1257
AN EFFICIENT MULTI-SCALE SEGMENTATION FOR HIGH-RESOLUTION REMOTE
SENSING IMAGERY BASED ON STATISTICAL REGION MERGING AND MINIMUM
HETEROGENEITY RULE
H. T. Li\ H.Y. Gu a,b , Y. S. Han 3 , J. H. Yang 3 , S. S. Han b
institute of Photogrammetry and Remote Sensing, Chinese Academy of Surveying and Mapping, Beijing, 100039,
P.R.China.
b Institnte of Surveying and Geography Science, Liaoning Technical University, 47 Zhonghua Road, Liao Ning 123000,
P.R.China.
Commission IV, WG IV/9
KEY WORDS: Multi-scale Segmentation, High-Resolution, Statistical Region Merging (SRM), Minimum Heterogeneity Rule
(MHR), Fractal Net Evolution Approach (FNEA)
ABSTRACT:
Multi-scale segmentation is an essential step toward higher level image processing in remote sensing. This paper presents a new
multi-scale segmentation method based on Statistical Region Merging (SRM) for initial segmentation and Minimum Heterogeneity
Rule (MHR) for merging objects where high resolution (HR) QuickBird imageries are used. It synthesized the advantages of SRM
and MHR. The SRM segmentation method not only considers spectral, shape, scale information, but also has the ability to cope with
significant noise corruption, handle occlusions. The MHR used for merging objects takes advantages of its spectral, shape, scale
information, and the local, global information. Compared with Fractal Net Evolution Approach (FNEA) eCognition adopted and
SRM methods, the results showed that the proposed method overcame the disadvantages of them and was an effective multi-scale
segmentation method for HR imagery.
1. INTRODUCTION
Image segmentation is the process of dividing an image into
homogenous regions, which is an essential step toward higher
level image processing such as image analysis, pattern
recognition and automatic image interpretation (Blaschke and
Strobl, 2001). So far, there are over 1000 kinds of segmentation
approaches developed (Zhang, 2001). General segmentation
methods include global behaviour-based and local behaviour-
based methods (Kartikeyan, et al., 1998). Global behaviour-
based methods group the pixels based on the analysis of the
data in the feature space. Typical examples are clustering and
histogram threshold. Local behaviour-based methods analyze
the variation of spectral features in a small neighbourhood.
Typical examples are edge detection and region extraction (Fu
and Mui, 1981).
However, not all of the segmentation techniques are feasible for
High-Resolution (HR) imagery due to the following facts:
(1) The HR imagery is multi-spectral and multi-scale, so both
the complexity and redundancy are increased obviously;
(2) The HR imagery provides the more details such as
spectral, shape, context and texture. The traditional
segmentation algorithm is only based on the colour
information and can not provide the satisfying results;
(3) Different class has its inherent feature in different scale.
For example, at coarse scales we may find fields, while at
finer scales we may find individual trees or plants. So the
segmentation model on one scale needs to be modified
when used on the other scale.
is important to segment imagery effectively with all kinds of
information and object character. Recently, several authors
have proposed multi-scale segmentation algorithms for HR
imagery (Chen, 2003; Cheng, 2005; Sramek, 1997). A majority
of image segmentation algorithms are based on region growing
methods which take some pixels as seeds and grow the regions
around them based on certain homogeneity criteria. The
commercial software, eCognition, adopts Fractal Net Evolution
Approach (FNEA) for segmentation. FNEA is a region growing
technique based on local criteria and starts with one pixel image
objects. Image objects are pairwise merged one by one to form
bigger objects. The merging criterion is that average
heterogeneity of image objects weighted by their size in pixels
should be minimized (Baatz and Schape, 2000; Benz, et al.,
2004).
However, the local region growing technique has some
limitations:
(1) It is not efficient in both computation and memory;
(2) It has some difficulties in gathering a set of seeds and an
adequate homogeneity criterion;
(3) It depends on the choice of starting point and the order in
which the pixels and regions are examined;
(4) It is hard to find coincident boundary because one pixel
image object merges with another without respect of
adjacent pixels.
In order to overcome the disadvantages of segmentation from
one pixel and get more accurate segmentation result, several
authors suggest merging bigger objects generated by initial
segmentation, which avoids of the disadvantages of region
growing method from single pixel.
Therefore, owing to the HR imagery is multi-spectral and multi
scale, and includes more details and information of the object, it