<p>Wildfire hazard assessment is essential for enhancing regional disaster prevention and mitigation capabilities. However, single Machine Learning (ML) algorithms often struggle to comprehensively and accurately capture the relationships between wildfires and their driving factors due to the complexity of wildfire hazard at the regional scale. To address this problem, the optimized Stacking ensemble framework was used in this study to construct the wildfire hazard assessment model in Xichang City. 11 mainstream ML algorithms were employed to construct the base model library, and 8 structurally diverse meta models were integrated. A dual optimization strategy combining Bayesian optimization and exhaustive search was adopted to optimize both model architecture and hyperparameters. Additionally, the influence of driving factors and their attribute intervals on wildfire hazard was analyzed in depth. Multi-metric evaluation showed that the optimal Stacking model, comprising Decision Tree, AdaBoost, eXtreme Gradient Boosting, and Categorical Boosting as base models, and Multilayer Perceptron as meta model, demonstrated the best performance in predictive accuracy and spatial zoning rationality. High and very high wildfire hazard zones were primarily located in mountainous regions flanking the central area, as well as in the south and northeast of Xichang. Temperature, wind speed, slope, NDVI, distance to roads, rainfall, and relative humidity were the most critical factors of their attribute intervals. The proposed optimal modeling method significantly improves wildfire hazard prediction and provides valuable insights for regional risk prevention and management.</p> Graphical Abstract <p></p> <p>This study presents an optimized Stacking ensemble framework for regional wildfire hazard assessment, applied to Xichang City. The framework systematically integrates data preparation, model optimization, wildfire mapping and assessment to overcome the limitations of single machine learning (ML) algorithms in capturing complex wildfire drivers. First, based on wildfire samples derived from VIIRS satellite data, the initial driving factors are discretized using the Jenks method combined with Goodness of Variance Fit and normalized via Certainty Factor (CF). Subsequently, 11 key driving factors are selected through multicollinearity testing and GeoDetector. Second, the modeling phase employs a dual optimization strategy combining Bayesian optimization and exhaustive search. This strategy is applied across 11 base ML algorithms and 8 meta-learners to evaluate structural diversity and determine the optimal architecture. The multi-metric evaluation reveals that the optimal Stacking model comprises Decision Tree, AdaBoost, eXtreme Gradient Boosting, and Categorical Boosting as base models, with a Multilayer Perceptron as the meta-model. Finally, the optimal model successfully maps wildfire hazards, indicating that high and very high hazard zones—accounting for 19.48% of the total area—are primarily concentrated in the mountainous regions flanking the central area, south, and northeast of Xichang. Comprehensive interpretability analyses using SHAP, GeoDetector, and CF identify temperature, wind speed, slope, NDVI, distance to roads, rainfall, and relative humidity as the most critical driving factors.</p>