A REMOTE SENSING IMAGE SEGMENTATION METHOD BASED ON SPECTRAL 
AND STRUCTURE INFORMATION FUSION 
Qinling Dai a Guoying Liu b Cancai Wang c Leiguang Wang* d 
a School of Wood Science and Interior Design, Southwest Forestry University, 300 Bailongsi Road, Kunming, Yunnan, 
China, 650224 
b College of computer and commmunication engneering, Changsha University of Science & Techology, 76Chiling 
Road,Changsha, China, 410079 
c School of Polymer Science and Engineering, Qingdao University of Science & Technology, Zhengzhou 
Road,Qingdao,Shandong,China,266042 
d National Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, 
129 Luoyu Road, Wuhan, China, 430079 
Commission VII, WG VII/6 
KEY WORDS: Segmentation, Mean Shift, Information Fusion 
ABSTRACT: 
The focus of this paper is on the segmentation algorithm for the high resolution multispectral imagery to get regions which are 
roughly corresponding to land cover types. We propose the combination of spectral features and spatial features by Gabor filter 
banks and a variable mean shift clustering algorithm in high dimension is employed to achieve feature fusion. Some issues on the 
measure metric of different features and standilzation are discussed. The whole algorithm is evaluated on the synthetic texture image, 
which are cropped from a QuickBird image with typical land-cover types. Compared with fixed bandwidth Mean Shift segmentation 
algorithm, our method has better performance in distinguishing the different land cover types with similar spectral surface. 
1. INTRODUCTION 
Image segmentation is a very important process in many image 
vision applications, which involved in partitioning an image 
into isolated regions, such that each region shares common 
properties and represents a different object.(Jung,2007).A large 
collection of literature on image segmentation has been 
proposed over the past decades. Since image segmentation 
represents the interface between image pre-processing and 
image understanding, segmentation algorithms are often 
application-dependent(Zhang et al.,2007).In our image 
understanding target, segmentation regions are as a basis for the 
following image classification mission. It is obvious that the 
quality of classification is directly affected by segmentation 
quality, so our segmentation algorithm focuses on partitioning 
HR multi-spectral remote sensing images into isolated regions 
which correspond to land-cover classes roughly. Compared 
with close-range images, the HR or very high resolution (VHR) 
remote sensing image is very rich in spatial detail and some 
land cover classes have similar spectral representation, such as 
glass and tree, road and roof. It means that the conventional 
spectral features will be inadequate for the VHR imagery 
segmentation. It seems evident that the VHR images do create 
additional challenges in terms of information extraction and 
classification (Dell'Acqua et al.,2004; LeiGuang et al.,2006; 
Puissant et al.,2005; Xin et al.,2007). 
Recently , a unified frame for gray and color image filter and 
segmentation based on the mean shift (Comaniciu and 
Meer,2002) presents good results for close-range images and 
the most important is robust in feature space. In this algorithm, 
a kernel in the spatial-range joint domain is defined, in which 
spatial location with 2-dimension and the range feature with 
image pixels in the CIE Luv color space is filtered, and then 
filtered pixels are clustered together according to the threshold 
of the minimum region defined by pixel count. Essentially, 
image filtering (also called image smoothing) based on mean 
shift is an instance of gradient ascent with an adaptive step size 
in joint color-position domain and the segmentation is the 
combination of seeking the mode of joint density in color- 
position feature space and merging of modes. One attractive 
property of image filtering based on mean shift is that the filter 
result is discontinuity preserving, which means that pixels in a 
region (mostly belonging to the same land-cover classes) will 
be smoothed and the variability in the region is deduced, while 
the boundary between regions has been preserved. 
In our remote sensing image segmentation tasks, decomposition 
of the multispectral image into homogenous tiles is desired, and 
one tile can correspond to one kind of ground object roughly. 
So our segmentation tasks are somewhat object-oriented. 
Considering the spectral similarity of some pairs of ground 
object, the general fixed-band mean shift based segmentation 
in (Comaniciu and Meer,2002), which only clusters color or 
gray level feature in image lattice, is not sufficient to getting 
desirable segmentation result. The spectral similarity of 
different land-cover types in HR imagery may lead to the 
convergence of different classes to the same mode in the 
spectral feature space. The discrimination performance of 
spectral feature is further reduced. Considering the 
* E-mail: corresponding author :wlgbain@gmail.com.