多源高分辨率遥感影像的多模态判别式稀疏学习与融合分类研究

Earth observation has entered the era of comprehensive high-resolution remote sensing, including that the multi-purpose application, multi-constellation cooperation and the multi-network collaboration for integrated service. Though high temporal-spatial-spectral resolution satellite images bring rich information to accurate, huge, and complex geographical patterns, it also raises new problems and challenges to the interpretation and fusion of multi-source images. The intelligent interpretation and fusion of high-resolution satellite constellation images is one of the most dynamic research topics in the field of remote sensing. This project is to research and develop robust interpretation methods for multi-source remote sensing images by adopting the latest theory in mathematics, statistics and optimization fields. Based on sparse representation classification paradigm, firstly, this project presents a uniform multi-source high resolution remote sensed data sparse representation framework to integrate the fusion and the interpretation phases. Secondly, aiming at importing the inter-class dissimilarity and the inter-source similarity to highlight the discriminability of training set, we designs a multi-modal discriminative class dictionary set learning algorithm in a compact way. Thirdly, by use of the correlation and distinctiveness of pixels in a spatial local region, we proposes the efficient object based multi-task joint sparse representation classification algorithm. The resulted algorithms can expand new theories and produce new methods for multi-source remote sensing image analysis, significantly enhance the potential applications of remote sensing data, therefore have important academic and practical significance.

对地遥感进入了一星多用、多星组网、多网协同、集成服务的综合观测时代，高分辨率卫星组网展现精细、海量、复杂的地理模式的同时，也对影像的智能化融合解译提出了新的问题和挑战。如何实现高分辨率多源卫星影像的智能化融合解译是国际遥感科学技术的前沿研究课题之一。针对这一问题，本项目拟引入数学、统计、最优化领域的最新理论，研究和发展多源高分辨率遥感影像的多模态判别式稀疏学习与融合分类理论与方法。基于稀疏表示分类理论，首先搭建多源高分辨率遥感稀疏融合解译框架，建立紧致多源判别性类别字典，实现超像素多源遥感影像稀疏表示与识别。本项目研究和发展多源遥感影像信息处理的新理论与新方法，可以大幅提升遥感数据的应用潜力，应对复杂多样应用需求、发挥多源卫星系统综合效能，具有重要的理论与应用意义。

对地遥感进入了一星多用、多星组网、多网协同、集成服务的综合观测时代，高分辨率卫星组网展现精细、海量、复杂的地理模式的同时，也对影像的智能化融合解译提出了新的问题和挑战。如何实现高分辨率多源卫星影像的智能化融合解译是国际遥感科学技术的前沿研究课题之一。针对这一问题，本项目在多源高分辨率遥感影像稀疏表示分类、判别式特征融合、多分类器集成等方面进行研究，开发原创性方法，并在全球城市多尺度产品制图与应用上取得了显著的进展。本项目部分方法和模型（如自动化房屋提取技术）在国家重点研发计划“基于国产遥感卫星的典型要素提取技术”得到了应用；环境监测成果已用于构建“深圳市生态安全监测系统”。在本项目的资助下，课题组共发表SCI论文9篇。依托本项目的部分研究成果，负责人获得2018年国家测绘科技进步一等奖(排名2)。负责人的研究被SCI和Google Scholar分别引用1000和1300余次；负责人担任国际著名SCI刊物Remote Sensing 和国产SCI期刊Geospatial Information Science(入选中国最具国际影响力学术期刊TOP5%)客座编辑，分别主持《多时序遥感监测》和《遥感图像信息提取》专刊。此外，项目组成员张依诺的研究获得2018年美国摄影测量与遥感协会John I. Davidson主席奖。