Modeling the spread of forest fires through cellular automata by leveraging deep learning to derive transition rules

Abstract

Predicting forest fire spread using simulation models is crucial for the effective management of forest fires. Cellular Automaton (CA) is a key model, and CA transition rules play a decisive role in the effectiveness of the simulation, highlighting the importance of accurately defining these rules. Traditional methods for extracting CA transition rules frequently neglect the intermediate stages of fire development, resulting in less effective outcomes. To overcome this limitation, our study introduces a deep-learning Transformer model to derive more accurate transition rules. The Transformer model excels in capturing fire-spread patterns owing to its robust feature extraction abilities and capacity to manage long-range dependencies, enabling the automatic generation of CA transition rules that more accurately reflect real fire behavior and ultimately improve the simulation of fire spread. Using forest fires in the back mountains of Wenbi Village, Dali City, Yunnan Province, and Sahai Village, Dongchuan District, Kunming City, Yunnan Province as case studies, we initially trained a Transformer model using historical fire data from these areas. We then extracted the CA transition rules from the training results and assessed the model performance using a least-squares support vector machine (LSSVM) model for comparison. The results revealed that the Transformer-CA model surpasses the LSSVM model for predicting fire spread, achieving simulation outcomes that closely align with real fire footprints and improving the overall accuracy, Kappa coefficient and IoU by 4.1 %,5.0 %, 5.5 %, and 3.8 %, 6.0 %,7.0 %, respectively, in the two study areas. This study demonstrated that the Transformer model is ideal for capturing the spatiotemporal evolution of forest fires and constitutes an effective technical approach for fire prevention and management.