Recursive Estimation and Prediction of Earth Deformation from SAR Image Time Series – REPED-SARIX

The systematic acquisition of and free access to Sentinel-1 A/B Synthetic Aperture Radar (SAR) images provide scientists with both opportunities and challenges for operational monitoring of Earth deformation by SAR image time series. In this project, the coordinator brings together, for the first time, Interferometry SAR, statistical learning, deep learning and geophysics skills in order to form an interdisciplinary team who aims to promote methodological development for operational Earth deformation monitoring and natural hazard prediction by SAR image time series. For this, we start from single look complex SAR image time series and deal with displacement estimation and natural hazard prediction problems by means of both statistics based learning and neural networks based learning approaches. First, we develop a recursive and robust multi-temporal InSAR method, allowing for efficient gradual integration of new arriving SAR images and considering non Gaussian properties of SAR image statistics, to estimate displacement velocity and displacement time series. Moreover, we propose a complete and original missing data imputation framework for SAR displacement time series based on statistical learning and deep learning. Second, we tackle, for the first time, the major issue of neural networks based physical parameters inversion and prediction from SAR derived displacement time series. We propose recursive neural network models, adapted to SAR displacement data specificity, in both supervised and semi-supervised learning frameworks. Prior physical knowledge will be incorporated into the neural networks in order to enhance the learning process and to improve the interpretability and explainability. A particular effort will be made on the understanding of the neural network functioning in order to ensure the accountability and then the actionability of the results for operational use. All previously developed methods will be applied to targets of geophysical interest, including volcanoes covered by Sentinel-1 A/B SAR images every 6 or 12 days and Alpine glaciers covered by Sentinel-1 A/B images every 6 days and by high resolution PAZ images every 11 days. The expected results consist of 1) advanced open access multi-temporal InSAR methods, providing complete and reliable displacement time series in line with the routine availability of SAR image time series 2) open access recursive neural network models with both linear and nonlinear functionalities, adapted to SAR displacement data specificity, providing temporal evolution of key geophysical parameters in line with the routine availability of SAR image time series 3) a dataset tutorial illustrating in detail the input and output of previously developed methods, together with open access to the datasets used for illustration.