基于多尺度特征的高光谱端元提取方法

摘要

为提升高光谱端元提取效率,避免同种地物端元的多次重复提取,根据高光谱同种地物光谱曲线主要特征近似、全体像元中端元个体两两具有极大差异的原则,提出了一种基于多尺度特征的高光谱图像端元提取方法,即多尺度特征像元纯度指数方法(MSPPI).首先利用一维离散小波变换获取多尺度的光谱信息,然后利用光谱角距离和欧式最小距离提取相应的多尺度的光谱特征,并利用像元和其邻域内像元点之间的关联,引入距离测度提取纯像元,最终实现端元提取.通过在高光谱数据库USGS和AVIRIS中的实验验证算法有效性,并与SPPI算法和N-FINDR算法进行对比.结果表明:MSPPI算法能够提取全部端元,且每种地物端元提取百分比低于5％,SPPI虽然能够提取全部端元但提取百分比均高于10％,端元重复提取现象严重,而N-FINDR不能有效提取小面积地物,说明MSPPI算法性能优于N-FINDR算法和SPPI算法.%In order to enhance the efficiency of the endmember extraction and avoid the repetitious extraction on those same object endmembers, according to the principle that the hyper spectral curves of same objects have similar fea-tures and the difference between endmembers is greatest in all pixels, a multi-scale hyperspectral image end-member extraction method, named as multi-scale pixel purity index(MSPPI), is proposed. Firstly, The one-di-mension discrete wavelet transform(1-DWT) is used to obtain multi-scale spectral information. Then, the multi-scale features are extracted from spectral angle distance (SAD) and euclidean minimum distance (EMD), and based on the correlation between the pixels and the pixels in the neighborhood, the distance measure is used to extract the pure pixels, which helps the endmember extraction finally to be realized. The validity of the algorithm is tested through the experiments on the hyperspectral database USGS and AVIRIS respectively, and compared with the N-FINDR algorithm and the SPPI algorithm. The experimental results show that the MSPPI can extract all the endmembers with the percentage of endmember extraction less than 5％of each object;the endmembers are also extracted completely by SPPI with a percentage higher than 10％and a serious endmember repeated ex-traction phenomenon has arisen;the endmembers in small areas can not be extracted effectively by N-FINDR. It illustates that the performance of MSPPI algorithm is superior to those of N-FINDR algorithm and SPPI algorithm.