Artificial neural networks for predicting soil water retention data of various Brazilian soils

Knowledge of the soil water retention (SWR) data is necessary for modeling soil water movement and assessing soil water holding capacity and availability. Since direct measurement is often time-consuming and costly, pedotransfer functions (PTFs) have been widely used to predict SWR data from basic soil physical properties. Considering the limited availability of PTFs derived from tropical soils, this paper developed artificial neural networks based on the pseudo-continuous approach (NN-PTFs) to predict SWR data for Brazilian soils. Natural logarithm of soil suction, ln (h), is considered as an extra input parameter in this approach. It enables to predict SWR data at any desired soil suction as it results in more extensive and useful database. The analysis was conducted on a previously compiled hydrophysical database for Brazilian soils representing a variety of soil compositions. The results demonstrated high accuracy and reliability in estimating SWR data, with an overall error of 0.045 cm³.cm^−³, when incorporating both soil texture (i.e., clay, silt, and sand fractions) and soil structure-related properties (i.e., soil density, particle density and organic matter content) as input parameters. Moreover, the proposed NN-PTFs outperformed PTFs developed for temperate climates, as well as equation-based PTFs derived for specific tropical locals, particularly for weathered soils. The results highlight not only the potential of using NN-PTFs to predict pseudo-continuous SWR curve in preliminary studies, but also their flexibility and the benefits of not limiting the SWR data to a pre-defined function.