Detection of dislocation-like signals in raw SAR images with deep learning

Over the last decades, synthetic aperture radar (SAR) images and SAR interferometry (InSAR) have revolutionized Earth observation, allowing for geophysical monitoring of Earth surface processes with centimeter-to-millimeter precision. Accurate measurement of ground displacement is essential for the comprehension of natural hazards, such as earthquakes, and the detection of the smallest ground (transient) displacement is of uttermost importance to better image the dynamics of active faults, especially in tectonic contexts that undergo low deformation rates. However, detecting small deformation signals in raw SAR images remains a significant challenge because of the significant noise level affecting the data (e.g., speckle noise, tropospheric and ionospheric perturbations). Multiple and successful InSAR mass processing methods, including state-of-the-art noise correction methods, have been developed over the last decade, but all rely on intensive computing of massive databases, a tedious procedure that cannot be applied yet at a global scale. Furthermore, because of the low probability of finding earthquakes in intraplate continental settings, automatic detection of such signals in such settings is currently out of the question with InSAR data.

Here, we leverage deep learning to enhance the detection of deformation (e.g., dislocation-like signals) directly from raw SAR images. Our deep-learning-based approach offers the potential to (1) retrieve potential deformation below the noise threshold, thus improving sensitivity, and (2) precisely localize regions of interest from full acquisitions to serve as input for InSAR pipelines, reducing the need to process entire datasets and significantly accelerating computation. Also, deep learning methods can process large-scale images much faster, enabling the creation of dense and extensive detection catalogs for subsequent analysis.