From Open Ocean to Estuary: Adapting Pre-Trained Ocean Graph Neural Networks for Estuarine Marine Heatwave Forecasting

Marine heatwaves (MHWs) threaten shallow estuarine systems that sustain critical coastal biodiversity, yet forecasting these events at sub-seasonal to seasonal lead times remains an open challenge due to complex local drivers. Graph neural networks (GNNs) trained on global open-ocean data can deliver skillful sea surface temperature (SST) forecasts at extended lead times. We propose to investigate here whether this capability can be transferred to estuarine environments where in-situ data data are limited to sparse monitoring locations across varying depth profiles. A preliminary case study transferring a pre-trained global ocean GNN to a New Zealand fjord shows the approach is promising outperforming persistence and climatology baselines. However, several limitations persist, namely underprediction of MHW peaks. We propose to address these limitations through depth-aware message passing, marine boundary forcing, and a coupled land-sea GNN that injects atmospheric and terrestrial drivers into estuarine nodes. The outcome would be an operational, transferable early-warning capability for estuarine heatwaves---a direct input to aquaculture management, and coastal ecosystem adaptation under climate change.