The Effectiveness of Methods Designed to Predict Forest Fire Risk: A Comparison of Machine Learning Algorithms

Forest fires represent a global environmental issue, posing severe threats to ecosystems, economic development, and social security. Accurate prediction of forest fires is crucial for formulating effective preventive measures and minimizing the associated losses. With the advancement of machine learning technologies, their application in forest fire risk prediction has become an emerging research focus. Diverse machine learning algorithms exhibit varying data processing capabilities and predictive accuracies; hence, comparing and selecting the most suitable algorithms is significant for enhancing the performance of predictive models. This study analyzed meteorological data and corresponding fire severity information from Guangxi Province between 1990 and 2019, employing 10 machine learning algorithms in experiments. Initial data preprocessing, including handling of missing and outlier values, ensured data quality. Subsequently, predictive performance across algorithms was assessed using accuracy, precision, recall, F1 score, and the Receiver Operating Characteristic (ROC) curve. To further examine the stability and robustness of the models, a 5-fold cross-validation was implemented. Results indicated that SVM, Bayesian classifiers, BP neural networks, logistic regression, AdaBoost, Gradient Boost, and XGBoost demonstrated superior performance in terms of AUC, while KNN and Random Forest algorithms showed advantages in precision and accuracy. The 5-fold cross-validation confirmed the stability and robustness of the models, revealing that most models maintained stable predictive performance across different datasets. The study suggests that integrating multiple algorithms can improve the accuracy and reliability of predictions and recommends that future research consider additional influencing factors and employ deep learning techniques to further enhance predictive performance.