investigation to quantify the magnitude of the vertical error component was conducted using SRTM data from two vegetation-free
areas in Iowa and North Dakota, revealing absolute errors of �4.0 and �1.1 m, respectively. It was also shown that the relative vertical error
due to phase noise can be reduced significantly through sample averaging. The relative error range for the Iowa site was reduced from 13 to 4
m and for the North Dakota site from 7 to 3 m after averaging of 50 samples. Following error reduction, it was demonstrated that the SRTM
elevation data can be successfully correlated via linear regression models with ground-measured canopy heights acquired during the general
mission timeframe from test sites located in Georgia and California. Prior to outlier removal and phase noise reduction, initial adjusted r2
values for the Georgia and California sites were 0.15 and 0.20, respectively. Following outlier analysis and averaging of at least 20 SRTM
pixels per observation, adjusted r2 values for the Georgia and California sites improved to 0.79 (rmse=1.1 m) and 0.75 (rmse=4.5 m),
respectively. An independent validation of a novel bin-based modeling strategy designed for reducing phase noise in sample plot data
confirmed both the robustness of the California model (adjusted r2=0.74) as well as the capacity of the binning strategy to produce stable
models suitable for inversion (validated rmse=4.1 m). The results suggest that a minimum mapping unit of approximately 1.8 ha is
appropriate for SRTM-based vegetation canopy height mapping.
Given that the absolute error associated with the SRTM
and reference DEMs varies at a relatively broad spatial scale
(thousands of kilometers), dv is expected to result in a
constant (i.e., trendless) offset for relatively small regions
and a spatially variable offset or trend for larger regions.
Following the relative and absolute error formulations
provided in Eqs. (4) and (5), respectively, Eq. (3) can now
be reformulated as:
hspc ¼ fveg vs; vmÞ þ dv þ erð ð6Þ
It follows from Eq. (6) that in order to accurately model
the functional relationship between vegetation character-
istics (vs and vm) and hspc, it is desirable to identify and
remove dv and minimize, to the extent possible, the error
component er, which is primarily dependent on the
reduction of epn.
5.2. Investigation of SRTM phase noise error and vertical
offset
In an attempt to quantify the potential range of dv and epnvalues to be expected within the SRTM coverage of the
conterminous United States, an analysis was carried out
using SRTM and NED elevation data from two large (c1.5
km2) agricultural fields, one in Iowa (IA) and one in North
Dakota (ND) (Fig. 2). Agricultural fields were selected
because they tend to be flat, and during the February 2000
timeframe of the SRTM mission, they would have been
devoid of vegetation. The selected fields were equal in size,
covering an area of approximately 1600 SRTM pixels.
Based on inspection of the SRTM coverage map (Fig. 2),
Field IA (41855VN, 94827VW) was located in a region
where only one SRTM data take was obtained. Conversely,
Field ND (48851VN, 101800VW) was selected from a region
where at least four data takes were acquired. The averaging
of multiple data takes, ranging from 2 to as many as 10 or
more, was performed during the NASA-JPL product
generation phase (USGS, 2003). Given the timeframe of
the SRTM mission and the northerly latitude of Iowa and
North Dakota, both fields are assumed to have been snow-
covered when the SRTM data were acquired. At northerly
latitudes, snow is assumed to be essentially dry in February
and, hence, the presence of snow was not expected to affect
the estimation of dv (Rignot et al., 2001). Given that several
more data takes (N4) were available for Field ND compared
with Field IA (Fig. 1), less phase noise was expected in the
SRTM pixel values for Field ND.
The first step in the analysis was the elimination of
existing topographic variation within the fields by subtract-
ing the NED DEM from the SRTM DEM. Under the
assumption that noise-free SRTM and NED data should
provide for consistent bald-Earth elevation estimates, it
follows that: (1) the residuals of the SRTM–NED difference
image can be attributed to epn, and (2) the mean of the
residuals, should it differ from zero, is an estimate of the
vertical offset (dv) between the SRTM and NED datasets.
