Compressive sampling: a new paradigm for seismic data acquistion and processing?

Seismic data processing and imaging are firmly rooted in the well-established paradigm of regular Nyquist sampling. Faced with a typical uncooperative environment, practitioners of seismic data acquisition make all efforts to comply to this theory by creating regularly-sampled seismic-data volumes that are suitable for Fourier-based processing flows. The current advent of new alternative transform domains–- such as the sparsifying curvelet domain, where seismic data is decomposed into localized, multiscale and multidirectional plane waves–- opens the possibility to change this paradigm by no longer combating sampling irregularity but by embracing it. During this talk, we show that as long as seismic data volumes permit a compressible representation–-i.e., data can be represented as a superposition of relatively few number of elementary waveforms–- Nyquist sampling is unnecessary pessimistic. So far, nothing new, we all know from the work on Fourier- or other transform-based seismic-data regularization methodologies that wavefields can be recovered accurately from sub-Nyquist samplings through some sort of optimization procedure. What is new, however, are recent insights from the field of "compressive sampling", which dictate the conditions that guarantee or, at least, in practice provide conditions that favor sparsity-promoting recovery from sub-Nyquist sampling. Random sub-sampling, or to be more precise, jitter sub-sampling creates favorable conditions for curvelet-based recovery. We explain this phenomenon by arguing that this type of sampling leads to noisy data, hence our slogan "Simply denoise: wavefield reconstruction via jittered undersampling", where we bank on separating incoherent sub-sampling noise with curvelet-domain sparsity promotion. During our presentation, we introduce you to what curvelets are, why random jitter sampling is important and why this opens a pathway towards a new paradigm of curvelet-domain seismic data processing. Our claims will be supported by examples on synthetic and field data. This is joint work with Gilles Hennenfent, PhD. student at SLIM.