基于mRMR特征优选算法的多光谱遥感影像分类效率精度分析

在遥感图像分类过程中,进行合理的特征优选操作,将有助于提高分类器的分类效率及精度。本文以淮南地区资源三号卫星多光谱遥感影像数据为例,采用二值离散化、直方图法及F统计法3种计算方法实现mRMR（minimal-Redundancy-Maximal-Relevance）算法特征优选过程。根据3种方法所得到的特征优选结果及全部特征信息,分别采用C5.0决策树和K近邻2种分类器进行图像分类实验,并利用目视解译方法对不同方法组合的影像分类结果进行精度验证。实验结果表明,利用3种计算方法实现mRMR特征优选算法对不同分类器的影响程度不同：在分类效率方面,C5.0决策树分类器可提高36.84%,而K近邻分类器可提高72.05%;在分类精度方面,C5.0决策树分类器能保证分类精度大致不变,总体分类精度可提高0.60%,Kappa系数可提高0.80%,而K近邻分类器总体分类精度可提高4.34%,Kappa系数可提高7.90%。

Efficiency and Accuracy Analysis of Multispectral Image Classification Based on mRMR Feature Selection Method

Image classification is a popular research topic in the field of remote sensing. This technology has been widely used in environmental protection, military, urban planning, and other fields. Interfering by the massive feature information of remote sensing image, applying the reasonable feature selection approach in the progress of image classification becomes critical for improving the efficiency and accuracy of classification. This paper extracts the image feature data from the ZY3 satellite multispectral image of Huainan region, and studies the mRMR (minimal-Redundancy-Maximal-Relevance) feature selection method. This algorithm has a simple core principle and low requirement of data. The core problem of this algorithm is the computation of mutual information. The mRMR algorithm is initially applied in the field of bioscience, such as the gene expression analysis, and it is not widely used in the field of remote sensing. This research uses three methods (the binary discretization, histogram method and F-statistic) to realize the computation process of mRMR algorithm. And two classifiers (the C5.0 decision tree and k-nearest neighbour) are used for the classification based on three types of feature selection results and the total feature information. Moreover, the visual interpretation is used to verify the image classification results from these different methods. The study shows that the results produced by different mRMR computation processes are distinct regarding to different classifiers. In terms of efficiency, all methods can improve the efficiency of C5.0 and KNN. The classification efficiency is increased by 36.84% for C5.0 and by 72.05% for KNN. In terms of accuracy, all method can maintain the accuracy of C5.0 while improve the accuracy of KNN. The total classification accuracy and Kappa coefficient are increased for C5.0 by 0.60% and 0.80%, respectively. The total classification accuracy is increased by 4.34% and the Kappa coefficient is increased by 7.90% for KNN. In summary, the feature selection method based on the mRMR algorithm is effective in the procedure of multispectral image classification.