SHAP-Based Identification of Grassland Fire Danger Thresholds and Temporal Transferability Validation: A Case Study of the Eastern Steppe in Mongolia

Based on the eXtreme Gradient Boosting (XGBoost) model and SHapley Additive exPlanations (SHAP) dependence plots, this study identified the fire-risk thresholds of the Normalized Difference Vegetation Index (NDVI) and the Dead Fuel Index (DFI), and analyzed their interaction using SHAP interaction values. The results showed that the suitable NDVI range was 0.1555–0.1943, and values outside this range were unfavorable. DFI ≥ 15.4008 was identified as a critical threshold, above which the probability of fire occurrence increased significantly, making it the dominant condition for fire occurrence. The interaction analysis showed that NDVI mainly exerted a positive regulatory effect on DFI, but the overall interaction strength was relatively weak. Comparison of threshold effectiveness showed that the single DFI threshold achieved the highest recall, reaching 98.57%, and was suitable for large-scale early warning. The combined threshold (0.1555 ≤ NDVI ≤ 0.1943 and DFI ≥ 15.4008) achieved the highest precision of 63.33%, but its recall was only 47.78%, indicating a relatively high missed-fire risk. Independent validation using 2022 samples showed that the high-risk area defined by DFI ≥ 15.4008 covered 96.49% of fire points, with a precision of 63.91%, an F1 score of 76.90%, and an accuracy of 71.00%. Overall, the single DFI threshold has good temporal transferability and can serve as a fire-risk monitoring indicator.