Fast and Accurate Sub-Pixel Displacement Estimation from Optical Satellite Images Using a New Hyper-Realistic Earthquake Database and U-Net Architecture

Estimating the ground displacement from non-rigid registration of two optical satellite images, separated from hours to months, is key in the study of natural disasters such as earthquakes. Compared to standard image registration and flow estimation tasks, a key challenge here lies in resolving very small displacements (typically cm- or m-scale) with sub-pixel accuracy and precision using coarser image resolutions (e.g. 15 m for Landsat-8). Traditional block matching/sliding window methods, employing local windowed correlation techniques, are unable to reduce the effects of long-wavelength noise arising from differences in image lightning, vegetation, or acquisition artifacts. By using both local and global scales, fully convolutional deep learning registration models (U-nets) are potentially able to better resolve ground displacements, less affected my multi-scale noise. Yet, no labelled database exists for ground deformation. Here, we develop a new synthetic database of 50,000 realistic satellite image pairs containing simulated earthquake displacements, along with their ground truth displacement maps, which are used to train state-of-the-art fully convolutional deep learning models (U-net). The inference shows good preliminary results, with a fast computation time (less than 1 second for a 256 × 256 image).