The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
terrain label. This procedure is not iterative. Finally the on- 
terrain points are interpolated to a raster image using the Natural 
Neighbor interpolation method to produce the Digital Terrain 
Model (i.e. DTM). Every pixel of DTM has a real coordinate in 
the object space and its value is proportional to the elevation of 
the terrain at that position. Fig3 shows the calculated DTM 
which is classified into 5 classes to show the slope aspect of the 
study area. The maximum and minimum elevations of DTM 
pixels are 25.94 and 23.17 meters respectively. 
Figure 3 - Calculated Digital Terrain Model (darker tones 
symbolize the higher pixels) 
3.3. Elevation analysis of nDSM 
Since buildings are highly elevated objects in a scene our 
system looks for high-rise objects in the study area in this step. 
A high-elevated pixel on a DSM may belong to any objects as 
well as a high region of terrain. To eliminate the effect of 
topography from a DEM, one should normalize the model. A 
surface model is normalized by the use of the corresponding 
terrain model. To normalize the DSM, we subtract the value of 
each pixel of DTM from the value of the corresponding pixel in 
DSM. The result is a normalized model called nDSM (i.e. 
normalized DSM). 
As shown in Fig. 4 on-terrain pixels have values (i.e. elevations) 
less than 10 centimeters. A 10 centimeters threshold classifies 
these pixels. This supports the fact that bare-earth segments of 
the scene have the same elevations in both DSM and DTM. A 
“Terrain” label is assigned to these pixels. The next pace is 
finding the locally highest pixels of the nDSM. These pixels 
represent objects which lie on the terrain. A threshold equal to 3 
meters derives the high-rise objects of the scene. As shown in 
Fig. 4 this threshold has classified the rest of nDSM pixels into 
two other classes which are “Low- rise” and “high-rise” objects 
classes. 
j | Q -0.1 Meters 
IS 0.1-3 Meters 
H|3-16.63 Meters 
Figure 4 - Classified normalized DSM (height thresholding) 
So we classified all nDSM pixels into three classes; “Terrain”, 
“Low-rise”, and “High-rise”. The pixels of “High-rise” class are 
the input to the last step of our building detection procedure. 
3.4. Roughness analysis of nDSM 
It is conspicuous that not all “High-rise” pixels belong to 
buildings. As explained earlier (section 3.1) some dense 
vegetation regions weren’t detected as “Rough” regions. So the 
presence of tree pixels in “High-rise” pixels is also expectable 
as shown in Fig. 5. The goal of this step is the detection of 
building pixels among the pixels of “High-rise” class. 
To fulfill this task we have used a simple concept of surface 
roughness that is the computation of the slope map of the nDSM 
image. The main motivation is that the slope of the roofs of 
buildings doesn’t often change abruptly. In addition, our study 
area contains a few flat-roofed buildings. Consequently we 
computed the slope map of the nDSM using the ESRI 
ArcGIS9.2 software. The amount of slope for each pixel is 
computed by this software using this formula [Burrough 1998]: 
sloperadians = ATAN ( V ([dz/dx] 2 + [dz/dy] 2 )) 
Where [dz/dx] and [dz/dy] are the rate of height change in X 
and Y directions respectively. These rates are computed for a 3 
* 3 cell neighborhood around every pixel. 
So the slope map of the nDSM image is computed and 
generated. Then a quick trial and error method leads a human 
operator to define an appropriate threshold by which the slope 
map of nDSM can be divided into two classes; “Severely 
variable” and “Slightly variable” regions. The members of the 
former class are those pixels representing high-rise vegetations 
and walls of buildings, and the members of the latter class are 
the representatives of building roofs and relatively flat areas on 
the terrain. Fig. 5 shows the results of the thresholding the slope 
map of nDSM where the threshold is set to 5% of the slope 
range of the slope map. 
Figure 5 - Classified normalized DSM (slope thresholding) - 
Dark and light tones represent “Severely” and “Slightly” 
variable areas respectively 
3.5. Detection of building pixels 
So far we have obtained an nDSM image with three classes; 
“Terrain”, “Low-rise”, and “High-rise” regions, as well as a 
slope map of nDSM with two classes; “Severely variable” and 
“Slightly variable” regions. 
As stated in section 3.3 buildings are highly elevated objects in 
a scene. In section 3.4 we also mentioned that the most of the 
buildings have smoothly sloped roofs. These facts suggest that 
the building pixels are those pixels of “high-rise” regions which 
their counterpart in the slope map belongs to the “Slightly 
variable” regions. In other words each pixel which belongs to 
both of these classes is a building point. This way our system 
detects building pixels and labels them as “Building”. All the 
remaining pixels of the scene are classified as “Non-building”. 
The resulting image is shown in Fig. 6(a). It is obvious that still 
some dense, high-rise vegetation are misclassified as building.