遥感影像单类分类的白化变换法

A Whitening Transformation Based Approach to One-class Classification of Remote Sensing Imagery

提出-种基于白化变换的单类分类方法.该方法仅需要兴趣类别的训练样本.首先,基于兴趣类别对原遥感影像作白化变换,使兴趣类别的分布在各个方向上的方差相同.然后,确定一个距离阈值实现单类分类,根据切比雪夫定理,选择不同倍数的标准差作为阈值进行单类分类试验.结果表明,各个地物类别都在3-4倍标准差的区间内获得最高的分类精度.最后,以3倍标准差作为阈值的单类分类结果,与单类支持向量机方法比较,两种方法的分类结果非常相近,而基于白化变换的方法阈值选择简单,鲁棒性强.

In this study ,a whitening transformation based approach to one‐class classification of remote sensing imagery isinvestigated .Only positive data are required to train the one‐class classifier .Firstly ,the image data is mapped to a newfeature space using the whitening processing with all directions of the class of interest .Then a threshold is selected to make a binary prediction .A heuristic method of threshold selec‐tion is performed in the experiment of one‐class classification .A series of values are set to the threshold based on standard deviation , and perform the one‐class classification with each threshold value .The experiment shows that high accuracy is achieved with the threshold rangefrom3 to 4 standard deviations of the mean .Finally ,the results of one‐class classification with the threshold of 3 standard deviations are compared to that of one‐class support vector machine .The results indicate that the proposed method provides nearly the same accuracy of one‐class classification as one‐class support vector machine .The advantage of the proposed method is that it can use a constant threshold to extract various land types .