Monitoring and modelling soil water dynamics using electromagnetic conductivity imaging and the ensemble Kalman filter

Abstract Understanding the soil water (θ) dynamics is important in irrigated agriculture. Due to the labour-intensive nature of determining θ, non-invasive electromagnetic (EM) induction techniques have been used. However, predicting depth-specific θ is a challenge because EM instruments measure the integral depth of the soil's apparent electrical conductivity (EC a ). Recently, “true” electrical conductivity with depth produced by inverse modelling of EC a has been employed to establish empirical models of θ. However, the potential to combine empirical models with soil physical models using data assimilation such as ensemble Kalman filter (EnKF) has not been explored. Along a 480-m transect with varying soil texture profiles, repeated EC a data were collected using a DUALEM-421S on 20 days over a 40-day period. A quasi-2d inversion algorithm was used to convert the temperature corrected EC a data to σ. An empirical model (artificial neural network) that predicted θ was calibrated using σ values, elevation, and topographic wetness index, as well as the mean and range of σ for each depth and each site over the whole study period. A physical soil-water tipping bucket model was also constructed using θ measured by soil moisture sensors. Afterwards, the EnKF approach was applied to 8 profiles along the transect separately combining both the empirical and the physical models. The results with the EnFK modelling (Lin's concordance = 0.89) were superior to the physical model and superior or equivalent to the empirical model for loam, clay and duplex soil profiles. In order to improve the prediction of θ dynamics, a more robust physical model could be used. In addition, correction of the diurnal effects of EC a data and redefining the model and measurement errors of the EnFK to account for the temporal dependence should be considered.