A new deep-learning approach for the sub-pixel correlation of optical images in the near-field of earthquake ruptures

Abstract

Precise estimation of ground displacement from correlation of optical satellite images is fundamental for the study of natural disasters. In the case of earthquakes, characterizing near-field displacements around surface ruptures provides valuable constraints on the physics of earthquake slip. Recently, image correlation has been used to investigate the degree of slip localization, and how it may vary as a function of geological parameters (such as fault structural maturity), raising the possibility that slip localization (vs distribution) may be predictable, with important implications for seismic hazard assessment. Current sub-pixel correlation methods (frequency or spatial domain) all rely on the same general approach: they work at a local scale, with small sliding windows extracted from a pair of co-registered satellite images acquired at different times, and they assume a rigid uniform shift between the two correlation windows. However, in the near-field of fault ruptures, where the correlation window spans the fault discontinuity, this hypothesis breaks down, and may bias the displacements. Additional smoothing associated with the correlation window further complicates the interpretation of sharp features in the displacement field, artificially shifting displacement to the off-fault region. We developed a U-net-based method to solve the sub-pixel displacement estimation problem at a global scale. Such architecture is able to retrieve full scale surface displacement maps, making use of both global and local features, and potentially tackling different noises of the input images. We trained our model with real satellite acquisitions, warped with ultra-realistic synthetic displacement maps representing realistic faults. The model exhibits promising preliminary results, showcasing its capability to retrieve full-scale surface displacement maps with high accuracy. While direct comparisons with other state-of-the-art approaches (COSI-Corr and MicMac) are pending, our findings suggest that our proposed U-net-based approach has the potential to compete or even outperform these correlators.