2.2 Reference Data 
The primary reference data were the “GPS on Bench Marks” 
dataset of geodetic control points from the National Geodetic 
Survey (NGS). These points represent NGS's best x-y-z control 
point dataset for CONUS, and they are used by NGS for gravity 
and geoid determination (Roman et al, 2004; Roman et al., 
2010). This set of control points is from NGS's latest US. 
geoid model, GEOID09. The points have millimeter to 
centimeter-level accuracies, so they are an excellent reference 
against which to compare DEMs across CONUS. For the 
accuracy assessment presented here, 18,207 points (Figure 1) 
were intersected with GDEM v2. The elevations of the GPS 
benchmarks are provided in the North American Vertical Datum 
of 1988 (NAVDSS8), and the elevations of GDEM v2 are 
referenced to the Earth Gravitational Model 1996 (EGM96) 
geoid. Therefore, before comparing the GDEM and the GPS 
points, the vertical referencing of the points was transformed to 
the EGM96 geoid. Over CONUS, the vertical offset between 
NAVD88 and the geoid averages about one-half meter 
(National Geodetic Survey, 2010). 
The 2006 update of the National Land Cover Database (NLCD) 
(Fry et al., 2011) was used to segment the accuracy assessment 
results by land cover class. NLCD includes land cover data in 
19 classes derived from 30-meter Landsat data. The GPS 
benchmarks used for validation of GDEM v2 fall into 14 of the 
NLCD land cover classes. 
  
  
  
  
  
Figure 1. GPS benchmarks (18,207 points) used as GDEM v2 
validation reference data. 
2.3 Absolute Vertical Accuracy 
The difference between the GPS benchmark elevation and the 
corresponding GDEM v2 elevation was recorded for each 
control point location. The recorded GDEM v2 elevation was 
derived through  bilinear interpolation at the precise 
latitude/longitude location of the GPS point. At each point, the 
difference was calculated by subtracting the GPS benchmark 
elevation from the GDEM v2 elevation, and these differences 
are the measured errors in GDEM v2. Positive errors represent 
locations where the GDEM v2 elevation was above the control 
point elevation, and negative errors occur at locations where the 
GDEM v2 elevation was below the control point elevation. The 
land cover associated with each control point was determined 
by intersecting the GPS benchmark locations with the NLCD. 
An ancillary data layer supplied with GDEM v2 indicates the 
number of individual ASTER scene (stereo pair) DEMs that 
were used to derive each elevation value. The individual 
ASTER DEMs are stacked and averaged to calculate the final 
elevation value for each pixel in GDEM v2. The "NUM" value 
(number of input individual DEMs) associated with each 
control point location was determined by intersecting the GPS 
benchmarks with the ancillary NUM layer. 
2.4 Comparison vs. Other DEMs 
GDEM v2 was differenced with NED and SRTM on a pixel-to- 
pixel basis across the full extent of their CONUS coverage. In 
the same manner as with the reference control points, the NED 
and SRTM were cach subtracted from GDEM v2. Thus, 
positive differences represent locations where the GDEM v2 
clevation was higher than the corresponding NED or SRTM 
elevation, and negative differences occur at locations where the 
GDEM v2 elevation was lower than the NED or SRTM 
elevation. Prior to differencing, the NED elevations were 
converted from the NAVDSS vertical datum to the EGM96 
geoid vertical reference frame. No such conversion was 
necessary for SRTM, as both GDEM v2 and SRTM are natively 
referenced to the EGM96 geoid. Difference statistics were 
calculated, and summary statistics were segmented by NLCD 
land cover class. 
3. RESULTS AND DISCUSSION 
3.1 Absolute Vertical Accuracy 
A plot of the GDEM v2 measured errors vs. elevations of the 
reference control points (Figure 2) indicates that there is no 
clear relationship of error with elevation. Also, it appears that 
there is no preference for positive or negative errors as the 
plotted GDEM v2 errors are uniformly distributed on both sides 
of the zero error axis. 
  
GDEM v2 error vs. elevation 
  
  
  
  
Enter meters) 
  
  
  
  
  
0 sco 1,000 1,500 2000 2,500 3000 3,500 
GPS benchmark elevation (meters) 
  
  
  
Figure 2. GDEM v2 measured errors plotted vs. elevation. 
Summary statistics of the measured GDEM v2 errors are 
presented in Figure 3 and Table 1. The error distribution 
approximates a normal distribution (smooth line in Figure 3). 
The Root Mean Square Error (RMSE) is an accuracy metric 
commonly used for elevation data, and the measured RMSE for 
GDEM v2 is 8.68 meters. This compares with the RMSE of 
9.34 m for GDEM vl (Table 1). Absolute vertical accuracy can 
also be expressed with a confidence level, in many cases 95%, 
or also referred to as “linear error at 95% confidence” (LE95). 
LE95 is derived directly from the measured RMSE (Maune et 
al, 2007). GDEM v2 exhibits an LE95 of 17.01 meters, 
compared with an LE95 of 18.31 meters for GDEM v1 (Table 
266