超像素低秩高光谱稀疏解混  

Superpixel Based Low-Rank for Hyperspectral Sparse Unmixing

作　　者：李璠[1] LI Fan(Jiangxi Province Key Laboratory of Water Information Cooperative Sensing and Intelligent Processing,Nanchang Institute of Technology,Nanchang 330099,China)

机构地区：[1]南昌工程学院江西省水信息协同感知与智能处理重点实验室,江西南昌330099

出　　处：《南昌工程学院学报》2021年第3期67-75,共9页Journal of Nanchang Institute of Technology

基　　金：江西省教育厅科学技术研究项目(GJJ190956,GJJ170992)。

摘　　要：近年来,稀疏解混在高光谱图像解混领域受到广泛关注。借助已知端元光谱库,稀疏解混规避了高光谱数据中纯像元缺失的问题和端元提取的过程,使解混简化为从端元光谱库中选择可以有效表达混合像元的最优光谱特征子集。这是一个组合优化问题,常采用稀疏线性回归算法解决,而利用图像的空间信息约束解空间可以有效提高求解精度,获得更好的解混性能。针对现有空间稀疏解混模型对数据空间结构描述不充分的问题,本文提出一种超像素低秩稀疏解混方法,该方法一方面采用超像素分割技术自适应生成同质区域,在传统稀疏解混模型中引入基于超像素的局部低秩正则项,保持图像内在的局部低维空间结构,促进图像的空间一致性,另一方面在稀疏正则项中引入光谱加权因子,诱导丰度矩阵的行稀疏性,促使图像中所有像元的丰度向量联合稀疏。模拟数据和真实数据实验结果表明,与同类算法相比,所提算法可以有效抑制噪声,在低信噪比情况下获得了更高的解混精度,能保留丰度图更精细的空间信息。In recent years,sparse unmixing has been widely concerned in the field of hyperspectral image unmixing.With the help of spectral libraries known in advance,sparse unmixing can ignore the issue of no pure pixels in the data and circumvent endmember extraction.The unmixing process is simplified to select the best spectral signature subset from the library that can effectively express the mixed pixels.This is a combinatorial optimization problem,which can be solved with sparse linear regression algorithms.The spatial information of the image is utilized to reduce the solution space of the optimization problem,which can effectively improve the accuracy of the solution and obtain better unmixing performance.The previous spatial sparse unmixing model described the spatial data structure insufficiently.To address this limitation,a superpixel based low-rank sparse unmixing model is established.In the proposed method,homogeneous regions are adaptively generated by superpixel segmentation,and then a local low-rank regularization term based on superpixels is introduced to maintain the inherent local low-dimensional spatial structure of the image and improve the spatial consistency of pixels.In addition,the spectral weighting factor is introduced into the sparse regularization term to induce the row sparsity of the abundance matrix and enhance the joint sparsity of the abundance vectors of all pixels in the image.The results of simulated and real data experiments show that the proposed algorithm is effective for noise suppression,retains more details of the abundance map and has higher unmixing accuracy in the case of low signal-to-noise ratio than other spatial sparse unmixing algorithms.

关 键 词：高光谱遥感 稀疏解混 超像素分割 低秩表示 光谱加权 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]