Boosting, bagging, and deep learning for soil-water characteristic curve prediction: a large-scale database study

Abstract

Several machine learning (ML) models have been developed in the past to predict the soil-water characteristic curve (SWCC). Most models utilize relatively small databases for training and testing. Under the emergence and growing adoption of new ML techniques, this paper presents a set of novel ML models for SWCC prediction, based on a combined database that is 4 times larger than the largest database used in previous models. Six databases with SWCC data from five continents were selected and combined for this study. In the end, two datasets were obtained from the unified, harmonized, and homogenized database: a) a complete dataset with 15,260 samples with 157,624 SWCC datapoints; b) a second dataset comprising 559 SWCCs that met specific quality criteria. The developed set of ML models was based on the grain-size distribution (GSD) as the only input variable and utilized artificial neural networks (ANN) with bagging techniques, and Gradient boosting (GB). The complete dataset was employed to establish a pseudo-continuous approach to predict the normalized volumetric water content, and the smaller dataset was used to predict the coefficients of the Gitirana-Fredlund (G-F) equation. The models using the pseudo-continuous approach proved superior when compared to the models to predict the G-F coefficients. The GB technique was superior to the use of the ANN ensemble, with a testing R2 value of 0.95 when using the pseudo-continuous approach. The developed models outperformed the Rosetta and NeuroFX ANNs, which showed an R2 for the test data of 0.68 and 0.61, respectively. Predictions for sands and silts were significantly superior to those for plastic clays or to bimodal soils that were underrepresented or absent in the dataset. Finally, the originality of the study lies in the development and use of an extensive database and the application of new ML techniques coupled with ANN and GB models for SWCC prediction. The main limitations are the difficulty in internalizing some important soil characteristics that affect SWCC, such as heterogeneity, anisotropy, plasticity, and volume variations.