基于端元字典稀疏解混的高光谱图像亚像元定位

Hyperspectral Image Sub-Pixel Mapping Based on Sparse Unmixing of Endmember Dictionary

针对高光谱图像中普遍存在的混合像元中各端元空间分布定位困难的问题,文中提出一种基于K-SVD的光谱解混算法,利用其解混结果进行亚像元定位。算法首先通过KNN分类来区分待处理图像中的混合像元和纯像元,然后借鉴基于冗余字典的稀疏分解相关理论,以标准光谱库为基础,通过基于K-SVD的字典训练算法训练产生最具代表性的地物光谱曲线,构建端元冗余字典,通过基于K-SVD的稀疏分解算法实现各端元丰度的求解。最后利用求得的丰度系数在两种空间性相关性约束下进行亚像元定位。实验结果表明,采用该算法进行模拟数据和真实数据的亚像元的定位可以取得不错的定位结果。

In general,it is difficult to locate the spatial distribution of each endmember in the hyperspectral image.To solve the problem,this study proposed a K-SVD-based spectral unmixing algorithm whose result of the demixing was further used for performing the subpixel location.First,the mixed pixel and pure pixel were distinguished by KNN classification,and then the sparse decomposition correlation theory based on redundant dictionary was used for reference.Based on the standard spectral library,the K-SVD based dictionary training algorithm was used to train the most representative material spectral curves,and the endmember redundancy dictionary was subsequently constructed.The abundances were solved by the sparse decomposition algorithm based on K-SVD.Finally,the obtained abundance coefficient was used to locate the sub-pixel under the two spatial correlation constraints.Experimental results showed that the proposed algorithm had reliable performance for the sub-pixel mapping effects of simulated data and measured data.