A Stacked Ensemble Model for Accurate Wildfire Prediction Using Meteorological Data and Bayesian Optimization

Wildfires have been and continue to be a growing threat to ecosystems, infrastructure and human lives. Over the past years there has been an increase in wildfire occurrences mainly due to climate change. This emphasizes the need for us to engineer highly accurate and early prediction models. Traditional wildfire forecasting methods rely heavily on data that is not very easily available like satellite imagery and fire specific indices like Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and Drought Code (DC). These methods often require specialized datasets that are not available in all regions. In contrast this paper explores a simpler and more accessible approach by using basic meteorological data (temperature, wind speed, precipitation, and seasonal variables) to predict the occurrence of a wildfire, hence making it more scalable and applicable to a wider range of regions. This study proposes a stacked ensemble learning model combining XGBoost, Random Forest, and LightGBM to give a better prediction accuracy compared to stand alone models. The model is fine-tuned using Bayesian optimization (Optuna) hyperparameter tuning, which searches for optimal hyperparameters in an efficient manner, outperforming traditional grid search methods. To enhance interpretability, SHAP (SHapley Additive Explanations) analysis is applied to explain individual model predictions and identify the most influential wildfire risk factors. Experimental analysis of the proposed model has also demonstrated high predictive accuracy, and that it also performs better than stand alone models.Integrating deep learning models like LSTMs and expanded global datasets are few of the future enhancements that can be made to the proposed methodology.