SEGMENTATION OF HIGH-RESOLUTION SATELLITE IMAGERY BASED ON 
FEATURE COMBINATION 
S. Wang 3 , A. Wang 3 ' 
a School Of Remote Sensing Information Engineering, Wuhan University, 129 Luoyu Road, Wuhan 430079, China 
wangsg@whu.edu.cn, wap531 @163.com 
Commission IV, WG IV/9 
KEY WORDS: Segmentation, Multispectral Imagery, Feature Distributions, Local Binary Patterns 
ABSTRACT: 
High resolution (H-res) satellite sensors provide rich structural or spatial information of image objects. But few researchers study the 
feature extraction method of H-res satellite images and its application. This paper presents a very simple yet efficient feature 
extraction method that considers the cross band relations of multi-spectral images. The texture feature of a region is the joint 
distributions of two texture labelled images that are calculated by its first two principal components (PCs) and the spectral feature is 
that of grayscale pixel values of its two PCs. The texture distributions operated by a rotation invariant form of local binary patterns 
(LBP) and spectral distributions are adaptively combined into coarse-to-fine segmentation based on integrated multiple features 
(SIMF). The performance of the feature extraction approach is evaluated with segmentation of H-res multi-spectral satellite imagery 
by the SIMF approach. 
1. INTRODUCTION 
High-resolution (H-res) satellite images have become 
commercially available and have been increasingly used in 
various aspects of environmental monitoring and management. 
The fine resolution satellite imagery makes it possible to detect 
the land cover/use type in detail. But on the other hand as the 
H-res satellite sensors increase the within field spectral 
heterogeneity and the traditional pixel-based image analysis 
method will produce many salt and pepper image areas. Object 
based image analysis (OBIA) method that makes it possible to 
get inferences based not only on spectral properties, but also on 
information such as object shape, texture, spatial relationship as 
well as human knowledge are proving to be useful in this high 
spatial resolution world. A necessary prerequisite for OBIA is 
successful image segmentation. 
As the traditional pixel-based segmentation/classification 
methods have some limitations, especially when they are 
applied to H-res satellite imagery. Recently, H-res satellite 
image segmentation has drawing considerable attention. Several 
new segmentation methods have been examined by a number of 
authors. The fractal net evolution approach (FNEA) is 
embedded in the commercial software environment (Hay et al. 
2003) and was thoroughly introduced by Baatz and Schape 
(2000). Various research projects have demonstrated the 
potential of this multi-scale segmentation approach (Hay et al. 
2003), yet it still suffers from some limitations, i.e., it cannot be 
fully exploited because of lack of a theoretical framework and 
users have to find useful segmentation levels by ‘trial and error’ 
(Hay et al. 2003). Pesaresi and Benediktsson (2001) proposed a 
new morphological multiscale segmentation method based on 
the morphological characteristic of connected components in 
images, which is however only suited for complex image scenes 
such as city area of H-res satellite images. Examples of more 
recent approaches include segmentation by the floating point 
based rain-falling watershed algorithm and a region adjacency 
graph based multi-scale region merging (Chen et al. 2004; Chen 
et al. 2006), multiscale object-specific segmentation (MOSS) 
(Hay et al. 2005), segmentation based on the Gaussian Hidden 
Markov Random Field model (Gigandet et al. 2005) and 
automatic segmentation of H-res satellite imagery by 
integrating texture, intensity and color features (Hu et al. 2005). 
There have been some research on segmentation method based 
on combining multiple features (Chen and Chen 2002; Hu et al. 
2005). Chen and Chen (2002) evaluated a color texture 
segmentation approach combining color and local edge patterns 
by constant weights, however the method only performs well on 
some simple color texture images and natural scenes and it is 
not suitable for complex H-res satellite images. The approach 
presented by Hu et al. (Hu et al. 2005) performs relatively well 
on H-res satellite imagery but the weights of three features are 
hardly to determine. As in our previous research work (Wang et 
al. 2007), SIMF by the features including texture and spectral 
distributions that are described in the paper and colour feature 
which is the Hue/Saturation histogram and by the weight 
combination similar to (Hu et al. 2005) performs well on few 
images. So we will make some comparison with segmentation 
approach combing the texture and feature distributions. 
In segmentation based on integrated multiple features (SIMF), 
the choices of highly discriminating features (Ojala and 
Pietikainen 1999) and how to combine the features are the most 
important factors for a successful segmentation. In this paper, 
we present a region-based unsupervised segmentation method, 
which utilizes features integrating texture and spectral 
distributions. The two features are then used to measure the 
similarity of adjacent image regions during the coarse-to-fine 
segmentation process (Chen and Chen 2002; Hu et al. 2005; 
Ojala and Pietikainen 1999). The main objective of this 
research is to examine the ability of the new feature extraction 
method in segmentation of H-res satellite images and easiness 
of SIMF by two features.