Spatiotemporal monitoring of soil moisture from EMI data using DCT-based Bayesian inference and neural network

Loop-loop electromagnetic induction (EMI) has proven to be efficient for fast and real-time soil apparent electrical conductivity (ECa) measurements. It is important to develop robust and accurate inversion strategies to obtain soil electromagnetic conductivity image (EMCI) from ECa data. Moreover, obtaining an accurate non-linear relationship between subsurface electrical conductivity (σ) and water content (θ) plays a key role for soil moisture monitoring using EMI. Here, we incorporated probabilistic inversion of multi-configuration ECa data with dimensionality reduction technique through the discrete cosine transform (DCT) using training image (TI)-based parametrization to retrieve soil EMCI. The ECa data were measured repeatedly along a 10 m transect using a CMD mini-Explorer sensor. Time-lapse reference data were collected as well to benchmark the inversion results and to find the in-situ relationship between σ and θ. To convert the inversely estimated time-lapse EMCI to the soil moisture, we examined two approaches, namely, Rhoades et al. (1976) model and artificial neural network (ANN). The proposed inversion strategy estimated the soil EMCI with an excellent agreement with the reference counterpart. Moreover, the ANN approach demonstrated superiorities than the commonly used petrophysical model of Rhoades et al. (1976) to obtain spatiotemporal images of θ from time-lapse EMCI. The results demonstrated that incorporation of the DCT-based probabilistic inversion of ECa data with the ANN approach offers a great promise for accurate characterization of the temporal wetting front and root zone soil moisture.