Remzi Eker;Kamber Can Alkiş;Abdürrahim Aydı

Graduate School of Natural and Applied Sciences, Izmir Katip Çelebi University, 35620, Izmir, Turkey;Department of Forest Engineering, Faculty of Forestry, Düzce University, 81620, Düzce, Turke;Department of Forest Engineering, Faculty of Forestry, Izmir Katip Çelebi University, 35620, Izmir, Turkey

disturbances;izmir;automl;disturbance;framework;rf

Abstract Disturbances such as forest fires, intense winds, and insect damage exert strong impacts on forest ecosystems by shaping their structure and growth dynamics, with contributions from climate change. Consequently, there is a need for reliable and operational methods to monitor and map these disturbances for the development of suitable management strategies. While susceptibility assessment using machine learning methods has increased, most studies have focused on a single disturbance. Moreover, there has been limited exploration of the use of “Automated Machine Learning (AutoML)” in the literature. In this study, susceptibility assessment for multiple forest disturbances (fires, insect damage, and wind damage) was conducted using the PyCaret AutoML framework in the Izmir Regional Forest Directorate (RFD) in Turkey. The AutoML framework compared 14 machine learning algorithms and ranked the best models based on AUC (area under the curve) values. The extra tree classifier (ET) algorithm was selected for modeling the susceptibility of each disturbance due to its good performance (AUC values > 0.98). The study evaluated susceptibilities for both individual and multiple disturbances, creating a total of four susceptibility maps using fifteen driving factors in the assessment. According to the results, 82.5% of forested areas in the Izmir RFD are susceptible to multiple disturbances at high and very high levels. Additionally, a potential forest disturbances map was created, revealing that 15.6% of forested areas in the Izmir RFD may experience no damage from the disturbances considered, while 54.2% could face damage from all three disturbances. The SHAP (Shapley Additive exPlanations) methodology was applied to evaluate the importance of features on prediction and the nonlinear relationship between explanatory features and susceptibility to disturbance.