From the frequency distribution and summary statistics for
the two difference images, a number of inferences can be
made (Fig. 7, Table 1). First, the phase noise within both
fields is observed to be Gaussian in nature. Second, the
mean value calculated from the difference images (IA=�4.0
m, ND=�1.1 m) is other than zero, reflecting a vertical
offset between the SRTM and NED datasets, which differs
between sites by approximately 2.9 m. Third, the narrower
noise range associated with Field ND (7 m) compared to
that of Field IA (13 m) confirms the hypothesis that greater
noise reduction occurs in locations (i.e., North Dakota)
where multiple data takes have been averaged. In order to
quantify the relationship between sample (i.e., pixel)
averaging and subsequent noise reduction, samples from
each of the difference images were block-averaged with a
window size ranging from 3�3 to 25�25 (i.e., 9–625
samples per block). Fig. 8 shows the results for each field.
In both cases, the results are shown after the removal (via
subtraction) of the respective dv offset. As expected, the
noise reduction is reflected in a decrease in the noise range
with increasing window size. This reduction is more
pronounced up to ca. 7�7 samples averaged, with a sharper
Fig. 7. Histogram of sample values extracted from Fields IA (Iowa) and ND (North Dakota) illustrating the presence of Gaussian phase noise (epn).
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358 349
decrease observed in Field IA. At 7�7 samples (~50 pixels)
averaged, the range for Field IA was reduced from 13 m to
approximately 4 m, and the range for Field ND was reduced
from 7 m to approximately 3 m. Beyond the 7�7 threshold,
a more gradual decrease is observed in both fields and the
maximum, minimum, and mean values begin to converge.
Table 1
Summary statistics for phase noise (epn) from Fields IA (Iowa) and ND
(North Dakota)
Field IA Field ND
41.918N, 94.458W 48.858N, 101.008W
SRTM Min 330.0 449.0
Mean 337.5 454.1
Max 346.0 457.0
Range 16.0 8.0
NED Min 337.7 451.4
Mean 341.5 455.1
Max 345.8 456.7
Range 8.1 5.3
SRTM–NED Min �10.1 �4.1
Mean �4.0 �1.1
Max 2.6 2.7
Range 12.7 6.8
At a threshold of 25�25 (=625) samples, the noise range is
reduced to less than 1 m.
Compared to the relatively flat, nonvegetated regions
studied here, the phase noise inherent to SRTM data from
forested terrain should be far less pronounced due to the
higher signal-to-noise ratio associated with InSAR back-
scatter from vegetation canopies. Hence, the noise curves
shown in Fig. 8 should represent worst-case scenarios with
respect to the relative error to be expected in the SRTM data,
particularly for Field IA where only one SRTM data take
was acquired.
5.3. Test site error reduction and vertical offset identification
An effort was made to apply the knowledge of SRTM
phase noise and vertical offsets gained above to the analysis
of the Jesup and Sierra Nevada test sites. The extent to
which these error sources were present in the datasets was
explored graphically via the assessment of elevation profiles
extracted from the SRTM, NED, and SRTM–NED differ-
ence images. A series of elevation profiles (A–C) was
extracted from the Jesup test site along selected transects
(Fig. 4b). In general, Profile A, representing a transect of
Fig. 8. Comparison of phase noise (epn) statistics (minimum, mean, and maximum) for Fields IA (Iowa) and ND (North Dakota) following a block averaging
procedure in which block size was increased linearly from 3�3 (nine samples) to 25�25 (625 samples). All values were adjusted by the mean absolute error
determined from each field.
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358350
approximately 15 km, demonstrates the close overall
correspondence between the SRTM and NED elevation
datasets. Further inspection of the profile reveals the
pronounced relative difference in elevation between the
two DEMs, which is best reflected in the SRTM–NED
difference (hspc) curve shown at the bottom of Profile A.
This difference in elevation reflects the presence of, and
variability associated with, vegetation and anthropogenic
features along the transect. Profiles B and C are approx-
imately 2–3 km in length and correspond to transects across
riparian (B) and upland (C) areas (Fig. 4b). In profile B, the
SRTM–NED difference curve reveals quite clearly the
presence of peaks associated with bands of riparian forest.
Similarly, profile C shows the sharp boundaries between
alternating plantations and clearcuts.
