Using Machine Learning Algorithms to Predict Forest Fire Probability in Mediterranean Region of Türkiye

Abstract

Determining the forest fire probability levels by analyzing the main fire factors can provide forest managers with the basis for making critical decisions on issues such as fire prevention strategies, fuel management, fire safety measures, emergency planning, and placement of firefighting teams. The main fire influencing factors, including vegetation factors, topographical factors, climate factors, and proximity to some features such as roads and residential areas, have been considered to generate forest fire probability maps. The machine learning (ML) algorithms have become an effective tool in predicting forest fire probability. This study aimed to generate a forest fire probability map by using two commonly used ML models, logistic regression (LR) and support vector machines (SVMs), integrated with Geographical Information System (GIS) techniques. The study was implemented in & Scedil;elale Forest Enterprise Chief (FEC) located in the Mediterranean city of Antalya in T & uuml;rkiye. In the study, the fire influencing factors were tree species, crown closure, tree stage, slope, aspect, and distance to roads. The forest fires that occurred from 2001 to 2021 in & Scedil;elale FEC was considered in the training stage of the models. The accuracy of the fire probability maps was verified using the area under curve (AUC) value. As a result of performing the ML models, estimations were made for 47 086 points on the map which were categorized into five fire probability levels (very high, high, medium, low, and very low). The results showed that the accuracy of the fire probability map generated by the LR model was better (AUC = 0.845) than the accuracy of map generated by the SVM model (AUC = 0.748). According to the probability maps, more than half of the forests had very high/high fire probability levels in the study area.