Deep Learning-Based Coseismic Deformation Estimation From InSAR Interferograms
Chuanhua Zhu, Xue Li, Chisheng Wang, Bochen Zhang, Baogang Li
Abstract
Accurate automated extraction of coseismic deformation from Synthetic Aperture Radar (SAR) data can be challenging owing to interference from inherent atmospheric noise. Particularly, the limited displacement of small-to-moderate earthquakes (Mw<6.5) can easily be obscured by phase errors and/or noise. To address this issue, we developed an autoencoder model based on a deep learning framework (i.e., Pytorch) to automate the accurate extraction of coseismic displacement from Interferometric SAR (InSAR) interferograms. We constructed a training dataset using simulated interferograms. Our trained model performed well for interferograms with real noise. When applied to worldwide real earthquakes of various rupture styles, the model produced clear coseismic displacement with less noise and a better fit to coseismic fault models compared to the differential InSAR method without noise correction. Additionally, it achieved co-seismic deformation similar to popular InSAR time series and GNSS methods. The approach will enhance the proceduralization and popularization of InSAR applications in earthquake monitoring, providing improved constraints on the kinematic characteristics of earthquakes.