Modelling of soil moisture retention curve using machine learning techniques: Artificial and deep neural networks vs support vector regression models

摘要

Soil water retention curve (SWRC) is of fundamental importance in analyzing both flow and contaminant transport in the vadose zone. Field and/or laboratory-based measurements of soil moisture and soil suction - the two main variables that are used to develop SWRC - is often time consuming and sometimes impossible. In this study, plausibility of various machine learning techniques to simulate SWRC of loamy sand are evaluated. Specifically, the machine learning techniques that are investigated include: three support vector regression (SVR) models (i.e. radial basis function (RBF), linear and polynomial kernels), single-layer artificial neural network (ANN), and deep neural network (DNN). The soil moisture and soil suction were measured using time-domain reflectometer (TDR) and tensiometer, respectively. The data were collected under both monotonic wetting and drying of a disturbed sample of loamy sand soil. These datasets were used to train and test the machine learning techniques. Results show that the RBF-based SVR outperforms all the other machine learning techniques in simulating SWRC for loamy sand subjected to either monotonic wetting or drying. The ANN and DNN models simulated soil water content with a RMSE of 0.004-0.009 cm(3)/cm(3) for monotonic wetting in the training phase; and 0.002-0.003 cm(3)/cm(3) for monotonic drying in training phase. In the testing phase, ANN and DNN models simulated soil water content with RMSE of 0.02-0.121 cm(3)/cm(3) and 0.003. The RBF-based SVR model - the best performing machine learning model - simulated soil water content with RMSE of 0.006 and 0.002 cm(3)/cm(3) for soil subjected monotonic wetting and drying, respectively. In the testing phase, the RBF-based SVRmodel simulated soil water content with RMSE of 0.02-0.033 cm(3)/cm(3) and 0.003-0.006 cm(3)/cm(3) for soil under monotonic wetting and drying, respectively. These machine models, therefore, provided plausible SWRC simulations, and the models do not require knowledge of physical soil parameters.