ie. final 
" value 
h each 
1e GPS 
ixel-to- 
age. In 
1e NED 
Thus, 
EM v2 
SRTM 
1ere the 
SRTM 
Is Were 
EGM96 
on was 
natively 
cs were 
NLCD 
s of the 
re is no 
cars that 
s as the 
rth sides 
  
  
ation. 
rrors are 
stribution 
igure 3). 
y metric 
MSE for 
RMSE of 
uracy can 
ses 95%, 
" (LE95). 
Maune et 
| meters, 
v1 (Table 
  
1). The accuracy statistics for GDEM v1 were derived from a 
comparison with a previous smaller set of GPS benchmarks 
(13,305 points) from NGS, which was the most recent dataset 
available at the time of the GDEM vl evaluation in 2009. 
However, most of these points are also included in the current 
GEOID09 GPS benchmark dataset used for GDEM v2 
validation. 
  
GDEM v2 errors vs. GPS benchmarks 
  
A» s 20 3 a AC Cc 16 n E: 46 « A) 
Fires meters 
  
  
  
Figure 3. GDEM v2 absolute vertical accuracy. 
  
  
  
  
  
  
  
  
DEM | Minimum | Maximum | Mean sd RMSE | LE95 
GDEM 23737 64.80 -0.20 | 868 | 868 | 1701 
NED 46.21 16.42 2033 | i31 | 134 | 361 
SRTM | 28.67 28.58 073 | 395 |. 401 |. 786 
Gory „127.74 105.41 -3.69 | 858 | 934 | 1831 
  
  
  
  
  
  
  
  
Table 1. Error statistics (meters) from an accuracy assessment 
vs. NGS GPS benchmarks. 
Another important descriptor of vertical accuracy is the mean 
error, or bias, which indicates if a DEM has an overall vertical 
offset (either positive or negative) from the true ground level. 
The GDEM v2 mean error of —0.20 meters is a significant 
improvement over the GDEM v1 mean error of —3.69 meters 
(Table 1). 
The absolute vertical accuracy testing also included evaluation 
of the NED and SRTM datasets over CONUS. Because NED 
and SRTM are both supplied at the same 1-arc-second posting 
as GDEM v1, and they have been extensively tested with many 
results reported in the scientific literature, summary statistics 
are provided (Table 1) to help give context for the GDEM v2 
results. The number of GPS benchmarks used for evaluation of 
SRTM was reduced to 16,865 points due to the deletion of 
points that fell in SRTM void or fill areas. 
3.1.1 Land Cover Analysis 
The absolute vertical accuracy assessment results, both mean 
error and RMSE, have been segmented by land cover to 
examine effects of cover types on measured errors. While the 
RMSE varies little across cover types, the mean error (bias) 
does appear to be affected by land cover, ranging from +5.00 to 
-2.27 meters across the 14 NLCD classes. Recall that ASTER 
images record the reflective surface, thus the derived elevations 
in GDEM v2 represent the height of those imaged surfaces. In 
areas with dense, taller vegetation or built structures, the 
derived ASTER elevation will represent the elevation of these 
features rather than ground level. The GDEM v2 mean errors 
by land cover class verify that the presence of aboveground 
features causes a positive elevation bias, as would be expected 
for an imaging system like ASTER. 
Figure 4 shows the results of aggregating into broad, 
generalized land cover classes. The GPS ground truth points 
were grouped into three broad land cover categories and the 
GDEM v2 mean error and RMSE were recalculated. The 14 
NLCD classes were grouped into forest (deciduous, evergreen, 
mixed, woody wetlands), developed (open space, low intensity, 
medium intensity, high intensity), and open (barren land, 
shrub/scrub,  grassland/herbaceous,  pasture/hay, cultivated 
crops, emergent herbaceous wetlands). The chart in Figure 4 
indicates the percentage of points that fell into each aggregated 
class. As with the individual classes, the RMSE varies little 
among the aggregated classes, but the mean error does appear to 
reflect the effects of land cover on the measurement of 
elevations by ASTER. As expected, the generalized forest class 
exhibits a noteworthy positive bias of about 3 meters. 
However, the aggregated open ground class should have a mean 
error at or very close to zero, which is not the case. It appears 
that GDEM v2 may have a “true” overall negative bias on the 
order of 1 meter. 
  
Absolute vertical accuracy: GDEM v2 vs. GPS benchmarks 
  
  
  
  
  
§ im e o. 3 [BGDEM v2 mean eror 
z i |BGDEM v2 RMSE 
  
  
  
-0.13| 
  
Forest (4^6) Developed (75%) Open (21%) All (100%) 
Land cover 
  
  
  
Figure 4. GDEM v2 mean error and RMSE by aggregated land 
cover class. 
Comparison of the mean errors by aggregated land cover class 
for GDEM v2, NED, and SRTM (Figure 5) reveals that in 
forested areas GDEM v2 is registering higher elevations than 
SRTM. Like ASTER, SRTM is a “first return” system, and 
elevations above ground level would be expected for areas with 
trees and/or built structures. It is likely that ASTER is 
measuring elevations at or near the top of the forest canopy 
while SRTM is recording elevations part way down into the 
canopy. Such performance of SRTM in recording elevations 
within the vegetation canopy rather than at the top has been 
previously documented (Carabajal and Harding, 2006; Hofton, 
et al., 2006). 
3.1.2 Scene Number Analysis 
The reference points were grouped into bins for each NUM 
value, and the measured GDEM v2 errors for the points in each 
bin were processed to calculate a mean error and average RMSE 
for each NUM bin. Figure 6 shows a plot of the mean error and 
RMSE associated with each NUM value. Note how both the 
mean error and RMSE improve rapidly as the NUM increases 
267