Knowing the centimeter- to meter-scale distribution of sand in clayey deposits is important for determining the dominating water flow pathways. Borehole information has a high vertical resolution, on the millimeter- to centimeter-scale, but provides poor lateral coverage. For highly heterogeneous deposits, such as glacial diamicts, this detailed borehole information may not be sufficient for creating reliable geological models. Crosshole ground-penetrating radar (GPR) can provide information on the decimeter- to meter-scale variation between boreholes, as the GPR response depends on the dielectric permittivity, electric conductivity, and the magnetic permeability of the subsurface. In this study, we investigate whether crosshole GPR can provide information on the material properties of diamicts, such as water content, bulk density, and clay content, as well as their structural relationships. To achieve ground truth, we compare the crosshole GPR data with geological information from both boreholes and excavation at the field site. The GPR data were analyzed comprehensively using several radar wave attributes in both time- and frequency domain, describing the signal velocity, strength, and shape. We found small variations in signal velocity (between 0.06-0.07 m/ns) but large variations in both amplitude and shape (either order of magnitude variation or doubling/tripling of attribute values). We see that the GPR response from wetter and more clayey diamicts have both lower amplitudes and lower centroid frequencies than the response from their drier and sandier counterparts. Furthermore, we find that the variation in amplitude and shape attributes are better correlated to the diamicts’ material properties than the signal velocity is.

Mapping high permeability sand occurrences in clayey till is fundamental for protecting the underlying drinking water resources. Crosshole ground penetrating radar (GPR) amplitude data have the potential to differentiate between sand and clay, and can provide 2D subsurface models with a decimeter-scale resolution. We develop a probabilistic straight-ray-based inversion scheme, where we account for the forward modeling error arising from choosing a straight-ray forward solver. The forward modeling error is described by a Gaussian probability distribution and included in the total noise model by addition of covariance models. Due to the linear formulation, we are able to decouple the inversion of traveltime and amplitude data and obtain results fast. We evaluate the approach through a synthetic study, where synthetic traveltime and amplitude data are inverted to obtain slowness and attenuation tomograms using several noise model scenarios. We find that accounting for the forward modeling error is fundamental to successfully obtain tomograms without artifacts. This is especially the case for inversion of amplitude data since the structure of the noise model for the forward modeling error is significantly different from the other data error models. Overall, inversion of field data confirms the results from the synthetic study; however, amplitude inversion performs slightly better than traveltime inversion. We are able to characterize a 0.4 - 0.6 m thick sand layer as well as internal variations in the clayey till matching observed geological information from borehole logs and excavation.