This study was conducted to evaluate ten closed‐form unimodal analytical expressions to describe the soil‐water retention curve, in terms of their accuracy, linearity, Akaike Information Criterion (AIC), and prediction potential. The latter was evaluated by correlating the model parameters to basic soil properties. Soil samples were taken in duplicate from 48 horizons of 24 soil series in Flanders, Belgium. All sample locations were under forest and hence the samples had, besides their difference in texture, a high variety in bulk density (ρb) and organic matter content (OM). The van Genuchten model with m as a free parameter showed the highest overall performance in terms of goodness‐of‐fit. It had the highest accuracy, the highest degree of linearity, and the lowest AIC value. However, it had a low prediction potential. Imposing the constraint m = 1 − 1/n and hence reducing the number of model parameters by one, increased the prediction potential of the model significantly, without loosing much of the model's accuracy and linearity. A high degree of accuracy and linearity was also observed for the two Kosugi models tested. Restricting the bubbling pressure to be equal to zero resulted in a rather high prediction potential, which was not the case when keeping the bubbling pressure as a free parameter. A major drawback of van Genuchten and Kosugi type models is that they do not define the soil‐water retention curve beyond the residual water content. We further demonstrated that the performance of all but one model in terms of their match to the data increased with increasing clay content and decreasing sand content, which is contradictory to the deterministic character of these models. Bulk density and OM did not have a significant effect on the accuracy of most models.

[1] Prediction of water and vapor flow in porous media requires an accurate estimation of the soil water retention curve describing the relation between matric potential and the respective soil water content from saturation to oven dryness. In this study, we modified the Kosugi (1999) function to represent soil water retention at all matric potentials. This modification retains the form of the original Kosugi function in the wet range and transforms to an adsorption equation in the dry range. Following a systems identification approach, the extended function was tested against observed data taken from literature that cover the complete range of water contents from saturation to almost oven dryness with textures ranging from sand to silty clay. The uncertainty of parameter estimates (confidence intervals) as well as the correlation between parameters was studied. The predictive capability of the extended model was evaluated under two reduced sets of data that do not contain observations below a matric potential of À1500 and À100 kPa. This evaluation showed that the extended model successfully predicted the water content with acceptable uncertainty. These results add confidence into the proposed modification and suggest that it can be used to better predict the soil water retention curve, particularly under reduced data sets.

The objective of this work was to evaluate eight closed‐form unimodal analytical expressions that describe the soil‐water retention curve across the complete range of soil water contents. To meet this objective, the eight models were compared in terms of their accuracy (RMSE), linearity (coefficient of determination, R2, and adjusted coefficient of determination, R2adj), and prediction potential. The latter was evaluated by correlating the model parameters to basic soil properties. Retention data for 137 undisturbed soils from the Unsaturated Soil Hydraulic Database (UNSODA) were used for the model comparison. The samples showed considerable differences in texture, bulk density, and organic matter content. All functions were found to provide relatively realistic fits and anchored the curve at zero soil water content for the coarse‐textured soils. The performance criteria were similar when averaged across all data sets. The criteria were found to be statistically different between the eight models only for the sandy clay loam soil textural class. An analysis of the individual data sets separately showed that the performance criteria were statistically different between the models for 17 data sets belonging to six different textural classes. We found that the Khlosi model with four parameters was the most consistent among different soils. Its prediction potential was also relatively good due to significant correlation between its parameters and basic soil properties.

Knowledge of soil hydraulic properties is of major importance for land management in dry‐land areas. The most important properties are the soil–water retention curve (SWRC) and hydraulic conductivity characteristics. Direct measurement of the SWRC is time and cost prohibitive. Pedotransfer functions (PTFs) use data mining tools to predict SWRC. Modern data mining techniques enable accurate predictions and good generalization of SWRC data. In this research we explore whether the use of support vector machines (SVMs) could improve the accuracy of prediction of SWRC. The novelty of our work is in the application of SVM data mining techniques, which are seldom used in soil research, to a limited dataset from Syria. The soil studied is calcareous and the climate is arid, for which no PTFs have been developed. Seventy‐two undisturbed soil samples were taken from four different agro‐climatic zones of Syria. The soil water contents at eight matric potentials were determined and selected as output variables. The data were split into two subsets: a training set with 54 samples for model calibration or PTF development and a test set with 18 samples for PTF validation. An overview of the theoretical foundation of this new approach and the use of specific kernel functions is given. Then, the model parameters were optimized with ninefold cross‐validation and a grid search method. The predictions of the SVM‐based PTFs were analysed with the coefficient of determination (R2) and root mean square error (RMSE). Our results showed that the accuracy of SVM was better in terms of RMSE and R2 than multiple linear regression (MLR) and the artificial neural network (ANN). The results support previous findings that the SVM approach performs better than MLR and the ANN. Furthermore, improvements in predictions of SWRC with the three data mining techniques were obtained by replacing the more conventional organic matter in the PTF with the plastic limit (PL). Therefore, SVM and PL markedly improved the accuracy of prediction of SWRC for calcareous soil. Highlights Pedotransfer functions (PTFs) to predict the soil–water retention curve of calcareous soil. Improved prediction of water retention with support vector machines (SVMs) The plastic limit (PL) appeared to be a significant predictor variable. The results suggest the use of SVMs and PL to improve and develop PTFs further.

In arid and semiarid areas, the availability of reliable data for water retention in relation to soil type, texture, and soil carbonate content is low. It is therefore desirable to explore the interaction between soil hydraulic properties and other physical and chemical properties to estimate the soil water retention curve (SWRC) from easily measured soil parameters. In this study, 72 soil samples were collected from rural areas throughout northwest Syria, covering most of its agroclimatic zones and soil types. Soil water content at different matric potentials and 11 chemical and physical soil properties were determined. A Pearson correlation matrix was computed on which principal component analysis was applied to three soil water contents, −1, −33, and −1500 kPa, and the 11 soil properties. Four principal components (PCs) explained 77% of the variation in the data set. The three soil water contents were highly linked to PC1, which was correlated with the plastic limit, texture, soil carbonate content, and specific surface area. In addition, the soil water content at −1 kPa was also linked to PC4, which was correlated with bulk density. Therefore, from the initial 11 soil properties, seven contributed to the three soil water contents (plastic limit, texture, soil carbonate, specific surface area, and bulk density); the remaining four (organic matter, gravel, cation exchange capacity, and hygroscopic water content) had a negligible influence. Consequently, pedotransfer functions might be estimated using the original seven, from the initial 11, soil properties or their corresponding PCs to estimate the SWRC.