Attenuation of linear noise based on denoising convolutional neural network with asymmetric convolution blocks

Source-generated linear noise is one of the most common types of noise in seismic data. One or more groups of slanted linear events with strong energy often appear in seismic records and severely degrade the quality of subsurface reflections. Therefore, how to effectively remove this noise is the key to improve the quality of reflections. Here, we use a method that combines a feed-forward denoising convolutional neural network (DnCNN) with asymmetric convolution blocks (ACB) to attenuate linear noise in seismic data. Compared with traditional filter methods, this method involves less assumptions concerning the signals and noise; we merely train the neural network to recognise the features of reflections in seismic data. The DnCNN is a supervised deep learning method. It needs sufficient training data to optimise network parameters. Therefore, we generate numerous pairs thereof - including synthetic and real seismic data - to feed to the network. This input of data enables the network to identify reflections in seismic data directly and thus obtain the denoised data. Based on the characteristics of linear noise in seismic data, we build an asymmetric network architecture by combining the DnCNN with ACB. This enables the network to develop an ability to automatically identify reflected signals in seismic data. To validate the performance of the proposed method, we apply it to synthetic and real seismic data. The results demonstrate the method can effectively identify signals from noisy data and obtain better results in attenuation of linear noise and preservation of signals compared with the four other methods.