The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
285 
edge information, however the elevation difference of ground is 
continuous without clear edge, meaning weak edge information, 
therefore, this paper selected edge information to refine the 
initial segmentation surface. First of all, the standard edge 
extracting method (e.g. Sobel method) was adopted to extract 
the edge of DSM, afterwards overlaid the edge image and the 
initial segmentation surface of buildings, and counted the total 
edge points of each overlaid segment at last. If the total edge 
points are less than the preset threshold n (e.g. n=10, 
determined by test), this segment is asserted as ground and 
removed from the initial building segmentation surface. 
2.2 Segmentation 
The purpose of segmentation is removing the vegetation and 
preserving building information. The existing methods usually 
use area threshold to remove small and separated vegetation, 
then based on the gradient or extent of surface undulation to 
remove those larger or building-closed vegetation 151 . Because 
vegetation almost grows together or near building, there is little 
small and separated vegetation. You may face two problems 
when adopt the existing segmentation sequence. One is the area 
threshold method can only remove small quantity of vegetation, 
the other is using gradient and the extent of surface undulation 
can not remove vegetation completely. For this, this paper 
proposed a different method of changing segmentation 
sequence as, firstly using gradient to remove vegetation or to 
separate vegetation from building, then using area threshold to 
remove the rest vegetation, finally processing segmentation 
result by neighbor iterative approaching to resume the removed 
edge information. 
1) Gradient segmentation 
From geometrical points, the roof shape of building and 
vegetation is completely different, that of building is usually 
plane or bevel but of vegetation is normally flexual. If not 
considering noise, the second derivative value of each point on 
building roof except for edge should be zero while that of point 
on vegetation surface is not zero. Moreover, either of 
vegetation or building edge, the second derivative value is very 
big. Therefore, we can use second derivative value to remove 
vegetation or separate vegetation from building. In processing 
of digital images, gradient is usually applied to replace 
derivative with the quadratic gradient being defined as follows: 
a 2 / 
+ 
d 2 f 
v 2 / 
+ 
v 2 / 
obc 2 
dy l 
dx 2 
dy 1 
Taking noise into account, we almost choose a smaller but 
non-zero value of quadratic gradient threshold, such as 0.03. 
2) Area Segmentation 
Because the quadratic gradient of most vegetation surface is big, 
after gradient segmenting, most vegetation information was 
removed leaving only a little but building information was 
better preserved. In this case, we can use area threshold to 
remove the residual vegetation information. 
3) Refining buildings 
As the building roof is not smooth enough, and the quadratic 
gradient is bigger at the edge, so, gradient segmenting will 
result in some absence on building roof or mis-removing edge 
information, which requires further refining. The absence can 
be refilled by local filling method of morphology and neighbor 
iterative method can be used to resume the edge information. 
The steps are, 
(1) Adopting the existing graph traversal method to 
judge neighbor of all points on the building 
segmentation, 
(2) Neighbor judgment: Comparing the absolute 
elevation difference Ah between the point and its 
eight neighboring points not in the segment. If 
Ah is less than the elevation difference threshold 
(such as Ah 3=0.6m), it shall be added into the 
segment. The elevation difference threshold is equal 
to grid cellsize, and its selection was based on the 
assumption of the max. building roof slope is 45°. 
(3) Count the increased points after graph traversal, if it 
is less than the assumed threshold (such as 10), then 
terminate iterating otherwise return to (1). 
3. TEST AND RESULTS 
3.1 Test data 
The LIDAR data used in this paper is an aerial image covering 
about 1200><420m 2 obtained by Optech 1210 LIDAR system in 
October 2000, which represents a large area of 3065x538m 2 in 
Wytheville, Virginia, U.S.A. The data resolution is 0.6m. All 
algorithms following were executed by Matlab. 
3.2 Results 
Fig.2a shows the DSM generated by nearest distance 
interpolation with grid space of 0.6m. Fig.2b is the sectional 
view of DSM. The section lies on the red line in Fig.2a from 
which we can see fewer buildings but more vegetation at the 
left, and several big buildings and complex ground surface but 
less vegetation at the right. From the sectional view we can see 
that this region is undulate. 
Fig.3a shows the DTM extracted by contour-line-based filtering 
method. Fig.3b is the sectional view of DTM with the same 
sectional position as that in Fig.2a. Comparing with the DSM, 
it is clear that most non-ground points as vegetation and 
building have been removed and have similar undulation, 
which indicates the DTM extracting is accurate. This filtering 
method can be applied on the undulate region.