Abstract: This study aims to quantitatively identify the relative hazard of different fire-inducing factors and to establish a metric framework applicable to large-scale screening and graded fire-risk management. Multi-source data including local gazetteers, official fire investigation reports, news records, and field surveys were integrated to construct a positive and negative sample database covering 226 ancient timber covered bridges and 285 disaster events. Fire occurrence (yes/no) was defined as the response variable, while eight fire-inducing factors—ritual fire use, fireworks setting off, electrical installations, combustible accumulation, tourism development, residential activities, lightning-prone areas, and important transportation nodes—were encoded as binary variables indicating presence or absence. Random Forest (RF), Support Vector Machine (SVM), XGBoost, and LightGBM classifiers were developed and compared, with model performance evaluated using accuracy (ACC) and the area under the ROC curve (AUC). The RF model exhibited the most balanced performance and was selected for subsequent analysis. Based on normalized feature importance derived from the RF model (MDI/Gini), hazard weights of fire-inducing factors were quantified, and a Fire-inducing Factor Hazard Index (FFHI) with a four-level classification scheme was further established. The results indicate that combustible accumulation (25.50%) and ritual fire use (21.20%) are the most hazardous factors, followed by residential activities (17.80%) and electrical installations (10.30%). Fireworks (8.30%) and important transportation nodes (8.20%) show comparable contributions, while tourism development (5.00%) and lightning-prone areas (3.70%) exhibit relatively lower weights. The FFHI ranges from 0 to 1 and enables rapid quantification of fire-inducing conditions at the individual bridge level, supporting classification into four hazard grades. The findings suggest that fire prevention for timber covered bridges should prioritize the control of combustible materials and ritual fire use, while simultaneously regulating residential activities and electrical installations. Context-dependent factors such as fireworks setting off, transportation nodes, and tourism peaks require strengthened time-specific management. The proposed FFHI and grading approach provide a quantitative basis for regional-scale screening, inspection prioritization, and intervention sequencing, and can be dynamically updated as new data become available.

Key words: ancient timber covered bridges, fire-inducing factors, fire hazard assessment, machine learning, fire protection of heritage buildings