表示模型框架下高光谱遥感影像分类若干技术研究

Sparse representation theory has been drawn more and more attention in hyperspectral image classification due to its ability to describe useful information of the image with only few samples and capture the differences among spectral signatures. Recently, sparse representation has been found to have the following drawbacks: it overemphasizes L1-norm penalty while ignoring the collaboration between samples; moreover, it cannot exploit the statistical structure of the data and satisfy the demand of smaller variance within groups while larger variance between groups in classification tasks. In this proposal, based on global and local texture information, a combination of sparse and collaborative representations is proposed. The proposed representation-based classification (RC) framework reveals the consistent coexistence of competition and collaboration. Main research contents include RC with texture features, structured RC, RC with multiple constraints, and ensemble RCs based on residue fusion. This project is expected to effectively impel the development of hyperspectral image analysis in theory and promote the applications of hyperspectral remote sensing.

稀疏表示理论用较少的数据量描述影像中的特征信息并捕捉不同地物之间光谱曲线的差异性，在高光谱遥感影像分类中受到越来越多的关注。随着研究的不断深入，稀疏表示在高光谱数据分析中的应用逐渐暴露了它的缺陷：由于稀疏表示过于强调L1范数约束项，忽略了数据间协作表示的重要性；同时，在分类的过程中，不能较好的反映数据集的内在分布结构，无法充分满足约束表示系数具有小的类内方差和大的类间方差等需求。本项目结合高光谱影像全局结构特征和局部纹理特征，提出基于特征空间的稀疏表示和协作表示相结合、竞争与协作关系共存的分类相关理论和技术。主要内容包括：1）研究高光谱遥感影像的空间纹理特征提取方法；2）构建结构化表示模型分类器；3）提出一种多约束条件的表示模型分类方法；4）开展基于重构残差融合的表示模型多分类集成研究。本项目成果将有力地推动高光谱遥感影像处理分析的基础理论研究的发展，促进高光谱遥感应用的深入。