Deep Learning Enhanced Fidelity InSAR Toolkit (DEFIT)

Lynntech and Southern Methodist University (SMU) Earth Sciences propose to leverage deep learning plus data resources to develop a tool useful for enhancing Interferometric Synthetic Aperture Radar (InSAR) interferograms. InSAR interferogram stacks with available ground truth of the phase due to ground deformation will be used to train a generative model that can be applied to other datasets. Automated spatialndash;temporal analysis of InSAR stacks to yield a high fidelity estimate of surface deformation remains as a challenging problem. nbsp;Spatio-temporal analysis tools for InSAR stack processing already exist on various platform, however are mainly the purview of researchers. With the advent of both increased quantity and quality of InSAR sources there is a dual need to (1) handle the big data problem of very fast revisit InSAR that covers the entire globe, and (2) make products more accessible to decision makers and industry. Existing methods and tools used by earth scientists to detect and mitigate the atmospheric anomaly that effects the time of flight of backscattered radar generally yield results with compromised fidelity and often require further interpretation or, if possible, correction that involves heuristics or incompletely modeled dependencies. Instead we will minimize multiple loss functions, inferred from the statistical properties of the training set, to train a generative network to reconstruct the deformation map without the atmospheric effect with high fidelity. The input would consist at least of the stack of raw interferograms and an initial estimate of the deformation map (done with new or existing tools) and will be modular to be able incorporate other information, e.g. weather data. The proposed work includes the development of a Deep learning Enhanced Fidelity InSAR Toolkit (DEFIT) that will be trained and tested on relevant datasets. This can be used for the development of new tools or to augment capabilities of existing tools.