Design of a support vector machine with different kernel functions to predict scour depth around bridge piers

Abstract

Scour depth is a vital subject in bridge pier design. The exact estimation of scour depth can prevent damage caused by bridge failure and facilitate optimal bridge pier design. In this article, the support vector machine (SVM) method is applied to predict scour depth around bridge piers. The SVM technique is developed using six kernel functions, including polynomial, sigmoid, exponential, Gaussian, Laplacian and rational quadratic. Scour depth is modeled as a function of three dimensionless variables, namely geometric characteristics, flow and bed materials. The performance of SVM (in training and testing) is evaluated using dimensionless variables gathered from a wide range of field datasets. The SVM designed using the polynomial kernel function produced the most accurate results compared with the other kernel functions (RMSE = 0.078, MRE = −0.181, MARE = 0.332, MSRE = 0.025). Sensitivity analysis is performed to identify the effect of each dimensionless parameter on predicting scour depth around bridge piers. The testing results of SVM-polynomial are compared with that of the artificial neural networks (ANN), adaptive neuro-fuzzy inference systems (ANFIS) and nonlinear regression-based methods presented in this study and in the literature. Evidently, SVM-polynomial predicted scour depth with higher accuracy and lower error than when using ANN, ANFIS and nonlinear regression-based equations. Moreover, as an alternative method, a simple program is presented by the SVM-polynomial to calculate scour depth around bridge piers.