The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1353 
2005). It is crucial to perform a full 3D comparison, because a 
2.5D analysis (generally done by most of the GIS commercial 
software) could overestimate the differences between two 
surfaces representing the same object. In fact, comparing 
heights instead of surface normals causes errors especially in 
cases where even small horizontal errors (georeferencing 
discrepancies or errors in point measurements) may lead to big 
height deviations, introducing methodological errors into the 
testing procedure. 
The DSMs generated with SAT-PP and PCI were co-registered 
to the reference DSM and between themselves. The shift values 
are reported in Table 3, 2 nd to 4 th columns. With respect to the 
reference data, the SAT-PP DSM shows a shift in East direction 
of about 0.5 pixels and a larger shift (about 1.4 ground pixels) 
in North direction, while PCI DSM has a shift of 1.3-1.5 ground 
pixels in planimetry. For both surfaces, the shift in Z direction 
is not significant. This results in an overall shift between SAT- 
PP and PCI DSM of 3 ground pixels in North direction, which 
is the satellite flight direction. This issue will be investigated in 
the future. 
After transforming the surfaces to a common system using the 
estimated shift parameters, the 3D residuals have been 
computed. From Figure 7 we see that the residuals are 
normally distributed with null bias. In Table 3 the mean values 
and RMSE of the residuals in Euclidean direction and in the 3 
components East, North and height, are reported (5 th to 9 th 
columns). The residuals in planimetry between the generated 
DSMs and the reference DSM are almost the same, while in 
height large differences exist. Figure 8 shows that the largest 
errors are located in urban areas. We suppose that some 
differences in height are due to surface changes between the 
acquisition time of the Cartosat-1 scenes and the aerial images 
used for the reference DSM generation. If we analyse the 
residual distribution between SAT-PP and PCI DSM (Figure 
9), we see that after co-registration the two surface models fit 
quite well (about half a pixel RSME). Again, the largest values 
are present in height. 
Figure 7. Histograms of residuals between reference DSM and Sat-PP DSM (left) and PCI DSM (right) after coregistration. 
DEM 
Shift (m) 
Residuals (m) 
E 
N 
H 
Mean 3D 
RMSE 3D 
RMSE E 
RMSE N 
RMSE H 
Ref vs. SAT-PP 
1.52 
-3.37 
-0.20 
0.02 
2.07 
0.39 
0.37 
2.00 
Ref vs. PCI 
3.16 
3.55 
0.59 
0.02 
2.02 
0.41 
0.44 
1.94 
SAT-PP vs. PCI 
2.10 
7.67 
0.86 
0.00 
1.39 
0.30 
0.30 
1.32 
Table 3. 3D comparison between PCI, SAT-PP and reference DSM for area 1: shifts between DSMs, residuals after co-registration. 
The first surface has been used as search, the second as template. 
S V Vv «1’ * •; 
- . v ’ * 
■v ; • T-V • , \ 
«■V\ .yf 
♦ v W* ■ 
‘S ~4 ' N> 
A" - - *' • V XT« P . *v - * . 
rJr *• ;>.j: .*• ■ *. . 
''? j, . ..*■£' 
* ■ .-'j -*.*■. .* V > 
•" Jyl 
~E*. , it \ ZST* 
-_i * S', : __ _ '• j 
Ref-PCI 
1° 
Figure 8. Residual distribution (in meters) after co-registration between SAT-PP and PCI DSMs (template surfaces) with respect to 
the reference DSM (search surface) in Euclidean direction. The black area is excluded from the analysis due to a gap in the reference 
surface.