Abstract: To identify the major drivers of lightning-caused fire and to develop predictive models in the Greater Khingan Mountains, we integrated multi-source data from 2021 to 2024, including lightning records, forest fire records, and multi-source environmental data. We identified candidate lightning events using a spatiotemporal mat-ching approach, and addressed the imbalance problem of sample type using the EasyEnsemble method to construct the model dataset. We then selected predictor variables through correlation analysis, collinearity diagnosis, and feature importance analysis, developed multiple statistical and machine learning models, and interpreted the final model by SHAP (SHapley Additive exPlanations). The results showed that a total of 103 lightning events associated with lightning-caused fires were identified. The average distance between lightning strikes and fire ignition points was 0.84 km, with a mean holdover time of 1.4 days. 83.5% of fires were detected within three days after lightning occurrence. The predictive model identified the duff moisture code, drought code, and hourly fire weather index as the primary predictors, with temperature and wind speed as the secondary predictors. The random forest model showed the best overall performance, with an AUC of 0.7959, sensitivity of 0.7229, and specificity of 0.7752. The optimal classification thresholds were mainly concentrated between 0.45 and 0.55. The occurrence of lightning-caused fire was jointly influenced by fire weather, meteorological, topographic, and lightning-related factors, exhibiting significant nonlinear and threshold effects.

Key words: lightning-caused fire, lightning, fire weather index, interpretable machine learning, Greater Khingan Mountains