Simulation-based high-resolution fire danger mapping using deep learning

Abstract

Wildfire occurrence and behaviour are difficult to predict locally for the next day. In the present work, we propose relying on fire spread simulations to provide a fire danger index representative of the potential for fire spread that includes not only weather but also surrounding vegetation and orography. This is achieved using an artificial neural network emulator called DeepFire, trained based on simulated fire sizes. To determine how relevant this index can be in the assessment of next-day fire danger, the application of DeepFire to fire danger mapping using actual weather forecasts is studied. DeepFire forecasts for 13 fairly large fires that occurred in Corsica are analysed and compared with corresponding forecasts using another fire danger index used in operational conditions, highlighting the differences in terms of precision and the expected results of such predictions. The weather forecasts from which the weather inputs of DeepFire are determined have high spatial resolution and high frequency, which also applies to the fire danger predictions. Additionally, input uncertainty is propagated through DeepFire, resulting in ensembles of emulated fire size. Several approaches are proposed to analyse the results and provide fire danger maps and ratings using this new simulation-based prediction system.