from 1 to 10 scenes. Beyond NUM values of about 10 to 15 
scenes, there appears to be little improvement in either mean 
error or RMSE values. 
that they are positive supports the previously described 
observation that GDEM v2 has proportionally higher elevations 
than SRTM in many forested areas. 
  
Absolute vertical accuracy: GDEM v2 vs. GPS benchmarks 
jncDEM 2 
juNED 
{0SRTM 
    
Mean eror (meteis] 
  
  
  
  
Develcped (75%) Open (21%) All (1C0%) 
Land cover 
  
  
GDEM v2 differences vs. other DEMs 
  
    
  
© GDEMv2- NEC 
| wODEMv2- SRTM 
  
RMSE (meters) 
  
Land cover 
  
  
  
Figure 5. GDEM v2, NED, and SRTM mean errors by 
aggregated land cover class. 
Figure 7. GDEM v2-NED and GDEM v2-SRTM RMSE by 
land cover class. 
  
GDEM v2 error as a function of number of scenes 
18.00 
16.00 
x 
1400 » Mean erar (bias) 
in NUM bin 
12.00 
10.00 
Coum 
Meters. 
  
0 5 10 15 0 25 30 35 4C 45 50 
Number of ASTER scenes 
  
  
GDEM v2 differences vs. other DEMs 
100 
  
  
  
  
Mean difference (meters) 
  
  
  
x a“ "d E SN sj 
X & " 
À $8 y o - e P = r $ SS SF = e aj 
e « e & SF x p S G S 
P d e # 9 © SF 
A s Fa 9 9 g? 
Cod + 
  
  
  
  
Figure 6. GDEM v2 mean error and RMSE vs. number of 
scenes used for elevation calculation. 
3.2 Comparison vs. Other DEMs 
The RMSE by land cover class (Figure 7) shows that in forested 
classes, GDEM v2 and SRTM generally agree better (as 
indicated by a smaller RMSE value) than GDEM v2 and NED. 
This is expected, as both ASTER and SRTM are first return 
systems that measure aboveground elevations in tall vegetation 
canopies. As land cover becomes more open (for instance, the 
six classes on the right side of the chart in Figure 7), the GDEM 
v2-NED RMSE and GDEM v2-SRTM RMSE are nearly 
equivalent as all three DEMs are measuring near ground level 
elevations. 
The chart of mean differences (Figure 8) supports previous 
observations from the absolute vertical accuracy assessment. In 
the forest classes (four classes on the left side of the chart in 
Figure 8), the GDEM v2-NED mean differences are larger than 
the GDEM v2-SRTM mean differences. Again, this is the 
expected condition as NED by definition is a "bare earth" 
elevation model (Gesch, 2007), and ASTER is a first return 
system that measures canopy elevations in forested areas. Even 
though the GDEM v2-SRTM mean differences for three forest 
classes (mixed, deciduous, woody wetlands) are smaller than 
the corresponding GDEM v2-NED mean differences, the fact 
Figure 8. GDEM v2-NED and GDEM v2-SRTM mean 
differences by land cover class. 
The negative mean differences for both GDEM v2-NED and 
GDEM v2-SRTM for the five open ground classes (shrub/scrub, 
pasture/hay, barren land, cultivated crops, 
grassland/herbaceous) on the right side of the chart in Figure 8 
provide further evidence that GDEM v2 has an overall true 
negative elevation bias. Both NED and SRTM exhibit a mean 
error very close to zero for open ground land cover classes, so if 
GDEM v2 was performing in the same way over those open 
ground conditions the mean differences would be at or much 
closer to zero. 
4. CONCLUSION 
The validation testing results reported here have raised several 
important observations about the quality of elevation 
measurements contained in GDEM v2: 
e There is an improvement in overall RMSE of nearly two- 
thirds of a meter (8.68 m vs. 9.34 m) when comparing the 
measured accuracies of GDEM v2 and GDEM vl. 
Likewise, there has also been an improvement in overall