Data matters most: improving fire activity forecasts through data-driven approaches

Recent advances in machine learning (ML) are transforming scientific applications, including weather and hazard prediction. In the context of wildfires, these methods enable a fundamental shift from forecasting fire weather conditions to predicting actual fire activity. In this study, we demonstrate that such a transition is not only feasible but can also be implemented in an operational forecasting framework.

Traditional fire danger indices, such as the Fire Weather Index (FWI), often overestimate risk, particularly in fuel-limited ecosystems, resulting in high false-alarm rates. By contrast, our data-driven approach, the probability of fire, integrates information on weather, fuel characteristics, ignitions, and observed fire activity to directly predict the probability of fire occurrence. This leads to substantial improvements in forecast reliability, reducing false alarms while maintaining sensitivity to high-risk conditions.

We show that model performance is driven more by the quality and completeness of input data than by the complexity of the ML architecture itself. In particular, fuel status emerges as the most critical predictor of fire activity. However, the lack of direct, real-time global observations of fuel remains a major limitation. To address this, we rely on physically based models to reconstruct fuel dynamics, highlighting the continued importance of process-based understanding in supporting ML applications.

Our results demonstrate that incorporating all components of the fire triangle, weather, fuel, and ignitions, can improve predictive skill by up to 30% compared to weather-only approaches. Furthermore, the probabilistic nature of the predictions enables direct verification against observed fire activity and opens new opportunities, such as reconstructing missing satellite detections and improving fire emission estimates.

Overall, this work underscores that meaningful progress in ML-based fire forecasting depends not only on algorithmic innovation, but critically on investment in high-quality, physically consistent datasets. Without this, the potential of data-driven approaches cannot be fully realised.