Revisiting the pseudo continuous pedotransfer function concept: Impact of data quality and data mining method

Abstract The pedotransfer function (PTF) concept has been widely used in recent years as an indirect way to predict soil hydraulic properties, particularly the water retention curve (WRC). The pseudo continuous (PC) approach allows us to predict water content at any predefined matric head, resulting in an almost continuous WRC. When combined with powerful pattern recognition approaches, a PC-PTF can be trained to learn the shape of WRC from a discrete set of measured points, unlike traditional parametric PTFs which follow a predefined WRC shape dictated by the selected soil hydraulic equations. The purpose of this study was to investigate the impact of two elements on the performance of a PC-PTF: (i) data mining method (neural network, NN, versus support vector machine, SVM) and (ii) distribution and density of the provided water retention data in the training phase. Two datasets from Turkey and Belgium, consisting of mainly fine and coarse-textured soils, respectively, were employed. Multiple scenarios containing different combinations of measured water retention points in the training phase were defined. The lower root mean square error (RMSE) on average (0.044 cm3 cm− 3) attained with the NN-based PTF shows that it is a better option than SVM (RMSE of 0.052 cm3 cm− 3) for deriving PC-PTFs. The accuracy of PC-PTF was firmly dependent on the presence of measured water retention points in the entire range of WRC. Applying different scenarios revealed that a well distributed set of measured water retention points in the training phase could result in up to 0.03 cm3 cm− 3 reduction in RMSE values.