SPARSE DICTIONARY LEARNING FOR NOISE ATTENUATION IN THE EXACTLY FLATTENED DIMENSION

摘要

Seismic noise attenuation is a long-standing and crucial problem in reflection seismic data processing community. In recently years, the dictionary learning based approaches have attracted more and more attention. Dictionary learning provides an adaptive way to optimally represent a given dataset. In dictionary learning, the basis function is adapted according the given data instead of being fixed in many analytical sparse transforms. The application of the dictionary learning techniques in seismic data processing has been popular in the past decade. However, most dictionary learning algorithms are directly taken from the image processing community and thus are not suitable for seismic data. Considering that the seismic data is spatially coherent, the dictionary should better be learned according to the coherency information in the seismic data. We found the dictionary learning performs better when the spatial correlation is stronger and thus we propose to use a flattening operator to help learn the dictionary in the flattened dimension, where the strong spatial coherence helps construct a dictionary that follows better the structural pattern in the seismic data. The presented dictionary learning in the flattened dimension (DLF) thus has a stronger capability in separating signal and noise. We use both synthetic and field data examples to demonstrate the superb performance of the proposed method.