Elevation profiles (A–C) extracted from the Sierra
Nevada test site illustrate the greater topographic variability
and elevation range that distinguishes this site from Jesup
(Fig. 6b). Whereas Sierra Nevada Profile (SP) A exhibits an
elevation range of approximately 400 m, Jesup Profile (JP)
A has a range that is closer to 25 m (Fig. 4b). In general, SP
A also suggests the more continuous nature of the forest and
the presence of taller trees. Sierra Nevada Profiles B and C
are approximately 2–3 km in length and are intended to
complement the corresponding Jesup profiles. Profile B
reflects a transect crossing three more or less parallel
riparian zones (Fig. 6b). Again, the SRTM–NED difference
curve (bottom) captures the presence of taller vegetation in
these moister, more hospitable areas. The transect associated
with Profile C crosses a region of alternating plantations and
clearcuts. The SRTM–NED difference curve (bottom) is
consistent with that of JP C, revealing the sharp boundaries
between these features.
In general, the elevation profiles presented in Figs. 4b
and 6b illustrate the marked sensitivity of the SRTM sensor
to vegetation canopy height. At the same time, however, the
profiles also reveal the presence of phase noise error. As a
result, and in order to improve subsequent scattering phase
center height calculations, efforts were undertaken to reduce
the observed noise component. In both test sites, this was
accomplished using a sample (i.e., pixel) averaging
approach.
For the Jesup site, sample averaging was carried out
within the boundaries of each of the 22 stands described in
Section 4.1. The field data (i.e., cluster plots) were
aggregated within these boundaries as well. The interior of
stand boundaries was buffered by 30 m (i.e., one SRTM
pixel) to minimize edge effects resulting from geocoding
errors. Following this procedure, the number of samples per
stand ranged from 2 to 574.
For the Sierra Nevada site, sample averaging was
accomplished by extracting all pixels within a radius of 45
m around the center point of each plot. This zone of
inclusion was selected as a basis for averaging all pixels
having a majority of their area within the boundary of the
56.4-m (1.0 ha) sample plot annulus. Given the relatively
low pixel-to-plot area ratio, the extraction procedure
resulted in the averaging of just 6–10 pixels per plot. Based
on the results presented in Fig. 8, this sample size was
deemed much too small to effectively reduce the phase
noise component. Operating under the assumption that
larger sample plots (e.g., coincident with stand-level
boundaries) would increase the pixel-to-plot area ratio, a
novel strategy was devised based on a linear binning or
partitioning of the observed canopy height variable. The
goal of this strategy was to effectively simulate stand-level
statistical averaging using plot level data. To implement the
approach, the original 227-plot dataset was divided among
20 bins, with the number of plots per bin ranging from 3 to
29. With the exception of the lowermost and two uppermost
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358 351
bins, the bin width was set to 1.5 m. Bin width was
determined based on a desire to obtain a minimum of 20
pixels per bin. Under this binning strategy, the number of
pixels per bsimulated standQ or bin ranged from 24 to 253—
a sample size much more suitable for reducing epn.Within each bin, the observed mean canopy height (hcan)
was computed according to Eq. (2), with NT representing
the sum of all stems in all plots within the respective bin,
and ht corresponding to the observed heights of these stems.
In general, the binning strategy can be equated to simulating
mean observed canopy and associated scattering phase
center heights within homogeneous (in terms of observed
height), contiguous stands, with the caveat that individual
sample plots are not co-located within contiguous stands but
are distributed across a larger forested region (see plot
locations in Fig. 5).
To determine if a vertical offset (dv) existed between the
SRTM and NED datasets for the test sites, regions
representing barren, nonvegetated areas were identified
from the National Land Cover Dataset (NLCD) (Vogelmann
et al., 2001) and recent Landsat ETM+ imagery. For the
Jesup site, the analysis revealed a small offset of �0.2 m
(i.e., bald-Earth SRTM elevations were, on average, 20 cm
lower than NED elevations). For the Sierra Nevada test site,
the offset was �1.3 m.
5.4. Test site scattering phase center height calculation
Following the procedures for error accounting described
above, the mean scattering phase center height (hspc) for a
vegetated region can be estimated within relatively narrow
error bounds if a large-enough sample population can be
averaged. Substituting the unknown functional dependency
on vegetation characteristics in Eq. (6) with the SRTM–
NED difference, the mean scattering phase center heights
for the Jesup and Sierra Nevada test sites were calculated
according to:
h¯ spc ¼1
Np
XNP
i¼1
hSRTM � hNEDÞ þ dvð ð7Þ
where hspc=mean scattering phase center height; hSRTM=
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358352
relatively large standard deviations in observed height (on
the order of 5 m and above), which points to the presence of
a dissimilar stand structure characterized by a more
heterogeneous canopy. Whereas this heterogeneity should
also be reflected in the corresponding standard deviations of
the SRTM hspc values, it is not (Table 2). This is due to the
fact that the variability in observed height within the four
stands exists at a scale less than the resolution of the SRTM
data. Visual inspection of the stands both on the ground and
Fig. 9. Plot of SRTM scattering phase center height (hspc) versus observed stand he
key differences in stand variables including stand size [e.g., triangles indicate sma
indicate stands with high height variability (S.D.N5)].
