Pier scour modelling using random forest regression

Abstract

This paper investigates the potential of random forest–based regression approach to predict the local scour around bridge piers using field data set. A data set consisting of 232 field measurements were used. Comparison of results with M5 model tree, bagged M5 model tree, back-propagation neural network and four predictive equations suggests an improved performance by random forest regression approach in modelling the pier scour depth with dimensioned data. The use of random forest regression was also explored to judge the importance of each input variable in predicting bridge pier scour. Results in terms of increase in mean square error with the removal of each input variable suggest the importance of pier width and depth and flow in predicting the pier scour depth with random forest regression.