Sub-seafloor Pore-fluid Salinity Estimation in the Canterbury Bight, New Zealand, based on Bayesian Inversion of Controlled-source Electromagnetic Data

Offshore groundwater systems have been suggested as unconventionalsources of portable water in islands and densely populated coastal regions, where terrestrial groundwater resources are over-extracted or contaminated. In this study, we evaluate how well pore-water salinity can be efficiently estimated from controlled-source electromagnetic (CSEM) resistivity models. Time-domain CSEM methods are an effective tool to explore offshore groundwater bodies since the electrical resistivity of the seafloor is primarily determined by the characteristics of the pore-water within the sub-surface sediments. We integrate offshore TD-CSEM data with borehole data and multichannel seismic reflection to identify an offshore groundwater system in the Canterbury Bight. The Canterbury margin is located off the eastern coast of the South Island of New Zealand and was previously investigated during IODP Expedition 317 in which a pore-fluid salinity anomaly was recorded in borehole U1353. By focusing on the low-salinity zone, we carry out synthetic modelling, implement Bayesian inversion to derive probability density functions for resistivity as a function of depth, and finally transform it to distribution of probability density for pore-fluid salinity using Archie’s equation. Subsequently, we use a trans-dimensional Bayesian inversion to interpret measured CSEM data, and consequently use the derived probability density functions to estimate uncertainty of pore-water salinity. Having done the same procedure for different waypoints in the Canterbury Bight, we extrapolate salinity values on a basin scale and estimate the probability density distribution for pore-water salinity as a function of depth. We show that using the Bayesian sampling algorithm provides us with a more precise estimation of hydrogeological model parameters with their uncertainties by generating an ensemble of models instead of inferring only one model using deterministic inversion approaches.