Presentation by Dr. Jules Bourcier.
Characterization of Strike-Slip Fault Offsets Using Deep Learning and Simulated Topographic Data
Abstract:
Tectono-geomorphic features, such as offset river channels and ridges, are typically used to estimate the surface displacements from past earthquakes along strike-slip faults. Offsets measurements are essential for constraining earthquake rupture models which inform seismic hazard assessments. Traditional offset measurements require extensive manual or semi-automatic analyses, which can be time-consuming and inconsistent. We present a deep learning approach for automatically quantifying strike-slip fault offsets from high-resolution topographic data. To address data scarcity, we generate a semi-synthetic dataset by simulating faulting and landscape evolution, enabling effective pretraining of our neural network. Thanks to a simulated-to-real transfer learning strategy, the obtained model reasonably approximates semi-automatic methods. These initial findings show promise for efficient, automated fault characterization.
