高光谱图像混合像元处理技术研究

Due to the current development trend and urgent requirement of mixed pixel processing techniques on hyperspectral imagery, in this project researches will be carried out on endmember extraction, spectral unmixing, sub-pixel mapping and color display for hyperspectral imagery. Some classical endmember extraction algorithms require massive calculation of volume as well as dimension reduction. The traditional linear spectral mixing models based spectral unmixing approach fails to accurately describe the intra-class spectral variation, and the cost function of the newly developed support vector machine based unmixing model is not consistent with the ultimate goal of spectral unmixing. In current sub-pixel mapping methods, the description of spatial dependence is quite rough, and the effectiveness of Markov random field based sub-pixel mapping model which has the advantage of simultaneously considering spatial and spectral constraints is limited by the single constraint. The typical color display approaches are processed based on the lowly accurate classification results. The goal of this project is to look for effective methods to solve these issues. Specifically, a fast endmember extraction algorithm will be developed which will transform the volume calculation to the distance calculation and more importantly, no dimension reduction will be required in the new algorithm. The unmixing error constraint will be incorporated into the support vector machine based unmixing model to improve the performance. In order to obtain more satisfying sub-pixel mapping results, a new spatial attraction model will be developed to fully consider the spatial dependence. Also, additional information will be added into the MRF model to provide multiple constraints. The results derived from spectral unmixing and sub-pixel mapping will be utilized to develop color display approaches with higher accuracy and more levels. The success of this project will be of great significance as well as application value to the effective exploration and utilization of hyperspectral imagery.

针对目前高光谱混合像元处理技术的发展现状和迫切需求，课题拟研究端元选择、光谱解混和亚像元定位及彩色显示等热点技术。主流端元选择方法包含大量体积计算，且需降维预处理；传统基于线性光谱混合模型的解混方法无法准确刻画类内光谱变化，而新兴的支持向量机解混模型与解混要求不完全一致；当前亚像元定位模型中空间相关性未能充分贯彻，而同时考虑光谱和空间约束的马尔可夫随机场模型约束条件单一；现有的典型可视化方法建立在低精度的分类结果之上。本课题通过研究新型距离测算方法建立免于降维预处理与体积计算的快速端元选择算法；在支持向量机解混模型中施加解混误差约束来提高光谱解混性能；建立完全贯彻空间相关性的新空间引力模型及多约束马尔可夫随机场模型以获得高精度的亚像元定位结果；利用光谱解混与亚像元定位的形数结合信息实现高精度多层次的可视化技术。本课题的成功研究对高光谱图像信息的有效挖掘和利用有着重要的理论意义和应用价值。

针对目前高光谱混合像元处理技术的发展现状和迫切需求，课题研究了端元选择、光谱解混和亚像元定位及彩色显示等热点技术。针对现有技术中存在的问题(即主流端元选择方法包含大量体积计算，且需降维预处理；传统基于线性光谱混合模型的解混方法无法准确刻画类内光谱变化，而新兴的支持向量机解混模型与解混要求不完全一致；当前亚像元定位模型中空间相关性未能充分贯彻，而同时考虑光谱和空间约束的马尔可夫随机场模型约束条件单一；现有的典型可视化方法建立在低精度的分类结果之上)，本课题通过研究新型距离测算方法，建立了免于降维预处理与体积计算的快速端元选择算法, 当所提取的端元数目为16时，新方法较之传统方法速度提高数十倍，并且端元数目越大，优势越明显；在支持向量机解混模型中通过施加解混误差约束。所提出新模型较之原始支持向量机、传统线性光谱混合分析解混方法和几种典型多端元光谱混合分析方法解混精度有明显提高；建立了完全贯彻空间相关性的新空间引力模型及多约束马尔可夫随机场模型，从而获得了高精度的亚像元定位结果，基于马尔可夫随机场的亚像元定位模型有着同时考虑空间和光谱信息且不依赖于软分类结果等优势；利用光谱解混与亚像元定位的形数结合信息，实现了高精度多层次的可视化技术。该模型产生的彩色图像具有良好的视觉效果及可分性，满足距离保持特性，同时，对局部地物的细节组成显示方案是可行的。本课题的研究结果对高光谱图像信息的有效挖掘和利用有着重要的理论意义和应用价值。