Comparison of Support Vector Machine Kernel Functions for Seismic Vulnerability Assessment

Earthquakes are considered the most destructive of natural disasters, which cause physical, human and environmental losses. Therefore, it is very important to identify areas vulnerable to seismic disasters in advance for sustainable management. In this study, a seismic vulnerability assessment was conducted in Gyeongju city, South Korea, where the M 5.8 earthquake occurred September 12, 2016. The main goal is to evaluate the predictive performance using four different kernel support vector machine (SVM) technique with for assessment. Eighteen seismic-related factors were used as independent variables for seismic vulnerability mapping including slope, elevation, groundwater level, distance to seismic epicenter, distance to fault, peak ground acceleration (PGA), age of children, age of elderly, population density, building density, construction materials, number of floors, age of buildings, distance to police station, distance to fire station, distance to hospital, distance to gas station, and distance to road network. Buildings damaged by the 9.12 Gyeongju earthquake were randomly classified as training and verification data and used as dependent variables (70:30). Next, the training data set generated model for each of the four kernels: the linear (LN), polynomial (PL), radial basis function (RBF), and sigmoid (SIG) functions of the SVM. Finally, each model was evaluated using receive operating characteristics (ROC) curves based on a validation data set for performance comparison. The results showed that radial basis function (RBF) was identified as the best kernel function with the highest predicted accuracy, followed by the polynomial (PL) kernel function. This study provides examples to select the appropriate type of SVM kernel functions for seismic vulnerability assessment and mapping.