The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Pari B3b. Beijing 2008 
443 
2.2 Extraction of Building and Trees by DSM 
Through the segmentation, we avoid to deal with pixel level but 
to deal with characteristic level. After the segmentation of 
images, now we deal with objects, not pixels. On this basis, 
there are many features can be used to extract buildings, like 
shape, form, orientation of the building and the ratio of 
length/width. But they often can not recognize buildings very 
well. DSM appears to solve the building extraction problems 
due to the data source requires much less model knowledge 
(Forstner, 1997). We use DSM as a key factor to extract 
buildings from surroundings. 
A Digital Surface Model (DSM) is an image consisting of 
height values including vegetation, buildings and other objects 
(Gerke et al.). The DSM supplies a good way to solve 3D 
problems in terms of building extraction, because the DSM has 
the information about the height, so it can distinguish buildings 
and roads, other low objects very well. 
Trees, buildings and other objects with large height, which in 
the DSM image appear high brightness, will be classified into 
the same group. Experiments are needed to find the optimal 
height parameter (in the DSM height usually appears as 
intensity) as threshold to classify. It works very well though 
some trees are also involved in. 
2.3 Building Extraction 
After using DSM to delineate the objects with high elevation, 
buildings and trees are grouped together. To reach a better 
delineation of the building areas, a ratio factor is needed. NDVI 
is the most useful factor to extract trees, e.g. in (G. Markus et 
al.). When the NDVI is not available (there is no NIR band), VI 
or other factors are involved. Through experiments, we found 
that the ratio of Green Band/Red Band is quite useful. Some 
other combinations are also available, e.g. the ratio of (Green 
Band layer value- Red Band layer value) / (Green Band layer 
value +Red Band layer value). With this, trees and buildings are 
separated. Then use mask to filter trees, only buildings are left. 
to exactly one image object. Each of the subsequent image 
object related operations like classification, reshaping, re 
segmentation, and information extraction is done within an 
image object level. Simply said, image object levels serve as 
internal working areas of the image analysis. 
3. RESULTS 
In our experiments, we use the software package eCognition to 
extract buildings. Definiens Developer is the product of 
Definiens Corporation with a powerful Integrated Development 
Environment (IDE) for rapid image analysis solution 
development. It has done quite well in object-oriented 
segmentation and classification. 
(d) 
Figure 3. Different scale parameters leading to different results, 
(a) scale parameter: 45; (b) scale parameter: 60; (c) result of a; 
(d)result of (b). The red area delineate building. (The original 
data supplied by Definiens Corporation) 
3.1 Multiresolution Segmentation 
For successful image analysis, defining object primitives of 
suitable size and shape is of utmost importance. As a rule of 
thumb, good object primitives are as large as possible, yet small 
enough to be used as building blocks for the objects to be 
detected in the image. Pixels are the smallest possible building 
block, however pixels have limited information. To get larger 
building blocks, different segmentation methods are available to 
form contiguous clusters of pixels that have larger property 
space. (Definiens, 2007). The segmentation scale parameter 
should be decided through the repeated practice.