1171 
Department of Geographical Information Science, Nanjing University, Nanjing 210093, P. R. China 
xiaopf@gmail.com 
KEY WORDS: Land Cover, Segmentation, Image Processing, Automation, Feature Extraction, Quickbird 
ABSTRACT: 
Image segmentation is a valuable approach that performs an object-based rather than a pixel-based analysis of high-spatial resolution 
satellite image. A multiscale approach for segmenting the pan-sharpened multispectral QuickBird-2 image based on vector field 
model is proposed. The edge features are obtained using the first fundamental form of the multispectral bands. The response of log 
Gabor bank filtering of each band is fused as multiscale texture features based on first fundamental form. Then, the image 
segmentation is implemented based on texture-marked watershed transform. The segmentation accuracy is assessed using dis 
crepancy measures between a reference map and the segmentation. The experimental results show that the proposed approach gives 
a better solution of integrating spectral and texture information for the segmentation of multispectral remotely sensed image. 
1. INTRODUCTION 
While remote sensing has made enormous progress over recent 
years and a variety of sensors, such as IKONOS-2, QuickBird-2, 
Orb View-4, deliver high resolution data on an operational basis, 
a vast majority of applications still rely on basic image 
processing concepts developed in the early 1970s: classification 
of single pixels in a multi-dimensional feature space. In most 
cases, information important for the understanding of an image 
is not represented in single pixels but in meaningful image 
objects. Procedures for image segmentation which are able to 
dissect images into sets of useful image objects are therefore a 
prerequisite for the successful automation of image inter 
pretation (Blaschke et al, 2000). 
Despite some early research activities (Kettig and Landgrebe, 
1976), image segmentation was established relatively late in the 
field of remote sensing (e.g. Ryherd and Woodcock, 1996), and 
has rarely featured thus far in image processing (Schiewe, 
2002). Although there has been a lot of development in the 
segmentation of remotely sensed image (e.g. Pal et al., 2000; 
Pesaresi and Benediktsson, 2001; Pekkarinen, 2002; Acharyya 
et al., 2003; Devereuxa et al., 2004; Li and Gong, 2005; Hu et 
al., 2005; Chen et al., 2006; Lia et al., 2007; Wang and Boesch 
et al., 2007), there has been little progress in the segmentation 
of multispectral image (e.g. Kartikeyan et al., 1998; Baatz and 
Schape, 1999; Evans et al., 2002; Sarkar et al., 2002; Li and 
Xiao, 2007). 
Remote sensing sensors are producing multispectral so that in 
contrast to the most often used greyscale image in the 
disciplines not only the complexity but also the redundancy 
increases. Early approaches to considering multispectral image 
attempted to combine the response of each bands. The way are 
combined is, in general, heuristic, and has no theoretical basis 
(Sapiro and Ringbach, 1996). A principled way to look at 
multispectral image is vector field model (Zenzo, 1986; Cumani, 
1991). The value of multispectral image at a given point can be 
regarded as A-dimensional vector in R v , and the difference of 
image values can be defined from the theory of surfaces in diff 
erential geometry. In the research, method of first funda-mental 
form (Sapiro and Ringbach, 1996; Scheunders, 2002) is applied 
to the definition of gradient and fusion of texture from multi 
spectral remotely sensed image. 
Watershed transform is a powerful morphological tool for 
image segmentation that is usually defined for greyscale image 
(Vincent and Soille, 1991). This paper presents an extension of 
the watershed algorithm for multispectral image segmentation. 
As shown in Figure 1, edge features are obtained using the first 
fundamental form of multispectral bands. For solving the over 
segmentation problem of watershed transform, The response of 
log Gabor bank filtering of each band is fused as multiscale 
texture features also based on first fundamental form. Then the 
image segmentation is implemented based on texture-marked 
watershed transform. Finally, The segmentation accuracy is 
assessed using discrepancy measures between a reference map 
and the segmentation. 
Figure 1. The scheme for multispectral image segmentation