基于结构性字典学习的高光谱遥感图像分类

摘要

In order to improve the classification accuracy of hyperspectral images,a new structured dictionary-based method for hyperspectral image classification was proposed. This method incorporates both spatial and spectral characteristics of hyperspectral images to obtain a dictionary of each pixel,the pixels in an identical pixel group have a common sparsity pattern;image sparsity representation coefficients are calculated in light of the dictionary to gain sparse representation features of hyperspectral images;the classification of hyperspectral images is determined using a linear support vector machine. Experiments on AVIRIS and ROSIS hyperspectral images were carried out. The experimental results show that compared with the common dictionary learning,the classification accuracy is respectively raised by 0 . 041 1 and 0 . 046 6 ,the Kappa coefficient is improved by 0 . 179 3 and 0. 056 3,respectively.%为提高高光谱遥感图像的分类精度，提出了一种新的结构性稀疏表示及字典学习的高光谱遥感图像分类方法．该方法能同时利用高光谱遥感图像像素间的空间及光谱关系得到表示每个像素的字典，被划分为同一像素组的像素具有通用的稀疏模式；由字典计算图像的稀疏表示系数获得遥感图像的稀疏表示特征；利用线性支持向量机算法实现对高光谱遥感图像的分类．对AVIRIS和ROSIS高光谱遥感图像进行的实验结果表明：提出的方法比普通字典学习分类精度分别提高0．0411和0．0466，Kappa系数分别提高0．1793和0．0563．