高分辨率遥感影像的自优化迭代分类方法

Self-optimizing iterative classification method of high-resolution remote sensing images

针对高分辨率遥感影像提出了一种面向像斑的自优化迭代分类算法,基于半监督聚类算法获取训练样本,以支持向量机为核心设计了自优化迭代分类器。使用分型网络演化算法获取像斑,并从中选取少量标记样本;结合标记样本,利用半监督模糊C均值算法对像斑进行聚类,并基于密集度筛选得到训练样本;设计了自优化迭代支持向量机分类算法,对所有像斑进行迭代分类直到满足分类要求,并在分类过程中对近邻分类结果进行统计得到高可信度样本以自主优化训练样本集。基于以上方法分别对武汉市Quick Bird和World View影像进行分类实验,分类总精度分别达到94.67%与92%,与基于人工选取训练样本情况下进行分类的分类总精度(82%与82.67%)、常规支持向量机分类总精度(87.33%与88%)、最小二乘支持向量机分类总精度(88%与89.33%)相比,精度有明显提升,分类效果较好。

A self-optimizing iterative classification method based on image segments which classifies high-resolution remote sensing images by acquiring training samples through semi-supervised fuzzy C-means and designing the self-optimizing iterative classifier based on support vector machine was proposed. Image segments could be obtained by fractal net evolution approach and a few labeled samples were selected; based on labeled samples,image segments were clustered by semi-supervised fuzzy C-means clustering method and then training samples could be obtained by intensity filtration from clustering results; the self-optimizing iterative support vector machine was designed to carry on classification iteratively until the classification requirements were met and during the classification process,training samples were updated and optimized to improve the performance of the classifier by statistical analyses of the two adjacent classifications. Quick Bird and World View images of Wuhan City were classified by the method proposed by this paper and the overall accuracy achieved 94. 67% and 92%. In comparison with the overall accuracy of the classification with training samples selected by manual work,the regular support vector machine classification method and the least squares support vector machine classification method,the accuracy of the suggested method is obviously higher and the classification effect is better.