Comparing Neural Networks, Linear and Nonlinear Regression Techniques to Model Penetration Resistance

Authors : Hossein BAYAT, Mohammad Reza NEYSHABOURI, Mohammad Ali HAJABBASI

Pages : 425-433

View : 14 | Download : 3

Publication Date : 2008-09-01

Article Type : Research Paper

Abstract :Penetration resistance insert ignore into journalissuearticles values(PR); is an important property of soils, and can be expressed as cone index insert ignore into journalissuearticles values(CI);. Because of high variability, there are no accurate and representative PR data in most cases. Variable PR is considerably affected by gravimetric soil water content insert ignore into journalissuearticles values(GWC); and bulk density insert ignore into journalissuearticles values(BD);. In this study, artificial neural networks insert ignore into journalissuearticles values(ANNs); were used to simulate relationship between BD, GWC, and CI. A data set of 381 samples was collected from 2 study sites, Hamadan and Maragheh. Pedotransfer functions insert ignore into journalissuearticles values(PTFs); were developed using ANNs and linear and nonlinear regression models to predict CI for the combined data set and each data set separately. For the combined and Hamadan data sets, ANNs produced a greater correlation coefficient insert ignore into journalissuearticles values(R = 0.85); and lower root mean square error insert ignore into journalissuearticles values(RMSE); compared with the linear regression model insert ignore into journalissuearticles values(R = 0.70);. For the Maragheh data set, however, the regression model yielded better results. Introducing TP and relative saturation insert ignore into journalissuearticles values(Qv/TP); into the models improved the prediction of CI. The results further showed that ANN models performed better than nonlinear regression models. Therefore, ANNs were recognized as powerful tools to predict CI by BD, GWC, TP, and Qv/TP as the independent variables under the very diverse conditions of the soils and treatments employed.
Keywords : Artificial neural networks, bulk density, cone index, regression models, water content