高光谱遥感影像特征学习-地物分类一体化建模

Hyperspectral remote sensing images, which have continuous spectrum, are able to finely recognize objects with high spectral diagnosis ability. This proposal aims at investigating the challenging problems in the classification of hyperspectral remote sensing images, which includes the space-time variety of spectrum, the complex spatial configurations of objects and the limited training samples. More precisely, this project will develop an integrative and robust model which combines the feature learning and object classification for hyperspectral remote sensing images under the theoretical support of representation learning. The main line of our research is from mining high-level intrinsic features to sparse coding based image classification via spatial-spectral fusion of superpixels, and finally to integrative coupling of feature learning and object classification. In particular, the three aspects for our main ideas are as follows: (1) on the basis of the massive hyperspectral image data, mining high-level but intrinsic feature representation in a data-driven fashion via unsupervised deep learning methods; (2) based on the spatial neighborhood homogeneity, building sparse representation based classifier for hyperspectral image on superpixel level; (3) with studying the intermodulation mechanism between low-level features and high-level semantic information of objects, proposing an united model which can be self-updating, self-learning and integrative coupling for feature mining and object classification. Our project proposes to explore and develop new theories and methods for information extraction in hyperspectral remote sensing images. The research findings can effectively promote the application of hyperspectral remote sensing image, and has significance both in theoretical and industrial application.

高光谱遥感影像光谱连续、图谱合一，能够以较高的光谱诊断能力对地物目标进行精细化识别。针对高光谱遥感影像分类中面临的光谱时空多变、地物空间分布复杂、训练样本数量有限等难点问题，本项目拟在表达学习理论框架下，以“特征深度挖掘—稀疏表达分类—联合模型一体化耦合”为研究主线，发展统一、稳健的高光谱影像特征学习与地物分类一体化模型与方法。其主要思路为：充分利用高光谱遥感影像的海量特性，以数据驱动的方式非监督深度学习影像高层抽象本征特征与表达；在影像空间邻域均质性驱动下，构建超像素水平的高光谱遥感影像稀疏表达分类器；研究影像特征与地物信息之间的双向互调机制，实现特征挖掘与地物分类联合模型的自更新、自学习与一体化耦合。本项目旨在研究和发展高光谱遥感影像信息提取的新理论和新方法，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。

高光谱遥感影像光谱连续、图谱合一，能够以较高的光谱诊断能力对地物目标进行精细化识别。针对高光谱遥感影像分类中面临的光谱时空多变、地物空间分布复杂、训练样本数量有限等难点问题，本项目拟在表达学习理论框架下，以“特征深度挖掘—稀疏表达分类—联合模型一体化耦合”为研究主线，发展统一、稳健的高光谱影像特征学习与地物分类一体化模型与方法。其主要思路为：充分利用高光谱遥感影像的海量特性，以数据驱动的方式非监督深度学习影像高层抽象本征特征与表达；在影像空间邻域均质性驱动下，构建超像素水平的高光谱遥感影像稀疏表达分类器；研究影像特征与地物信息之间的双向互调机制，实现特征挖掘与地物分类联合模型的自更新、自学习与一体化耦合。本项目旨在研究和发展高光谱遥感影像信息提取的新理论和新方法，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14篇，其中SCI期刊论文11篇， EI检索论文2篇，北大中文核心期刊论文1篇，获批国家发明专利2项，申请国家发明专利3项，荣获国际性竞赛冠军1项、省部级奖励2项；学术交流方面，参加国际学术会议12人次，国内学术会议44人次；人才培养方面，项目负责人破格晋升教授、入选教育部“长江学者奖励计划”青年学者，培养毕业博士生3名、硕士生12名，在读博士生7名、硕士生17名，指导学生荣获全国性竞赛奖励2项、校级奖励4项。