Seismic data processing with the parallel windowed curvelet transform

The process of obtaining high quality seismic images is very challenging when exploring new areas that have high complexities. The to be processed seismic data comes from the field noisy and commonly incomplete. Recently, major advances were accomplished in the area of coherent noise removal, for example, Surface Related Multiple Elimination (SRME). Predictive multiple elimination methods, such as SRME, consist of two steps: The first step is the prediction step, in this step multiples are predicted from the seismic data. The second step is the separation step in which primary reflection and surface related multiples are separated, this involves predicted multiples from the first step to be matched with the true multiples in the data and eventually removed. A recent robust Bayesian wavefield separation method have been recently introduced to improve on the separation by matching methods. This method utilizes the effectiveness of using the multi scale and multi angular curvelet transform in processing seismic images. The method produced excellent results and improved multiple removal. A considerable problem in the seismic processing field is the fact that seismic data are large and require a correspondingly large memory size and processing time. The fact that curvelets are redundant also increases the need for large memory to process seismic data. In this thesis we propose a parallel approach based windowing operator that divides large seismic data into smaller more managable datasets that can fit in memory so that it is possible to apply the Bayesian separation process in parallel with minimal harm to the image quality and data integrity. However, by dividing the data, we introduce discontinuities. We take these discontinuities into account and compare two ways that different windows may communicate. The first method is to communicate edge information at only two steps, namely, data scattering and gathering processes while applying the multiple separation on each window separately. The second method is to define our windowing operator as a global operator, which exchanges window edge information at each forward and inverse curvelet transform. We discuss the trade off between the two methods trying to minimize complexity and I/O time spent in the process. We test our windowing operator on a seismic denoising problem and then apply the windowing operator on our sparse-domain Bayesian primary-multiple separation.