高光谱遥感表示模型与分类器动态集成方法

There are many open issues in hyperspectral remote sensing classifier ensemble classification, such as classifier diversity, multiple features, small sample etc. As the new machine learning algorithms, representation model (sparse & collaborative representation, collaborative graph and lower-rank representation) provides new tools for hyperspectral imagery dynamic classifier ensemble. In this project, the efficient dynamic classifier ensemble and high-reliable classification approaches for hyperspectral remotely sensed imagery based on representation models are proposed. the sparse representation and collaborative representation classifiers with Tikhonov regularization is proposed based on representation model, specifictly, the representation model classifier set is designed. Furthermore, the classifier diversity evaluation model, which taken intrinsic geometrical structure of pixel relationship into consideration, is proposed based on Laplacian regularized collaborative graph model for classifier dynamic ensemble. Considering the multiple feature representation, the low-rank subspace representation method is proposed for hyperspectral subspace feature extraction, and a dynamic ensemble method based on multiple features is designed. Consider small size samples situations of hyperspectral images, classifier dynamic selection algorithm based on adaptive shape for samples spectral-spatial features is proposed. With the support of classifier diversity, multiple features, and sample selection, the high-relieble hyperspectral image classifier dynamic ensemble classification system can be built. The experiment results of this project are expected to provide a new method and technique for hyperspectral image classification, and will expand the development of ensemble learning technology for hyperspectral remote sensing, and promote the applications of hyperspectral remote sensing.

高光谱遥感分类器集成在分类器差异性、多特征利用、小样本约束等方面仍存在许多问题；新型机器学习算法如协同表示、协同图和低秩表示等，可为高光谱遥感分类器动态集成研究提供新工具。本项目拟基于表示模型理论，以“分类器层动态集成—特征层动态集成—样本层动态集成”为主线，研究高光谱遥感分类器动态集成分类方法。基于表示模型分类器机理，研究吉洪诺夫正则化稀疏/协同表示分类器，提出基于拉普拉斯正则化协同图的分类器差异性评价模型及其引导的动态集成方法；考虑多特征，研究基于低秩子空间表示的高光谱遥感子空间特征表达方法，建立多特征驱动的分类器动态集成方法；针对小样本问题，研究基于协同表示的自适应形状样本邻域提取方法，提出样本邻域信息约束的分类器动态集成方法；实现分类器差异性引导、多特征驱动、小样本约束的高光谱遥感分类器动态集成分类。项目成果有望拓展高光谱遥感分类理论方法，推动高光谱遥感集成学习的发展与深入应用。

针对高光谱遥感分类器集成分类中存在的分类器静态固定集成、分类器差异性评价难、特征利用匮乏、小样本等问题，课题组在协同表示和集成学习理论的基础上，以“分类器层动态集成—特征层动态集成—样本层动态集成”为主线，研究了高光谱遥感分类器动态集成分类方法。项目基于表示模型分类器机理，研究了吉洪诺夫正则化稀疏/协同表示分类器，提出了基于拉普拉斯正则化协同图的分类器差异性评价模型及其引导的动态集成方法；考虑多特征的差异性，研究了基于低秩子空间表示的高光谱遥感子空间特征表达方法，建立了多特征驱动的分类器动态集成方法；针对小样本问题，研究了基于协同表示的自适应形状样本邻域提取方法，提出了样本邻域信息约束的分类器动态集成方法；实现了分类器差异性引导、多特征驱动、小样本约束的高光谱遥感分类器动态集成分类方法。本课题提出的方法具有较好的性能，进一步拓展了高光谱遥感影像分类的理论方法，提升了高光谱遥感分类精度的稳定性和可靠性，为高光谱遥感的深入应用奠定了坚实理论基础。围绕项目研究，共发表学术论文21篇，其中在ISPRS Journal of Photogrammetry and Remote Sensing, IEEE Geoscience and Remote Sensing Magazine，IEEE Transactions on Geoscience and Remote Sensing等国际权威期刊发表SCI论文16篇，在《遥感学报》等国内核心期刊发表研究论文4篇，在国际会议发表论文1篇，培养博士生3人，独立培养硕士生13人，授权国家发明专利1项、软件著作权1项。