We used neural networks (NNs) to model the drying water retention curve (WRC) of 204 sandy soil samples from particle-size distribution (PSD), soil organic matter content (SOM), and bulk density (BD). Neural networks can relate multiple model input data to multiple model output data without the need of an a priori model concept. In this way a high performance black-box model is created, which is very useful in a data exploration effort to assess the maximum obtainable prediction accuracy. We used a series of NN models with an increasing parametrization of input and output variables to get a better interpretability of model results. In the first two models we used the nine PSD fractions, BD, and SOM as input, while we predicted the nine points of the water retention curve. These NNs had 12 input and 9 output variables, predicting WRCs with an average root-mean-square residual (RMSR) water content of 0.020 cm3 cm-3. After a few intermediary models with increasing parametrization of PSD and WRC using (adapted) van Genuchten  equations we arrived at a final NN model that used six input variables to predict three van Genuchten  parameters resulting in a RMSR of 0.024 cm3 cm-3. We found saturated and residual water contents to be unrelated to the PSD, BD, or SOM, therefore the saturated water content was considered to be an independent input variable, while the residual water content was set to zero. Sensitivity analyses showed that the PSD had a major influence on the shape of the WRC, while BD and SOM were less important. On the basis of these sensitivity analyses we established more explicit equations that demonstrated similarity relations between PSD and WRC and incorporated effects of SOM and BD in an empirical way. Despite the fact that we considered a large number of linear and nonlinear variants these equations had a weaker performance (RMSR: 0.029 cm3 cm-3) than the NN models, proving the modeling power of that technique.
ASJC Scopus subject areas
- Water Science and Technology