Global Data-Driven Prediction of Fire Activity

In recent years, newly available observations, and modelling systems as well as advancements in machine learning have transformed the capabilities of fire danger prediction systems. The European Centre for Medium-Range Weather Forecasts (ECMWF) has set out to forecast wildfire probability on a global scale up to a week in advance. A key milestone was the development of the SPARKY-Fuel Characteristics dataset, released in 2024, which provides the first long-term, high-resolution record of real-time fuel status.

This study evaluates ECMWF’s operational data-driven fire prediction system over its first year. Through analysis of major wildfire events, including the extensive fires in Canada in 2023 and the fires in Los Angeles in 2025, we demonstrate the potential of data-driven methods to outperform traditional fire danger metrics. The results highlight the role of dynamic, global fuel assessments and machine learning in improving the accuracy and timeliness of fire probability forecasts.

Our findings underscore the importance of integrating both innovative data-driven approaches and key variables into operational forecasting systems, providing critical support for fire management and mitigation efforts worldwide.