Our proposed solution introduces a data-driven Probability of Fire (PoF) model designed to shift wildfire management from monitoring “danger” to predicting actual fire activity. Unlike traditional systems (e.g. fire weather indices) that rely primarily on meteorological indicators, this approach uses a Machine Learning (ML) architecture—specifically Extreme Gradient Boosting (XGBoost)—to estimate the daily likelihood of fire occurrence at ~9 km resolution (scalable to 1 km). At its core, the system integrates the full “fire triangle” as predictive features: Weather: Near-real-time inputs including temperature, dew point, wind speed, and precipitation. Fuel: High-quality satellite observations combined with physical models to generate daily estimates of live and dead fuel loads and fuel moisture, enabling the model to account for both short-term drying and long-term fuel accumulation. Ignitions: Spatial representations of human drivers (e.g. population density, road networks, socio-economic factors) alongside natural triggers such as lightning forecasts. Trained on decades of satellite-derived fire observations, the model learns where fire is structurally inhibited—such as in barren deserts, thereby substantially reducing the false alarms typical of traditional fire danger indices. It also captures patterns linked to agricultural practices and human behaviour, improving predictive realism. Importantly, the system moves beyond a “black-box” paradigm by incorporating explainable AI techniques to identify the dominant physical and socio-environmental drivers behind each predicted risk level. The final product is delivered both as an operational data stream and through a modular PoF-Toolbox framework, enabling users to develop tailored probability-of-fire products using local datasets for enhanced regional fidelity.
This development aims to fundamentally transform wildfire management by shifting the paradigm from monitoring abstract “fire danger” to predicting actual fire activity. Rather than relying on traditional flammability indices such as the Fire Weather Index, the ambition is to deliver a high-resolution, data-driven Probability of Fire (PoF) framework that estimates the likelihood of real-world fire occurrence at daily ~9 km resolution, with scalability to finer spatial grids. At the core of this ambition is the full integration of all key drivers of fire occurrence, the “fire triangle”: weather conditions, fuel status, and ignition sources, both natural and human-induced. By simultaneously accounting for these interacting components, the system addresses a long-standing gap in global fire forecasting, where fuel dynamics and anthropogenic ignitions are often insufficiently represented. By learning where fires are structurally inhibited, the model enhances realism and strengthens decision relevance and responds to the growing socio-economic and environmental impacts of wildfire. The proposal further seeks to empower regional ownership of fire forecasting. Through the modular PoF-Toolbox, institutions will be able to train, adapt, and deploy models using their own local datasets, transitioning from one-size-fits-all products toward regionally tailored, high-fidelity systems. Beyond prediction, the ambition extends to advancing fire attribution. By applying explainable AI methods such as SHAP values, the framework identifies the dominant physical and socio-environmental drivers underlying each forecast, linking risk levels to processes such as fuel accumulation, dewpoint thresholds, lightning occurrence, or human pressure. In a rapidly changing climate, this framework offers an adaptive, scalable, and scientifically transparent system capable of evolving alongside emerging fire regimes and shifting patterns of human activity worldwide.
The strength of innovative approaches in modern wildfire forecasting lies in four key advances: 1. Superior Accuracy and Reduced False Alarms Traditional indices such as the Fire Weather Index (FWI) frequently overestimate risk in fuel-limited biomes. A data-driven Probability of Fire (PoF) model learns directly from observed fire activity, enabling it to recognise where fire is structurally inhibited. This eliminates barren-area false alarms and significantly improves discrimination skill. 2. Inclusion of Long-Term Effects in Forecasting Unlike meteorology-only indices, PoF integrates weather, fuel dynamics, and ignition sources. By incorporating fuel abundance and moisture estimates, the system captures long-term accumulation processes and compound effects such as “hydroclimate whiplash,” where wet periods promote vegetation growth that becomes highly flammable during subsequent drought. Ignition layers include both lightning and human drivers (e.g. population and road density), accounting for triggers typically absent from purely physical forecast models. 3. Probabilistic Reliability for Operations In this framework, accuracy is defined through reliability: predicted probabilities are statistically consistent with observed fire frequencies. A forecast of 0.3% fire probability corresponds to real-world occurrence rates, enabling agencies to allocate resources based on quantified risk rather than deterministic “danger” levels. This strengthens early warning credibility and reduces costly over-preparedness. 4. Local Ownership and Explainability Through a modular PoF-Toolbox, regional institutions can replace generic global inputs with high-resolution local datasets, increasing fidelity and ownership. Crucially, explainable AI methods such as SHAP values link predictions to specific physical drivers, ensuring transparency and scientific interpretability rather than opaque “black-box” outputs.
Wildfire management agencies face a critical operational challenge: large areas can simultaneously experience elevated fire danger while surveillance and response resources remain limited. Traditional fire danger indices primarily rely on meteorological conditions and often highlight broad regions of risk without clearly identifying where fires are most likely to occur. This limits the ability of authorities to prioritise monitoring efforts and allocate prevention resources effectively. The Probability-of-Fire (PoF) framework addresses this gap by providing probabilistic forecasts of fire occurrence, combining meteorological conditions with indicators of fuel state, vegetation dynamics and ignition likelihood. By identifying areas where ignitions are most probable, the system enables more targeted surveillance, earlier detection and more efficient deployment of firefighting resources, improving preparedness during high-risk periods. The solution benefits from the operational infrastructure of the European Centre for Medium-Range Weather Forecasts, which provides reliable forecasting systems, global environmental observations and near-real-time updates. Additional support is available through collaborations with national meteorological services, forest agencies and civil protection authorities, as well as through the ARISTOTLE consortium, which provides scientific advice to the Emergency Response Coordination Centre during major disasters. Together, these elements ensure that the PoF framework can be operationally integrated into existing wildfire preparedness and response systems, providing actionable information to decision-makers.
Our team includes leading fire scientists at ECMWF, Dr Francesca Di Giuseppe and Dr Joe McNorton, supported by hydrologist, Dr Fredrik Wetterhall and high-level technical expertise from Christopher Barnard. Together, they have developed an operational global Probability of Fire (PoF) system, running in real time since 2023. Our AI-driven expertise delivers robust, probabilistic forecasts designed to support informed decision-making in global wildfire management.
Our team includes leading fire scientists at ECMWF, Dr Francesca Di Giuseppe and Dr Joe McNorton, supported by hydrologist, Dr Fredrik Wetterhall and high-level technical expertise from Christopher Barnard. Together, they have developed an operational global Probability of Fire (PoF) system, running in real time since 2023. Our AI-driven expertise delivers robust, probabilistic forecasts designed to support informed decision-making in global wildfire management.