Global Short-term Daily Wildfire Forecasting and Predictability Attribution using a new Spatio-temporal Deep Learning Framework

Reliable short-term wildfire forecasting is essential for early warning, timely air-quality management, and mitigating wildfire-related health impacts and economic losses. However, global prediction remains difficult because wildfire occurrence is rare and highly heterogeneous across fire regimes. The Fire Weather Index (FWI) is widely used as a benchmark, but it mainly reflects weather-driven fire danger and does not explicitly represent fuel or fire dynamics, limiting predictive accuracy. Physics-based coupled models can resolve fire–atmosphere interactions, yet they typically require prescribed ignition information and are too computationally expensive for global deployment. Data-driven methods enabled by satellite and reanalysis data offer an efficient alternative. However, many conventional ML approaches treat grid cells as independent samples, which limits learning of neighborhood interactions and multi-day preconditioning. Recent DL studies improve representation learning, but many remain regional and lack unified spatiotemporal dependency modeling. Thus, global spatiotemporal frameworks tailored to the rare and sparse nature of wildfire occurrence remain scarce.

Here we present the STA-Net, a novel global daily wildfire forecasting framework built on a harmonized multi-source dataset spanning 2013–2024. The dataset integrates meteorology, vegetation, lightning, and topography information on a unified 0.5° global grid. Through modeling of spatiotemporal dependencies and imbalance-aware training, The STA-Net learns coherent features that capture multi-day environmental preconditioning and neighborhood-driven fire evolution, which enables accurate next-day wildfire forecasts at the global scale. It also supports short-range forecasts at 1–7 day lead times, although predictive skill decreases progressively as lead time increases.

The STA-Net outperforms the FWI and representative data-driven baseline models, including XGBoost (non-spatiotemporal), LSTM (temporal-only), and 2D-CNN (spatial-only). On an independent global test set, the STA-Net achieves an AUC of 0.97 and maintains stronger discrimination than FWI across all 14 GFED fire regions. Two 2024 case studies in Bolivia and Canada further show that the STA-Net captures the spatial footprint and concentrated high-risk cores of catastrophic outbreaks, supporting event-level generalization beyond aggregate metrics. Using F1 as the primary rare-event metric, the STA-Net achieves the highest score among the data-driven baselines (F1 = 0.65). An ignition–spread–persistent (I–P–S) stratification attributes the largest improvement to spread fire, where neighborhood propagation is central, providing direct evidence for the effectiveness of the STA-Net’s spatiotemporal modeling.

Beyond forecasting, we perform predictability attribution across fire types and regions. SHAP analyses under an IPS stratification show that persistent fire prediction is dominated by prior fire states, spread fires depend on coupled fuel–environment conditions, and ignition is driven mainly by vegetation and land-surface properties with a stronger role of soil moisture. Region-aggregated attribution further indicates that FRP and NDVI are consistently influential predictors, while secondary drivers vary by region and fire regime, with meteorological controls shifting in importance and lightning density contributing more strongly in regions with frequent lightning-driven ignitions.

Overall, the STA-Net provides a high-skill and scalable approach for global short-term daily wildfire forecasting together with transparent attribution of predictive drivers, supporting wildfire risk management and emission forecasting.