The ability to predict wildfires - such as those that recently devastated Los Angeles and Canada - is advancing rapidly with the help of ML–driven high-quality data. A new paper, published today (Tuesday 1 April, 16:00 BST | https://www.nature.com/articles/s41467-025-58097-7 ) in Nature Communications, highlights how the collection and integration of higher-quality data can significantly improve the accuracy and reliability of wildfire predictions.
The paper evaluates how ECMWF's new data-driven fire danger forecasting model, the Probability of Fire (PoF) , performed in 2023 and in recent extreme events. ECMWF has been producing fire danger forecasts since 2018 as part of the Copernicus Emergency Management Service (CEMS) led by the Joint Research Centre of the European Commission. In recent years, it has developed innovative approaches using machine learning methods. This effort has moved ECMWF from predicting fire danger – a measure of landscape flammability – to forecasting fire activity. These new products are distributed to the Copernicus Emergency Management Service, and accessible to ECMWF Member States.
"Wildfire prediction as a field of research has been active for decades, which led to the establishment of early warning systems in the 1970s," explains lead author Dr Francesca Di Giuseppe. "Our new Probability of Fire model incorporates multiple data sources beyond weather to refine predictions. Thanks to a machine learning algorithm, it takes a more holistic approach. Traditional weather-based fire danger indices often fail to pinpoint areas at risk of ignition with enough specificity. This is where ML can help."
She explains that in the case of the Los Angeles fires, for example, traditional weather forecasts identified broad areas as very flammable but did not accurately target the most probable ignition hotspots. However, by incorporating additional parameters beyond weather - such as human presence, development indices such as road density, and most importantly vegetation abundance and its dryness, Dr Di Giuseppe says by being able to spot the most likely area to experience ignition this helps avoid over-predicting and provides a more accurate and targeted fire risk. "Being able to add all these elements with ML helps refine predictions. For example, we can exclude areas that are hot and dry but unlikely to experience ignition - either because people are not present or there is no fuel to burn," states Di Giuseppe.
This comprehensive approach has proven to be more accurate in identifying fire-prone areas, as evidenced by the Southern California fire danger forecast for 7 January 2025. In this event, leveraging this Probability of Fire model was able to provide a far more localised and accurate assessment of high fire danger in areas where fires occurred than the traditionally used Fire Weather Index. By accounting for additional parameters, the model captured the intricate dynamics that drive fire risk.
Florence Rabier, Director-General of ECMWF, comments: "As witnessed in recent years with some devasting fires, like those in Portugal, Greece and Canada, improving fire forecasts through better data and AI integration will be a game-changer in the years to come. The new Probability of Fire tool has benefited from ECMWF's expertise in AI and ML for medium-range weather predictions (3 to 15 days), and the experts involved have made significant advances in fire prediction using similar data-driven methods. Although fire prediction is a challenging subject as ignition remains an unpredictable process, agencies in charge of providing information, like the European Commission's Joint Research Centre, have now access to improved tools to help better protect lives, livelihoods, and ecosystems."
The Los Angeles fires provide a stark example of the increased wildfire threat. The period leading up to the fires in 2024 saw unusually wet conditions facilitating rapid vegetation growth, followed by an exceptionally dry autumn and early winter. This pattern, known as 'hydroclimate whiplash', is being amplified by climate change. In LA, it created an abundance of dry and flammable vegetation, creating the perfect conditions for the catastrophic fires.
"Understanding these patterns is crucial to accurate fire prediction," says Joe McNorton, another ECMWF expert contributor to the study. "In this research, we found that high-quality data, such as information on vegetation moisture and fuel availability, is the most important factor for improving forecast accuracy. The Probability of Fire model's ability to capture these changes such as the 'whiplash' effect demonstrates how ML-driven models are increasingly crucial for accurate wildfire predictions and emergency preparedness."
The study found that one of the most critical elements in fire prediction is knowledge about fuel availability. Including all data sources improves up to 30% the model's predictive skill from only including weather. Fuel information is unlike weather data in that it is not easily obtained through direct observations or prediction systems. However, ECMWF has used its weather forecasting expertise and data from the EU's Copernicus Atmospheric Monitoring Service (CAMS) to construct a modelling framework to extract this critical information.
The research findings also reveal that the integration of ignition sources and fuel status is more critical for fire prediction than complex ML algorithms. This insight provides an opportunity for centres with less computational capacity to implement their own predictive systems, provided they have access to reliable data for training.
Lead Author Francesca Di Giuseppe concludes: "By focusing on high-quality data, even smaller agencies can implement effective fire prediction systems. This is a key takeaway for the global community in our ongoing efforts to combat wildfires."
With wildfires becoming increasingly frequent and severe, driven by climate change, ECMWF's work is positioning Europe at the forefront of ML-driven natural hazard prediction. As wildfires continue to intensify globally, this research emphasizes the importance of better data integration and AI tools in forecasting, helping to reduce the catastrophic impacts of these extreme events.
ECMWF's work in fire prediction is funded by the EU Commission's Joint Research Centre (JRC), which coordinates emergency preparedness for fires across Europe.
Scientists from the European Commission' s Joint Research Centre (JRC), which is in charge of fire danger prediction services for wildfires in Europe, concluded that: "The integration of AI and data-driven models is key for the advancement of fire prediction systems across Europe. ECMWF and the JRC share a commitment to using the best available science and technology to protect lives and manage the growing wildfire threat in an increasingly volatile climate."
