Deep Learning Prediction of Wildfire Probability Occurrence Using Meteorological Factors from Khenchela Province Forests

Wildfire prediction is crucial for mitigating ecological and economic losses during forest management. This study presents a three-phased AI-driven methodology for enhanced wildfire risk assessment in Khenchela Province. First, explainable AI (XAI, LIME, and SHAP) analyzes eight machine-learning models (CatBoost, XGBoost, SVM, RF, AdaBoost, LR, LightGBM, and DT) to identify 12 key wildfire predictors from a historical dataset. Second, deep learning models (LSTM, RNN, and Prophet) forecast meteorological conditions to improve temporal resolution. Finally, a Bayesian model integrates these forecasts with moisture codes (FFMC, DMC, and DC) and indices (ISI, BUI, and NDVI) to predict wildfire likelihood. Validation using Accuracy, Precision, Recall, F1-score, and AUC demonstrated superior performance, with CatBoost achieving a 95.93% accuracy. This integrated approach enhances wildfire prediction reliability, enables proactive mitigation strategies, and shows the synergistic benefits of combining machine learning, deep learning, and Bayesian modeling for improved wildfire risk management.