高光谱图像多核字典学习与分类研究

Multiple kernel dictionary learning (MKDL) aims at constructing an over-completed and discriminant dictionary, which can promote the mining the sparsity of nonlinear structure information in HSI, so as to improve the performance of MKDL for complex scenes. In order to handle the challenge of insufficient support of current sparse representation technique to hyperspectral image (HSI) processing for complex scenes, this project will construct a multiple kernel dictionary learning and classification framework，firstly, a multiple kernel dictionary learning algorithm with multiple kernel combination pattern is proposed, which can effectively mine the diversity nonlinear structure information. Secondly, by investigating the effect of spatial neighborhood information for MKDL, a spatially joint MKDL is proposed by fully mining the spatial homogeneity and heterogeneity information of HSI. Thirdly, based on dictionary update, a MKDL algorithm is proposed for learning a over-completed dictionary from the spatial structure information of HSI, and the learned dictionary can enhance the representation ability to nonlinear structure information of HSI. Depending on online dictionary learning, the proposed MKDL has the ability to process large-scale HSI. The outcomes of this research will promote the application of sparse interpretation technology for HSI classification task with refined interpretation requirement. Last but not least, the project will further promote the application of HSI in the fields of urban construction areas, agriculture areas and forestry areas.

多核字典学习旨在构建具有完备性和判别性的字典，充分挖掘高光谱图像中非线性结构信息的稀疏性，以提升对复杂场景精细化解译的性能。面向当前稀疏解译技术难以满足复杂场景高光谱图像处理研究需求的挑战，本项目拟构建基于多核字典学习的高光谱图像分类框架。首先，构建基于多种核模式的多核字典学习算法，以实现算法对多样性非线性结构信息的有效挖掘。其次，探究空间邻近效应对多核字典学习算法的影响，充分挖掘局部空间同质与异质性信息，提出空间联合多核字典学习算法。第三，构建基于字典更新的多核字典学习方法，充分挖掘高光谱图像蕴含的空间结构信息学习获得完备的字典，以增强对非线性结构信息的稀疏表示能力，同时借助在线字典学习技术，拓展多核字典学习算法对大范围星载高光谱图像的处理能力。研究成果将有利于促进稀疏表示技术在高光谱图像精细化分类中的应用，进一步推动高光谱遥感在复杂城市建筑区域、农业、林业等领域的行业应用。

面向当前稀疏解译技术难以满足复杂场景高光谱图像处理研究需求的挑战，本项目充分地挖掘高光谱图像多样性的非线性结构信息、局部空间邻域信息并构建完备字典，系统地构建了基于多核字典学习的高光谱图像分类框架，提出了基于多种核模式的多核字典学习算法、空间联合多核字典学习算法和基于字典更新的多核字典学习算法，研究成果将有利于促进稀疏表示技术在高光谱图像精细化分类中的应用。依托本项目在重要国际刊物、国际会议及国内重要期刊上发表论文合计7篇，研制1套高光谱图像多核稀疏表示与分类系统，培养研究生4名。