Firefront Forecasting in Boreal Forests: Machine Learning Approach to Predict Wildfire Propagation

Abstract

Wildfires have become increasingly prevalent worldwide due to climate change, posing significant threats to human lives, property, and natural ecosystems. The rapid progression of wildfires necessitates predictive computational models to assist firefighters in effectively developing strategies to control firefronts. However, existing models often face challenges in computational complexity as the firefront expands. This study aims to develop a faster, more computationally efficient, deep-learning-based model for predicting wildfire spread. We hypothesise that firefront propagation can be modelled using stochastic cellular automata and that a deep-learning model can mimic this approach. With this in mind, we will first introduce our in-house stochastic cellular automata model, which is being validated with data from a known Finnish wildfire. After that, we propose a novel deep-learning model which uses the data generated by our cellular automata. The deep-learning-based model was based on Unet architecture, and it is capable of predicting firefront progression accurately and efficiently one time-step at a time. The model provided realistic simulations of firefronts with high computational efficiency, leaving future development needs to longer time series. One potential application of the developed model is in UAV-based real-time wildfire management systems.