Rapid measurement of water diffusivity using a short horizontal soil column implementing a time-domain reflectometry (TDR) soil moisture sensor

Abstract

To clarify the effectiveness of machine learning approaches such as linear and random forest (RF) regressions in predicting apparent permittivity (ε ) in clayey soils, we obtained long-term ε and meteorological datasets from a reclaimed agricultural field and constructed prediction models of the ε at 6, 12, 24 and 48 h using these approaches. The predicted ε values from the linear regression model were generally consistent with the observed data, except during rainfall events. Although datasets for at least the last 72 h must be included as explanatory variables in such models, we confirm that the RF regression model could provide more accurate forecasts at the specified times than linear regression. A machine learning approach with RF regression would facilitate the autonomous prediction of ε values in clayey soils exhibiting structural changes based on the availability of long-term ε and meteorological datasets at the locations.