Dynamic Spatio-Temporal Fire Pressure Modelling for Short-Term Wildfire Forecasting

Accurate short-term wildfire forecasting is essential for effective early warning, operational planning, and resource allocation. This study proposes the Dynamic Spatio-Temporal Fire Pressure Model (DST-FPM), a leakage-controlled forecasting framework that integrates wildfire memory, spatial connectivity, cumulative fire pressure, and seasonal variability using historical satellite-derived active fire detections. The framework combines an Active Cell Framework (ACF), Dynamic Fire Pressure (DFP), the Fire Connectivity Index (FCI), Five-Day Fire Pressure (FFP), and the Operational Fire Risk Pressure (OFRP) index within an Extreme Gradient Boosting (XGBoost) model to predict wildfire occurrence over three-day and five-day forecasting horizons, with the five-day horizon adopted as the primary operational scenario. The methodology was evaluated across Bosnia and Herzegovina, Croatia, and Montenegro using 3,591,054 grid-cell-day observations collected between January 2020 and December 2025. Independent chronological training, validation, and testing datasets were combined with temporal, spatial, and spatio-temporal validation procedures to assess model robustness. For the primary five-day forecasting horizon, the proposed framework achieved a ROC AUC of 0.773, a PR AUC of 0.147, a balanced accuracy of 0.678, and a Matthews correlation coefficient of 0.135 on the independent testing dataset, while maintaining stable performance across all validation procedures. The fitted XGBoost model consistently assigned high predictive importance to the proposed fire pressure indicators, while Top-K analysis showed that 13.5% of future wildfire occurrences were identified within only 1% of the highest-priority grid-cell-day observations. These findings indicate that integrating wildfire memory, spatial connectivity, cumulative fire pressure, and seasonal variability provide complementary predictive information for short-term wildfire forecasting while preserving interpretability, robustness, and operational applicability.