关联图像融合的高光谱亚像元分类多任务联合稀疏表示方法研究

Hyperspectral image subpixel classification is a hot topic in the field of remote sensing information processing, and its classification accuracy can be improved significantly by fusing a high spatial resolution auxiliary image. Until now, there have three problems needed to be solved in this study. The first one is that image fusion and subpixel classification are considered individually. The second one is that in the degradation cases, the relationship between the two images used in the fusion process is unreliable. And the last one is that there exist the propagation errors caused by unmixing in the subpixel classification process. Based on multitask joint sparse representation, this project studies the subpixel classification methods coupled with image fusion, and the details are described as follows. Firstly, by exploiting the dependences among the structures of the auxiliary image, the dictionaries that can describe the image in different views are constructed for the multitask learning. Secondly, by taking advantage of the regularization theories and methods, we model the hyperspectral image fusion problem by integrating with the unmixing problem, and thereby build the multitask joint sparse representation model that can deal with the propagation errors. Finally, by using the structural low rank representation based image construction method, we propose a set of methods for subpixel classification coupled with image fusion, as well as the fast optimization algorithms of their models. This project will improve the accuracies of hyperspectral remote sensing information processing and its quantitative interpretation, and promote the development of hyperspectral remote sensing applications in the fields of national economy, military affairs and etc. Therefore, it is of great theoretical and practical significance to study the proposed project.

高光谱图像亚像元分类是遥感信息处理领域的研究热点，与高空间分辨率辅助图像进行融合是提高其分类精度的一种经济且有效的手段。目前此类研究还需解决三个方面的问题：图像融合与亚像元分类过程相分离，退化情形下融合图像间的相互转化关系具有不确定性，亚像元分类过程中存在混合像元分解传播误差。本项目以关联图像融合的亚像元分类为具体科学问题，以多任务联合稀疏表示为方法依据，首先深入挖掘辅助图像内的结构相关性，构造具备不同表示能力的多任务联合字典；然后利用正则化理论与方法，联合混合像元分解进行高光谱图像融合建模，构建有利于控制误差传播的多任务联合稀疏表示模型；最后基于结构化低秩表示图像重构，提出一套关联图像融合的亚像元分类方法，并设计相应模型的快速优化算法。本项目将为提高高光谱遥感信息处理和定量解译的精度奠定基础，并将推动高光谱遥感在国民经济和军事等领域的实际应用，具有重要的理论和实际意义。

高光谱遥感图像具有较高的光谱分辨率，能够区分传统遥感中不可探测的物质，是当前对地观测和定量遥感领域研究的国际前沿。亚像元分类是高光谱遥感图像后续分析和应用的关键，与高空间分辨率辅助图像进行融合是提高其分类精度一种经济且有效的手段。项目以关联图像融合的高光谱亚像元分类为具体科学问题，以多任务联合稀疏表示为方法依据，首先深入挖掘辅助图像内的结构相关性，构造具备不同表示能力的多任务联合字典；然后利用正则化理论与方法，联合混合像元分解进行高光谱图像融合建模，构建有利于控制误差传播的多任务联合稀疏表示模型；最后基于结构化低秩表示图像重构，提出一套关联图像融合的亚像元分类方法，并设计相应模型的快速优化算法。项目研究过程中取得重要成果包括：就多任务学习中的参数选择问题，提出一种参数自动选择的核协同表示分类方法；采用张量描述多任务特征，提出一种面向高光谱图像分类的张量迹回归方法；考虑采用超像素分割提取高光谱图像的多尺度空间信息，继而提出一种多核协同表示分类方法；探索非负性先验在核分类中的作用，提出一种面向高光谱图像分类的全约束最小二乘方法；采用典型相关性分析挖掘多重特征间相关性，提出一种基于改进组合核的高光谱图像分类方法；采用多任务描述高光谱图像全色锐化过程，提出一种面向高光谱图像超分辨的多任务联合稀疏表示方法等。本项目的研究将丰富并拓展高光谱图像融合及亚像元分类问题理论和算法的发展，对于推动高光谱图像后续分析与应用具有重要的理论意义。同时，本项目提供的方法与技术促进高光谱遥感在国民经济和军事等领域的实际应用，具有重要的理论和实际意义。