GIS-based seismic vulnerability mapping: a comparison of artificial neural networks hybrid models

Abstract

Earthquake hazards cause changes in landforms, economic losses, and human casualties. Seismic Vulnerability Mapping (SVM) is key information to prevent and predict the damage of earthquakes. The purpose of this study is to train and compare the results of the Classification Tree Analysis (CTA) learner model with three Gini, Entropy, Ratio split algorithms, and Fuzzy ARTMAP (FAM) model by the development of hybrid models for SVM. The Seismic Vulnerability Conditioning Factors (SVCFs) such as environmental, physical, and social were selected using experts' opinions and experience. Thirteen factors were edited and prepared as the seismic vulnerability conditioning factors (SVCFs) used in this study. In order to seismic vulnerability mapping and models training, a database of training sites was created by the Multi-Criteria Decision Analysis-Multi-Criteria Evaluation (MCDA-MCE) hybrid process. Then, 70% of the points were used for training and 30% were used to validate the models' results based on the holdout method. Moreover, Relative Operating Characteristics (ROC), Seismic Relative Index (SRI), and Frequency Ratio (FR) were used to validate the results. The Area under the curve (AUC) for the algorithms Gini, Entropy, Ratio, and FAM model are 0.895, 0.890, 0.876, and 0.783, respectively. The results of the three validation methods show the highest performance for the Gini splitting algorithm. Accordingly, the percentage of social and physical vulnerability of Sanandaj city was determined based on the MCE-Gini optimal model: 27% of the area and 62% of the population of Sanandaj are under high vulnerability to earthquakes. So that, various factors such as worn urban texture, high population density and environmental factors were among the most important factors affecting seismic vulnerability.