“学习型”干涉SAR单/多基线相位解缠方法研究

InSAR技术已被广泛应用于地面高程信息获取和地质灾害形变探测中。相位解缠是InSAR中的关键步骤直接关系到地表高程或形变提取的精度。然而，单基线相位解缠无法突破相位连续性假设和多基线相位解缠噪声鲁棒性差的两大难题，已经成为制约InSAR获取地面高程或形变信息的关键因素。随着国内外SAR卫星的迅猛发展，更多的InSAR数据为“学习型”相位解缠算法设计提供可能。本项目将深度卷积神经网络（DCNN）引入相位解缠，通过构建基于InSAR成像机理的干涉图数据集，将单基线相位解缠问题转换成DCNN可学的相位模糊数分类问题，建立面向相位模糊数分类的DCNN，利用数据驱动突破相位连续性假设；将多基线相位解缠问题转换成DCNN可学的截距项分类问题，建立面向截距项分类的DCNN，利用特征抽象能力增强算法的噪声鲁棒性。本项目的研究成果将为国产SAR卫星进行复杂地形测绘和地质灾害形变监测提供可靠的相位解缠方案。

InSAR technology has been widely applied to digital elevation model (DEM) generation and geo-hazard deformation analysis. Phase unwrapping (PU) is one of the key procedures in InSAR, which is directly related to the accuracy of the DEM estimation and deformation monitoring. However, single-baseline PU cannot violate the phase continuity assumption and multibaseline PU suffers poor noise robustness, which have become two key factors restricting the performance of the InSAR product. With the rapid development of domestic and foreign SAR satellites, more and more InSAR datasets make it is possible for transforming the model-based PU algorithm into the learning algorithm. This project will abridge the gap between deep learning and PU by introducing deep convolutional neural networks (DCNN) into the field of PU. To be specific, a massive datasets of interferograms based on InSAR mechanism will be established. Then, the single-baseline PU problem will be transformed into the ambiguity classification problem, and the DCNN used for ambiguity classification will be designed to break through the limitation of the phase continuity assumption due to its data-driven methodology. Also, the multibaseline PU problem will be transformed into the intercept classification problem, and the DCNN used for intercept classification will be designed to enhance the noise robustness owning to its feature generalization ability. The results of this project will provide reliable PU solutions to complex terrain mapping and geo-hazard deformation monitoring, especially for domestic SAR satellites.