Enhancing land seismic data with compressive sensing and processing

Abstract

Active-source seismic methods can provide a wealth of information about the structural and stratigraphic makeup of the subsurface as well as about physical and reservoir properties of rocks. However, high-resolution imaging subsurface techniques are challenging to apply for land seismic data. Unlike in marine acquisition where the source signal propagates through a near-homogeneous acoustic water layer, land acquisition places sources and receivers on a free-surface boundary between air and often poorly consolidated sediments. This near-surface layer traps most of the source-generated energy and gives rise to surface and guided waves and scattering noise that can propagate with extremely slow velocities. Wavefields triggered in the near surface have short wavelengths, and thus are difficult to acquire non-aliased which is the main source of discrepancy between marine and land seismic data quality. In this thesis, I research compressive sensing approaches to reduce the number of sensors required for non-aliased recordings of land wavefields and methods to improve regularly sampled aliased data. I consider a multi-channel extension of compressive sensing using both signal and its spatial derivative with a common sparse support constraint but conclude that due to noise and sub-optimal recovery algorithm such an approach yields negligible benefits over a single-channel variant. I establish that complex wavelet domain is an optimal choice for sparsifying highly non-stationary land wavefields for single-channel compressive sensing and develop thresholding techniques that can be used for a sparsity-promoting data reconstruction and for interpolation beyond aliasing. To demystify the complex wavelet domain, I represent complex wavelet coefficients on their idealized Fourier domain support with frequency-wavenumber octave bands representing different scales and orientations. Such approach provides the direct link between complex wavelet scales and orientations and phase velocities, enabling straightforward definitions of velocity filters that are localized in time, space, frequency and wavenumber and can therefore yield better signal and noise separation than the traditional approaches based solely on the Fourier domain.