On the Value of Graph Attention Network for Interpretable Drought Forecasting

Machine learning is now widely used for environmental forecasting. Although predictive skill often varies only modestly across architectures, interpretability remains a persistent challenge, reducing transparency and limiting stakeholders’ ability to understand model behavior, build trust, and apply forecasts in practice. Balancing accuracy and interpretability are therefore essential for scientific credibility and real-world decision-making. In this context, Graph Attention Network (GAT) is particularly promising. Graph representations encode spatial dependencies and capture complex non-Euclidean relationships, such as upstream–downstream hydrological connectivity or large-scale teleconnections, that conventional grid-based models often struggle to represent. Attention mechanisms then adaptively weight information from different neighbors, helping the model focus on the most informative signals while offering a transparent view of which connections drive each prediction. Here, we evaluate the transferability and representational capacity of GAT for soil-moisture drought forecasting by modeling hydrological response units (HRUs) as nodes in a soil-moisture interdependence graph that preserves connectivity between locations. Beyond predictive accuracy, our analyses show that the model learns stable, physically meaningful relationships and yields interpretable hydrological insights. Feature-importance results reveal consistent links between key predictors and drought dynamics across both space and time. Attention diagnostics indicate pronounced seasonality: weights respond to the relative variability of source-node inputs, producing alternating dominance of high- and low-elevation sources between winter and summer. Spatially, the model consistently prioritizes same-elevation connections, suggesting that it internalizes distinct hydrological regimes in its learned representation. We also highlight three ongoing efforts: 1) extending evaluation to additional climatic regions to test transferability; 2) exploring hybrid GAT–sequence architectures to better capture temporal dynamics, while carefully assessing potential trade-offs in systematic, physically meaningful interpretability; and 3) developing an easy-to-use, open-source codebase to support broader use and reproducibility.