When I started working on fire prediction many years ago, what really annoyed me about traditional weather-based prediction systems was the unrealistically high fire danger in barren areas. This was an obvious consequence of the fact that deserts experience very hot and dry conditions, which are theoretically fire prone. However, it was clear that fires could not occur there—because there was nothing to burn. The idea of applying a predetermined mask to exclude such areas from fire danger maps, rather than publishing clearly nonsensical predictions, made me think that we could certainly do better—starting with incorporating information on fuel availability.
Including fuel in fire danger prediction proved to be far from simple, as very little information was available, either from observations or models. It took us a year of work and a substantial amount of creativity—leveraging ECMWF’s infrastructure—to produce daily global estimates of fuel availability and dryness as a part of ECMWF fire model (SPARKY). While these estimates can still be improved with more specific observations, they marked a crucial step forward.
Fast forward a few years, and the advent of new machine learning (ML) methods arrived like a storm in weather forecasting. For me, ML is, first and foremost, an opportunity—not just to reproduce what a physical model can do, but to go beyond its limitations. Physical models do not account for human behaviour, which is often a key driver of fire ignition. This realization led to the development of the Probability of Fire (POF) model, built on three pillars:
Scientific understanding of fire drivers—thanks to decades of research by fire scientists.
Earth observation programs—enabled by space agencies and European initiatives like Copernicus, which provide data on fuel evolution and fire activity (special thanks to ESA and NASA, among others).
Machine learning algorithms—which allow us to integrate these data sources and produce actionable forecasts with ease (thanks to a vast community across science and tech companies).
From the very first events we analysed using POF, the quality of the forecasts was astonishing. The system finally excluded vast regions that were never going to burn due to a lack of fuel. More importantly, it learned where people were—and where the probability of ignition was highest due to human activity.
However, none of this would have been possible without a full understanding of the physical processes and without relying on a physical model to generate the data needed to train the machine learning model. Looking back at the improvements in prediction quality and this journey, I am happy to say that—for fire forecasting, at least—30% of the accuracy globally comes from data beyond weather. In some regions, weather is not even the most relevant factor determining fire predictability, so the time spent on studying physics and producing SPARKY-fuel was well spent after all ...
A video explainer can be found here
These products are available to ECMWF users through eccharts now being tested for inclusion in the Global Wildfire Information System (GWIS), and we hope they will soon see widespread adoption.