Explainable global wildfire prediction model using graph neural networks

• Correlation-defined graph encodes climate-fire co-variability and long- range dependencies • Addresses structural limitations of convolutional models for global wildfire prediction. • GCN–LSTM model produces 12-month global burnt-area forecasts. • Outperforms convolutional baselines on 30 years of withheld JULES- INFERNO data. • Interpretability analysis links accuracy gains to long-range message passing mechanisms. Wildfire activity is shaped by long-range dependencies and teleconnections that couple distant regions. However, existing image- and convolution-based prediction models, constrained by limited effective receptive fields, exhibit structural limitations for global-scale wildfire prediction. This study reformulates global burned-area forecasting as graph-based sequence learning. Specifically, instead of geographic adjacency, we construct a correlation-defined graph where edges encode climate–fire co-variability, enabling explicit long-range message passing. On this graph, we couple graph convolutions for spatial encoding with LSTM modules for temporal dynamics in a sequence-to-sequence architecture for 12-month forecasts. Evaluated on 30 years of unseen JULES–INFERNO ensemble data, the proposed approach consistently outperforms convolutional baselines across all metrics and shows reduced degradation under distribution shifts associated with strong climate anomalies. Interpretability analysis links the observed gains to the graph-based mechanism, revealing patterns consistent with established teleconnection behaviour in the climate–fire literature. Overall, the results demonstrate that embedding a teleconnection-aware inductive bias materially improves global burned-area prediction and provides a scalable alternative to convolutional formulations for spatio-temporal wildfire forecasting.