Drivers of forest fires under fire management policies: A comparison of machine learning models based on multi-temporal MODIS fire data

Forest fires across China are influenced by various factors due to the different forest fire management policies. This study utilized MODIS fire point data (2001−2020) from Guizhou Province. The datasets were divided into two periods based on the implementation of the new Forest Fire Prevention Regulations (2001–2008 and 2009–2020), and the spatiotemporal dynamics of forest fire points were analyzed. By integrating meteorological, topographic, vegetation, anthropogenic, and socioeconomic driving factors, the differences in the main drivers across periods were analyzed. Multiple machine learning models were built to determine the optimal model to generate regional fire risk maps. Our analysis reveals that over 80 % of forest fires occurred from January to April from 2001 to 2008, and fire counts trended upwards. Conversely, from 2009 to 2020, approximately 85 % of fires remained concentrated during these months, but the frequency of fires showed a steady decline. Meteorological, socioeconomic, and vegetation factors were the main drivers of fire occurrence in both periods. The Random Forest model achieved optimal performance in both periods, with an accuracy over 88.30 % and Area Under Curve value ≥0.953, significantly outperforming eXtreme Gradient Boosting, Support Vector Machines, and Artificial Neural Networks. Across the two periods, the probability of forest fire occurrence was highest in spring and lowest in winter. This study revealed the main drivers of forest fires in Guizhou across different periods and built an optimal prediction model, thereby providing a scientific basis for forest fire management departments to conduct forest fire prevention, control, zoning, and other management work. These findings are vital for protecting forest resources, maintaining ecological and environmental security, and safeguarding human lives. • Main drivers showed significant differences pre- vs. post-implementation of the Forest Fire Prevention Regulation. • Forest fires increased (2001–2008) then declined steadily (2009–2020). • Identified January – April as peak forest fire season. • Random Forest model achieved optimal predictive performance. • Offers practical prevention strategies for forest fire management policies.