Deep learning-based surface deformation tracking with interferometric fringes: A case study in Taiwan

• A deep learning model (FLM) detects deformation, supporting an FDM model that classifies displacement from interferograms. • The FDM showed strong transferability, detecting deformation across different times and regions using a single training area. • The approach enables scalable, continuous monitoring by automatically analyzing new D-InSAR images for evolving deformation. Monitoring surface deformation is critical for understanding and mitigating natural and anthropogenic hazards, such as landslides and subsidence. Although Multi-Temporal Interferometric Synthetic Aperture Radar (MT-InSAR) provides detailed displacement measurements, its application in continuous monitoring remains constrained by high computational demands and complex data processing, often interrupting observation continuity. To address these challenges, this study proposes a deep learning-based method that processes wrapped Differential InSAR (D-InSAR) interferograms to directly detect surface deformation patterns . A Fringe-Labeling Model (FLM) was developed to identify deformation regions, followed by a Fringe-Detection Model (FDM) using Faster Region-based Convolutional Neural Networks (Faster R-CNN) to classify deformation magnitudes. The method achieved an average mean Average Precision (mAP) of 83.9% in Central Taiwan. Temporal transferability was validated by detecting deformation one year beyond the original MT-InSAR observation period. Spatial transferability was confirmed by applying the model to Northern Taiwan, where an F1 score of 78.74% was achieved while effectively identifying both uplift and subsidence. By enabling deformation detection across different magnitudes, time periods, and regions, the proposed framework offers a scalable and transferable solution for extending MT-InSAR-based surface hazard tracking.