Authors: Hamza Kaddour, Baya Cherif, Marouane Temimi – Stevens Institute of Technology
Title: Deep Learning-Based Segmentation of Sentinel-1 SAR satellite imagery for River Ice detection with Weak Optical Supervision
Presentation Type: Poster Presentation
Abstract: Collaborations with the APRFC and field campaigns to develop data-driven streamflow and stream temperature models for Alaska
Abstract: Reliable river-ice mapping is critical for hydrological monitoring and mitigating ice-jam flood risk in high-latitude regions. Sentinel-1 SAR enables all-weather observation, yet operational products such as NOAA SARRIS remain largely threshold-driven and can degrade during freeze-up and break-up. We present a SAR-only river ice segmentation framework trained with weak optical supervision: river ice labels are automatically derived from Sentinel-2 Scene Classification Layer (SCL) and rigorously aligned to Sentinel-1 geometry via date coupling, reprojection, and Sentinel-1 footprint clamping. A U-Net model is trained using Sentinel-1 VV/VH backscatter and polarization-derived features. Across multiple Arctic/sub-Arctic river reaches and seasons, global percentile normalization and a VV-VH difference channel improve robustness, increasing mean accuracy to 0.736 and reducing Water to Ice misclassification from 50.52% to 35.43%. The proposed approach demonstrates systematic improvements over SARRIS.