Kernel methods for pier scour modeling using field data

Three kernel-based modeling approaches are proposed to predict the local scour around bridge piers using field data. Modeling approaches include Gaussian processes regression (GPR), relevance vector machines (RVM) and a kernlised extreme learning machine (KELM). A dataset consisting of 232 upstream pier scour measurements derived from the Bridge Scour Data Management System (BSDMS) was used. The radial basis kernel function was used with all three kernel-based approaches and results were compared with support vector regression and four empirical relations. Coefficient of determination value of 0.922, 0.922 and 0.900 (root mean square error, RMSE = 0.297, 0.310 and 0.343 m) was achieved by GPR, RVM and KELM algorithm respectively. Comparisons of results with support vector regression and Froehlich equation, Froehlich design, HEC-18 and HEC-18/Mueller predictive equations suggest an improved performance by the proposed approaches. Results with dimensionless data using all three algorithms suggest a better performance by dimensional data.