Estimation of soil water retention in conservation agriculture using published and new pedotransfer functions

Abstract

Conservation agriculture has been developed as a means to improve the sustainability of agricultural systems and reduce drawbacks of conventional agricultural practices. Cropping practices can have a large influence on soil properties such as water retention. Proper tools are needed to assess and model effects of conservation agriculture on soil properties. As measuring soil water retention is expensive and time consuming, pedotransfer functions (PTFs) are now commonly used to predict them. The objectives of this study were to (i) present a new dataset of conservation agriculture data, (ii) assess performances of existing PTFs in predicting soil water retention of soils under conservation agriculture and (iii) develop new specific PTFs to predict water retention in conservation agriculture more accurately. We used data collected only in fields under conservation agriculture in France to evaluate several published PTFs with three evaluation criteria (RMSE, prediction bias (ME) and Nash-Sutcliffe Efficiency (EF)). We then developed new PTFs using three methods ― multiple linear regression, regression tree and random forest ― to predict soil water content at matric heads of -100 (θ100, field capacity for sandy soils), -330 (θ330, field capacity for other soils) and -15 000 cm (θ15 000, wilting point). Soil tillage, presence of a cover crop, rotation length and previous reduced/no tillage were used as predictors in addition to basic soil properties for regression trees and random forests. The quality of prediction (RMSE, ME and EF) was calculated for each new PTF using a cross-validation procedure. Generally, predictions of wilting point had lower absolute error than those of sandy-soil field capacity (RMSE = 0.044 and 0.066 cm3/cm3, respectively). EF was usually negative for all water contents. The cross-validation performance of the new PTFs was similar for multiple linear regression (RMSE: 0.028, ME: 0.000, EF: 0.34 for θ100) and random forest (RMSE: 0.027, ME: 0.000, EF: 0.36 for θ100), and generally worse for regression tree (especially EF). Multiple linear regression that did not consider cropping practices performed as well as random forest and thus did not identify any major influence of agricultural management on predicted water content. Future research on developing PTFs should focus on identifying more relevant predictors.