> DATA 
t is fast method 
re for building 
as obtained by 
gical filter for 
nto building or 
1entation of the 
on plane fitting 
g the roof faces, 
e buildings and 
k for all terrain 
f 2-3 points per 
alts for building 
etric and height 
f the presented 
s. Buildings are 
y due to human 
Is. This leads t 
graphic objects 
d to obtain 3D 
n of the sensor, 
rmine the rang 
« day and night 
)SM as obtained 
ofs can be done 
in section 2. The 
d the analysis d 
nhancements d 
y TopoSys last 
n, Public Work 
  
Michel Morgan 
  
and Water Management Directorate, The Netherlands. We obtained a standard deviation of the elevation values of 0.10 
meters by analyzing the data. The average density (resolution) of the DSM is 2-3 points per square meter. The ground 
truth data that we used for reference is a vector map (lines) of the buildings, parcels, and roads in the area and check 
points with 3D coordinates. For the sake of exploring the data, we first produced shaded relief and contour maps from 
the DSM. The buildings can be easily seen in both maps. It has to be mentioned that we had no complete 3D data for 
the buildings at our disposal for a comprehensive evaluation of the results. Moreover, the accuracy and the degree of 
reliability of the ground truth data were not known. 
3 THE BUILDING DETECTION AND EXTRACTION PROCEDURE 
3.1 Data Re-sampling 
Processing of irregularly distributed points as obtained from laser scanning takes more time than processing of regularly 
distributed ones. The suggested procedure starts by re-sampling the “laser data” into a regular raster, i.e., producing a 
two-dimensional array of elevation values. To this end the cell size and the interpolation method should be chosen 
sensibly. 
3.1.1 Cellsize 
Unless many points are eliminated in the filtering process of the raw laser data the DSM from laser will have a fairly 
regular distribution of points. For a regularly distributed set of points, the cell size can easily be chosen such that each 
resulting cell contains only one point, thus avoiding information loss and redundancy. The cell size must be equal to the 
reciprocal of the square root of the “point density". We have applied this formula to our data, considering the highest 
density available in our original DSM (3 points/ni). It has to be mentioned that the regular raster has the same values of 
the regular grid whose grid points are located in the centers of the raster cells. In this research, when analyzing elevation 
values and the elevation difference the grid points are considered, and when dealing with areas, we consider the regular 
raster, assuming constant elevation throughout a raster cell. 
3.1.2 Interpolation method 
Interpolation of the DSM from laser scanning points using linear or higher order degree of surface curvature gives a 
chance to have erroneous height values of the cells along the boundaries of the non-terrain objects, which leads to some 
difficulties in building detection. Therefore nearest neighbor interpolation is used, although it leads to shift in the 
building boundaries. However, as the grid size becomes smaller, this error will be smaller. Another drawback of using 
nearest neighbor interpolation occurs for inclined roof faces where the interpolation error has horizontal and vertical 
components. Therefore, two interpolation methods will be employed for the same data using the same grid size; nearest 
neighbor interpolation will be used for building detection and linear interpolation for the classification between 
buildings and vegetation and for building extraction. 
3.2 Building Detection 
3.2.1 Distinction between terrain and non-terrain segments 
3.2.1.1 Morphological filter 
Separation between terrain and non-terrain pixels is done by applying a morphological filter (Hug, 1997), (Hug and 
Wehr, 1997) and (Kilian et al, 1996). Using the morphological filter, different sizes of moving windows are used. For 
each window, the “deepest pixel” and all other pixels which are “higher” than the deepest pixel within a certain range 
(band width) are detected. Weights are assigned to these pixels according to the chance of being terrain pixels. The 
weights of the pixels in a large window should be larger than those in smaller windows. Moreover, among the deep 
pixels, the deeper the pixel, the larger its weight. Equation (1) is the formula for computing weights. The idea behind 
this equation is to scale the weights from 0 to 1 linearly between the smallest and the largest window sizes and also 
along the band width above the deepest pixel inside the window. It has to be mentioned that the morphological filter is 
used to threshold the DSM data in a window relatively smaller than that of the whole data set in order to compensate for 
the sloping/variable terrain. 
window. size - min window size band width— (point. ht —deepest point_ht) (1) 
Pixel _ weight= 
band width 
max window size— min window. size 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 617