In:crautional Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BL. Istanbul 2004 
  
3.1 Single Images 
‘The HRS stereo data as well as the THR supermode scene were 
included into the geolocation accuracy analysis. For these data 
only the orbital information was utilized but not the detailed 
attitude parameters. Instead, constant values were extracted 
from the header data and initially used for the sensor attitude 
angles. This results in an a-priori geolocation accuracy of tens 
to even hundreds of input pixels. Hence, also ground control 
points had to be used in order to optimise the sensor models of 
the Spot 5 image data. 
The sensor models of these images were optimised using a least 
squares parameter refinement procedure — an equivalent to 
photogrammetric bundle adjustment — as implemented in the 
RSG software. The RMS, minimum and maximum point 
residuals, resulting after the sensor model optimisation are 
summarized in Table 1. Sub-pixel location accuracy was 
achieved for each of the images, being represented by RMS 
values in the order of 0.7 pixels for the HRS images and of 0.9 
pixels RMS for the THR scene. 
  
116 control points Along | Across | Length 
RMS | 072 | 673 |] 102 
HRSI MIN | -158 [-143 | 0.01 
MAX | 1590 | 156 | 213 
RMS | 479 | 971 | i06 
HRS2 MIN. |.-L83 1-136 1 005 
MAX | 106 |. ].66 | 27% 
RMS | 091 | 091 | 1.29 
THR MIN | -1.78 | -1.96 | 0.09 
Max [iW 10% | 251 
Table 1: Statistical results of parameter optimisation. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
3.2 Stereo Models 
To evaluate the stereo mapping performance of Spot 5, the 
standard HRS stereo pair as well as multi-sensor stereo pairs 
were considered, which comprise the THR supermode scene as 
well as one of the HRS scenes. For these 3 stereo models the a- 
priori stereo mapping accuracy was determined. Therefore, 
ground coordinates are calculated for stereo control points 
measured in both images of the stereo model using a least 
squares point intersection algorithm, and 3D point residuals are 
determined through comparison with given control point 
coordinates. The RMS, minimum and maximum point residuals 
being achieved in East, North and Height are summarized in 
Table 2. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
116 control points East | North | Height | Length 
RMS 6.6 3.9 4.0 8.6 
HRSI-HRS2 | MIN 152 [117 -9.7 1.4 
MAX 122 77 12.0 16.2 
RMS 2.4 2.3 8.4 9.0 
THR-HRSI | MIN -4.5 -4.6 | -18.0 1.6 
MAX 5.7 5.5 16.0 18.4 
RMS 2.2 2.3 7.9 8.5 
THR-HRS2 | MIN -4.5 -48 | -172 1.0 
MAX 5.0 5.4 19.3 19.9 
  
  
Table 2: A-priori stereo mapping accuracy. 
For the HRS stereo model with a base-to-height ratio of 0.72 a 
height accuracy of 4 meters was achieved, while the planimetric 
484 
accuracy is about 8 meters and hence worse by a factor of 2. 
For the multi-sensor models with a base-to-height ratio of 0.36 
the achieved height accuracy of some 8 meters is worse by a 
factor of about 2 in comparison to the HRS model. On the other 
hand, a planimetric accuracy of less than 2.5 meters in East and 
North is achieved for both models, which is significantly 
superior (by more than a factor of 2) in comparison to the HRS 
model. Hence, planimetric accuracy is improved at the cost of 
height accuracy for the THR-HRS image pairs. The overall 
residual length shows an RMS value of about 9 meters is 
roughly the same in either case. 
3.3 Image Block 
The benefit to merge the HRS scenes with the THR scene to an 
image triple for 3D data extraction was further investigated. 
Considering an image block formed by these 3 images, point 
intersection of homologue control points was performed to 
evaluate the 3D location accuracy. The overall statistics of 
resulting point residuals are summarized in Table 3. An RMS 
height accuracy of 3.4 meters is achieved, i.e. slightly superior 
to what is achieved from the pure HRS stereo data, while the 
planimetric accuracy is 2 to 3 meters in East and North, i.e. 
significantly better than for the HRS image pair and close to the 
one achieved from THR/HRS multi-sensor image pairs. 
  
East North | Height | Length 
  
  
  
RMS 3.0 2.3 3.4 54 
MIN -7.2 -4.7 -9.0 1.1 
MAX 8.8 4.9 8.3 12.3 
  
  
  
  
  
  
  
Table 3: A-priori image block mapping accuracy. 
4. DSM GENERATION 
Three of the detailed study areas were selected to apply the 
DSM generation procedure and to investigate the performance 
of algorithms as well as the quality of achieved results. 
Selection was made upon land cover and morphology as 
follows: 
e  Rural/hilly area, being partly covered by forests 
e  Mountainous terrain 
e — Urban area, represented by the city of Barcelona 
Anaglyph presentations of the HRS stereo images of these test 
areas are shown in Figure 1. 
4.1 DSM from HRS stereo pair 
First, surface models were extracted for selected test areas 
using the HRS stereo pair. The procedure comprises matching 
of the stereo images, calculation of ground coordinates from the 
matching result, and interpolation of a regular surface elevation 
raster. 
For stereo matching, the widely used cross correlation approach 
was applied. The performance of this image matching approach 
with respect to these stereo data is summarized in Table 4, 
which shows the percentage of pixels where no matching was 
possible. The matching failures in general are caused by 
homogeneous areas where discrimination of individual pixels is 
difficult if not impossible (similarity too high), and by major 
geometric differences (similarity too low). 
  
  
Interna 
  
  
While 
the ent 
achiev 
paralla 
are pr 
height 
The qi 
matchi 
referet 
differe 
repres 
exclud 
The di 
as we 
detern 
in Fig 
differe 
comm 
Hilly 
Moun 
pa 
co 
Le 
3C 
m 
Urbar 
th