A comparative study of machine learning and convolutional neural network approaches for forest fire occurrence prediction

Abstract The increasing frequency and severity of forest fires, driven by climate change and human activities, highlight the urgent need for accurate forest fire prediction models. This study developed a deep learning model based on a convolutional neural network (CNN) that integrates topographic and meteorological information to predict forest fire occurrences in South Korea. The CNN’s performance was compared against four traditional machine learning (ML) models: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF) and XGBoost. Using a dataset of Korean forest fire records from 2011 to 2021, monthly meteorological data, and seven topographic variables, the models were validated through fivefold cross-validation. The results demonstrate that the CNN model, which effectively learns spatial patterns from gridded data, achieved superior predictive power with an accuracy of 0.962 and an F1-score of 0.962. However, ensemble models such as RF and XGBoost also showed excellent performance with accuracies of 0.931 and 0.937, respectively, while SVM and LR showed relatively low performance. Variable importance analysis with LR, RF and XGBoost consistently identified meteorological factors, particularly relative humidity, precipitation and temperature, as the primary drivers of fire occurrence. Among topographic variables, the Topographic Wetness Index (TWI) was the most influential. This study demonstrates that a prediction model considering both spatial patterns and climatic factors can significantly improve accuracy. These findings provide a robust basis for developing advanced early warning systems and targeted fire prevention strategies, which are the most direct methods for mitigating fire-related carbon emissions.