-ffW 
ÍS5WS 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part Bl. Beijing 2008 
id-' 
MS 
Figure 3. Catalonia image. Left original, right after preprocessing. 
3. IMAGE QUALITY AND PREPROCESSING 
In Baltsavias et al. (2007), we have reported about image 
quality. The images in this test, which are later than the images 
commented in the previous investigations, are generally better. 
Pattern noise is less and the edge jitter of horizontal edges is 
almost invisible. The interlacing noise still remains, and the 
Fore channel images compared to the Aft ones are less sharp, 
partly to the different ground pixel resolution. 
The preprocessing performed was applied to the original 10-bit 
images, along the same line as mentioned in Baltsavias et al. 
(2007). First, an adaptive noise filtering was applied to reduce 
noise without smoothing edges, then Wallis filtering to enhance 
contrast, especially in dark regions, and to radiometrically 
equalize the images used for matching. Finally, the images were 
reduced to 8-bit for further processing, since many software 
packages for matching including our SAT-PP (see Section 5) 
can not handle images with more than 8-bit. An example of 
preprocessing is shown in Figure 3. 
Method 
GCP no 
RMSE 
X(m) 
RMSE Y 
(m) 
RMS 
EZ 
(m) 
Sigma 0 
(pixel) 
RPC-1 
All 
(70) 
2.12 
1.84 
4.23 
1.07 
RPC-1 
6* 
2.15 
2.01 
4.25 
0.78 
RPC-1 
6 ** 
2.12 
2.04 
4.53 
0.32 
RPC-2 
All 
(70) 
0.94 
1.31 
1.26 
0.50 
RPC-2 
6* 
0.98 
1.43 
1.49 
0.41 
RPC-2 
6 ** 
1.23 
2.02 
1.60 
0.21 
* GCPs well-distributed over the images. 
** GCPs cover approximately l/4' h of the image area. 
Table 3. Triangulation results of the Catalonia dataset. 
SENSOR ORIENTATION AND ACCURACY OF 3D 
POINT MEASUREMENT 
Sensor orientation is performed using RPCs provided by ISRO. 
We have applied correction to the RPCs by two shifts (RPC1) 
and by an affine transformation (RPC2). Six and all GCP 
versions are tested in both models. For the 6 GCP case, we have 
applied two different distributions by covering the whole and 
Method 
GCP 
no 
RMSE 
X(m) 
RMSE 
Y(m) 
RMSE 
Z(m) 
Sigma 0 
(pixel) 
RPC-1 
All 
(61) 
1.90 
3.42 
5.02 
1.46 
RPC-1 
6 * 
1.97 
3.47 
5.27 
0.95 
RPC-1 
6 ** 
2.00 
3.95 
5.77 
0.65 
RPC-2 
All 
(61) 
1.43 
1.31 
1.42 
0.62 
RPC-2 
6 * 
1.64 
1.55 
1.77 
0.40 
RPC-2 
6 ** 
2.73 
2.50 
2.10 
0.41 
1 /4 th of the image area. The results are reported in Table 3 and 
! GCPs well-distributed over the images. 
'* GCPs cover approximately l/4' h of the image area. 
Table 4 for both datasets. 
Table 4. Triangulation results of the Sakurajima dataset. 
The estimated line and sample shift values in Catalonia dataset 
are approximately (49, -34) and (51, -14) pixels for the aft and 
fore images, respectively. For the Sakurajima dataset, shift 
corrections are approximately (-87, -41) and (-104, -8) pixels in 
the line and sample directions of the aft and fore images, 
respectively. The values are obtained from the RPC2 method 
using all GCPs. These values give an indication of absolute 
geolocation accuracy of Cartosat-1. 
In both datasets, the RMSE values for RPC2 are at sub-pixel 
level. The accuracy, for both RPC1 and especially RPC2, which 
is necessary for higher accuracy, deteriorates with poor GCP 
distribution. The RPC1 method (only shift correction) gives 
much inferior results compared to RPC2, especially in height. 
For the DSM generation, the RPC2 and all GCPs versions have 
been used. 
1365 
. ■ '