Predicting fire susceptibility: Integrating environmental factors with knowledge graph and neural network

Global climate change and human activities have increased the frequency of wildfires, posing significant threats to ecosystems and human safety, including in regions traditionally less prone to wildfires. Accurately predicting fire susceptibility is therefore crucial for comprehensive fire management strategies. In this study, we propose a novel approach to fire susceptibility analysis, focusing on the Edinburgh and Lothians region over two fire seasons from 2018 to 2021. We formalize fire scenarios as heterogeneous graphs and develop a knowledge graph-based fire susceptibility assessment model. By leveraging graph neural networks, we perform representation learning on these fire scenarios, analyze and visualize the results of fire susceptibility assessments, and evaluate the importance of environmental factors. The experimental results demonstrate that this approach not only effectively predicts fire susceptibility but also identifies the four most significant environmental factors in the study area, including land use, land cover, path, and elevation through repeated experiments, providing a more reliable foundation for fire management and prevention decisions. • Multi-source data integration in a knowledge graph for accurate fire susceptibility. • Graph Attention Network application enhances prediction accuracy and interpretability. • Key environmental factors identified as critical to assessing fire susceptibility. • Model validation across seasons demonstrates generalizability and time-based reliability.