UNet-SAR: Deep Learning for Simulating Synthetic Aperture Radar Sea Surface Roughness in Northwest Pacific Western Boundary Region

Abstract

This study presents UNet-SAR, a deep-learning framework based on a U-shaped network (UNet) encoder–decoder architecture for simulating sea-surface roughness represented by vertical–vertical (VV)-polarized normalized radar cross section (NRCS) under complex sea states in the Northwest Pacific western boundary current region. A total of 5213 dual-polarized VV/vertical–horizontal (VH) Sentinel-1 (S-1) synthetic aperture radar (SAR) scenes acquired during 2024–2025 were collocated with hourly hindcast fields from the Coupled Ocean–Atmosphere–Wave–Sediment Transport (COAWST) model on a 0.05° grid. Independent validation against in situ observations from National Data Buoy Center (NDBC) and National Marine Data and Information Service (NMDIS) buoys yielded root-mean-square errors (RMSEs) of 2.01 m/s, 0.51 m, and 0.15 m/s for wind speed, significant wave height, and surface-current speed, respectively, with corresponding correlation coefficients (r) of 0.90, 0.95, and 0.81. The dataset was partitioned at the scene level into 3213 training scenes and 2000 independent test scenes. Shapley additive explanations (SHAP) identified nonBragg roughness and wind speed as the dominant predictors, supporting the physical consistency of UNet-SAR. On the independent test dataset, UNet-SAR achieved an RMSE of 2.32 dB and an rof 0.80, outperforming the theoretical model incorporating nonBragg roughness (RMSE = 3.44 dB, r= 0.63) and achieving accuracy comparable to extreme gradient boosting (XGBoost; RMSE = 2.34 dB, r= 0.78). Despite their comparable overall statistical performance, UNet-SAR better preserved localized roughness variations and spatial continuity than XGBoost. Evaluation at output grid spacings of 500 m, 1 km, 3 km, and 5 km showed that coarser grids slightly improved pointwise statistics through speckle suppression, whereas finer grids better retained spatial structural similarity and localized NRCS variability. Overall, UNet-SAR provides a physically interpretable framework for generating spatially coherent synthetic NRCS fields, with potential applications in ocean monitoring and data assimilation.