Comparison of artificial neural network and regression pedotransfer functions for prediction of soil water retention and saturated hydraulic conductivity

Merdun H., ÇINAR Ö., Meral R., Apan M.

SOIL & TILLAGE RESEARCH, vol.90, no.1-2, pp.108-116, 2006 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 90 Issue: 1-2
  • Publication Date: 2006
  • Doi Number: 10.1016/j.still.2005.08.011
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Page Numbers: pp.108-116
  • Keywords: artificial neural network, hydraulic parameters, pedotransfer function, prediction, regression, soil properties
  • Ondokuz Mayıs University Affiliated: No


Modeling water flow and solute transport in vadose zone requires knowledge of soil hydraulic properties, which are water retention and hydraulic conductivity curves. As an alternative to direct measurement, indirect determination of these functions from basic soil properties using pedotransfer functions (PTFs) has attracted the attention of researchers in a variety of fields such as soil scientists, hydrologists, and agricultural and environmental engineers. In this study, PTFs for point and parametric (van Genuchten's parameters) estimation of soil hydraulic parameters from basic soil properties such as particle-size distribution, bulk density, and three different pore sizes were developed and validated using artificial neural network (ANN) and multiple-linear regression methods and the predictive capabilities of the two methods was compared using some evaluation criteria. Total of 195 soil samples was divided into two groups as 130 for the development and 65 for the validation of PTFs. Although the differences between the two methods were not statistically significant (p > 0.05), regression predicted point and parametric variables of soil hydraulic parameters better than ANN. Both methods had lower accuracy in parametric predictions than in point predictions. Accuracy of the predictions was evaluated by the coefficient of determination (R-2) and the root mean square error (RMSE) between the measured and predicted parameter values. The R-2 and RMSE varied from 0.637 to 0.979 and from 0.013 to 0.938 for regression, and varied from 0.444 to 0.952 and from 0.020 to 3.511 for ANN, respectively. Even though regression performs insignificantly better than ANN in this case, ANN produces promising results and its advantages can be utilized by developing or using new algorithms in future studies. (c) 2005 Elsevier B.V. All rights reserved.