Premium
Machine learning to estimate soil moisture from geophysical measurements of electrical conductivity
Author(s) -
Moghadas Davood,
Badorreck Annika
Publication year - 2019
Publication title -
near surface geophysics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.639
H-Index - 39
eISSN - 1873-0604
pISSN - 1569-4445
DOI - 10.1002/nsg.12036
Subject(s) - petrophysics , electrical resistivity tomography , water content , soil science , artificial neural network , electrical resistivity and conductivity , transect , geology , pedotransfer function , geophysics , soil water , environmental science , machine learning , geotechnical engineering , hydraulic conductivity , porosity , computer science , engineering , oceanography , electrical engineering
ABSTRACT Soil water content (θ) is a key variable in different earth science disciplines since it mediates the water and energy exchange between the surface and atmosphere. Electrical and electromagnetic geophysical techniques have been widely used to estimate soil electrical conductivity (σ) and soil moisture. However, obtaining the σ − θ relationship is not straightforward due to the non‐linearity and also dependency on many different soil and environmental properties. The purpose of this paper is to determine if artificial neural network is an appropriate machine learning technique for relating electrical conductivity to soil water content. In this respect, time‐lapse electrical resistivity tomography measurements were carried out along a transect in the Chicken Creek catchment (Brandenburg, Germany). To ensure proper retrieval of the σ and θ, reference values were measured near the beginning of the transect via an excavated pit using 5TE capacitance sensors installed at different depths. We explored robustness and pertinence of the artificial neural network approach in comparison with Rhoades model (as a commonly used petrophysical relationship) to convert the inversely estimated σ from electrical resistivity tomography to the θ. The proposed approach was successfully validated and benchmarked by comparing the estimated values with the reference data. This study showed the superiority of the artificial neural network approach to the Rhoades model to obtain σ − θ relationship. In particular, artificial neural network allowed for more accurate estimation of the temporal wetting front than the petrophysical model. The proposed methodology thus offers a great promise for deriving spatiotemporal soil moisture patterns from geophysical data and obtaining the in situσ − θ relationship, taking into account the non‐linear variations of the soil moisture.