Detection of linear deformation features in arctic landfast ice using InSAR and deep learning

Interferometric Synthetic Aperture Radar (InSAR) provides a powerful tool for observing landfast-ice deformation. Linear deformation features (LDFs) are narrow structures associated with fracture development in landfast ice. However, studies on the automated detection of LDFs from InSAR data remain limited. In this study, a framework integrating differential InSAR and deep learning–based semantic segmentation is proposed to detect LDFs from Sentinel-1 InSAR data. A Unet-based semantic segmentation method, termed LDF-UNet, was developed for detecting sparse, narrow, and elongated LDFs from InSAR coherence imagery. Dice loss was incorporated to improve the segmentation of spatially sparse LDF regions. Experimental results show that the proposed LDF-UNet improves deformation feature detection compared with several conventional segmentation networks, with a maximum improvement of approximately 14% in the F1-score. LDF-UNet achieves an overall accuracy of 0.97 and an F1-score of 0.63 for LDF detection. Experiments across multiple Arctic regions demonstrate the robustness of the proposed method and enable large-scale mapping of LDFs at 60 m resolution. In the Nares Strait region, detected LDFs correspond well with fracture boundaries associated with the 2017 ice-arch breakup and appear several days before collapse, indicating that LDFs detected from InSAR may serve as early indicators of landfast-ice instability.