In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
367 
was conducted at a fine level of segmentation. On the other 
hand, non-shadow areas were classified at an appropriate 
segmentation level, different from the levels for shadow 
detection and classification. After all classes were extracted, 
impervious surfaces in both shadow areas and non-shadow 
areas were aggregated to a single class, i.e. impervious surface. 
2.1 Multi-level segmentation 
In this study, an improved watershed transformation method (Li 
et al., 2010), was adopted for high resolution multispectral 
image segmentation. However, other image segmentation 
methods can also be used to produce segmentation results. 
In the image segmentation method by Li et al. (2010), 
multispectral gradient proposed by Li and Xiao (2007) was first 
used to extend the watershed transformation to multispectral 
image segmentation, and then the dynamics of watershed 
contours proposed by Najman and Schmits (1996) was adopted 
to reduce the oversegmentation in initially segmented image 
and produce multilevel segmentation results. 
After the dynamics of watershed contours (Najman and 
Schmits, 1996) were obtained, a threshold is applied to the 
values of the contour dynamics, in order to remove the 
watershed lines that have less significance and to produce a 
final segmentation result. The details for algorithm for 
computation of contour dynamics can be found in (Najman and 
Schmits 1996, Lemarechal et al. 1998, Schmitt 1998). 
2.2 OCSVM 
The OCSVM is a recently developed one-class classifier and 
has been widely used in ecological modeling (Guo et al. 2005), 
and remote sensing classification (Sanchez-Hemandez et al. 
2007) as well as change detection (Li and Xu 2009). In the 
OCSVM training process, only samples from the target class 
are used. Thus, it is suitable for the situations where only one 
class or some classes (but not all classes) are of interest and 
easy to sample or measure; the other class might be very 
difficult or expansive to measure. Therefore, the boundary 
between the two classes has to be estimated from data of the 
only available target class. The task is to define a boundary 
around the target class, such that it encircles as many target 
examples as possible and minimizes the chance of accepting 
outliers (Tax 2001). 
Scholkopf et al. (1999) developed an OCSVM algorithm to deal 
with the one-class classification problem. The OCSVM may be 
viewed as a regular two-class SVM where all the training data 
lies in the first class, and the origin is taken as the only member 
of the second class. The OCSVM algorithm first maps input 
data into a high-dimensional feature space via a kernel function 
and then iteratively finds the maximal margin hyperplane, 
which best separates the training data from the origin. 
2.3 Impervious surface mapping 
After multilevel segmentation was carried out, a hierarchical 
classification strategy was adopted using the OCSVM and 
multilevel segmentation results. Since shadows are common in 
the VHR images, the OCSVM was first used to extract the 
shadow areas at a coarse level of segmentation. After that the 
shadow areas and non-shadow areas were separately classified 
using the OCSVM and different levels of segmentation to 
extract the impervious surface. Finally, the impervious surface 
from both shadow areas and non-shadow areas were merged to 
produce a final impervious surface map. In each step, only 
samples from the class of interest are used to train the OCSVM. 
For example, in the stage of shadow extraction, only samples 
from shadow areas were used in the training process. This is 
different from traditional classification methods, where samples 
from all classes are required. 
2.4 Result evaluation 
In order to validate the proposed impervious surface mapping 
method, a method based on the use of traditional SVM and 
multilevel segmentation results was also used to extract the 
impervious surface. After multilevel segmentation results were 
obtained, the SVM classifier was first used to classify the 
image of the study area into several land cover classes, such as 
grass, tree, soil, impervious surface and shadow. The obtained 
shadow areas were then further classified using the SVM to 
several land cover classes, including impervious surface. 
Finally, the classification result from both shadow and 
non-shadow areas were merged to an 
impervious/non-impervious surface map, where the classes, 
such as tree, grass and soil were merged to a class 
non-impervious surface. 
3. DATA AND STUDY AREA 
A Quickbird image of Beijing urban area, acquired in 
September of 2003 was used in the experiment. The Quickbird 
imagery contains four multispectral bands with 2.44m