Background <p>Forest fires exhibit distinct regional patterns. Identifying their driving factors and developing prediction models are crucial for early prevention and control. While previous studies have explored driving factors in southwestern China, discrepancies in model predictions across spatial scales and insufficient validation of their reliability have limited their application in fine-scale management and risk zoning. Based on this, this study utilized small-scale regional forest fire point data (2001–2020) in Guiyang City to analyze spatiotemporal distribution patterns. By comprehensively considering meteorological, topographical, vegetation, and socioeconomic factors, we identified the main driving factors of forest fires and subsequently developed data-driven probabilistic fire prediction models, namely random forest model (RFM) and logistic regression model (LRM). Subsequently, the data were split into a 60% training set and a 40% test set, with this process repeated five times. The training set was used for parameter tuning and model construction, while the test set was employed for model validation. Performance evaluation metrics (e.g., accuracy, precision, recall) were utilized to assess the model’s robustness and generalization ability. In addition, the optimal model was validated temporally and spatially using an additional dataset (2021–2023). Finally, the forest fire hazard zoning map for Guiyang was generated.</p> Results <p>The results show that fire incidents in Guiyang initially increased and then decreased; 97% of forest fires mainly occurred from January to April. Meteorological factors (monthly average relative humidity, temperature, and monthly accumulated rainfall) were the primary drivers of forest fire occurrence, with their importance (%IncMSE) significantly surpassing that of topographic and anthropogenic factors. In the five random samples, the RFM outperformed the LRM significantly in predicting fire probability. On the full sample, the RFM achieved an accuracy of 95.96% and an area under the curve (AUC) of 0.998 while maintaining high precision and recall. The monthly average relative humidity, temperature, and accumulated rainfall all demonstrated feature importance scores exceeded 9%. Furthermore, the RFM achieved robust performance on the additional dataset, with an accuracy of 82.61% and an AUC of 0.724, demonstrating strong predictive capability for forest fire occurrence. The finally generated forest fire zoning map indicated that spring had the highest fire risk, followed by summer, with lower risk in autumn and winter.</p> Conclusions <p>This study clarifies the mechanism by which spatial scale influences fire-driver identification and model predictive performance, addressing a gap in previous research regarding multi-scale comparisons and mechanistic interpretation. Beyond providing a theoretical explanation of scale effects, the study also delivers a validated forest fire hazard zoning map. These findings are of great significance for scientific forest fire management in the region and serve as a reference for examining the driving factors of forest fires at different scales. And can guide relevant departments in implementing targeted prevention measures during high-risk periods, as well as optimizing monitoring infrastructure in key areas.</p>