Creation of wildfire susceptibility maps in the Mediterranean Region (Turkey) using convolutional neural networks and multilayer perceptron techniques

Considering the natural disasters that have developed in the world in recent years, it is known that there is an increase in wildfire disasters with the effects of climate change. In this study, wildfire susceptible areas were determined in the provinces of Mugla, Antalya, Mersin, Adana, Osmaniye, and Hatay in the Mediterranean re-gion (Turkey). Within the scope of this purpose, Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP) methods, the most widely used deep learning techniques in the literature in recent years, were preferred to create Wildfire susceptibility models. Seventeen environmental variables were used in the analyses, and these variables were grouped as topographic factors, anthropological and environmental factors, climatic factors, and vegetation factors. In addition, the number of fire inventory data has been balanced with the help of the Syn-thetic Minority Oversampling Technique (SMOTE) used to increase the model result performance of the scarce inventory data. In the maps obtained by CNN and MLP methods, 17% and 28% of the study area were deter-mined as high and very high susceptible areas, respectively. The results demonstrated that the CNN model had superior performance in Wildfire susceptibility assessment with accuracy (%85.8), precision (%98.7), sensitivity (%85.5), F-Score (%91.6), and ROC curve (%78.6). This model was followed by the MLP model with slightly lower accuracy values, which indicates that the CNN models can reach considerably better prediction capability than the MLP models. Finally, the wildfire susceptibility maps produced by deep learning methods could aid decision-makers and government organizations in the Mediterranean region in preventing future natural disasters.

Automated summary

Considering the increase in wildfire disasters due to climate change, this study used Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP) methods to create wildfire susceptibility models in six provinces in the Mediterranean region of Turkey. Seventeen environmental variables were analyzed, and the Synthetic Minority Oversampling Technique (SMOTE) was used to balance the limited fire inventory data. The CNN model showed superior performance in wildfire susceptibility assessment, indicating its better prediction capability compared to the MLP model. The produced susceptibility maps can assist decision-makers and government organizations in preventing future natural disasters in the Mediterranean region.