A new approach to classification of ground background using MSS images

A new approach to classification of ground background using MSS images

Wen Jian-Xian
First Scientific Research Institute
of Corps of Engineers of General Staff
Wuxi, China

Shi Peng-Fei
Institute of Image Processing & Pattern
Recognition, Shanghai Jiao-Tong University
Shanghai, China

Abstract
Starting from the research of the topical characteristics of ground background, A new approach of classification using MSS Image and algorithms of segmentation on the basis of the region growing principle and the pyramid data structure is presented in this paper. Some results of classification of ground background are also given.

Introduction
The background is a set of types of surface features in spatial domain according to the spectral characteristics, the time characteristic and the combined laws between them, for instance, to divide areas in the southeast of China into forest, shrub, cultivated area and the area of river and lake to classify the northwest of China into desert, grassland, forest and cultivated area etc… In fact, the background is sometimes more meticulous, so that the wilderness may also be divided into desert, Gobi, soil desert and salt desert; the grassland can be divided into grassy marshland, prarie and wilder grassland and so on. Segmentation of the background plays an important role in the research of background optical distributive characteristics of background, but is also the prime way to obtain parameters of background optical characteristics. It not only represents the spatial combined laws of background and provides the region distributive characteristics of background, but is also the prime way to obtain parameters of background optical characteristics. With a vast territory of our country and the facts of various kinds of landforms and the very complicated natural geographical conditions, it is extremely difficult to obtain background optical characteristics only through on-the-spot investigation and measurement by hand. And the existing background information has many limitations.

Now first of all, we classify the background into various kinds of regions, according to the available geographic and remote sensing information. Then for the same kind of region, we select some training areas and measurement spots. The background information can be obtained. It will greatly save manpower and material resources, and shorten the cycle of research. The strategy of background segmentation is to classify the background using the method of auto-classification with segmentation the background into independent regions with their own features which is based on the principle and pyramid data structure.

Features Extraction an Classification
According to the similarity of spectral features, there are two classification methods : supervised and unsupervised classifications. Considering the former, we must supply triaging samples, i.e. the typical region with known type-training areas. however, because we don't know much about the background, it is difficult to provide the correct surface samples. From technical point of view, it is unnecessary to know what the various training features are. When the latter is considered the information, of MSS is so large a datum matrix. Because we need to process a whole even several images at the same time, it will be restricted by the capacity of memory and time. In view of above mentioned reasons, we use the method of combining supervised and unsupervised classifications. The idea of this method is first to choose several sample areas in the whole image, and to unsupervisedly classify the background of sampling areas, then to analyse and process te above classification to obtain the clustering centers of the entire image in accordance with those centers to supervisedly classify the whole image.

Selection of sample Areas.

We choose the sample area according to available geographic information such as maps of vegetation, soil, topography and geomorphology as well as initial satellite images. The primary principles are : (a) a sample area should hold its own background features, that is , the sample areas we choose should possess there distinguishable background features. (b) a simple area should have representations. We select an area which covers only a very small proportion of background, as a sample area. (c) a sample, area should contain as many types of background as possible. As a result, we can guarantee the complete types of background, reduce the number of sampling areas and increase efficiency. The size of sample depends on the practical situation. Generally, 256 by 256 pixels are responsible, sometimes 128 by 128 by 512 pixel are possible.
Classification of sample areas.

The classification of a sample area depends on the multispectral information of resource satellite (MSS information or TM information). A satellite image offers the data of spectrum brightness that represent the energy pf the reflected light from background. The experimental results show that it is reasonable to use MSS image as parameters of area classification. In order to implement this process, we consider Euclidean distance as the discriminating function of classification.

The threshold C_k of the interest similarity coefficient and the threshold D₀ of the interest distance are used. They satisfy the following relations.

Background Region Partition
The background segmentation is fulfilled by making use of the region growth principle and the pyramid data structure. We support the original image is N by N, and N = 2^m, it can divide into m+1 levels, each of which is also divided into some small regions using the same data structure. As dealing with the kth level, the membership of a subregion to the known category i is.

where
J_k (x,y,i) is the membership of the subregion in the kth level, the coordinate of whose center is (x,y) to the ith category,
J_m (x,y,i) is the average membership of all pixels in the original image to the ith category,
J_o (x,y,i) is the average membership of all pixels in the original image to the ith category,
The procedure for calculating the membership of each pixel in the original image to the sets I(x,y), and

(a) Calculating the sets I(x,y) and

Where
C: the number of category in the original image,
j: 1,2,……….C
g(x,y) ; the grey value of the (x,y) pixel in the original image
V_j: the average grey value of the jth category in the original image.
d_j: the distance between the (x,y) pixel and the clustering center of the jth category
(b). calculating the membership U_i(x,y)
if I(x,y) = {f}, then

Where
i = 1,2,.........C
U_i(x,y): the membership of the (x,y) pixel in the original image to the ith category.

Conclusion
A method of the background region segmentation is presented in this paper. All algorithms written by C language have been run on the Micro Vax - II computer system. A result is shown in Fig. 1 which satisfied the practical use in the different terrain classification.

Acknowledgement
We would like to thank Prof. W.T. Wu for his useful help and comments of the earlier draft.

References

W.K. Pratt, Digital Image Processing, John Wiley & sons, Imc. 1978
Tou, Pattern Recognition Principles. Addison-Welsely Publishing company, 1974.
Swain, S.M. Davis, Remote Sensing: THE Quantitative Approach, McGram-Hill company, 1987.