The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1141 
Figure 3. DSM (left) of a test region and corresponding slope 
edges (right) 
DTM in ERDAS LTF format are generated and compared with 
manually digitized baseline data, as in Table 1. 
Data Type 
Point # 
DTM 
Check # 
RMSE 
Urban (Hamburg) 
155014 
170 
(meter) 
1.51 
Rural (Quasco) 
551492 
65 
1.48 
Mountain (Mexico) 
195097 
100 
1.82 
Table 1. ATE accuracy for frame sensors 
Figure 4. Status at elevation levels 277.4m, 272.4m, 265.4m, 
252.2m, ordered from top to bottom. Legend: white 
(terrain), black (under water), yellow (slope edge), 
k red/green (new isolated regions). Right: white 
(extracted terrain), red (filtered objects) 
This object filter can also be used to filter lidar point cloud. 
Preliminary test shows a good performance. 
Object filtering can effectively remove most spikes as well as 
points on buildings and trees. It can cut off approximately 60% 
of manual editing time. It can also improve bare-earth quality: 
for an urban dataset (Hamburg), RMSE drops from 4.84m 
before object filtering to 1.51m after object filtering. 
Fig. 5 shows one example: the majority of buildings on a flat 
terrain are removed except for an over-size building and a high- 
rising road that are beyond threshold. 
Figure 5. DSM before (above) and after (below) object filter 
Fig. 6 shows another example where buildings on a slope are 
removed, yet small terrain variation is still preserved. 
3. EXPERIMENTAL RESULTS 
3.1 Frame Sensors 
We tested adaptive ATE on frame images with three typical 
scenes: urban, rural, and mountainous areas. 2m-resolution 
Fig. 7 shows a DSM from rural area. Points are located on both 
ground and trees and contours reveal the coverage of trees. 
After object filtering points on trees are removed and contours 
become much smoother.