One Backbone, Two Flood Hazards: Cross-Basin CNN Forecasting of Tropical Cyclone Rainfall and Storm Surge

Foundation weather models have improved medium-range synoptic forecasting, but tropical cyclone rainfall and coastal storm surge are still conventionally modelled through separate pipelines, despite sharing the same atmospheric driver. We test whether a shared ResNet-50 backbone with two task heads can support both hazards from one storm-centred ERA5 input, where the heads differ only by a bathymetry channel for surge. We train the model on six years of Pacific typhoons and benchmark it against two independent per-task networks. On held-out Pacific storms it reaches $R^2 = 0.82$ for 3-hour accumulated rainfall and $R^2 = 0.90$ for 1-hour peak significant wave height (SWH), matching the baseline at half the parameter count. We then evaluate both heads on 18 Atlantic 2024 hurricanes, a basin excluded from training. Both retain $R^2 \geq 0.80$ and exceed the independent baseline by 0.08-0.09 on each task. SHAP attribution maps show the two heads attend to physically distinct regions. The rainfall head depends on inner-core specific humidity and eyewall winds. The SWH head depends on lower-tropospheric wind fetch and bathymetry. This indicates that transfer is driven by portable physical mechanisms rather than Pacific-specific patterns. The framework trains on a single GPU using freely available ERA5 and IBTrACS data.