Artificial neural networks for the prediction of the soil-water characteristic curve: an overview

Abstract

Several models have been developed for predicting the soil-water characteristic curve (SWCC). Artificial Neural Networks (ANN), which are machine learning systems based on biological neural networks, are among the techniques used for SWCC prediction. This article presents a literature review of the main SWCC prediction models created using ANNs, including a detailed overview and comparison of the different architectures and their selected hyperparameters. The paper demonstrates the progress made in the field, indicates the next challenges for SWCC prediction using ANNs, and provides a benchmark for the construction of future networks. It is shown that three main approaches have been adopted when modeling SWCC using ANN: predicting the water content corresponding to suction values predefined by the ANN developer or to suction values specified by the user and predicting the coefficients of common fitting equations. Most models used volumetric water content, while some used normalized volumetric water content, and gravimetric water content to quantify water storage. These models usually have a single hidden layer, with the number of neurons per layer varying between 1 and 400. The main input variables are the complete grain-size distribution, soil texture parameters, bulk density, and porosity. The databases used for the development of each model varied in size, comprising between 23 and 3850 SWCCs and between 794 and 13230 water content data points. The developed models have an average R2 of 0.75 and an RMSE of 0.05 cm3 cm−3. Most models were developed using databases of unimodal materials from temperate regions, with only one model having considered bimodal and tropical soils. Current ANN designs seem compatible with the size of existing databases and input data but will require revisions as more soil information becomes available.