面向多视角卫星的三维城市场景理解与动态监测研究

In China, along with the rapid urbanization, cities are developing vertically to expand the city space without consuming limited land resources. The large-scale, high-precision three-dimensional (3D) monitoring of the urban environment has become an urgent task. Thanks to the launch of Zi Yuan 3 (ZY-3) domestic stereo mapping satellite system, high-resolution stereo and multispectral urban images can be acquired and used as a major data source of the 3D city monitoring. Limited by the common non-stereo remotely sensed imagery, however, conventional studies, including feature extraction, scene understanding, and change detection, mostly focus on constructing two-dimensional (2D) models and extracting 2D information. Studies about 3D information extraction and stereo scenes understanding are therefore urgently needed. By breaking through the traditional 2D image interpreting system, this study aims to establish a 3D framework of feature representation and spatiotemporal monitoring on the basis of ZY-3 stereo imagery. The framework is developed with the “geometry-feature-scene” as the key point, and the proposed methods include accurate 3D reconstruction based on geometry information, 3D morphological feature representation framework, 3D primitive co-occurrence matrix, multi-level stereo grid representation, transfer learning of stereo scenes, etc. The framework established in this study not only breaks through the limitations of 2D remote sensing information processing in theory, but, in practice, provides new technical support and solutions for the 3D city expansion, smart cities, air quality monitoring, livability assessment, commercial potential evaluation, population-economic analysis, the disaster simulation and early-warning, etc.

随着我国城市化水平的迅速提高，城市正在朝垂直方向发展，从而扩展空间，应对土地资源的紧张。我国资源三号(ZY-3)立体测绘双星的成功发射和组网，为大尺度、高精度的城市立体化监测提供了数据支持。然而，当前遥感影像的特征提取、场景理解和变化监测大多采用二维模型，缺乏影像三维信息挖掘与立体场景理解的研究。针对这一问题，本课题拟在ZY-3多视角卫星的数据支持下，突破传统的二维影像解译框架，建立三维的遥感影像特征表达与动态监测体系，以“几何-特征-场景”为主线，包括ZY-3多视影像的几何-信息精确三维重建、三维形态学特征表达框架、三维基元共生矩阵、多级立体格网、三维场景迁移学习等原创性方法。本课题所构建的三维场景理解框架，在理论上突破了二维遥感信息处理的局限，在应用上，为城市三维扩张、智慧城市建设、空气质量监测、城市宜居评价、商业潜力估计、人口-经济分析、灾害模拟预警等提供新的技术支撑和解决方案。

本项目旨在为高分辨率三维场景识别、分析与监测提供新的理论方法，为高分辨率遥感在智能计算、景观规划、地理国情、城市环境等领域提供基础和关键技术。经过4年的研究，项目组基于国产ZY-3、GF-7多视角卫星数据，开展了基于多视角卫星特征提取、机器学习、城市应用等研究。主要完成情况如下：.1）在特征提取方法上，提出了多角度特征ADF、多角度张量纹理、多角度房屋指数等原创性算法。.2）针对多视角卫星的观测特点，提出了融合多光谱与多角度的多分支3D卷积网络，实现精确的城市土地/场景分类；提出了M3Net，利用ZY-3多视角卫星，估计了我国42个主要城市的房屋高度，并测试了该模型向国外区域的迁移效果。.3）制作了多个有影响力的遥感数据产品，包括：我国50个城市的高分辨率三维地表覆盖数据集ULCM，全球多个城市的高分辨率居民区数据集，1972~2020全球30米不透水面数据集GISA（已被下载6万次），1980~2020我国30米土地覆盖数据集CLCD（已被下载11万次）等。.4）城市遥感应用：在高分辨率2D/3D城市热岛解译、全球城市热岛分析、城市2D/3D规划与新冠确诊、城市规划与城市噪声、城市规划与空气污染等方面开展了创新性研究。. 在本项目支持下，课题组已在RSE、TGRS、ISPRS Journal、Science China Earth Sciences、Science Bulletin等国内外著名刊物上发表SCI论文38篇；参加了2018西班牙IGARSS、2019日本IGARSS、2020年SPIE、2020年中国城市遥感大会等国内外会议；2018~2021年，项目组共有8位博士生毕业，17位硕士生毕业。

内容获取失败，请点击重试