Identifying Fault-Related Deformation in Large InSAR Datasets Using Transfer Learning

The rapid expansion of SAR archives (e.g., Sentinel-1) offers new opportunities to detect fault-related deformation signals, such as weak coseismic-like patterns and interseismic shallow fault creep. However, their low signal-to-noise ratio and coherence variability pose challenges for large-scale screening. To address this, we develop a transfer learning–based convolutional neural network for automatic recognition of fault-related deformation in single wrapped interferograms. Built on SarNet and fine-tuned with globally curated Sentinel-1 data, the model achieves 92.25% validation accuracy (AUC 0.9689) with a low false-alarm rate. Cross-mission tests on ALOS-2 (L-band) and ENVISAT (C-band) confirm strong portability (accuracy 87.24%, precision 95.54%, recall 78.13%, F1 85.96%, AUC 0.9686), though recall decreases due to sensor-specific domain shifts. Grad-CAM visualizations indicate that the classifier relies on meaningful interferometric fringes rather than noise artifacts, supporting physical interpretability. These results show that transfer learning enables robust and scalable automated recognition of fault-related deformation, thereby supporting systematic, large-scale mining of multi-sensor InSAR archives.