The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
The first step is multi-resolution segmentation. The right scale 
of parameter is crucial. When segmenting in Color mode with 
high radiometric resolution, it’s better to choose a larger scale 
parameter for generating image objects. In our example, a high- 
resolution aerial photograph was used for segmentation. Image 
objects like buildings, roads and types of land covers required 
large scale parameters. If the scale is too small, the 
segmentation will generate many fragmentations while the scale 
is too large to omit some useful details. Through trials, we 
chose 60 as final decision. The edge of DSM usually is quite 
rough, so it did not involve in the segmentation step. 
Nevertheless, the information contained in the elevation image 
can be well used for discriminating elevated rooftops from low 
lying roads. Therefore, DSM just can be utilized in the 
classification. In this step, its weight was set to 0. 
The result of multi-resolution is shown in figure3. From figure3 
we can see, the large-scale segmentation can describe buildings 
well. 
3.2 Classification 
In this step, we take DSM as thematic information to extract 
buildings and trees by the classification function of eCognition. 
Height information is used as a filter. Based on this, image 
objects can be analyzed according defined criteria and assigned 
to classes that best meet these criteria (Definiens, 2007). In 
“Class Hierarchy” view we define a new class named con&tree, 
which mainly contains buildings and trees. In eCognition, in the 
“Feature View” window, take layer means->layer /(this is the 
DSM layer), and then find the right scale parameter. Use “class 
description” to insert the classification expression, then classify. 
After the procedure of classification, each image object is 
assigned to “con&tree” or not—so classes connected with the 
class hierarchy. 
Figure 4. The elevation information in DSM (Green represents 
high value while blue represents low) 
3.3 Building Extraction 
In this step we separate trees from “con&tree”. As we know, 
kinds of ratio index are quite useful to extract vegetation. For 
the lack of NIR band, we use the ratio Green Band layer value / 
Red Band layer value instead of NDVI to extract trees. The 
building’s expression uses “similar to” tree but in the “invert 
to” form. The entirety of image objects is organized into a 
hierarchical network, and the “tree” and “building” classes 
inherit from “con&tree”. 
Figure 5. Class Hierarchy window 
In the end, as we can imagine, the green area represents the 
trees, and the red area represents buildings. The exhausted work 
of distinguishing buildings, roads and other similar objects are 
avoided in this approach. 
3.4 Evaluation 
Two measurements for a detection evaluation as described in 
(Lin et al, 1998) were made: 
Detection percentage =100 ■ TP/(TP + TN) 
Branch factor =100 ■ FP/(TP + FP) 
The two measurements are calculated by making a comparison 
of the manually detected buildings and the automated results, 
where TP(True Positive) is a building detected by both a person 
and the program, FP(False Positive) is a building detected by 
the program but not a person, and TN (True Negative ) is a 
building detected by a person but not the program. A building is 
considered detected if any part of the building is detected; an 
alternative could be to require that a certain fraction of the 
building be detected. The analysis is based on the image of 35 
buildings. Among these, 35 buildings are detected with 
lmistaken involved, and 1 omitted. The DP is 97.1% and the 
BF is also 97.1%. The figure is relatively high in contrast with 
the data obtained by other methods. 
4. CONCLUSIONS 
This paper proposes an object-oriented method to extract 
building information by DSM and orthoimage in eCogniton. It 
extracts buildings in only two main steps and reaches quite high 
accuracy. Also there are some points should be paid attention to: 
The object-oriented method proposes a way leading us to think 
in a new manner. Within this, the spectral, semantic and 
contextual information can be combined together to mine the 
image information. Although many methods to extract 
buildings are developed in recent decades, the most useful 
information in the extraction is spectral information. With 
segmentation, the structure of level and inheritance is 
constructed, and the time consuming of computation has been 
cut down. 
Building extraction with DSM has quite a lot of limitations. It is 
required flat terrain, and when the area is large, the extraction 
result is not good. To solve this problem, we can use a 
normalized DSM (nDSM) which is the difference DSM-DTM. 
Here, a DSM is an image consisting of height values including 
vegetation, buildings and other objects. A DTM consists of only 
those points lying directly on the terrain. Another method is to 
444