特征优选随机森林的土地利用分类

Land use classification using random forest method based on feature optimization

针对遥感影像土地利用分类中特征空间选择对分类结果产生重要影响的问题,以Sentinel-2A影像为数据源,基于其丰富光谱信息和空间信息提取多维特征空间,利用平均不纯度减少方法排序特征重要性,采用特征优选的随机森林算法优化分类土地利用的特征空间,提高分类精度。结果表明,特征重要性差异较大,优选排名前14位特征变量组成特征空间,特征的贡献程度由大到小依次为红边指数、光谱特征、纹理特征、植被指数、水体指数。特征优选的随机森林分类结果总体精度为94.03%,Kappa系数为0.9179,优于原始随机森林算法,一定程度上提高了遥感影像土地利用分类精度。

This paper is concerned with the important influence produced by the feature space selection on the result of the classification,as occurs in the remote sensing image classification of land use.The study involves extracting the characteristics of multidimensional space using Sentinel-2A image as the data source and based on its abundant spectral information and spatial information;sorting features importance using the method of mean decrease impurity and thereby optimizing the feature space;achieving land use classification using random forest algorithm based on feature optimization;and ultimately improving classification accuracy.The results show that there is a great difference in the score of feature importance,and the feature space consists of the top 14 feature variables,and the contribution degree of the feature is ranked in the order from the largest to the smallest:red edge index,spectral feature,texture feature,vegetation index,and water index;and the random forest classification results based on feature optimization enables the overall accuracy of 94.03%,and the Kappa coefficient of 0.9179.The proposed method superior to the original random forest algorithm enables an improvement in the accuracy of land use classification based on remote sensing image.