Data-driven fire cause identification and prediction for forest fires in Heilongjiang Province, China

Forest fires are usually caused by lightning or by human-beings, which are called lightning-caused fire or human-caused fires. Fire cause identification is crucial for fire case studies and subsequent forest fire prevention and control. a large amount of historical forest fire records are classified as with “unknown fire cause,” which hinders effective fire management. To address this issue, we collected eight categories of 27 variables related to meteorology, vegetation, topography, and human activities in Heilongjiang Province, China, and such variables will be used for fire cause identification. We analyzed the peak values of high-incidence variables for each fire cause and examined the spatial clustering of different fire causes using spatial analysis techniques. Eight machine learning algorithms were evaluated via the entropy weight-TOPSIS method to identify the optimal model for fire cause prediction, which was then applied to predict unknown fire causes. The results indicate that lightning-caused fires occur predominantly in the Greater Khingan Mountains, smoking-related fires cluster near the border of Heihe and Hegang, and agricultural/forestry fires are distributed across multiple fire-prone regions in Heilongjiang. The entropy weight-TOPSIS analysis identified the Multilayer Perceptron (MLP) algorithm as the optimal model, with a comprehensive score of 0.91. Key drivers for MLP included DME, elevation, AWCE, and vegetation type, all with SHAP values exceeding 0.15. The MLP model demonstrated high prediction accuracy in major fire-prone areas such as the Greater Khingan Mountains, addressing regional fire cause prediction challenges. Furthermore, for approximately 52.4% of unknown fires, MLP provided two probable causes. and the sum of the predicted probabilities of these two fire causes exceeds 80%. This study proposes a comprehensive framework for forest fire cause prediction, capable of identifying causes for historically unknown fires and assisting in future fire cause determination. The approach provides a scientific basis for fire management, enhances prevention efficiency, and offers critical insights for forest fire risk mitigation.