FIELD ASSESSMENT OF FDEM AND GPR TO RESOLVE SOIL MOISTURE VARIATIONS USING MACHINE LEARNING
On this page:
Methods to monitor the soil moisture distribution in landfills using non-invasive methods are desirable. Certain near-surface geophysical tools indirectly sense soil moisture variations, but these methods (1) require calibration to link interpreted data with soil moisture, and (2) suffer from smoothing and/or artifacts introduced from inversion. Machine learning methods address some of these challenges through an ability to extract a potentially complex (e.g. nonlinear) relationship between soil moisture and measured geophysical (and/or other) properties, and possibly circumvent the need for a traditional inversion step entirely. In this study, we were specifically interested in developing a relationship between frequency electromagnetic induction data (FDEM) and soil moisture. A field experiment was performed at a field site adjacent to the Connecticut River in Haddam, CT. At the site, about 3 m of unsaturated materials overlie approximately 30 m of saturated unconsolidated materials over crystalline bedrock. An initial characterization of the site with FDEM, electrical resistivity imaging (ERI), ground penetrating radar (GPR), and soil moisture probes was conducted. The site was then irrigated with 1000 µS/cm water, and another FDEM, ERI, GPR, and soil moisture dataset was collected. Data were divided into a training and test set, with the training set used to develop the relationship between geophysical data and soil moisture, and the test set used to evaluate the predictive power of that relationship. Two approaches were compared. In the first approach (approach 1), FDEM and ERI data were individually inverted to obtain electrical conductivity, which were then converted to soil moisture using Archie’s Law. Diffraction hyperbolas within GPR common offset data were fit to dielectric permittivity assuming low-loss conditions and were converted to average soil moisture above the hyperbolas using Topp’s equation. The soil moisture estimates from the different methods were combined through kriging, and the accuracy of this combined soil moisture image was evaluated by comparing to the test data set. In the second approach (approach 2), geophysical data (quadrature from multiple FDEM frequencies, apparent conductivity pseudosections from ERI data, and log absolute amplitude from common offset GPR data) from the training data set were fed directly into machine learning algorithms to predict soil moisture, and predictions were evaluated by comparing to the test soil moisture data set. Overall, the machine learning methods (approach 2) helped to automate data processing and develop site-specific relationships between soil moisture and geophysical data that may not be adequately captured by generalized petrophysical models (approach 1).