Self-Supervised Seismic Random Noise Suppression With Higher-Quality Training Data Based on Similarity Differences

Suppressing random noise and improving the signal-to-noise ratio of seismic data holds immense significance for subsequent high-precision processing. As one of the most widely used denoising methods, self-learning-based algorithms typically partition the large zone into several smaller zones for individual training and processing, thereby achieving lower training costs. However, as the volume of seismic data that needs to be processed continues to increase, the cost advantage of this method becomes less apparent. This is because a larger data volume necessitates more independent training, ultimately increasing the overall training cost. Therefore, we propose a denoising method based on self-supervised learning to overcome the aforementioned problem. This method can directly acquire higher-quality training data from large zones by leveraging similarity differences, decreasing the need to divide the large zone into smaller parts for individual processing. As a result, it can effectively reduce the times for individual processing, leading to a decrease in the overall training cost. Compared to traditional denoising methods and self-supervised learning methods, the experimental results on both synthetic and field data demonstrate that the proposed denoising method exhibits superior performance in random noise attenuation and reduction in training costs.