A pseudo-continuous neural network approach for developing water retention pedotransfer functions with limited data

Abstract

Summary In this study, a new approach, which we called pseudo-continuous, to develop pedotransfer functions (PTFs) for predicting soil–water retention with an artificial neural network (ANN) was introduced and tested. It was compared with ANN PTFs developed using traditional point and parametric approaches. The pseudo-continuous approach has a continuous performance, i.e. it enables to predict water content at any desirable matric potential, but without the need to use a specific equation, such as the one by van Genuchten. Matric potential is considered as an input parameter, which enables to increase the number of samples in the training dataset with a factor equal to the number of matric potentials used to determine the water retention curve of the soil samples in the dataset. Generally, the pseudo-continuous functions performed slightly better than the point and parametric functions. The root mean square error (RMSE) of the pseudo-continuous functions when considering local data for training and testing, and with both bulk density and organic matter as extra input variables on top of sand, silt and clay content, was 0.027 m 3 m −3 compared to 0.029 m 3 m −3 for both the point and parametric PTF. The increased number of samples in the training phase and the selection of matric potential as an input variable enabling to predict water content at any desired matric potential are the most important reasons why pseudo-continuous functions would need more intention in the future. Uniformity in the training and test dataset was shown to be important in deriving PTFs. We finally recommend the use of pseudo-continuous PTFs for further improvement and development of PTFs, in particular when datasets are limited.