The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1114 
workstation. Operators of the Photogrammetry and Remote 
Sensing Department of Shanxi Bureau of Surveying and 
Mapping have manually collected 1489 well-distributed 
checking points in study area. We computed the differences as 
these points minus the interpolated heights from our generated 
DTM. The accuracy of the DTM is between 3.6 - 4.5m 
depending on the terrain relief and landform. Since the DTM 
generation and point measurement procedure share the same 
image orientation results, the errors of the DTM would be from 
ME derived from stereo image matching, which is directly relate 
to the matching accuracy. Detailed accuracies are shown in 
Table 1 downside. 
According to the experiment results, the following conclusions 
can be made: (a) DEM at 1:50,000 scale can be generated 
automatically from SPOT-5 HRS/HRG stereo/triplet-stereo 
images by using the automatic DTM generation approach 
proposed in this paper. The elevation accuracy can roughly be 
50% better than requirements of the Chinese national surveying 
regulation; (b) simultaneously matching SPOT-5 HRS/HRG 
triplet stereo strips and matching both feature points and 
feature-lines are quite necessary for deeply incensed 
mountainous area and rugged terrain, such as arid and semi-arid 
broken mountains in northern part and rugged mountain ranges 
in southern part of the study area; (c) the proposed automatic 
DTM generation approach can largely increase the working 
efficiency, however, careful manually stereoscopic checking and 
editing are still necessary to remove some small matching 
blunders, which are in most cases within shadow and gully/rill/ 
steep-valley well-developed areas. 
Table 1: DTM accuracy evaluation results with SPOT-5 HRS/HRG imagery 
Reference 
data 
Map 
sheets 
Num. of 
check points 
Height accuracy 
(RMSE) 
Accuracy requirements of 
1:50,000 scale l st /2 nd level DTM 
Type of terrain 
GPS points 
3 
99 
6.7 m 
14.0 m/19.0 m 
Mountainous area 
1 
26 
3.6 m 
8.0 m/11.0 m 
Mountainous area 
2 
64 
1.7 m 
5.0 m/ 7.0 m 
Hilly area 
3 
127 
2.8 m 
3.0 m/ 4.0 m 
Flat/hilly area 
Manually 
measured 
points 
47 
937 
3.6 m 
5.0 m/ 7.0 m 
Hilly area 
21 
367 
4.5 m 
8.0 m/11.0 m 
Mountainous area 
8 
185 
4.4 m 
14.0 m/19.0 m 
Mountainous area 
3.2 DSM Generation from 23 IRS-P5 Stereo Images over 
Beijing Test-field 
The test-field is an area around the city of Beijing, China. It 
consists of a steep mountainous region in the north-western part 
and flat regions in the middle and southern parts. The city of 
Beijing is located in the lower part of the study area. The whole 
area is about 70 x 210 km 2 . The site has an average terrain 
height of 300m and an elevation range of more than 1100 m. 
Totally there are 23 
scenes of IRS-P5 
stereo images 
involved in the 
Beijing test field, 
after the re-mosaic 
procedure which 
mention in the text, 
3 long IRS-P5 
stereo strips are 
used for subsequent 
processing. 
Fig. 6: In Beijing test-field, there are 23 scenes of IRS-P5 
stereo images. In order to facilitate the subsequent 
processing, we use the re-mosaic procedure based on 
metadata to combine these 23 stereo images into 3 long 
stereo strips. 
Over the test area, totally 23 scenes of IRS-P5 stereo images 
were acquired. The ground resolution of IRS-P5 images is 2.5m, 
and each scene covers about 30 x 30km 2 . In order to precisely 
georeference these images, with the cooperation of the l rd 
Institute of Surveying and Mapping of Heilongjiang SBSM, 
about 66 well-distributed GCPs were collected with differential 
GPS in 2006. The measurement accuracy was better than 0.5m 
in planimetry and lm in height. The GCPs are well-distributed in 
the test area, their ground intervals are about 30km, and half of 
them are located at the center of road intersection which can be 
precisely measurement both in image and object space. 
Normally, we could commercially get so-called IRS-P5 
“standard scene” of images, which have 12000 x 12000 pixels 
and cover 30 x 30km 2 . Basically the standard scenes could be 
sub-images of a long IRS-P5 strip. They are just subdivided 
from a long IRS-P5 strip for commercial reason and normally 
they have 5%-15% overlap between adjacent scenes. This 
situation will cause difficulties or complicate the sub-sequent 
processing. To avoid these problems, we develop a procedure to 
re-mosaic adjacent IRS-P5 standard scenes into a long stereo 
strip (Fig. 6) in condition that these scenes are sub-images of the 
original long IRS-P5 strip. After this re-mosaic procedure, 23 
IRS-P5 scenes of the Beijing test-field are re-mosaic into only 3 
IRS-P5 long stereo strips. 
With aids of 66 GCPs, block-adjustment was made and the 
orientation accuracy is about 1.7m in planimetry and 2m in 
height. After the block-adjustment, the proposed DSM/DTM 
generation approach was applied to 3 re-mosaic IRS-P5 stereo 
image strips simultaneously. As a result, 12.5m grid-spacing 
DSM of the study area has been generated automatically. Since 
our DSM/DTM generation approach not only generates a large 
number of mass points but also produces line features, which are 
necessary for the modeling the rugged mountainous terrain. 
Parts of DSM results are shown in Fig. 7, the resulted 12.5 m 
DSM reproduced quite well not only the general features of the 
terrain relief but also small geomorphological and other features 
visible in the IRS-P5 images. The DSM shows many 
topographic details and features such as small valleys in the 
mountains, detailed patterns related to streets and buildings in 
suburban and urban areas, linear features related to highways 
and main road networks, sparse trees, small clusters of houses 
and forest areas. Checking by GCPs measured by differential 
GPS shows that the accuracy of the DSM is 3.7 m and it meet 
the requirements from the Chinese Surveying and Mapping 
regulations for 1:50000 topographic maps.