High-resolution SH-wave crosshole seismic full-waveform inversion in the glacially overdeepened Tannwald Basin

During the Quaternary, the Rhine Glacier formed several overdeepened valleys, including the Tannwald Basin (ICDP site 5068_1, Germany) about 45 km North of Lake Constance. These structures form sedimentary climate archives and thus help to understand climate dynamics in the Alps.

To obtain very high-resolution images of the sediment, seismic crosshole data was acquired using a high-frequency borehole source that predominantly generates SH-waves. The source was excited very meter at 78 to 143 m depth, and the wavefield was recorded at a depth of 105 to 134 m using an eight-station three-component geophone string in a second borehole 28 m away. Given the receivers are spaced 2 m apart along the receiver string, it was moved by 1 m after shooting at all positions. The orientation of both the source and the receivers was done manually, though a calibrated compass attached to the receiver string facilitated this procedure, in constrast to the source orientation.

The SH-data is characterized by a high level of complexity, despite the lithology from a core obtained from one of the boreholes suggesting a predominantly homogeneous material, i.e., fine glaciolacustrine sediments. Additionally, the high-frequency, large-amplitude, long-coda P-wave masks the SH-wave arrivals.

In preparation for full-waveform inversion (FWI), we mute the trigger peak at time zero, perform data reorientation to account for misaligned sources and receivers, apply a 3D-to-2D spreading correction, delay the wavefield by 0.1 s to ensure convergence of the source-time-function inversion, and normalize the data shot-wise. In a previous study, we derived a traveltime tomography model from an additionally acquired SV-wave dataset, which we use as a starting model.

We apply 2D elastic mono-parameter (v_SH) time-domain FWI using the finite-difference method to invert the transverse component data. To mitigate the non-linearity of the problem, we use the multi-stage approach with frequencies starting at 100 Hz. To reduce the effects of source and receiver coupling, the global correlation norm is chosen as the misfit function. The misfit is minimized iteratively by means of an optimization through the quasi-Newton l-BFGS algorithm, which reduces the memory requirements and provides faster convergence. Furthermore, to reduce short-wavelength artifacts, the gradients are smoothed with a Gaussian filter. Source-time-function inversion is performed by a stabilized Wiener deconvolution in the frequency-domain using the Newton method with Marquardt-Levenberg regularization. Additionally, we apply frequency-adaptive time-windowing to precondition the data.

Despite the limited parameter space in the isotropic SH-case, the FWI does not yield convincing results. In this study, we explore the potential factors contributing to this outcome, including the data quality and properties, as well as our FWI approach.