in high-resolution imagery confirms that these stands exhibit
a very heterogeneous structure including windfalls and
small clearings, which contribute to the large variance in
observed height. Thus, these stands were deemed not
representative of the remaining sample and would have to
be treated with an alternative approach (e.g., elimination of
the windfall/clear areas from both the ground and image
sample). Since the ground survey did not record exact plot
locations, this was not possible retrospectively. Removal of
ight (hcan) for all stands (22) in the Jesup, Georgia, test site. Symbols reflect
ll stands with few (b20) pixels per stand] and stand structure [e.g., squares
Fig. 10. Plot of SRTM scattering phase center height (hspc) versus observed stand height (hcan) for 15 stands in the Jesup, Georgia, test site.
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358 353
the four stands resulted in a significant improvement in the
correlation between SRTM hspc and hcan. The adjusted r2
value of the model with seven outlier stands removed was
0.79 with an rmse of 1.1 m (Fig. 10).
Under the assumption that additional sample averaging
could further reduce the influence of phase noise on model
performance, a minimum threshold of 50 pixels per stand was
explored since (1) according to Fig. 8, the noise range is
reduced significantly for stands of this size, and (2) a sufficient
number of stands (i.e., 12) remained available with which to
develop the regression model (Table 2). The resulting adjusted
r2 value was 0.86 with an rmse of 1.0 m (Fig. 10).
Table 3
Summary statistics for 20 bsimulated standsQ (i.e., bins) based on data from 227
Bin
number
Bin range
[m]
Number
of plots
per bin
Plot
size
[m2]
Number
of SRTM
pixels per bin
Mean canop
height [m]
(hcan)
1 b5.00 3 707 26 3.5
2 5.00–6.49 3 707 25 5.9
3 6.50–7.99 3 707 26 7.4
4 8.00–9.49 11 707 85 8.8
5 9.50–10.99 29 707 253 10.2
6 11.00–12.49 24 707 214 11.8
7 12.50–13.99 20 707 173 13.3
8 14.00–15.49 22 707 191 14.7
9 15.50–16.99 19 707 158 16.1
10 17.00–18.49 21 707 184 17.8
11 18.50–19.99 10 707 87 19.2
12 20.00–21.49 16 707 139 20.9
13 21.50–22.99 7 707 56 22.0
14 23.00–24.49 13 707 114 23.8
15 24.50–25.99 5 707 44 24.9
16 26.00–27.49 4 707 35 27.1
17 27.50–28.99 6 707 47 28.5
18 29.00–30.49 4 707 35 29.4
19 30.50–33.99 4 707 36 33.0
20 34.00–38.00 3 707 24 37.3
The parameters of the linear regression models associated
with the 20- and 50-pixel thresholds indicate that the height
of the scattering phase center increases moderately
(slope=1.2–1.3) with increasing canopy height and lies
approximately 6+ m (intercept=5.8–6.6) below the canopy
surface.
7.2. Test site Sierra Nevada, California
An overview of biophysical parameters, including hspcand observed hcan values, corresponding to the 20 simulated
Sierra Nevada stands (i.e., bins) is provided in Table 3. Prior
plots surveyed within the test site Sierra Nevada, CA
y S.D.
height
[m]
Basal area
per bin
[m2/ha]
Number
of trees
per bin
SRTM–NED
Height [m]
(hspc)
S.D.
