利用SVD卷积神经网络的多极化时序SAR非监督变化检测方法研究

Time series SAR remote sensing images can truly represent the dynamic changes of surface features over a long period of time. There are some significant drawbacks in most existing methods for SAR timing analysis: i) It is easy to generate a lot of false alarms and missed detections by using the decomposition processing — “multi-difference images—multi-thresholds” with low efficiency; ii) The comprehensive treatment methods ignore the correlations between the polarimetric scattering characteristics of SAR images at different times, resulting in easily-missing detections for changes of ground objects with weak scattering features; iii) Due to the influence of SAR inherent speckle noises, there are still a lot of isolated spots in the generated difference image (DI), which easily produce false alarms. Deep learning technology can more abstractly and better expresses SAR images with speckle noises. But there also faces some serious problems, i.e., large amount of data and less training samples. In order to solve the aforementioned problems, this project will focus on the researches about three aspects: i) a generalized mathematic model of the DI generation method for multi-polarimetric SAR time series analysis; ii) an automatic method for high-quality training samples selection based on the combination of GGMM-MRF and the similarity measurement from time series SAR; iii) time series SAR change detection with high precision based on SVD networks. Through the achievement of those three critical technologies, the time series multi-polarimetric SAR change detection method will be formed to realize the complete detection of the dynamic changes of the observed surface features within the time series. And the developed method will provide scientifically and technologically support for many applications, such as land monitoring, urban dynamic change monitoring, crop growth status monitoring, and disaster assessment, etc.

时间序列SAR影像能够较为真实地反映地表地物在一个长时间范围内的动态变化情况。现有SAR时序分析方法中1）分解处理方式采用“多差异图——多阈值”易产生检测虚警和漏检、且效率低；2）综合处理方式忽略了不同时相极化散射特征的相关性，导致弱散射地物之间的变化易漏检；3）受SAR固有相干斑噪声的影响，生成的差异图中仍存在野点，易产生伪变化。深度学习技术能够更抽象、更好地表达具有相干斑噪声的SAR影像。但利用深度学习进行时序分析通常面临着数据量大、训练样本较少的问题。为解决上述问题，本项目将分别开展多极化通用模型的时序SAR差异图生成方法、基于GGMM-MRF和相似性测度统计分布的高质量训练样本自动选择方法、基于SVD卷积神经网络的时序SAR变化信息高精度提取方法等关键技术研究，从而形成多极化时序SAR变化检测技术，为国土监测、城镇动态变化监测、农作物生长状况监测、灾害评估等应用提供科学技术支撑。

全球变化、土地利用/覆盖变化、城市发展、自然灾害频发等关系到人民生活和人类社会可持续发展的一系列问题，集中起来可归结为地表地物的动态变化问题。时间序列SAR影像能够较为真实地反映地表地物在一个长时间范围内的动态变化情况。现有SAR时序分析方法常采用时间序列中连续两两影像对比分析的方式，易产生误检和漏检、且效率低；而且，受SAR固有相干斑噪声的影响，易产生误检。深度学习技术通过系列非线性组合能够更好地学习表征SAR影像特征，有利于克服相干斑噪声的影响。为此，本项目以“时序差异图计算”——“差异图分析”——“时序变化检测”为研究思路，突破了多极化通用模型的时序SAR差异图生成方法、基于GGMM-MRF的高质量训练样本自动选择方法、基于SVD卷积神经网络（SVD-CNN）的时序SAR变化信息高精度提取方法等关键技术研究。并且，利用位于武汉地区的2个典型区域的两组Radarsat-2、高分三号（GF3）时序全极化SAR数据进行算法验证，结果表明本项目提出的相似矩阵Shannon熵时序差异图计算方法能够实现时间序列差异图的一次性整体计算，产生的差异图能够有效反映地表覆盖的整体变化情况；同时，结合变化地面参考图，证明了本项目GGMM-MRF算法自动选取的变化类和未变化类样本的有效性和可靠性；并验证了本项目SVD-CNN时序变化检测方法的检测性能优于传统两两影像对比分析的时序变化检测方法、综合相似矩阵Shannon熵和MRF的时序变化检测方法以及Omnibus检测方法。项目研究成果进一步丰富了遥感影像地表覆盖变化监测的手段和技术，为自然资源调查监测、土地利用/覆盖变化、农作物生长状况监测、灾害评估等应用提供有力的技术支撑。