The European Centre for Medium-Range Weather Forecasts (ECMWF) has developed a machine learning system called the Probability of Fire (PoF) model that predicts where wildfires are likely to ignite by analyzing vegetation, human activity, and weather data. Unlike traditional danger indices that estimate the likelihood of fire-prone weather, this model focuses on actual fire activity—offering a more precise early warning tool for wildfire response and planning.
Detailed in a Nature Communications article published April 1, 2025, the PoF model uses a combination of satellite-based vegetation metrics, weather conditions, and ignition data to assess daily wildfire probability across the globe.
It was tested on historical wildfire data using several machine learning approaches, including neural networks and random forests. The team found that XGBoost consistently delivered the most accurate predictions. However, model choice was less important than the quality and completeness of the input data—especially data on vegetation abundance and ignition patterns. “By using data on fuel characteristics, ignitions and
observed fire activity, data-driven predictions reduce the false-alarm rate of high-danger forecasts, enhancing their accuracy. This is made possible by high quality global datasets of fuel evolution and fire detection.,” the study’s authors wrote.
Model Proves Effective in Real Wildfire Events
One of the PoF model’s strengths lies in its real-world validation. It correctly forecasted wildfire patterns ahead of the May 2023 fires in Canada, offering valuable lead time of up to ten days in some cases. The model also demonstrated precision during the Los Angeles wildfires in January 2025, outperforming traditional fire danger metrics that often overestimate risk in dry but fuel-limited regions.
The AI model, known as the Probability of Fire (PoF), correctly predicted where fires would break out during the January 2025 Los Angeles wildfires with more precision than traditional models.
Rather than sounding broad regional alarms based on temperature and wind alone, the PoF model evaluates how flammable the landscape is—and whether an ignition is likely to occur.
It identifies areas where vegetation is both dry and abundant, then overlays human and natural ignition data to refine its probability assessments. Rather than providing a fire danger rating, the model estimates the actual probability of fire occurrence.
The model has been operational since 2023 as part of ECMWF’s Copernicus Emergency Management Service, delivering daily updates that guide national agencies and emergency responders. According to ECMWF, incorporating high-resolution vegetation and ignition data improved predictive skill by up to 30% compared to weather-only models.
Forecasting Fire With Minimal Computing Power
Despite its accuracy, the PoF model is designed to be computationally efficient. It doesn’t rely on the supercomputers typically required for high-resolution weather simulations. Instead, it can run on relatively modest systems, making it accessible to smaller agencies or countries with limited infrastructure.
This accessibility opens the door for broader global adoption of advanced fire prediction tools without the burden of extreme hardware requirements.
At the core of the PoF framework is ECMWF’s SPARKY model—a vegetation monitoring tool that estimates live and dead fuel moisture content based on recent weather conditions and satellite data. According to ECMWF’s wildfire analysis blog, SPARKY plays a crucial role in determining whether vegetation is ready to burn, enhancing the reliability of fire risk predictions in both densely forested and scrubland areas.
Human activity is also a major component of the PoF model. It ingests data on population density, road networks, lightning activity, and other ignition proxies to simulate the probability of a fire starting—whether through a discarded cigarette, a spark from a powerline, or a lightning strike.
ECMWF’s Ecmwf’s AI Strategy Spans Wildfire, Weather, and Climate Forecasting
The PoF model is part of a broader AI push by ECMWF, which has become a key player in AI-enhanced forecasting. In 2024, ECMWF partnered with Google to launch NeuralGCM, a hybrid forecasting model that merges machine learning with traditional atmospheric physics. NeuralGCM demonstrated superior performance in tracking cyclones and reducing temperature and humidity forecast errors by up to 50%.
Later that year, Google DeepMind introduced GenCast, an ensemble-based AI system that generated 15-day forecasts in just minutes using a generative diffusion model. GenCast surpassed ECMWF’s own ENS system in 97.2% of scenarios tested.
ECMWF also contributed to the development of Aardvark Weather, a new model co-created with the University of Cambridge, Microsoft Research, and The Alan Turing Institute. Unlike physics-based simulations that require heavy computing resources, Aardvark bypasses those requirements entirely. It uses deep learning on real-time satellite and radar data, allowing the system to run on standard desktops—enabling accurate forecasts in regions with limited infrastructure.
AI Accuracy vs. Data Availability and Interpretability
Despite promising performance, AI models come with limitations. Their accuracy is only as good as the data they ingest. In regions lacking timely or detailed vegetation, ignition, or meteorological inputs, prediction quality may suffer. The PoF model, for instance, depends on up-to-date satellite feeds and ground station observations to maintain accuracy. In less-monitored environments, its advantages may diminish.
Another challenge is interpretability. Traditional physics-based models offer meteorologists clear physical explanations for their outputs, while machine learning systems often behave like black boxes. This can make it harder for analysts and policymakers to understand or justify decisions based solely on AI-driven probabilities.
The PoF model attempts to address some of these concerns by relying on well-understood variables—like fuel moisture and ignition likelihood—and making its outputs publicly available via the Copernicus Emergency Management Service. It is also designed to integrate feedback and improve over time as it accumulates real-world performance data.
As fire seasons grow longer and wildfire intensity increases, early-warning tools like PoF are becoming essential. With a flexible design and tested performance in major fire events, the model is already contributing to global disaster readiness. Rather than waiting for flames to appear, it helps agencies plan for where and when the spark might happen.
