A hybrid modeling approach for wildfire danger assessment: combining data-driven ignition and fire spread models

Wildfire activity is influenced by a wide range of factors, meteorological, topographical, vegetation-related, and anthropogenic, making its modeling a highly complex task. In this work, we present a methodology that integrates two distinct modeling approaches within a single tool: a Machine Learning-based ignition model and a physical fire spread model. 

Outcome of our approach are event-based burn probability maps, derived by aggregating the outcomes of many fire-spread simulations initialized from stochastic ignition events generated by a Machine Learning ignition model. This model is trained on historical ignition records and integrates meteorological, vegetation, and anthropogenic variables to yield daily ignition probability maps. From each daily map, we sample stochastic ignition events and run the fire spread model for each, generating an ensemble of plausible outcomes whose aggregated footprint yields the final event‑based burn probability map. 

This combined approach enables us to address separately two critical wildfire processes: ignition and spread. Utilizing a data-driven model allows us to account for anthropogenic influences on ignition through variables such as proximity to roads, power lines, and land use. Meanwhile, the complexity of fire spread is handled by a physical propagation model that considers key factors such as fuel continuity, terrain, and processes like spotting. 

The tool is currently under development within the UNICORN project, funded by the EU Horizon Europe Programme (grant agreement No 101180172), and is being tested in the cross-border region of Northwest Spain and Northern Portugal, one of Europe’s most wildfire-prone areas.