PNNL

Abstract

Geophysical data can provide 3D, spatially extensive estimates of soil moisture. However, point-to-point comparisons to measured moisture data are frequently unsatisfactory, resulting in geophysics being used for qualitative purposes only. This is because geophysics (1) requires models to link interpreted data with soil moisture, (2) suffer from uncertainties, smoothing and/or artifacts introduced from processing and inversion, and (3) are not easily combined within a single moisture estimation framework. We performed an irrigation experiment where we monitored soil moisture through time and collected several surface geophysical datasets indirectly sensitive to soil moisture before and after irrigation: ground penetrating radar (GPR), electrical resistivity tomography (ERT), and frequency domain electromagnetics (FDEM). Data were exported in both raw and processed form, as well as evaluated in terms of soil moisture using common calibrations. Post-irrigation datasets were snapped to a common 3D grid to facilitate analysis and testing in multivariate regression and machine learning (support vector regression and random regression forests). Results were explored in terms of advantages/limitations of each geophysical method and their respective sensitivity and resolution to subtle soil moisture changes; however, for this site raw ERT pseudosection data were the most informative in terms of accurately predicting soil moisture using a random regression forest model (R2 from 0.49 – 0.96). This model was supported by both cross-validation and a separate training/test set which used data from physically separated portions of the study area. Machine learning was conducive to a semi-automated model selection process that could be used for other sites and datasets to locally improve accuracy.