Prediction of Forest Fires Based on Hotspots and Climate Using the Random Forest Algorithm

Abstract

Forest fires pose a serious threat to ecosystems, biodiversity, and surrounding communities, making early prediction important for supporting mitigation and management efforts. This study applies the Random Forest algorithm to predict forest-fire likelihood using climate variables, including rainfall, solar radiation, humidity, and temperature. The study utilizes climate and hotspot data collected from BRIN and BMKG for the period from January to December 2023. The integrated dataset is evaluated using three training and testing split ratios, namely 70:30, 80:20, and 90:10. The experimental results show that the Random Forest model achieves accuracies of 85.8%, 85.9%, and 86.4%, respectively. The 90:10 split produces the highest accuracy. The confusion matrix for the 70:30 split records 1,082 correctly classified No Fire samples and 38 correctly classified Fire samples, while 21 No Fire samples and 163 Fire samples are misclassified. Feature importance analysis identifies solar radiation as the most influential variable with an importance value of approximately 0.48, followed by temperature at 0.26, humidity at 0.20, and rainfall at 0.06. These findings demonstrate that Random Forest can provide consistent forest-fire classification performance using climate-based variables. The results also indicate the need to improve the identification of actual fire occurrences, particularly for samples that the model classifies as No Fire.