In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
The shape of DDASF is determined by the two parameters, 
slope( S ) and searching scope( /). The slope can be estimated 
by adjacent points and the searching scope can be determined 
by the biggest size of objects in the landscape. If / is not big 
enough, a tall and big object may remain some non-ground 
points near the break line which is shown as Figure 5. 
(a) Diagram of point clouds 
(c) some non-ground points are remained after the union of 
the two filtering results 
Figure 5. Filtering errors resulted from insufficient searching 
scope 
2.3 Procedure of DEM generation by DDASF 
The flow chart of DDASF is shown in Figure 6. Firstly the 
outliers of LiDAR point cloud are removed. Normally they are 
extremely higher or lower than adjacent points and isolated 
from other points. The outliers can be removed from raw 
LiDAR data using the above conditions, the remaining points 
are then structured for constructing a spatial relationship 
between points. Triangle irregular network and regular voxels 
are both common methods to structure scattered points. In this 
paper, the regular voxel is used. The complete scattered spatial 
is divided into voxels of the same size. DX,DZ denote the 
length and height of a voxel respectively. Since the ground 
points are normally located on lower voxels, the lowest voxels 
which contain at least one point are searched. The initial ground 
surface can be therefore obtained by those points contained in 
the lowest voxels. Once the initial ground surface is determined, 
the local plane of a voxel is calculated with its adjacent voxels. 
The rough filter refers to remove the points which exceed a 
threshold from the local plane to its position. The remaining 
points are again filtered by DDASF. And then the procedure 
iterates until a condition is satisfied. The final output is 
therefore used to generate DEM. 
Figure 6. Flow chart of DEM generation by DDASF 
3. EXPERIMENTS AND ANALYSES 
Three DEM generation methods were compared. One is the 
adaptive slope-based filter(ASF) developed by Tseng, et 
al.(2004). The second one is using the commercial software, 
TerraScan. The last is our developed method. The LiDAR data 
used in this experiment is provided by ISPRS Commission III 
(http://www.itc. nl/isprswgIII-3/filtertest/index.html). Fifteen 
samples chosen as reference data are generated by manual 
filtering. To test the effectiveness of filtering in terracing fields 
and cliff areas by DDASF, Sample 23 and Sample 53 are 
chosen because disconnected terrain exists in both data. 
(a) sample 23 (a) sample 53 
Figure 7. Test data 
3.1 Filtering results of Sample 23 
Sample 23 is located at urban areas. The buildings are complex, 
large and some of them are even connected. The lower left of 
630