Research On Segmentation Of Remote Sensing Image Based On Fuzzy Theory

Segmentation of remotely sensed image is a synthetic research task which is relate to many research fields such as image processing, pattern recognition, computer vision, neural works and so on. It has many important applications and has given rise to more and more interests for researchers, and it is important to deal with the progressively remote sensing data. Analyzed the development states and current developing of remote sensing image segmentation, this paper does more research on segmentation with spectrum, segmentation with texture, and both information fusion segmentation by fuzzy algorithm, and many valuable achievements are obtained.As to segmenting with spectrum, FasART neural networks based on fuzzy logical system is used. In this part, firstly analyses the uncertainty of remote sensing data and the characteristic of explaining the image with man vision, then summarizes formed mechanism of the networks from adapt resonate theory, discusses networks structure characteristic, puts forward a simple model. In order to input feature pattern combine perfectly with the networks, the feature pattern is described with lingual variable, finally does comparative experiment with Maximum Likelihood and fuzzy ARTMAP. The results indicated that this algorithm has higher precision.Segmentation with texture, it includes extracting feature, transforming and segmentation. In the section of extracting feature of multispectrum remote sensing image, summarizing the usual algorithms, and analyzing the characteristic of multispectrum image, this paper makes use of fuzzy texture analysis method to extract the feature, defines spatial fuzzy texture spectrum, and validates by classification with samples. In the section of feature transforming, thinking of worse classification characteristic of texture feature, the feature is represented by lingual variable. In the section of segmentation, analyzing the feature with spatial fuzzy texture spectrum, and discussing the problems of split-and-merge and multiresolution image pyramid algorithm, aiming at uncertainty of the feature which is extracted form little measure region, this paper presents a new quadtree-stuctural based on region algorithm ?split-and-expand algorithm. Comparative experiment between split-and-merge and split-and-expand showed the algorithm is effective.As to segmenting with fusion multispectum and texture information, first, this paper summarizes the knowledge of data fusion, analyses development states of application of data fusion technology in the remote sensing segmentation, and discusses briefly Dempster-Shafer theory. Then, according to characteristic of these feature, proposes a new fuzzy integral fusion algorithm based on decision fusion, and evaluates theinfluence of feature dimension to the output of classifier. Finally, validates by experiment and achieves higher segmentation precision.The general work of this paper is summarized and the further research direction is pointed in the last chapter.