An intelligent IoT-machine learning framework for wildfire detection and prediction using a hybrid RF-XGB model.

Forest fires in Turkey have received comparatively limited scholarly attention despite the country's high seasonal susceptibility, particularly during summer due to adverse climatic conditions. For real-time detection and risk assessment, this research suggests an integrated intelligent wildfire monitoring and prediction framework that integrates a unique weighted-voting RF-XGB hybrid model with Internet of Things (IoT)-based wireless sensor networks (WSNs). The adaptive weighting approach, which goes beyond traditional majority-voting ensembles, combines Random Forest and Extreme Gradient Boosting to take use of complementary variance-reduction and boosting processes. This is the methodological innovation. A multi-season Turkish forest fire dataset that included environmental sensor data, including temperature, relative humidity, and carbon monoxide concentration, was used to train the model. Distribution-preserving sampling and stratified k-fold cross-validation were used to alleviate class imbalance. With an accuracy of 0.9631, F1-score of 0.9627, and ROC-AUC of 0.994, the suggested hybrid model outperforms the others when compared to RF, XGBoost, KNN, Decision Tree, MLR, SVM, and ANN. Larger improvements were shown over KNN (10.4%) and Decision Tree (18.3%), while Relative Improvement (RI), as determined by the AUC measure, reveals a 4.6% increase over XGBoost and 5.7% over Random Forest-the strongest baselines. When compared to MLR, SVM, and ANN, improvements of over 50% were seen, demonstrating the hybrid model's greater robustness and discriminating capabilities. At the system level, a lightweight Multiple Logistic Regression (MLR) model was deployed on Arduino Nano-based sensor nodes to enable edge-level probability estimation and reduce communication overhead. Nodes operate using hourly duty cycling and transmit only when fire probability exceeds a predefined threshold, achieving an analytically estimated lifetime of up to 11 months. The framework was implemented in Zeytinpark using 80 sensor nodes deployed via hybrid grid and K-means clustering, achieving 95.58% coverage. Real-time detections are verified at the sink node using the RF-XGB model before triggering multi-level alerts, including local alarms, cloud updates, Telegram notifications, and mobile-based fire localization. The results demonstrate that the proposed contribution lies in the adaptive hybrid ensemble design, hierarchical edge-cloud intelligence distribution, and validated real-world deployment. The framework provides a robust, energy-efficient, and scalable solution for rapid wildfire detection and forecasting.