The added value of CoreVESS score and penetration resistance in predicting the soil–water retention curve

Abstract

Soil hydraulic properties, such as the soil–water retention curve (SWRC), play a crucial role in simulating water transport within the vadose zone. They can be directly measured in the laboratory or field, or indirectly predicted by using Pedotransfer Functions (PTFs). It has been advocated that accounting for soil structure could marginally improve the prediction accuracy of PTFs in its wet range. The aim of this study was to test whether the use of an easy-to-determine soil structure-related variable could effectively increase the prediction accuracy of water retention curve PTFs. Additionally, we explored whether including soil strength, another soil property that is easy and quickly to determine in the field in many replicates, could improve PTF performance. Our investigation involved extensive sampling of 252 soil horizons across 42 cropped fields, each exhibiting within-field variations in soil structural degradation. Soil structure was represented by a soil structural quality score (Sq) taken from a semi-quantitative visual soil assessment (CoreVESS), which was included in the PTFs as predictor variable or as discriminator (Sq) for data grouping in addition to or instead of ‘classical’ predictors such as soil organic carbon content, clay and sand content, bulk density and soil layer. Penetration resistance (PR) was taken as a variable for soil strength. Various regression methods from classical regression to machine learning were evaluated on Train and Test subdatasets. To evaluate the effect of data grouping, the dataset was split in two based on structure and on texture. Both point and parametric PTFs were developed, with the latter predicting the parameters of the Van Genuchten (VG) water retention equation. Overall, the k-nearest neighbors PTFs (KNN-PTFs) achieved the best performance showing the highest R2 and lowest RMSE values. Moreover, the point PTFs showed much better and more stable predictive ability than parametric PTFs. The inclusion of PR as an additional predictor variable slightly improved the prediction performance in the wet range of the SWRC. Interestingly, Sq proved beneficial when substituted for bulk density but not when used in addition to it. Grouping data based on texture and structure improved the prediction accuracy of the PTFs, particularly evident in textural grouping and MLR, but not in KNN.