Wildfire risk is shaped by local conditions: climate, vegetation, land management, and human activity all play a role. While global fire-danger products provide valuable situational awareness, they cannot always capture the specific drivers and thresholds that control fire activity at regional or national scales. This can make it difficult to translate large-scale fire-risk information into locally relevant assessments.
To address this gap, we introduce the Probability of Fire (PoF)-Toolbox, a lightweight, reproducible framework designed to help users build, train, and deploy local machine learning (ML)-based fire-danger models using their own data.
From global signals to local fire risk
Probability of Fire models estimate the likelihood that fire activity will occur under a given set of environmental conditions at a specified spatiotemporal resolution (for example, daily predictions at 1 km grid scale). Traditional fire-danger approaches often rely on weather-based indices calibrated for specific vegetation types, which can struggle to transfer across regions with different climates or land-use practices.
ML-based PoF models offer a flexible alternative, learning relationships directly from data. Crucially, these approaches allow the incorporation of novel data sources, such as remotely sensed vegetation metrics, human activity indicators, or socio-economic proxies, which cannot easily be represented within physical or empirical fire-danger equations. Here we provide a core dataset to help you develop your own PoF model; however, we strongly recommend incorporating additional high-resolution data to produce forecasts with greater spatial granularity and fidelity.
Building such models from scratch can be technically demanding. Data preparation, feature selection, model architecture, and validation all require careful choices to ensure results are robust and reproducible. The PoF-Toolbox lowers this barrier by providing a ready-to-use set of notebooks that guide users through the full modelling workflow with generalised features provided as a starting point.
A reproducible, modular workflow
The framework is built around reproducible Jupyter notebooks, each focusing on a specific step of the modelling process. Users can start from available meteorological, vegetation, and socio-economic indicators, or replace and extend these with with locally available datasets. The notebooks are modular by design, making it easy to experiment with different predictors, spatial domains, or training periods without rewriting the entire pipeline.
The workflow is designed for experimentation and analysis using past forecast data, and is not intended to reproduce the operational PoF system currently running at ECMWF.
The workflow (illustrated in Figure 1) typically follows three main stages:
- Build – Prepare input data, define the spatial and temporal domain, and assemble predictors describing weather conditions, fuel availability, and longer-term environmental memory.
- Train – Fit an ML-based PoF model using historical fire occurrence data, applying cross-validation and diagnostics to assess performance and avoid overfitting.
- Deploy – Apply the trained model to new data to generate probabilistic fire-risk estimates that can be used for analysis, scenario exploration, or downstream applications.
At each stage, the notebooks emphasise transparency and traceability, ensuring that modelling choices are clearly documented and results can be reproduced.
Figure 1: The PoF-Toolbox workflow. Graphic generated by ChatGPT and edited by ECMWF.
Designed for local ownership
A key goal of the PoF-Toolbox is to support local ownership of fire-risk modelling. Rather than distributing a single “one-size-fits-all” product, the framework enables institutions, researchers and practitioners to train models that reflect their own environments. This is particularly valuable in regions where local datasets provide information that is unavailable at global scale.
We have long argued that incorporating additional data sources can substantially enhance predictive skill, and that training models solely on weather variables significantly limits their ability to anticipate fire events. A clear illustrative example is provided by the Evros fire near Alexandroupolis in Greece, which ignited on 19 August 2023. To date, this fire burned the largest recorded area (around 900,000 hectares) in Europe since records began.
A PoF model trained using only weather variables would give a low probability of ignitions, indicating that the event was not strongly driven by meteorological conditions on the day of ignition (Figure 2). When information on fuel availability and source of ignitions is included, the estimated probability increases substantially, providing a more consistent explanation of the event. Adding locally available, high-resolution data on fuel and forest management is expected to further improve the forecast. We hope the framework could also support comparative analysis. Local PoF models can be evaluated alongside traditional fire-weather indices or global fire-risk products, helping users understand where different approaches agree, diverge, or complement each other.
Figure 2: The PoF-Toolbox forecasts for the Evros region on 20 August 2023. The top panel includes weather-, fuel- and ignition-related predictors, while the bottom panel uses weather variables only. Colours show probability levels from absent to extreme, with dots indicating observed fire activity.
Supporting science-informed decisions
By making ML-based PoF modelling more accessible and reproducible, the PoF-Toolbox supports a wide range of applications, from scientific research to operational fire-risk assessment and long-term planning. The framework is designed to evolve, allowing new predictors, model architectures, and evaluation strategies to be added as understanding of fire–climate–vegetation interactions advances.
Ultimately, the goal is not just to predict fire activity, but to provide tools that help users explore why fire risk changes, how sensitive it is to different drivers, and how it may respond under future conditions. By design, the PoF-Toolbox encourages users to look inside the modelling system, linking predictions to the physical processes that drive fire risk and avoiding the “black-box” effect often associated with AI models.
Access the PoF-Toolbox
The PoF-Toolbox includes all the tools, data access, and workflows needed to explore fire activity prediction, from input preparation to model evaluation. The Jupyter notebooks for PoF-Toolbox are available on GitHub.
Top banner image: © Toa55 / iStock / Getty Images Plus