A Deep-Learning Neural Network for Postseismic Deformation Reconstruction From InSAR Time Series

作者：Chenglong Li, Xi Xi, Guohong Zhang, Xiaogang Song, Xinjian Shan · 发表于：IEEE Transactions on Geoscience and Remote Sensing · 年份：2024 · DOI：10.1109/tgrs.2024.3389772 · 被引用次数：8 · 研究领域：earthquake and tectonic studies、Earthquake Detection and Analysis、Seismology and Earthquake Studies

Residual noise effect in Interferometric Synthetic Aperture Radar (InSAR) time series data typically complicates the characterization of deformation behaviors on tectonic faults, especially postseismic deformation caused by multiple physical mechanisms. To reconstrue denoised postseismic deformation signal in noisy InSAR time series data, we proposed a deep-learning-based neural network and trained it on noisy InSAR time series dataset, which was generated by adding Sentinel-1 InSAR time series noise onto synthetic postseismic deformation. The application results on experimental dataset validated that our neural network efficiently removed residual noise and automatically recovered clean postseismic deformation in noisy InSAR time series with any space and time sizes. Moreover, when applied to real InSAR time series data following the 2021 Maduo earthquake, our network provided accurate reconstructions of postseismic deformation in both time and space, demonstrating its generalization performance on actual data. In comparison to the original data, the denoised postseismic deformation exhibits greater consistency with the observations from continuous Global Positioning System (cGPS). Interestingly, our denoised postseismic slip reveals spatial juxtaposition between the geometrically complex coseismic rupture segments and the heterogeneity of the early afterslip along the Maduo rupture. Overall, these successful applications highlight the network’s generalization performance to...