SRTM–NED
height [m]
0.9 0.2 17 3.3 3.4
2.8 2.6 91 3.7 2.2
4.0 3.7 122 8.7 1.8
3.9 10.3 220 7.5 1.8
5.0 47.2 990 13.0 2.9
6.3 57.0 805 12.2 2.8
7.5 55.5 663 14.1 2.6
7.5 68.2 688 16.1 2.6
9.3 63.0 483 12.8 2.7
8.1 80.2 684 18.1 2.9
10.5 46.0 289 18.3 1.9
10.3 60.5 331 19.5 3.0
10.4 25.8 172 20.8 2.4
11.1 59.7 321 21.3 3.4
11.0 18.1 78 19.7 4.0
8.2 20.2 100 18.6 2.4
10.7 33.7 139 19.9 3.0
12.4 20.0 69 13.5 3.0
12.8 18.3 52 22.1 2.9
10.1 15.1 38 26.3 2.8
Fig. 11. Plot of SRTM scattering phase center height (hspc) versus observed plot height (hcan) for all plots (227) in the Sierra Nevada, California, test site.
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358354
to binning, a simple linear regression analysis was con-
ducted with observed hcan regressed on hspc (Fig. 11). Data
from all 227 plots were included in model development. The
result was an adjusted r2 value of 0.20 with an rmse of 5.8
m (Table 4), indicating the presence of only a slight positive
correlation. This relatively weak relationship was attributed
to the presence of residual phase noise resulting from the
fact that only 6–10 pixels were averaged per plot under the
45-m buffering strategy. This number was observed to be
quite low in light of the findings presented in Fig. 8 as well
as in comparison to the Jesup site where the average stand
size was significantly larger and hence the number of pixels
extracted was markedly greater (i.e., N20 pixels following
outlier removal) . The binning approach described in
Section 5.3 was therefore undertaken in an effort to simulate
stand-level statistics using plot-level data. Under this
strategy, the number of pixels being averaged per observa-
tion was increased from a minimum of 6 to a minimum of
20 (Table 3). Fig. 12 shows the hspc values plotted against
observed hcan values for the resulting 20 height bins. The
regression analysis based on this strategy resulted in a
significantly improved adjusted r2 value of 0.75 with an
rmse of 4.5 m (Fig. 12; Table 4). Based on the Jesup results,
Table 4
Results of simple linear regression analyses performed on the test sites Jesup, Ge
Dataset Observed Adjusted r
Jesup test site: linear least squares regression
Number of pixels N20 19 0.42
Number of pixels N20 and hcan S.D. b5 m 15 0.79
Number of pixels N50 and hcan S.D. b5 m 12 0.86
Sierra Nevada test site: linear least squares regression
No pixels threshold (based on all plots) 227 0.20
Number of pixels N20 (based on all plots) 20 0.75
Number of pixels N50 (based on all plots) 11 0.86
Number of pixels N20 (based on 126 training plots) 20 0.74
an attempt was made to determine whether any remaining
phase noise could be removed. Again, a minimum threshold
of 50 pixels per bin was tested, which resulted in 11 height
bins (Table 3). The least squares regression based on these
data resulted in an adjusted r2 value of 0.86 with an rmse of
1.7 m (Fig. 12; Table 4). Inspection of Fig. 12 reveals the
influence that bins with fewer than 50 samples had on the
20-bin model prior to their removal. In the initial 20-bin
model, the slope and intercept were 1.336 and �1.892,
respectively. The regression based on the remaining eleven
bins resulted in slight decreases in both the slope (1.088)
and intercept values (�0.940).
7.3. Model inversion and validation
In the context of this research, validation efforts were
focused on the bin-based modeling approach developed and
applied as part of the Sierra Nevada pilot study. Validation
was limited to the Sierra Nevada site given (1) the limited
amount of field data available for the Jesup site (22
georeferenced stands) compared to the Sierra Nevada site
(227 georeferenced plots), and (2) a desire to conduct a
formal and independent assessment of the bin-based
orgia, and Sierra Nevada, California
2 F statistic p value Intercept (b0) Slope (b1) rmse [m]
14.30 0.00149 9.785 0.784 1.8
54.93 N0.0001 6.619 1.160 1.1
67.88 N0.0001 5.778 1.261 1.0
58.81 N0.0001 10.883 0.370 5.8
58.15 N0.0001 �1.892 1.336 4.5
63.72 N0.0001 �0.940 1.088 1.7
55.44 N0.0001 �1.775 1.351 4.6
Fig. 12. Plot of SRTM scattering phase center height (hspc) versus observed binned plot height (hcan) for 20 bsimulated standsQ (i.e., bins) in the Sierra Nevada,
California, test site.
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358 355
modeling strategy employed as part of the Sierra Nevada
analysis.
