Publications

Iban, MC; Sekertekin, A (2022). Machine learning based wildfire susceptibility mapping using remotely sensed fire data and GIS: A case study of Adana and Mersin provinces, Turkey. ECOLOGICAL INFORMATICS, 69, 101647.

Abstract
In recent years, the number of wildfires has increased all over the world. Therefore, mapping wildfire susceptibility is crucial for prevention, early detection, and supporting wildfire management decisions. This study aims to generate Machine Learning (ML) based wildfire susceptibility maps for Adana and Mersin provinces, which are located in the Mediterranean Region of Turkey. To generate a wildfire inventory, this study uses active fire pixels derived from MODIS monthly MCD14ML composites. Furthermore, as a sub aim, the performance of seven ML approaches, namely, stand-alone Logistic Regression (LR), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and ensemble algorithms, namely Random Forest (RF), Gradient Boosting (GB), eXtreme Gradient Boosting (XGB), and AdaBoost (AB), was evaluated based on wildfire susceptibility mapping. The capabilities of the corresponding ML methods were assessed using thirteen wildfire conditioning factors, which can be grouped into four main categories: topographical, meteorological, vegetation, and anthropogenic factors. The Information Gain (IG) approach was used to assess their importance scores. A multicollinearity analysis was also performed to assess the relationship between conditioning factors. To compare the predictive performances of ML algorithms, five performance metrics, namely average accuracy, precision, recall, F1 score, and area under the curve, were used. To test the significance of the generated wildfire susceptibility maps and to detect similarities and differences among the output of these ML algorithms, McNemar's test was implemented. In the end, the ML-based models were locally interpreted using the Shapley Additive exPlanations (SHAP) technique. The AUC values of seven methods varied from 0.817 to 0.879, and the accuracy scores ranged between 0.734 and 0.812. The results showed that the RF model provided the best results considering all performance metrics. The accuracy score and AUC values of the RF model were equal to 0.812 and 0.879, respectively. On the other hand, stand-alone algorithms (LDA, SVM, and LR) represented lower performance than tree-based ensemble methods. Both the IG and SHAP analyses showed that elevation, temperature, and slope factors were the most contributing factors. The RF model classifier found that 7.20% of the study area has very high wildfire susceptibility, and the majority of the wildfire samples (68.84%) correspond to the very high susceptible areas in the RF model. The outcomes of this study are likely to provide decision-makers with a better understanding of wildfires in the Eastern Mediterranean Region of Turkey.

DOI:
10.1016/j.ecoinf.2022.101647

ISSN:
1878-0512