Evaluation of Supervised Machine Learning Methods for Forest Fire Severity Classification

Forest fires are among the most destructive natural disasters, causing substantial ecological, economic, and human losses. Accurate assessment of fire severity is crucial for preparedness, rapid response, and efficient resource management. This study evaluates three supervised machine learning (ML) algorithms—Linear Discriminant Analysis (LDA), Kernel Naive Bayes (KNB), and Fine Gaussian Support Vector Machine (Fine Gaussian SVM)—to classify forest fire severity using a real-world dataset from Türkiye. The dataset includes over 15,000 fire incidents (2010–2024) and 36 initial features. To improve predictive performance and reduce dimensionality, feature selection was performed using the Chi-square test. Fire severity was reclassified into three levels (low, moderate, high) based on burned area (hectares). Models were trained and validated with 10-fold cross-validation. KNB achieved the highest accuracy (82%), followed by Fine Gaussian SVM (79%) and LDA (65%). The advantage of KNB likely stems from its ability to capture nonlinear class boundaries and probabilistic structures typical of complex environmental data. Overall, the results suggest that nonlinear, kernel-based classifiers outperform linear methods for forest fire severity classification. The proposed national-scale, interpretable framework can support policymakers and disaster-management authorities in developing intelligent early warning systems and optimizing suppression resource allocation in high-risk areas.