No abstract

Traditional least-squares full-waveform inversion (FWI) suffers from severe local minima problems in case of the presence of strongly dispersive surface waves. Additionally, recorded wavefields are often characterized by amplitude errors due to varying source coupling and incorrect 3D-to-2D geometrical-spreading correction. Thus, least-squares FWI is considered less than suitable for near-surface applications. In this paper, we introduce an amplitude-unbiased coherency measure as a misfit function that can be incorporated into FWI. Such coherency was earlier used in phase-weighted stacking (PWS) to enhance weak but coherent signals. The benefit of this amplitude-unbiased misfit function is that it can extract information uniformly for all seismic signals (surface waves, reflections, and scattered waves). Using the adjoint-state method, we show how to calculate the gradient of this new misfit function. We validate the robustness of the new approach using checkerboard tests and synthetic data contaminated by random noise. We then apply the new FWI approach to a field dataset acquired at an archaeological site located in Ostia, Italy. The goal of this survey was to map the unexcavated archaeological remains with high-resolution. We identify a known tumulus in the FWI results. The instantaneous-phase coherency FWI results also establish that the shallow subsurface under the survey lines is quite heterogeneous. The instantaneous-phase coherency FWI of near-surface data can be a promising tool to image shallow small-scale objects buried under shallow soil covers, as found at archaeological sites.

No abstract

High‐resolution reflection seismics is a powerful tool that can provide the required resolution for subsurface imaging and monitoring in urban settings. Shallow seismic reflection data acquired in soil‐covered sites are often contaminated by source‐coherent surface waves and other linear moveout noises (LMON) that might be caused by, e.g., anthropogenic sources or harmonic distortion in vibroseis data. In the case of shear‐wave seismic reflection data, such noises are particularly problematic as they overlap the useful shallow reflections. We have developed new schemes for suppressing such surface‐wave noise and LMON while still preserving shallow reflections, which are of great interest to high‐resolution near‐surface imaging. We do this by making use of two techniques. First, we make use of seismic interferometry to retrieve predominantly source‐coherent surface waves and LMON. We then adaptively subtract these dominant source‐coherent surface waves and LMON from the seismic data in a separate step. We illustrate our proposed method using synthetic and field data. We compare results from our method with results from frequency–wave‐number (f‐k) filtering. Using synthetic data, we show that our schemes are robust in separating shallow reflections from source‐coherent surface waves and LMON even when they share very similar velocity and frequency contents, whereas f‐k filtering might cause undesirable artefacts. Using a field shear‐wave reflection dataset characterised by overwhelming LMON, we show that the reflectors at a very shallow depth can be imaged because of significant suppression of the LMON due to the application of the scheme that we have developed.

For high-dimensional data analysis, dimensionality reducing is a common optimization means. A number of traditional multivariate statistical based approaches are applied and proposed recently, but cannot be solving dimensionality reduction problem well. The difficulty is caused by the fact that high-dimensional data generally do not have specific distribution or enough prior information. Aiming at the problem, an optimized dimensionality reduction model based on Restricted Boltzmann Machines (RBM) is presented. The model was optimized through adjusting the RBM hidden layer structure dynamically. Data distribution and prior information are not required in this model. Tests revealed the model performed well for handwritten digits data (get from the MNIST datasets) dimensionality reduction.