基于SVM的高光谱数据解混技术研究

针对高光谱图像混合像元处理技术的现状和需求，课题拟研究SVM多端元光谱解混方法统一框架下的以下五个方面：1)多类解混器选择。在1-a-r型多类解混器的基础上，研究1-a-1型多类解混器在解混中的拓展应用方法；2)全约束处理。遵从最小二乘准则，符合物理意义，建立解混丰度的快速全约束调整方法；3)SVM解混模型的改进。研究最小二乘SVM对于解混的特殊意义和不足，在此基础上对相应解混模型进行改进；4)非平行边界拓展。将孪生SVM拓展应用到光谱解混中，建立由分类结果到解混分量的映射，并进一步完善该模型；5)不确定性分析。重点分析多端元解混中的两种不确定性问题产生的根由，从而为降低不确定性寻求科学方法。研究有效的光谱解混技术，对于提高高光谱图像的分析效果和科学决策意义重大。

Aiming at the current situation and demand of hyperspectral image processing technology, this project studies the following five aspects under the unified framework of SVM-multi-endmember spectral unmixing method: 1) multi-class soft-classifier selection. On the basis of 1-a-r multi-class soft-classifier, the extended application method of 1-a-1 multi-class soft-classifier in unmixing is studied; 2) full constraint processing. According to the least square criterion and the physical meaning, a fast total constraint adjustment method for the solution mixing abundance is established; 3) the improvement of SVM spectral unmixing model. In this project, we study the special significance and shortcomings of the least square SVM for the spectral unmixing, and then improve the corresponding spectral unmixing model; 4) non-parallel boundary expansion. The twin SVM is extended to spectral unmixing, the mapping from classification results to unmixing components is established, and the model is further improved; 5) uncertainty analysis. This project focuses on the analysis of the causes of two kinds of uncertainty problems in the multi- endmember spectral unmixing, so as to seek scientific methods to reduce the uncertainty. It is of great significance to study effective spectral unmixing method for improving the analysis effect and scientific decision-making of hyperspectral imagery.