Evaluating typhoon cloud morphology as a candidate early satellite proxy for storm-surge risk under ERA5 physical benchmarking

Storm surge is driven by wind, pressure, tide, bathymetry, and coastal geometry; satellite cloud structure is therefore evaluated here as a candidate proxy for cyclone organization rather than a direct forcing. We matched Hong Kong Observatory residual-surge records for 72 tropical-cyclone events (2015–2025) with Himawari-8/9 B13 cloud metrics at 24, 12, and 6 h before the observed surge peak. Seventy-one events had complete sequences, including 19 events with peak residual surge \(\ge 0.70\) m. Under leakage-free event-grouped cross-validation, the seven-variable cloud Transformer showed no reliable stand-alone discrimination (ROC-AUC = 0.518, 95% CI 0.380–0.673; permutation \(p = 0.4209\) ) and yielded station RMSE = 0.420 m (95% CI 0.284–0.574). An untrained ERA5 wind-stress/pressure rank index achieved AUC = 0.853 and Spearman \(\rho = 0.741\) , while the ERA5 wind-pressure baseline achieved AUC = 0.860 and station RMSE = 0.313 m. Adding seven cloud variables changed ERA5 event AUC by \(-0.005\) (95% CI \(-0.023\) to 0.012). Cloud morphology provides reproducible observational descriptors, but in this sample it does not support a validated stand-alone early-warning model and is best treated as a contextual diagnostic alongside atmospheric and hydrodynamic guidance.