Self-adaptive denoising net: Self-supervised learning for seismic migration artifacts and random noise attenuation

Abstract

Seismic noise attenuation is essential for seismic interpretation and reservoir characterization. Recently, many researchers have applied convolutional neural network (CNN) to attenuate seismic noise. We propose a novel self-supervised CNN based approach to attenuate seismic random noise and migration artifacts simultaneously, termed as self-adaptive denoising net (SaDN). In this approach, we modify the loss-function of denoising CNN (DnCNN) based on the assumption that the synthetic noise with mixed Gaussian-Poisson distribution can simulate random noise and migration artifact. Moreover, we develop a new framework for the training and testing procedure. By adding the synthetic noise to seismic data, we can use the proposed CNN model to learn the characteristic of signal-dependent noise. The additive signal-dependent noise is self-adapted to the original seismic data according to the statistical property. Then, the well-trained SaDN model is applied to the original 3D seismic survey to recognize and extract the original random noise and migration artifact. The SaDN is training in the original seismic data with synthetic noise and testing in the same targeting seismic data without synthetic noise. Besides, our method does not require other denoising method for preprocessing in the training data. Thus, our proposed SaDN can be trained in a self-supervised way and adapted to different field data. Different experiments have been tested and compare with the method of CWT to illustrate the robustness and superiority of our proposed SaDN. The experiment results indicate that our proposed SaDN is able to remove the seismic random noise and migration artifacts without harming the effective signal.