Diffusion‐Based Reconstruction of Western North Pacific Surface Winds From CYGNSS for Tropical Cyclone Analysis

Abstract

Sea surface wind retrieval under extreme weather conditions is constrained by the sparse and irregular spatiotemporal sampling of Global Navigation Satellite System Reflectometry (GNSS‐R). This study presents a probabilistic wind field reconstruction framework for the Western North Pacific using Score‐Based Data Assimilation (SDA). A diffusion model trained on historical wind analyses serves as a statistical prior, and sparse Cyclone Global Navigation Satellite System (CYGNSS) observations constrain the reverse diffusion process to draw samples from an approximate posterior. Track‐segment withholding (80% input, 20% verification), additionally excluding verification cells containing input observations, yields a case‐averaged Root Mean Square Error (RMSE) of 1.637 m/s. In Tropical Cyclone (TC) scenarios, for cases with analysis‐box CYGNSS coverage above 10%, the reconstruction reduces the inner‐core intensity bias of both gridded references (the fifth‐generation European Centre for Medium‐Range Weather Forecasts reanalysis by 78.4%, and the Cross‐Calibrated Multi‐Platform analysis—its own training prior—by 70.1%) with lower RMSE than the reanalysis but slightly higher RMSE than the multi‐platform analysis. Furthermore, the framework provides informative pixel‐level uncertainty estimates that are under‐dispersed relative to withheld CYGNSS observations, with the spread correlating with the reconstruction error. These results indicate improved mean intensity estimates in the higher‐coverage subset, with performance still dependent on inner‐core sampling.