Estimation of the Soil Water Characteristics from Dielectric Relaxation Spectra -- a Machine Learning Approach

TL;DR

This paper presents a machine learning approach to estimate soil water characteristics from dielectric relaxation spectra, demonstrating a physical relationship and the potential of broadband electromagnetic sensors for soil analysis.

Contribution

It introduces a multivariate machine learning method to derive soil water properties from dielectric spectra, outperforming empirical and theoretical models.

Findings

MV-approach reveals a physical link between dielectric relaxation and soil water properties

Sensor techniques can estimate soil parameters from dielectric spectra

Machine learning improves soil water characteristic estimation

Abstract

The frequency dependence of dielectric material properties of water saturated and unsaturated porous materials such as soil is not only disturbing in applications with high frequency electromagnetic (HF-EM) techniques but also contains valuable information of the material due to strong contributions by interactions between the aqueous pore solution and mineral phases. Hence, broadband HF-EM sensor techniques enable the estimation of soil physico-chemical parameters such as water content, texture, mineralogy, cation exchange capacity and matric potential. In this context, a multivariate (MV) machine learning approach (principal component regression, partial least squares regression, artificial neural networks) was applied to estimate the Soil Water Characteristic Curve (SWCC) from experimentally determined dielectric relaxation spectra of a silty clay soil. The results of the MV-approach were compared with results obtained from empirical equations and theoretical models as well as a novel hydraulic/electromagnetic coupling approach. The applied MV-approach gives evidence, (i) of a physical relationship between soil dielectric relaxation behavior and soil water characteristics as an important hydraulic material property and (ii) the applicability of appropriate sensor techniques for the estimation of physico-chemical parameters of porous media from broadband measured dielectric spectra.