Monthly fire probability prediction integrating lightning and multivariate predictors across China using machine learning

This study develops a Monthly Fire Probability Index (MFPI) framework for China using Random Forest and XGBoost, driven by 17 predictors spanning meteorology, vegetation/fuel properties, topography, human activity, and lightning, with a one-month lead to represent antecedent conditions. Predictor–target pairs from 2015 to 2022 were used for model fitting, those from 2023 to 2024 for temporal hold-out selection, and selected models were refitted using all valid 2015–2024 samples; 2025 was retained for independent testing. RF achieved stronger probability and broad-pattern performance in the test period (ROC-AUC 0.858, Brier score 0.143, log loss 0.412, Expected Calibration Error 0.263, and Pearson spatial correlation 0.204) than XGBoost (0.845, 0.174, 0.483, 0.288, and 0.195, respectively). Systematic multi-lag experiments indicate that the 1-month lag represents an appropriate trade-off among predictive accuracy, spatial fidelity, and operational early-warning timeliness. In 2025, the RF predictions adequately capture China's seasonal fire spatial migration patterns, including winter aggregation in South and Southwest China, spring northward expansion to Northeast and North China, summer contraction to northern peripheral margins, and autumn resurgence in Southwest and South China. These findings provide an interpretable framework with national-scale applicability for fire risk assessment, monthly early warning, and adaptive ecosystem management.