Indirect models for SWCC parameters: reducing prediction uncertainty with machine learning

Abstract

The soil–water characteristic curve (SWCC) is crucial for modelling the transport of water and hazardous materials in the vadose zone. However, measuring SWCC is often cumbersome and time-consuming. This paper introduces indirect models that predict SWCC parameters in probabilistic distributions using easily measurable quantities such as particle-size distributions and porosity. This paper starts with building a joint normal model and the derived conditional probability from it serves as a predictive model. However, this model had extremely high prediction uncertainty. To reduce such uncertainty, various machine-learning techniques were explored, including introducing the dependence of variation scale on predictors, using artificial neural networks (ANN) to model nonlinear dependence, incorporating additional predictive features, and generating a larger dataset. The final machine-learning model successfully reduces prediction variability and has been rigorously tested on a separate set of samples to prevent overfitting.