The Accurate Extraction Of Frozen Lake Water Based On SVM-Neighboring Filter Assimilation Model

This article is based on the program servey of Lakes in China, which is the sub-topics of the program Survey of Water Quality,Volume and Biological Resource of Lakes in China which is belonged to National Fundamental Research Program of China,and adopts SVM (Support Vector Machine) to extract multi-spectral remote sensing image of frozen lake water. This method is learning from small sample, good at noise proof and quite efficient. The most prominent feature of SVM is to improve the generalization ability of the learning machine as much as possible according to the Structure Risk Minimization Principle, which means if the machine gets small errors from limited training sets sample it can also be insure to keep small error in the case of individual training set. The method of SVM possesses great generalization ability in the field of remote sensing information extraction, especially when we are lack of a priori knowledge.This article analyzed spectral signature of frozen lake water, and explained the reason why they can't get higher accuracy when using the traditional remote sensing image classification methods. On this basis, writer established the SVM extraction model by utilizing the MATLAB-SVM toolkits, and executed the extraction experiment on CBERS multi-spectral remote sensing image in Lake Balapan and Heaven Pool. The experiment shows that the extraction accuracy increases 3.93% when using SVM method than using Maximum Likelihood. But SVM method also has its weak points, which are smaller extraction result, more marks and the incomplete classification. So this article added the Neighboring Filter Assimilation method to the existed model, which eliminates small marks efficiently.After combining the research proceed of extracting the frozen lake water by means of SVM and Neighboring Filter Assimilation, conclusions come as follow: (1) The surface of Lake Balapan is frozen completely. According to the field survey and relative documents, the Heaven Pool doesn't get frozen in winter owing to the hotspring in the bottom. In the images, the navy blue part in the west section is the lake water without frozen. This SVM-Neighboring Filter Assimilation model not only can extract the completely frozen lake water precisely, and also work in ice-water mixed lake water. Therefore, the extraction model in this article has vast application in extracting precisely the complicate-situation lake water.(2) Comparing to that in Heaven Pool, the degree of fragmentation in the frozen surface of Lake Balapan is higher. Because the pixel spectra in the broken part are quite complicate, Maximum likelihood method can't work out high extraction accuracy. In the condition of same sample, SVM can increase the accuracy about 5.36% than Maximum likelihood. It shows that SVM is better at small sample learning ability, and has higher extraction accuracy.(3) In order to eliminate the affection from shallow marsh around the lake and the snow on the lake surface in ice period, this article also provided extraction experiment in Neighboring Filter Assimilation method .This is also the biggest innovation of this article. The result shows that the image is smaller, marks are less and classification is complete.It still needs development in the future in the following two aspects:(1) Heaven Pool is in the mountain area, and Figure 4.17 shows that the affection from the shadow of mountain in south hasn't been eliminated. Therefore, during extracting the frozen lake water in mountain area, it still needs research on how to eliminate the affection of mountain shadow.(2) When choose the size of Assimilation filtering template, it should be decided dependently according to the practical situation. If the size is too big, the classification will be distorted; if it's too small, the interferent cannot be eliminated efficiently. So it still needs research on how to choose suitable Assimilation filtering template.