Michel Morgan 
Prior knowledge about buildings is used in determining the window sizes. The minimum and the maximum size of the 
windows used are selected to be slightly larger than the expected minimum and maximum building size respectively, 
The maximum and the minimum building size can be chosen according to the specifications of the application. Th; 
band width is the expected range of terrain elevation in the window. Some factors play a role in band width selection 
such as the terrain variability, the window size and the minimum building height. As the terrain roughness (or the 
standard deviation of the terrain elevation) increases, the band width increases. Moreover, as the window size increases 
the band width increases as far as the terrain has a certain slope. The band width has to be smaller than the minimum 
building height in order not to classify building roofs as terrain. All pixels should be classified by thresholding the 
weight values to distinguish between terrain and non-terrain pixels. 
3.2.1.2 Connected component labeling 
Segmentation is applied after having identifying the non-terrain pixels. Segmentation is based on the adjacency 
relationships among pixels. For this reason, connected component labeling is used to identify the connected pixel; 
(based on 8 directions) and assign a unique label for each segment. To reduce of the noise and the misclassification, 
median filtering is used before connected component labeling, Also median filtering is used to eliminate thin lines ani 
to disconnect some wrongly connected segments. The resulting segments are candidates to be building segments and 
will be further classified as described in the following section. 
3.2.1.3 Distinction between building and vegetation segments 
The distinction between building and vegetation segments is done based on the minimum building area and the chang 
in elevation and slope/orientation within each segment. 
Minimum building area 
Using the minimum building area as a threshold value for the areas of the segments, we exclude and eliminate the smal 
segments. Concerning a certain application, the threshold can be set according to the specifications of the required 
buildings. For example, if the building area is specified as larger than 50 nt to have a certain function as housing, thi 
value is used as the threshold. Moreover, the maximum building area can be used also as another threshold. 
Change in elevation 
The change in elevation will be obtained by applying a Laplacian filter—with coefficients as presented in equation (J 
—to the DSM. Within only the planer building faces, the pixel values in the resulting image are (ideally) zeros, whik 
for vegetation and along edges the pixel values differ from zero. Segment shrinkage of one pixel width is done tt 
exclude the boundary pixels because they disturb the classification. 
-1 -1 -i 
Laplacian Filter Coefficients 2| -1 8  -1 0 
-l -1 -1 
The classification of the whole segment into building or vegetation is difficult due to the contribution of the edge pixel 
inside the building segment (i.e. the pixels between building faces). Therefore, the classification will be done pixel bj 
pixel inside each segment. It is assumed that errors of the laser data are random errors and have a vertical componell 
only. Moreover, it is assumed that there are no errors due to the interpolation (using bilinear interpolation). The standart 
deviation of the change in elevation is computed based on the coefficients of the Laplacian filter by applying the em 
propagation rule. A threshold value can be chosen based on the standard deviation of the elevation change. Each pix 
will be classified in a binary classification into face pixel or vegetation (or edge) pixel. By counting the number of fa 
pixels in comparison to the total number of pixels for each segment, the segment will be classified into building ¢ 
vegetation. For example, if the number of interior face pixels is larger than 50% of the total number of pixels in te 
segment, the segment is classified as building. The selected ratio (threshold) depends on the face size, the size d 
chimneys, and the pixel size. Prior knowledge about building faces is required for selecting this threshold value. It ha 
to be mentioned that the previous ratio can be used in determining whether the segment contains a building and adjac? 
vegetation detected as one segment in the building detection process. 
  
618 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
Chan; 
The s 
in tw 
on the 
image 
menti 
the cl 
by ap 
the st: 
again 
each s 
know] 
It has 
the pi 
be av. 
segme 
33 ! 
331 
3.3.1.1 
The st 
the sai 
the ch 
standa 
into su 
Havin; 
image 
the cas 
based | 
the ses 
orient: 
standa 
examp 
of the : 
where 
3.3.1.2 
After 2 
correct 
used. '] 
the pix 
same r 
more c 
boundz 
segmeı 
applyir 
height 
faces. 
reclass;