Analysis of neural network based pedotransfer function for predicting soil water characteristic curve

• Published in: Geoderma Vol. 351; pp. 92 - 102
• Main Authors: Pham, Khanh, Kim, Dongku, Yoon, Yuemyung, Choi, Hangseok
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2019
• Subjects: algorithms; Bayesian learning; Bayesian theory; data collection; engineering; information processing; Neural network; neural networks; neurons; pedotransfer functions; prediction; soil types; soil water; soil water characteristic; SWCC; UNSODA; SWCC; Neural network; UNSODA; Bayesian learning
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/j.geoderma.2019.05.013

Approximate estimates employing the pedotransfer function (PTF) for predicting the soil-water characteristic curve have received general agreement among the geotechnical engineering society. Notably, the machine-learning-based PTFs (ML-PTFs) offer robust approaches with high prediction accuracy. This study analyzed the potential factors governing the prediction accuracy of the neural-network-based PTF (NN-PTF), which is one of the most popular ML-PTFs. Multiple analysis scenarios for the NN structure, learning algorithm and data processing were presented to evaluate the influence of these components. The analyses were performed on the UNsaturated Soil hydraulic DAtabase, which consists of a broad range of soil types. It is noted that employing the Bayesian regularization significantly improved the prediction accuracy for the same NN structure and optimizing algorithm when compared to using the early stopping, i.e., the maximum reduction in root mean squared error (RMSE) was 0.014. Architectural selection of the network worked most efficiently in case of the Bayesian regularization neural network (BRNN), i.e., RMSE dropped by 36% when the number of neurons increased from 9 to 54. Contrarily, an insignificant variation of RMSE indicated that increasing the NN complexity did not affect the performance of NN-PTF with the conjugate gradient descent and the early stopping. In addition, training the NN-PTF with a well-processed dataset could improve the prediction accuracy, i.e., the maximum reduction in RMSE was 0.004. Overall, the three-hidden-layer BRNN trained by the processed dataset outperformed the other scenarios in consideration, with RMSE = 0.028 and R2 = 0.977. Consequently, the data pre-processing and Bayesian regularization are strongly suggested for deriving the NN-PTF. •Explored potential factors governing the NN-PTF performance for predicting SWCC•Strong correlation between NN-structure and learning algorithm•Strongly suggested the Bayesian regularization for implementing the NN-PTF•Significant effect of the database properties on the prediction accuracy of NN-PTF•Optimal database for deriving the NN is defined by both volume and distribution.