Seismic Random Noise Attenuation Based on M-ResUNet

Abstract

Suppressing random noise and improving the signal-to-noise ratio of seismic data are of great significance for subsequent high-precision processing. As one of the most popular denoising methods, the algorithms based on deep learning usually utilize the network's poor generalization ability for denoising processing. As a result, such methods usually face the problem of high training set construction costs and high computational costs. In addition, the widely used network UNet's processing of multiscale features is limited to its U-shaped processing path. Therefore, we proposed a denoising method based on unsupervised learning, including a new denoising strategy based on spatial correlation, the corresponding training method, and a new network M-ResUNet. The training method allows networks to be trained directly on the test area and conducts independent training and processing for subtest areas with different geological characteristics. This not only effectively solves the problems caused by the method using generalization ability for denoising but also improves the denoising accuracy to a certain extent. In addition, M-ResUNet breaks through the limitation of the U-shaped processing path and effectively improves the training effect by controlling the bias degree of the network output to features with different scales. Compared with the traditional denoising methods and networks, the results' test on synthetic and field data indicates that the proposed denoising method has superior performance in random noise attenuation, and M-ResUNet is effective.