Patch classification based EPLL with mixed GMM for nonstationary seismic random noise attenuation

Abstract

We propose an expected patch log likelihood (EPLL) based denoising method for nonstationary seismic random noise which has spatiotemporally varying variance, by incorporating the patch classification and mixed Gaussian mixture model (GMM), named PC-EPLL. In this method, the overlapping patches extracted from the noisy data are clustered into several classes based on a binary-tree, so that random noise in each class remains stationary. Thus the EPLL can denoise each class respectively by assigning proper parameters related to the noise characteristics of the class. Moreover, we combine the generic GMM learned from the nature image database and the target GMM learned from the seismic data by utilizing the weights related to the structural features of seismic data. In this way, the mixed GMM effectively captures the complex seismic features in a flexible learning way, and alleviates the over-fitting of the target GMM that is induced by the limited seismic training samples. Therefore, the PC-EPLL is capable of reconstructing the seismic signals from nonstationary seismic random noise. Synthetic and field seismic data are used to evaluate denoising performance of the proposed method. The results show that the proposed method achieves better performances in seismic event preservation and nonstationary random noise attenuation than the EPLL with the generic GMM and is superior to several state-of-the-art seismic denoising methods.