高光谱遥感图像的张量低秩表达与非监督学习方法研究

Hyperspectral image records the signal from both of the spatial and spectral domains and with high feature dimensionality in spectral bands, therefore, the conventional pixel based and spectral channel based remote sensing feature representation and processing methods could not well suit for the information extraction tasks any more due to their common limitation of low order signal processing. In this research, we propose to target the key issues in hyperspectral remote sensing image information extraction based on tensor representation and tensor analysis, and the project includes the following three packages: (1) the local tensor low rank representation for hyperspectral remote sensing image, (2) the discriminative multi-linear projection for tensor observations, and (3) the unsupervised tensor learning towards remote sensing knowledge. The expected achievements of this project could overcome the drawback of the current vector based approaches which neglect the structure of spatial and spectral features, thus would greatly improve the accuracy and robustness of the hyperspectral remote sensing image analysis algorithms, which have important theoretical significance and practical value in many aspects including geoscience information updating, military target detection, urban planning, and resource surveying.

高光谱遥感图像图谱合一、波段数目巨大，传统的逐像素光谱特征分析、逐波段空间特征提取等基于低维信号处理的方式难以有效解译，无法实现自适应数据表达的智能化高光谱遥感图像信息提取。本研究拟利用张量分析理论，围绕高光谱遥感图像的张量信息提取中的关键性问题展开研究，主要内容包括：（1）高光谱遥感图像的局部张量低秩表达；（2）张量流形空间中的判别多维线性投影；（3）从张量表达到遥感信息的非监督抽象模式。项目研究成果能从根本上克服基于向量处理方式割裂了光谱-空间特征的缺陷，提供自适应数据表达的可扩展性高光谱图像信息提取模型，显著提高高光谱遥感图像的解译精度和鲁棒性，为更智能化的高光谱图像研究打下基础，在地理信息更新、军事目标识别、城市规划、资源勘查等方面具有重要的应用价值。

高光谱遥感图像图谱合一、波段数目巨大，传统的逐像素光谱特征分析、逐波段空间特征提取等基于低维信号处理的方式难以有效解译，无法实现自适应数据表达的智能化高光谱遥感图像信息提取。本研究利用张量分析理论，围绕高光谱遥感图像的张量信息提取中的关键性问题，研究了高光谱遥感图像的局部张量低秩表达方法、张量流形空间中的判别多维线性投影方法以及从张量表达到遥感信息的非监督抽象模式等内容，显著提高高光谱遥感图像的解译精度和鲁棒性。项目资助发表高水平论文26篇，其中SCI 1区论文4篇、2区论文8篇，CCF A类会议论文10篇，ESI高被引论文8篇，入选国际学术会议优秀论文2篇。本项目的研究成果在地理信息更新、军事目标识别、城市规划、资源勘查等方面具有重要的应用价值。