2D seismic random noise attenuation by self-pace nonnegative dictionary learning

Seismic random noise always exists in seismic data acquisition. Dictionary learning methods are the effective tools to find a typical sparse representation. It can be introduced to reduce the seismic random noise. According to the geometric structures of 2D seismic data, the effective seismic data are also low-rank in time-space domain. Therefore, the effective seismic data can be models as a linear combination of a few elements from a learned dictionary. In this paper, we propose a new dictionary learning method to reduce the 2D seismic random noise in time-space domain by incorporating self-pace (SP) learning method with the nonnegative dictionary learning. This method is implemented by sequentially including matrix elements into nonnegative matrix factorization (NMF) training from easy to complex. It can void bad local minima and obtain a high SNR seismic data. The effectiveness of the proposed self-pace nonnegative matrix factorization (SPNMF) method is demonstrated by the synthetic data and the field data.