SVM核函数对南汇潮滩湿地高分二号卫星数据分类精度的影响分析  

Influence of different SVM kernel functions on the classification accuracy of GF-2 image in Nanhui tidal flat

作　　者：王文柳 韩震[1,2] 李静 郭雨桐[1] 崔艳荣 WANG Wenliu;HAN Zhen;LI Jing;GUO Yutong;CUI Yanrong(College of Marine Sciences,Shanghai Ocean University,Shanghai 201306,China;Engineering Research Center on Estuarine and Oceanographic Mapping,Shanghai Municipal Ocean Bureau,Shanghai 201306,China)

机构地区：[1]上海海洋大学海洋科学学院,上海201306 [2]上海市海洋局河口海洋测绘工程技术研究中心,上海201306

出　　处：《海洋湖沼通报》2020年第2期78-89,共12页Transactions of Oceanology and Limnology

基　　金：国土资源部公益性行业科研专项(201211009)资助。

摘　　要：针对支持向量机在高分二号卫星遥感图像分类中的核函数选择问题,以长江口南汇典型潮滩湿地为研究区,进行了支持向量机不同核函数分类结果的对比分析。首先,根据实地测量以及无人机航拍影像标记六类地物共计1800个优质样本点,然后将标记样本点的像元值投影到三维空间中,分析了支持向量机分类过程中最优分类超平面的构建以及误差来源,最终从标记样本点中选取训练样本和测试样本,进行了支持向量机不同核函数的分类训练,得到分类结果和分类精度。研究结果表明,在训练样本数量相同的情况下,线性核函数支持向量机的分类结果好于RBF核函数和Sigmoid核函数的分类结果。三种核函数的分类精度都随着训练样本数量的增加逐渐增高并趋于稳定,RBF核函数和Sigmoid核函数支持向量机分类精度变化趋于对数分布。Considering the selection of SVM(support vector machine) kernel functions in GF-2 images, we analyzed the classification results using different kernel functions, and chose Nanhui tidal flat, a typical sample of tidal flat and wet land, as the object of study hoping to seek an optimal solution. Firstly, the experimental sample points were annotated by field measurements and the image of UAV(unmanned aerial vehicle) comprising 1800 sample points of high quality, which belong to the pinpointed six categories. Then, the pixel digital number of experimental sample points were projected to a three-dimensional space. As a result, the creation of optimal separating hyperplane and source of error were analyzed in the process of SVM classification. Finally, training samples and test samples which were selected from the experimental sample points were trained to get the classification results and classification accuracy. Experiment manifests that the classification result of linear kernel function is better than RBF(Radial Basis Function) kernel function and Sigmoid kernel function. With the increase of the quantity of training samples, the classification accuracy of three kernel functions also enhances gradually and tends to be stable at last. At the same time, the classification accuracies of RBF kernel functions and Sigmoid kernel functions present a trend of logarithmic distribution.

关 键 词：支持向量机 核函数 分类 高分二号卫星 最优分类超平面 训练样本 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]