A validation effort based on independent training and
testing datasets was carried out in order to test the robustness
of the binning strategy summarized in Table 3. Here, a 20-
pixel minimum threshold was established to pool the plot
data into 20 bins (height classes), thereby simulating 20
stands of relatively homogeneous height. To properly
validate the binning approach, sufficient data were needed
within each bin to allow for the 20-pixel minimum threshold
to be maintained in both training and testing datasets. Thus,
Table 5
Summary statistics for bsimulated standsQ (i.e., bins) used in regression model de
Bin
number
Bin range [m] Training (126 plots)
Number
of plots
per bin
Number
of SRTM
pixels per bin
Mean canopy
height [m]
(hcan)
SR
hei
(hs
1 b5.00 3 26 3.5 3.
2 5.00–6.49 3 25 5.9 3.
3 6.50–7.99 3 26 7.4 8.
4 8.00–9.49 6 47 8.8 8.
5 9.50–10.99 14 119 10.3 12.
6 11.00–12.49 12 108 11.9 11.
7 12.50–13.99 10 90 13.3 13.
8 14.00–15.49 11 99 14.6 16.
9 15.50–16.99 10 85 16.1 12.
10 17.00–18.49 10 88 17.8 18.
11 18.50–19.99 5 45 19.0 18.
12 20.00–21.49 8 68 20.8 20.
13 21.50–22.99 4 32 22.2 19.
14 23.00–24.49 6 51 23.8 20.
15 24.50–25.99 3 27 24.9 20.
16 26.00–27.49 4 35 27.1 18.
17 27.50–28.99 3 23 28.6 17.
18 29.00–30.49 4 35 29.4 13.
19 30.50–33.99 4 36 33.0 22.
20 34.00–38.00 3 24 37.3 26.
for all bins containing 40 (i.e., 2�20) pixels or more (see
Table 3), half of the sample plots (102 plots) and associated
pixels were randomly selected for model development (i.e.,
training), while the remaining half (101 plots) were reserved
for inversion and validation (i.e., testing) (Table 5). Bins
containing fewer than 40 pixels (24 additional plots) were
used exclusively for model development. Thus, whereas the
model was trained on the entire range of height classes (0–38
m) represented by the 20 bins (126 plots), validation was
conducted on height classes ranging from 8 to 26 m and from
27.5 to 29 m represented by 13 bins (101 plots) (Table 5).
velopment (20 training bins) and validation (13 testing bins)
Testing (101 plots)
TM–NED
ght [m]
pc)
Number
of plots
per bin
Number
of SRTM
pixels per bin
Mean canopy
height [m]
(hcan)
SRTM–NED
height [m]
(hspc)
3
7
7
3 5 38 8.7 6.5
0 15 134 10.1 13.9
4 12 106 11.8 13.1
5 10 83 13.3 14.8
2 11 92 14.7 16.0
6 9 73 16.1 12.9
2 11 96 17.8 17.9
6 5 42 19.3 17.9
0 8 71 21.0 19.0
9 3 24 21.9 21.9
1 7 63 23.8 22.3
2 2 17 24.8 18.9
6
2 3 24 28.3 22.5
5
1
3
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358356
The regression analysis applied to the training dataset
resulted in an adjusted r2 value of 0.74 with an rmse of 4.6 m
(Table 3). These results are consistent with the adjusted r2
and rmse values (0.75, 4.5 m) reported for the original 20-
bin model developed using data from all 227 plots.
Furthermore, the slope (b0) and intercept (b1) parameters
associated with the original 20-bin model (b0=1.336,
b1=�1.892) and the 20-bin training model (b0=1.351,
b1=�1.775) are virtually identical, indicating that the
original model parameters were appropriate. When the
training model is inverted to predict observed canopy height
values for the independent 13-bin testing dataset, an rmse of
4.1 m is achieved (Table 4), a result consistent with those
reported above for both the original 20-bin dataset (rmse=4.5
m) and the 20-bin validation dataset (rmse=4.6 m).
