Fire risk assessment using machine learning techniques: a case study of Jinan City, China

Effective fire risk assessment is crucial in disaster prevention and mitigation, facilitating optimized resource allocation and minimizing potential losses. This study conducted a comprehensive fire risk evaluation in Jinan City using five machine learning models—Random Forest (RF), Support Vector Machine (SVM), Maximum Entropy (MaxEnt), Light Gradient-Boosting Machine (LightGBM), and Convolutional Neural Network (CNN)—based on historical fire occurrence data and 15 environmental factors. A comparative analysis of these models revealed that the CNN model exhibited the highest performance, achieving an accuracy of 82.7% and an Area Under the Curve (AUC) value of 89.1%, closely followed by the LightGBM model. An evaluation of environmental drivers indicated that Land Use/Land Cover (LULC) and Normalized Difference Vegetation Index (NDVI) were the most significant contributors to fire risk. Additionally, spatial analysis demonstrated a characteristic risk distribution pattern featuring “three high-risk zones and two risk belts”, with high-risk areas predominantly clustered in central and southern urban regions and their peripheries. Seasonal variation analysis indicated significantly elevated fire risks during spring and summer. Furthermore, district-level prioritization identified Huaiyin District as requiring the highest-priority intervention, followed by Tianqiao, Gangcheng, and Zhangqiu Districts. These findings provide empirical evidence to inform spatially differentiated fire risk management strategies within Jinan City.