Prediction of forest fires using machine learning methodology: A case study in Türkiye

Forests play a crucial role in mitigating soil erosion and act as a barrier against various natural disasters, especially drought. Forest fires, which endanger millions of hectares of forest areas and cause the destruction of natural resources, have serious consequences for the environment and pose significant threats to human lives. The prediction of forest fires is of great importance in terms of preventing the occurrence of fires. Prediction of forest fires is based on many meteorological, topographical and spatial factors. Türkiye has a high susceptibility to forest fires and comprehensive studies on the forest fire prediction in Türkiye are needed. The increase in forest fires along the Aegean and Mediterranean coasts of Türkiye in recent years has increased the importance of forecasting forest fire risks. In this study, it is aimed to predict forest fires temporally and spatially with machine learning methods by using meteorological data and forest fire data from Adana, Antalya, Aydın, Balıkesir, Burdur, Çanakkale, Hatay, Isparta, İzmir, Kahramanmaraş, Mersin, Muğla and Osmaniye provinces between 2013-2018. The periods with the highest forest fire risk were examined and time series forecast model was applied based on the variation in the number of monthly forest fires. The influence of meteorological factors such as wind speed, relative humidity, temperature and precipitation was studied. Several machine learning methods were tried and random forest, support vector machine and decision tree algorithms were selected to process the data. The selected models were evaluated using accuracy, F1-score, The Receiver Operator Characteristic (ROC) curve, error rate, precision and recall performance metrics. Results showed that, random forest has the highest accuracy with 86.3%, whereas support vector machine has 86% and decision tree has 80.1% accuracy. The highest value of area under (AUC) ROC curve was obtained in the random forest model with 0.93. In this study, the highest performing model was random forest with an accuracy of 86.3% and an AUC value of 0.93. The results of this study are believed to contribute to the assessment of forest fire risks and to the planning and preparation against forest fires to mitigate the damage caused by them.