8. Discussion
In this study, two test sites representing a range of
vegetation structure, moisture, and topographic character-
istics were selected for use in determining the feasibility of
deriving vegetation canopy height from SRTM digital
elevation data. Whereas the Jesup test site is dominated
by heavily managed, relatively homogeneous plantations of
slash pine on moderately rolling terrain, the Sierra Nevada
test site is characterized by less heavily managed, relatively
heterogeneous tracts of coniferous forest of variable
composition and structure on highly variable terrain. All
else being equal, these structural-, topographic-, and
associated moisture-related differences have a direct and
differential influence on the height of the scattering phase
center (i.e., the vertical position within the canopy from
where the majority of backscattered electromagnetic energy
is being returned). As expected, different empirical models
are needed to describe the differing relationships that exist
between the two test sites with respect to observed and
remotely sensed estimates of vegetation canopy height.
Preliminary inspection of the data indicated that simple
univariate linear models would be sufficient for describing
these relationships. It follows that efforts to extend this
research to regional, continental, or global scales will
necessarily require development of a more comprehensive
region-specific and/or multivariate approach. To be most
effective, such an approach would be conducted in the
context of an ecoregional framework where regional
ecological drivers of vegetation composition and structure
including climate and topography can be appropriately
controlled. Within individual ecoregions, remotely sensed
descriptors of horizontal structure (e.g., canopy density,
cover type, etc.) can then be used together with SRTM
scattering phase center height information as part of a
multivariate and structurally hierarchical approach to
regression model development.
Whatever the spatial scale at which the SRTM data are
used, based on the research presented here, it is clear that
accurate estimates of vegetation canopy height cannot be
expected at the level of individual SRTM resolution cells
(c900 m2). Due to the relative error component resulting
largely from phase noise, sample averaging is a critical
prerequisite of any effort to derive canopy height informa-
tion from the SRTM data. The results from the Jesup and
Sierra Nevada test sites suggest that 20 pixels is perhaps the
minimum acceptable threshold for sample averaging (Figs.
10 and 12). The results of the Iowa and North Dakota field
studies support this finding but also indicate that a higher
averaging threshold may be warranted depending on what
level of residual relative error is deemed acceptable (Fig. 8).
Regardless of the final threshold selected, thought needs to
be given to the specific method used to aggregate pixels
prior to averaging. The stand-based approach applied in the
Jesup case is generally ideal assuming that (1) survey data
are collected on a per-stand basis; (2) stand boundaries are
available for use in the analysis, and (3) stands exhibit
relatively homogeneous structural characteristics. Since in
the general case one or more of these conditions are likely
not to be met, an alternative averaging strategy with more
universal application is desirable. Future work in this regard
will focus on the use of knowledge-based segmentation,
which provides for the automatic and optimal delineation of
local homogeneous regions reflecting a partitioning of the
landscape into units of similar height and overall structure.
This approach is superior to kernel-based methods (i.e., the
simple buffer employed in the Sierra Nevada case) for
several reasons. First, image segments can be highly
adaptive to the irregular shapes of homogeneous regions
(e.g., bands of riparian forest). Second, ancillary information
including other high-resolution remote sensing imagery
such as Landsat-derived vegetation indices or land cover
information can be incorporated in the retrieval of image
segments in order to refine the process of boundary
delineation. National-scale data of this sort will soon be
available in the United States through the ongoing release of
the National Land Cover database 2001 (Homer et al.,
2004).
Despite the presence of phase noise and other contrib-
utors to the overall relative and absolute height errors (see
Eqs. (3) and (4)), efforts undertaken as part of this research
to both understand and minimize errors sources have been
quite successful. In most cases, the errors themselves are
relatively minor. Some of these errors such as those related
to soil moisture and roughness are nearly impossible to
quantify given the complex nature of SRTM data acquisition
and processing (e.g., averaging of multiple data takes). As a
result, and given their minimal influence on the height
estimation process, no effort has been made to account for
them explicitly. On the other hand, more significant sources
of error such as those attributed to phase noise and vertical
offsets are reasonably well understood and can be mini-
mized, if not removed, using the techniques (e.g., sample
averaging) described here. Future work in the arena of error
characterization will focus on development of methods for
J. Kellndorfer et al. / Remote Sensing of Environment 93 (2004) 339–358 357
minimizing this suite of errors across broader geographic
extents. For example, regional-scale efforts to estimate
vegetation canopy height require techniques for spatially
modeling the absolute error associated with vertical offsets
or trends between the SRTM and NED datasets. Results
from the Iowa/North Dakota investigation of phase noise
indicated a difference in vertical offset amounting to almost
3 m. In the context of regional-scale analyses, spatially
varying vertical trends such as that observed in the Great
Plains can be identified using trend analysis together with
SRTM–NED difference data from vegetation-free areas
(e.g., snow-covered fields, pastures, meadows, etc.). The
trend can then be modeled using geostatistical techniques
such as kriging and accurately removed from the regional
estimates of scattering phase center height.
