Abstract: Forest fire prediction is essential for sustainable forest management, as wildfires cause significant ecological, environmental, and economic damage. This paper introduces a hybrid stacking model that improves the precision and resilience of forest fire risk prediction by merging Random Forest and Neural Network classifiers with Logistic Regression acting as a meta-learner. To solve the problem of class imbalance, new data samples were introduced and the SMOTE was used to add more high-risk fire events to the dataset, which was originally sourced from Kaggle. Additionally, feature engineering was used to create new variables that captured intricate connections between environmental and meteorological aspects. The suggested stacking framework achieves an accuracy of 87.2% on tests and an Area Under the Curve of 0.925 by combining the interpretability of RF with the nonlinear learning strength of NN. Strong dependability in differentiating between low-risk and high-risk fire incidents is demonstrated by these data. Overall, this approach provides an effective, data-driven foundation for intelligent wildfire monitoring and proactive forest management.

Keywords: Forest Fire Prediction, Random Forest, Neural Network, SMOTE, Machine Learning, Stacking, Feature Engineering.

Downloads: | DOI: 10.17148/IJIREEICE.2025.131113