基于多任务协同自表达学习的高光谱遥感非监督地物识别方法

Hyperspectral remote sensing imagery can acquire nearly continuous spectra of ground materials, which supports fine object recognition thanks to its high spectral discriminant capability. Faced to the difficulties of obtaining prior knowledge for objects in the remote no-man area, this project aims at developing a robust and efficient unsupervised hyperspectral remote sensing object recognition algorithm based on multi-task collaborative self-representation, within the representation learning framework. The main research line of this project includes joint self-representation modeling, multi-task collaborative learning, and multi-feature fusion clustering. The main idea is to construct joint sparsity and low-rank self-representation model to extract the local and global structure information of the hyperspectral pixel set simultaneously. We then build multi-task collaborative self-representation learning model to overcome the limits of the spectral separability of different land materials by using the method of multi-feature collaborative learning, in order to efficiently obtain the self-representation coefficients and weights of multi-features to accurately describe the relative contributions of multi-features and to accurately separate the hyperspectral pixel set. The project aims at exploring and developing new theories and methods for hyperspectral remote sensing information extraction to give technical supports for solving the difficulties of fine object recognition in the remote no-man area without any prior knowledge. The research results can effectively promote the applications of hyperspectral remote sensing image, which has significance both in theoretical and industrial application.

高光谱遥感影像可以捕获地表近乎连续的光谱曲线，能够以较高的光谱诊断能力对目标地物进行精细化识别。针对边远无人区地物先验信息难以获取的难题，本项目拟在表达学习框架下，以“联合自表达建模—多任务协同学习—多特征融合聚类”为研究主线，发展稳健、快速的高光谱遥感影像多任务协同自表达非监督地物识别方法。其主要思路为：构建稀疏低秩联合自表达学习模型，同时挖掘高光谱遥感像素点集的局部、全局结构信息；建立多任务协同自表达模型，以多特征协同学习的方式，突破不同地物的光谱区分度限制；快速求解多元特征的自表达系数、权重因子，准确表征多特征的相对贡献，实现高光谱像素集的准确划分。本项目旨在研究和发展高光谱遥感信息提取的新理论和新方法，为有效解决边远无人区无先验信息条件下的精细化地物识别难题提供技术支撑，其研究成果可以有效提升高光谱遥感影像的应用潜力，具有重要的理论与应用意义。

高光谱遥感影像可以捕获地表近乎连续的光谱曲线，能够以较高的光谱诊断能力对目标地物进行精细化识别。针对边远无人区地物先验信息难以获取的难题，本项目拟在表达学习框架下，以“联合自表达建模—多任务协同学习—多特征融合聚类”为研究主线，发展稳健、快速的高光谱遥感影像多任务协同自表达非监督地物识别方法。其主要思路为：构建稀疏低秩联合自表达学习模型，同时挖掘高光谱遥感像素点集的局部、全局结构信息；建立多任务协同自表达模型，以多特征协同学习的方式，突破不同地物的光谱区分度限制；快速求解多元特征的自表达系数、权重因子，准确表征多特征的相对贡献，实现高光谱像素集的准确划分。本项目旨在研究和发展高光谱遥感信息提取的新理论和新方法，为有效解决边远无人区无先验信息条件下的精细化地物识别难题提供技术支撑，其研究成果可以有效提升高光谱遥感影像的应用潜力，具有重要的理论与应用意义。.按照既定的研究计划，项目圆满地完成了课题设定的研究任务，实现了预期目标，并在研究的深度和广度上都比既定计划有了进一步的扩展。在项目资助下，项目组发表科研论文26篇，其中SCI期刊论文21篇，ESI热点论文1篇、高被引论文6篇，EI检索论文4篇，北大中文核心期刊论文1篇，获批国家发明专利2项，荣获国际性竞赛冠军1项、国际性竞赛亚军3项、省部级奖励1项；学术交流方面，参加国际学术会议9人次，国内学术会议15人次；人才培养方面，项目负责人当选英国工程与技术学会会士(IET Fellow)，2021-2022年连续入选斯坦福大学全球前2%顶尖科学家、Elsevier中国高被引学者和全球学者库前10万顶尖科学家等榜单，培养毕业博士生4名、硕士生13名，在读博士生8名、硕士生10名，指导学生荣获全国性竞赛奖励3项、校级奖励5项。