A number of additional research areas are currently being
explored in the context of efforts to extend SRTM-derived
canopy height estimation to larger scales. Perhaps most
important among these is the need to acquire accurate and
consistent field survey information for use in larger-scale
applications of the SRTM data. One potential source of such
data is the Forest Inventory and Analysis (FIA) program of
the USDA Forest Service (Gillespie, 1999). This federal
program represents the only nationwide source of timely
and reliable forest inventory and monitoring information.
The inventory consists of some 150,000 permanent plots
nationwide; 20% of these are surveyed annually. Generally
speaking, the FIA sampling design is quite similar to that
employed as part of the Sierra Nevada field survey. The
encouraging results obtained from the Sierra Nevada test
site suggest that FIA survey data may be useful for scaling
the SRTM-based height estimation process up to the
conterminous United States and beyond.
A second key area of research involves developing a
more complete physical understanding of the interferometric
scattering phase center height and determining what this
measure best reflects in the context of traditional field
metrics for describing vegetation canopy height. In this
study, the functional relationship between hspc and the
observed mean canopy height was explored exclusively.
However, it is possible that other alternative height metrics
(e.g., median, maximum, or dominant height) may in fact be
more highly correlated with the hspc. In areas like the Jesup
test site, dominated almost exclusively by even-aged pine
plantations, variability in canopy height is very low,
resulting in a situation where very little difference exists
between values of the mean, median, maximum, or
dominant height. In this case, the mean represents an
entirely appropriate measure of the observed height.
However, in regions like the Sierra Nevada where variability
in vertical structure tends to be significant, mean canopy
height is a very different measure from that of maximum or
dominant height. Thus, in the presence of heterogeneous
structural conditions, it may be that hspc is a much better
predictor of one of these alternative height metrics,
especially in stands of moderate to high density.
9. Conclusions
The research presented here demonstrates that elevation
data from the NASA JPL SRTM holds considerable potential
for developing statistically significant inversion models for
deriving estimates of vegetation canopy height. Various
relative and absolute SRTM error sources were identified as
part of this work and broadly applicable methods have been
proposed for removing or significantly reducing these errors.
Most important among these relates to the minimization of
the SRTM-inherent phase noise. Using elevation profiles and
outlier analysis together with an analysis of noise statistics in
vegetation-free areas, it was established that averaging a
minimum of 20 SRTM pixels yields stable estimates of the
mean scattering phase center height. This corresponds to a
minimum SRTM mapping unit for SRTM-based vegetation
canopy height estimation of approximately 1.8 ha.
The possibility of expanding this work to the scale of the
conterminous United States is now being explored. Com-
plementary field and ancillary data sources including the
national FIA database and the soon-to-be released NLCD
2001 represent unprecedented sources of information for
enabling the production of the first-ever nationwide map of
vegetation canopy height. Such a map would undoubtedly
prove invaluable for a wide range of applications including
biomass estimation and carbon accounting as well as timber
management and habitat characterization.
In summary, SRTM elevation data exhibit a significant
sensitivity to the vertical structure of vegetation. Although
part of the user community views vegetation as amajor source
of noise and nuisance in what would otherwise be highly
accurate SRTM-derived bald-Earth elevation data, it is shown
that for the purposes of vegetation height estimation, the
SRTM dataset is an invaluable resource of near-global scope.
Acknowledgements
The collection of ground data within the Jesup test site
was supported by the NASA EOCAP-SAR program and the
Plum Creek Timber. Analysis of the Jesup data was
supported by a NASA Solid Earth grant titled bVegetationCorrected SRTM-Derived Digital Elevation Models.Q Datacollection within the Sierra Nevada test site was supported
by a NASA Earth Science grant titled bForest Structure fromMultispectral Fusion.Q The analysis was supported, in part,
by the aforementioned grant as well as a NASA Earth
System Science Fellowship awarded to the second author.
References
Bamler, R. (1999). The SRTM Mission: A world-wide 30 m resolution
DEM from SAR Interferometry in 11 days. In D. Fritch, & R. Spiller