基于贝叶斯模型的鲁棒高光谱解混方法研究

Hyperspectral unmixing is the foundation of developing many hyperspectral application systems. However, most of the existing unmixing methods can not satisfy the real world applications in terms of noise robustness,scalability and accuracy. In hyperspectral images, there are always noise bands. To address this problem, this project aims to develop a robust hyperspectral unmixing model based on the mechanism of the mixed pixels and noise bands. Meanwhile, in order to improve the stability and accuracy of the proposed unmixing model, the intrinsic structure information hidden in hyperspectral images and the adaptive sparsity prior are embedded into the robust hyperspectral unmixing model based on local learning methods and sparse representation. To this end, we integrate all the above proposed formulations into a unified model based on bayesian framework. This will eventually provide a theoretical tool for developing effective optimization algorithms to solve the constructed unmixing model. To guarantee the above goals to be achieved, we will develop an experimental hyperspectral unmixing system based on the proposed algorithms. Upon this system, related algorithms will be analyzed comprehensively with hyperspectral data of several sources, which helps to improve their accuracy, scalability and robustness.

高光谱解混是构建多种高光谱应用系统的基础。现有的高光谱解混方法在鲁棒性、普适性以及精度方面还不能满足实际应用的需求。针对高光谱图像普遍存在噪声波段的问题，本项目从噪声波段与混合像元产生机理的出发，拟重点研究噪声鲁棒的高光谱混合像元分解模型。同时，项目将利用局部学习方法和稀疏表达，挖据高光谱数据的本质结构信息和自适应稀疏先验，并将其嵌入到高光谱解混模型中，提高解混算法的稳定性和精度。最后，通过集成高光谱数据的结构信息、自适应的稀疏先验和鲁棒的混合像元分解模型，项目将研究如何构建统一的贝叶斯解混框架，并在此框架基础上提供高效的高光谱解混模型求解算法。另外，本项目还将以研制的高光谱解混算法为基础，构建一个高光谱解混原型系统，并利用此系统对不同的高光谱数据源进行测试研究，以验证本项目所提出的算法，进一步完善和提高解混方法的准确性、普适性和鲁棒性。

本项目在贝叶斯框架下研究高光谱解混问题。高光谱解混是构建多种高光谱应用系统的基础。现有的高光谱解混方法在鲁棒性、普适性以及精度方面还不能满足实际应用的需求。针对高光谱图像普遍存在噪声波段的问题，本项目从噪声波段与混合像元产生机理的出发，研究噪声鲁棒的高光谱混合像元分解模型，提出了基于相关熵度量和L21范数的鲁棒高光谱解混方法。同时项目利用局部学习方法和稀疏表达，挖掘高光谱数据的本质结构信息和自适应稀疏先验，并将其潜入到高光谱解混中，提出了基于结构稀疏约束和基于数据引导稀疏的解混方法。同时，项目中提出的算法在遥感图像处理系统中得到应用。