基于改进模糊ISODATA算法的遥感影像非监督聚类研究

Study on the Supervised Classification of Remote Sensing Image Based on a Modified Fuzzy-ISODATA Algorithm

F-ISODATA是一种有效的遥感图像非监督聚类算法。但是,最优迭代次数很难设定;一般遥感图像的数据量大,若迭代误差限取极小值,分类也很难实现。本文以某次迭代中"合并"和"分裂"都为零为求最优分类数的迭代条件,而不是预先设定迭代次数;取最大和最小隶属度取代每一个隶属度为比对特征值,提高了分类速度和精度;利用等效转换研究隶属度矩阵的迭代误差变化规律,得出变化速度趋于稳定时为求解最优隶属度矩阵的智能迭代控制,减少人为事先干预。最后,进行实验分析,结果显示整个改进的算法提高了分类的智能化,整体效果较好。

The Fuzzy-ISODATA （FISODATA） is an effective non-supervised classification algorithm of remote sensing image. But it is difficult to set optimal number of iteration in advance. The size of RS image usually is large, if the iterative error limiteis set a minimum, then FISODATA is also difficult to operate. In this paper, it is designed that when there is no ＂merger＂ or ＂split＂ at certain iteration, instead of the iteration time set in advance, the classification number is regarded as the optimal one. The variation trend is studied by equivalent transformation,the optimal membership Matrix is determined intelligently by the membership error variation rate instead of the threshold set in advance. Finally, the experiment result is worked out, the analysis in theory and the experiments indicate the algorithm not only improves the intelligence of the algorithm, but also can get more accurate classification results.