本文依托于国家科技基础性工作专项重点项目“中国湖泊水质、水量和生物资源调查”（2006FY110600）子课题“中国湖泊卫星遥感调查”，采用小样本学习、抗噪声性能好、学习效率高的支持向量机(SVM，Support Vector Machine)方法来提取多光谱遥感影像冰冻湖泊水体。支持向量机最大特点是根据结构风险最小化原则,尽量提高学习机的泛化能力，即由有限的训练集样本得到小的误差能够保证对独立的测试集仍保持小的误差，SVM在遥感信息提取方面，特别是在缺少先验知识的情况下，具有较高的推广能力。 本文分析了冰冻湖泊水体的光谱特征，说明了传统遥感影像分类方法不能达到较高精度的原因，在此基础上利用MATLAB支持向量机工具箱建立了SVM提取模型，并选择了东北地区的查干湖和长白山天池CBERS多光谱遥感影像进行了提取试验。试验表明，两湖提取精度支持向量机较最大似然法提高了3.93%。但仅应用SVM方法提取冰冻湖泊水体，提取结果小图斑多、分类不完整，因此本文在原来模型的基础上加入了邻近同化滤波方法，有效剔除了小图斑。 综合SVM邻近同化滤波模型提取冰冻湖泊水体的研究过程，本文得出的结论如下： （1）查干湖湖面整体冰冻；据实地考察和相关文献，长白山天池西部由于湖底受温泉影响，湖水冬季不封冻，从影像上看天池西部呈深蓝色的部分就是未冰冻的湖泊水体。应用SVM邻近同化滤波提取模型，不仅实现了对全部冰冻湖泊水体的精确提取，对冰水混合的湖泊水体也同样适用，所以本文的提取模型对冰冻情况复杂的湖泊水体精确提取有着广阔的应用前景。 （2）与长白山天池相比，查干湖结冰湖面破碎度较高，由于处在破碎部分的像元光谱复杂，最大似然法难以达到较高的提取精度。在同样训练区样本下，与最大似然法相比，SVM提取精度提高了5.36％，由此可以看出SVM具有较强的小样本学习能力，提取精度高。 （3）为剔除湖边浅水沼泽和冰冻期湖面雪等干扰地物的影响，本文提出邻近同化滤波方法，这也是本文的创新之处。经邻近同化滤波模型处理后，提取结果小图斑较少，类别完整。 本文在如下两方面还需要进一步研究： （1）长白山天池处在山区，从图4.17可以看出，其南部的山体阴影影响并未有效去除掉，因此提取山区型冰冻湖泊水体时，如何有效剔除山体阴影的影响有待于进一步研究。 （2）选择邻近同化滤波模板的大小时需要根据实际情况人为设定，模板选择太大会导致分类失真，选择太小又不能有效除去干扰地物，因此如何选择合适的同化滤波模板有待于进一步研究。

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.