The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1112 
success rate and less mismatches. These parameters include the 
size of the correlation window, the search distance and the 
correlation threshold values. This is done by analyzing the 
matching results at the previous image pyramid level and using 
them at the current level. 
(4) High matching redundancy: With our matching approach, 
highly redundant matching results, including points and edges 
can be generated. Highly redundant matching results are suitable 
for representing very steep and rough terrain and allow the 
terrain microstructures and surface discontinuities to be well 
preserved. Moreover, this high redundancy also allows for 
automatic blunder detection. 
(5) Efficient surface modeling: The object surface is modeled 
by a triangular irregular network (TIN) generated by a 
constrained Delauney triangulation of the matched points and 
edges. A TIN is suitable for surface modeling because it 
integrates all the original matching results, including points and 
line features, without any interpolation. It is adapted to describe 
complex terrain types that contain many surface microstructures 
and discontinuities. 
(6) Coarse-to-fine hierarchical strategy: The algorithm works 
in a coarse-to-fine multi-resolution image pyramid structure, and 
obtains intermediate DSMs at multiple resolutions. Matches on 
low-resolution images serve as approximations to restrict the 
search space and to adaptively compute the matching 
parameters. 
Svtp'Efattvdlma«* Hie Rtfcrtaee Stnp-1 K*kwdiaa*t 
t'orvfwrd teNadir Ss»s*$t Baefcwwd 
(a) 
Fig. 3: GC 3 matching with 6 high-resolution airborne linear 
array images (ca. 5cm footprint) from 2 strips with changing 
flight directions for solving multiple solution problems. The 
individual NCC functions and the SNCC function within the 
search range determined by height increment of ±10.0 meters 
is shown in (b) 
3. Performance Evaluation 
The height accuracy (or to be more precise the vertical accuracy) 
of DSMs/DTMs usually results from the quantitative and 
statistical evaluation of the DSMs/DTMs and it is determined by 
its root-mean-square error (RMSE), the square root of the 
average of the set of squared differences between height values 
of the DSM/DTM being evaluated and height values from an 
independent source with much higher accuracy. According to 
(McGlone, 2004), there are at least 3 major sources of error 
when DSMs/DTMs are generated by using the optical imaging 
systems, i.e. the Photogrammetric Modeling Error (PME), the 
Measurement Error (ME) and the Surface Modeling Error 
(SME). 
These errors can be estimated empirically or estimated using 
error propagation. For instance, we could manually measure a 
sample of randomly spaced points using the stereo model with 
the same image orientation parameters and then compare them 
with their interpolated heights from the generated DTM. In this 
case, the height errors mainly come from ME, but also from 
SME in cases of very rough terrain. However, if we measure 
these points with a different method such as the traditional field 
surveying, the estimated errors may include all the errors 
mentioned above. 
In order to evaluate the performance of our approach for 
DSM/DTM generation it has been verified extensively with 
several HRSI datasets, such as IRS-P5 and SPOT-5 HRS/HRG 
images, over different terrain types, which include hilly and 
rugged mountainous areas, rural, suburban and urban areas. In 
the following, we will report in detail about 2 experiments. The 
first involves the evaluation of SPOT-5 HRS/HRG triplet images 
over a testfield in Zone of headstream of Three rivers, eastern 
Tibet Plateau, China with accurate GCPs, more than 2500 m 
height range and variable land cover. The accuracy study was 
based on the comparison between as many as 160 accurate GPS 
check points, more than 1400 manually measured check points 
and the automatically extracted DTMs. In the second test, the 
proposed approach has been also applied to 23 IRS-P5 stereo 
pairs over Beijing city. Other processing and evaluation results 
of IKONOS and SPOT5 HRS/HRG can be found in Zhang and 
Gruen, 2004; Poli et al., 2004; Baltsavias et ah, 2006 and Poon 
et ah, 2005. 
3.1 Automatic DTM generation from SPOT-5 HRS/HRG 
Images over Test-field in Zone of headstream of Three rivers, 
Eastern Tibet Plateau, China 
The test area in Zone of headstream of Three rivers, Eastern 
Tibet Plateau, China covers 250 topographic maps at 1:50,000 
scale with the area of about 12,000km 2 , where contains 
large-area of seasonally and perennially frozen soil, 
mountain/valley glacier and perennial snowfield and large area 
of unman area. The test-field is the headstream of Yangtze River, 
Yellow River and Lancangjiang River, and the QingZang 
railway and national road cross the region from north-east to 
south-west. The study area consists of steep arid/semi-arid 
mountainous region in the northern part (transition zone between 
Kunlun Mountain and Tsaidam Basin), smooth hilly regions in 
the middle parts (plateau mountains and intermountain basins 
are well-developed) and high-plateau mountain ranges in the 
southern part (mountain/valley glacier, glacier canyon and 
knife-edge crest are well developed). The average elevation is 
4000m in test-field. Main geological structures are in trend of 
nearly east-west direction (Fig. 4). The various landforms in 
study area provides better environment for DTM automatic 
generation. 
Over the test area, totally 11 pairs of 10 x 5m SPOT-5 HRS and 
nearly twenty 5m HRG images were acquired. These images 
were used to generate DTM over the whole test area (Fig. 4). In 
particular, for DTM accuracy analysis, 2 SPOT-5 HRS satellite 
image pairs imaged in Nov 2003 and 6 HRG images, which can 
form the SPOT-5 stereo triplets, have been selected. The images 
have the fine quality and have no cloud coverage, which provide