A denoising method of seismic data based on self-supervised learning

Random noise adversely affects seismic signal resolution, subsequent interpretation, and reservoir prediction accuracy. In actual seismic exploration, it is often difficult and costly to obtain clean labels to train a denoising network. Compared with supervised learning, self-supervised learning does not need clean labels, but constructs supervised information according to the data itself. This paper presents a denoising network of seismic data based on self-supervised learning, which mainly includes four parts: data processing module, encoder, decoder, and residual noise separation module. The data processing module performs Bernoulli sampling on the input single 2D seismic signal to construct the supervision information. The encoder consists of four parts: partial convolution, dilated convolution, residual learning block, and down sampling. Dilated convolution can increase receptive fields and make the encoder better capture the features of useful signals. The decoder consists of up-sampling and standard convolution with a dropout strategy. The encoder and decoder use skip connections between the layers of the same height to realize the feature fusion of deep and shallow layers. The residual noise separation module obtains the predicted noise by calculating the residual between actual seismic data and predicted useful data, then uses the noise prior information as the regularization constraint to avoid the phenomenon of overfitting during training. The experimental results of synthetic and real seismic data indicate that our network not only suppresses random noise with effect, but also does have high fidelity.