基于支持向量机的扎龙湿地遥感分类研究  

Remote Sensing Classification for Zhalong Wetlands Based on Support Vector Machine

作　　者：张策[1] 臧淑英[1] 金竺[1] 张玉红[1] 

机构地区：[1]哈尔滨师范大学地理科学学院,黑龙江哈尔滨150025

出　　处：《湿地科学》2011年第3期263-269,共7页Wetland Science

基　　金：国家自然科学基金重点项目(41030743)资助

摘　　要：湿地遥感分类是实现湿地动态监测、管理与利用的重要手段之一。由于湿地具有独特的生态环境,获取实测样本点相对困难,因此,研究小样本、高精度的湿地遥感分类方法十分必要。以扎龙国家级自然保护区为研究区,采用支持向量机方法进行了研究区湿地遥感分类研究,初步剖析了样本数量与特征维度对分类结果的影响,并同传统的最大似然分类方法进行了比较。研究结果显示,支持向量机分类结果一般优于最大似然分类结果,尤其在小样本、高维度下,支持向量机方法具有较大优势。当每类样本数为100时,支持向量机高维分类结果总精度最高,达到88.125%,分类获得的扎龙保护区湿地总面积为90307.17hm2,其中水体面积为8301.15hm2,积水沼泽面积为33063.57hm2,无积水沼泽面积为48942.45hm2。研究结果表明,支持向量机方法是湿地遥感分类的有效手段。Wetlands are integral parts of the global ecosystem as they can prevent or reduce the severity of floods, feed ground water, and provide unique habitats for flora and fauna. Wetland remote sensing classification is one of the most important means to realize wetland dynamic monitoring, management and utilization. But spectral uncertainty or vagueness caused by spectral confusion between-class and spectral variation with- in-class remains a challenge to the wetland remote sensing classification. Exploring small sample size, high accuracy classification methods are hence necessary, due to the complex aquatic-terrestrial environments of wet- lands, which are relatively difficult to acquire training samples by field survey. Traditional remote sensing clas- sification methods usually require large training data sets, for instance, maximum likelihood classification, a widely used classical supervised classification method in remote sensing. The method of maximum likelihood classification assumes that the training statistics for each class have a normal distribution. Training of a maxi- mum likelihood classification aims at a complete description of each class. To achieve this goal, the training set should be sufficiently large. Furthermore, the training data instances should be typical and representative of the classes in order to derive appropriate training statistics on which the classification can be based. It is hard to find such training samples in wetland remote sensing classification due to the wetland＇ s complexity and ecological vulnerability. New approaches which need small training data sets and could deal with high feature dimensions remain to be investigated. Primary analysis was made with respect to the impact of sample size and feature dimensionality on the classification accuracy of the support vector machine, in comparison with classical maximum likelihood clas- sifier. Training samples were partitioned into five sample sizes per class 20, 40, 60, 80 and 100, based on strat- ified random sampling st

关 键 词：扎龙国家级自然保护区 支持向量机 湿地遥感分类 样本数量 特征维度 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]