Estimating properties of a homogeneous bounded soil using machine learning models

TL;DR

This paper investigates machine learning techniques to accurately estimate soil properties from water moisture data, using simulated infiltration data and assessing model robustness under various data conditions.

Contribution

It introduces a framework for parameter estimation in soil models using ML, highlighting the effectiveness of SVMs and NNs in noisy and limited data scenarios.

Findings

SVMs and NNs outperform other models in accuracy and robustness.

Diffusivity D is predicted more accurately than hydraulic conductivity K.

Models maintain high performance even with limited or noisy data.

Abstract

This work focuses on estimating soil properties from water moisture measurements. We consider simulated data generated by solving the initial-boundary value problem governing vertical infiltration in a homogeneous, bounded soil profile, with the usage of the Fokas method. To address the parameter identification problem, which is formulated as a two-output regression task, we explore various machine learning models. The performance of each model is assessed under different data conditions: full, noisy, and limited. Overall, the prediction of diffusivity $D$ tends to be more accurate than that of hydraulic conductivity $K.$ Among the models considered, Support Vector Machines (SVMs) and Neural Networks (NNs) demonstrate the highest robustness, achieving near-perfect accuracy and minimal errors.