DeepOT: A Deep Learning Framework for Pixel-Level Ground Surface Displacement Estimation From SAR Amplitude Imagery

Abstract

Monitoring ground surface displacement is critical for understanding geophysical processes and mitigating natural hazards, yet conventional Synthetic Aperture Radar (SAR) techniques are often limited by decorrelation, complex terrain, and heterogeneous motion. We present deep learning based-offset tracking (DeepOT), an adaptable deep-learning framework for estimating pixel-level ground surface displacement directly from SAR amplitude image pairs. The framework is enabled by a synthetic-to-real training strategy in which controlled displacement fields are embedded into real SAR imagery, allowing large-scale supervised training without reliance on ground truth displacement measurements or offset-tracking-derived labels. We evaluate DeepOT using multiple deep-learning models and apply it to contrasting landslide settings, including the Slumgullion landslide in Colorado and the Barry Arm landslide in Alaska. The framework supports time-series displacement construction and is evaluated using independent extensometer measurements at Slumgullion. Results show that DeepOT recovers spatially coherent displacement patterns under challenging conditions where interferometric coherence is limited and conventional offset tracking is sensitive to surface heterogeneity. Qualitative comparisons in earthquake case studies further indicate applicability to large-scale, high-gradient deformation. DeepOT is designed as a modular and extensible framework, providing a foundation for future advances in data-driven SAR-based displacement monitoring.