基于结构稀疏模型的高光谱影像亚像元级分类和超分辨率制图技术研究

The young scientist fund of NSFC (Grant No. 60902070) chaired by the applicant has intensively studied the key problems of hyperspectral imagery, including denoising, feature extraction/selection, unmixing and classification, through probabilistic model-based nonnegative matrix factorization, graphical model-based affinity propagation clustering, Gabor wavelet transform, Memetic hybrid search, etc. Several papers have been published on international top journals and conferences, among which, six are indexed by SCI and fifteen indexed by EI..The project mainly focuses on structured sparse model-based hyperspectral imagery sub-pixel classification and super-resolution mapping techniques. Based on the investigation of model construction and optimization algorithms of structured sparsity, the structured sparse expression mechanism of hyperspectral imagery is examined. The structured sparse model-based sparse transformation/decomposition, structural feature selection/extraction, structure sparse classification approaches are developed. The co-expression and analysis mechanism between pixel classification and spectral unmixing are established, implementing joint structured sparse sub-pixel classification. Further, multi-modal structured sparse super-resolution mapping is accomplished by integrating the complementary information among spectra, the relevant information and a priori knowledge of space. Then systematic structured sparsity theory and methods can be formulated, which could provide new theory and tools for the sub-pixel classification and super-resolution mapping of hyperspectral imagery.

申请人主持的青年科学基金项目(编号：60902070)通过基于概率模型的非负矩阵分解、基于图模型的仿射传播聚类、Gabor小波变换、Memetic混合搜索等方法对高光谱影像的去噪、特征提取/选择、解混、分类等关键问题进行了比较深入的研究，在国际权威期刊和会议上发表论文多篇(SCI收录6篇，EI收录15篇)。.本项目重点研究基于结构稀疏模型的高光谱影像亚像元级分类和超分辨率制图技术。在探讨结构稀疏的模型构建及优化算法基础上，研究高光谱影像的结构稀疏表达机理，设计基于结构稀疏模型的稀疏变换/分解、结构特征选择/提取、结构稀疏分类方法，建立像元分类与光谱解混之间的联合表达和分析机制，实现联合结构稀疏的亚像元级分类；进一步融入光谱间的互补信息、空间的相关信息和先验知识，实现多模态结构稀疏的超分辨率制图。形成比较系统的结构稀疏理论和方法，为高光谱影像的亚像元级分类和超分辨率制图提供新的理论和工具。

近年来，高光谱成像技术得到了学术界和工业界的广泛关注。相比于全色、彩色和多光谱影像，高光谱影像数据提供了更加丰富的地物特征信息，为地物的精确分类和识别提供了重要的数据基础。但同时也带来了维数高、波段相关性强、计算复杂度高的难题，高光谱影像处理为传统模式识别和计算机视觉技术提出了新的挑战。本项目针对高光谱影像内在的结构稀疏特性，重点研究基于结构稀疏表达的高光谱影像数据的特征提取/选择、识别和分类等技术，形成比较系统的结构稀疏理论和方法，为高光谱影像数据处理提供新的分析工具。本项目取得的成果包括：1）研究了高光谱影像的结构稀疏表达模型，提出了基于空谱信息融合（超像素）的高光谱影像稀疏分类、基于低秩稀疏表示的高光谱影像图割，和基于密度排序聚类算法的高光谱影像波段选择等方法。2）研究了面向Gabor特征的高光谱影像结构稀疏模型，提出了基于协同表示和学习的高光谱影像分类和基于条件互信息的Gabor特征选择等方法。3）研究了高光谱影像的多任务稀疏模型，提出了基于Fisher特征选择的多任务联合稀疏表示分类和基于卷积神经网络的高光谱影像分类等方法。项目组在IEEE地球科学与遥感汇刊等顶级学术期刊和会议上发表论文20余篇，在高光谱影像信息处理领域具有一定的国际学术影响力。