Forest fires occur frequently in the southwest of China. It is crucial to construct forest fire prediction models and explore the driving factors of fire occurrence for effective fire management. We employed six machine learning models to explore the optimal model and important driving factors for predicting forest fires in different seasons in Yunnan Province, China. The results indicated that the BRT was the best model for predicting forest fires, and meteorological and human factors were the important driving factors for fire occurrence. The XGBoost was the optimal model for predicting fires in summer and autumn, mainly influenced by meteorological and soil vegetation factors. We also found that the areas with a higher probability of forest fire occurrence were mainly concentrated in the southwest, southeast, and northwest. This study can provide useful reference for the formulation of forest fire prevention strategies in specific seasons in the research area.

Frequent and intense grassland fires represent a significant threat to the stability and sustainability of grassland ecosystems. Therefore, understanding the driving factors of grassland fire and the occurrence of fire is key to formulating effective fire management policies and management plans. Based on the fire dataset (manually recorded data, satellite remote sensing data) from 2001 to 2022, this study uses six models to analyze the differences in grassland fire driving factors in different regions and fire prevention periods in the study area, determine the relative importance of fire driving factors, and draw a probability map of grassland fire. The results show that both types of data selected the Boosted Regression Trees (BRT) model as the optimal model for predicting grassland fires in the Inner Mongolia Autonomous Region. Meteorological factors are the main driving factors of grassland fire in the Inner Mongolia Autonomous Region, and topographic factors and socio-economic factors are important factors. The number and probability of fires gradually decreased from east to west, and fires were mainly concentrated in the northeast and middle of the study area. Therefore, our study functioned to explore the spatio-temporal pattern of grassland fire, accurately predict the probability of grassland fire at different scales, and provide a scientific basis for the rational allocation of grassland fire prevention resources in the study area.