基于地物光谱可分性的CHRIS 高光谱影像波段选择及其分类应用

本文以黄河口湿地为研究区,应用覆盖该区域的CHRIS高光谱遥感影像,提出了一种基于地物光谱可分性的滨海湿地高光谱影像波段选择方法。该方法利用研究区的7种典型地物的110余条现场实测地物光谱曲线,通过分析比对地物两两之间的光谱可分度,确定地物类型之间的光谱可分区间,基于此选取CHRIS高光谱影像的地物分类特征波段,应用三种经典的监督分类方法(支持向量机法SVM、人工神经网络法ANN和光谱角制图法SAM)开展利用全波段的和利用本文方法选择特征波段的分类对比实验。结果表明:(1)基于光谱可分性特征波段的方法较全波段分类精度有所提高,其中ANN分类精度最高,为82.52%,较全波段分类精度提高了约为5.1%;(2)芦苇、水体、黄河水和裸滩4种地物的识别能力高,生产者精度都在80%以上;(3)碱蓬的用户精度提升最为明显,约在7%。

A CHRIS hyperspectral band selection method based on spectral separability and classification application

In this paper, we proposed a hyperspectral image band selection method based on spectral separability feature with the covering CHRIS hyperspectral remote sensing of the Yellow River estuary wetland. This method was used to determine the spectral dividing range of featured types by using the spectra of more than 110 fields of seven kinds of typical objects analyzing and comparing the spectral features. Based on these, we can select classification features bands of CHRIS hyperspectral images. Three classic applications of supervised classification methods (SVM, ANN and SAM) were used in selective features bands classification and comparison experiments with the whole bands and this method. The results showed that: (1) The classification overall accuracy of the spectral separability method has increased, among which, the classification accuracy of ANN has the highest accuracy of 82.52%, improved about 5.1%. (2) The producer accuracy of reed, water, Yellow River and nude beach are as high as more than 80%. (3) The user's accuracy of Suaeda was significantly enhanced about 7%.