基于时序InSAR协方差矩阵估计的城镇化场景分类方法研究

In last two decades, the great development of radar remote sensing provides the excellent opportunities for land cover mapping at larger scale. However, the relatively few studies investigate the use of SAR datasets over richly textured areas where heterogeneous land covers exist and intermingle over short distances. One of main difficulties is that the shapes of the structures in a SAR image cannot be represented in detail as mixed pixels are likely to occur when conventional SAR parameter estimation methods are used. To solve this problem and further extend previous research into remote monitoring of urban environments, this project will develop a methodology to refine SAR-based input variables and simultaneously preserve the resolution of SAR images, by fully exploring multi-temporal InSAR covariance matrix. The research will prove that in contrast with classifiers that are frequently investigated in optical remote sensing, the classification accuracy of SAR instruments is dependent on the quality of the input variables and spatial resolution. We further extend this conclusion from single-polarimetric SAR to quad-polarimetric SAR. Based on accurate estimation of the SAR features, an information fusion will be finally presented to resolve the ambiguities for land cover classification in the built environment by taking radiometric, interferometric, spectral, temporal and spatial-contextual signatures into account. The results will reveal that this methodology can break through the bottleneck of previous studies on land cover mapping from resolution preservation, surface structures and classification reliability. The technology to be developed is providing a more comprehensive perspective of radar remote sensing research. The project also provides a strong scientific basis for the maintenance of native ecosystems and sustainable land use and therefore is of both significant scientific and practical values.

近年来，合成孔径雷达（SAR）系统的迅猛发展为大尺度地表分类研究提供了契机。然而这一新兴遥感技术在城镇化复杂场景的分类精度并不高，主要原因是当前的研究思路多从改善分类器的角度出发，未兼顾SAR图像分辨率和参数估计精度，因而无法准确描述精细化的结构特征。为拓展SAR技术在城镇环境的应用，本项目在时序InSAR协方差矩阵估计的框架下，论证保持特征层的分辨率和参数估计精度是决定SAR图像分类精度和可靠性的关键，并推广至全极化SAR，以获取高精度、全分辨率的极化参数。针对城镇化场景地貌复杂性以及SAR有限的辐射分辨率，项目将进一步融合时间信息和光学资料，以增强信息量，并顾及空间上下文背景信息，力求在维持空间分辨率、地表几何和分类可靠性方面取得突破，解决当前SAR技术在城镇化场景分类中精度低的难题。项目成果将为城市雷达遥感提供更为全面的视角，也为土地表层系统空间分布和精细化场景研究提供新的探测手段。

针对城市飞速发展的背景下，面向智能化城市规划、国土资源调查、未来土地变更预测和环境治理等方面的迫切需求，本项目从探讨不同类型地物目标在时间序列 SAR 影像中的散射机理及统计特性入手，提出时序 InSAR 协方差矩阵估计方法，将当前主流的SAR分类理论基础由空间扩展到时空一体化求解上来，通过融合SAR 辐射特征、全极化参数、InSAR 相干性、光学谱特征、时间、空间-上下文(Spatial-contextual)信息等，在空间分辨率、地表几何和分类可靠性方面寻求最优解，从而提升当前基于SAR的空间对地观测技术在城镇化以及复杂场景地表分类的监测精度。项目资助相关成果已发表在国际和国内权威刊物上，共发表国内外高水平期刊论文 4篇，包括 SCI 论文 3 篇和 EI 论文1 篇，申请（授权）专利3项，获奖2项。本研究对现代InSAR数据处理理论的完善和时序InSAR技术可靠性的提高具有理论价值和科学意义。本项目的研究成果有利于提升对复杂地貌的精细化场景理解，促进国土资源管理精准化水平和国家治理体系和治理能力现代化，服务和支撑国民经济和社会发展，并为揭示人文景观的空间格局提供更为全面的视角。