Monitoring Seismic Noise for Groundwater Dynamics Using Machine Learning

Climate change has led to more frequent and widespread droughts motivating robust monitoring of groundwater resources. Ambient seismic noise interferometry allows to derive relative seismic velocity changes (Δv/v) over time and space in the subsurface. Δv/v correlates well with groundwater fluctuations. Traditional datasets used to monitor groundwater changes, such as groundwater level data from wells and GRACE satellite gravimetric data, are either spatially sparse or limited in spatial resolution. Seismic velocity changes offer an additional, high-resolution measure of groundwater changes. Here, we aim to enhance groundwater monitoring in central Scandinavia, which experienced severe droughts in 2018 and 2022, and increase understanding on how groundwater levels decrease during droughts and recharge during periods of higher precipitations. One challenge of the ambient seismic noise interferometry method is the assumption of uniform noise sources, which rarely applies to seismic stations in Norway and Sweden. In this study, we test several denoising and spatial inversion robustness methods, including denoising autoencoders, convolutional neural networks, and variational inference. Through the integration of seismic and hydrological data, complex signal enhancement, and probabilistic inversion, we develop a robust method for monitoring groundwater in areas with heterogeneous station spacing and non-uniform noise sources.