Self-Supervised Attenuation Method Based on Similarity Comparisons for 3D Seismic Random Noise

Abstract

The improvement of the signal-to-noise ratio of seismic data is crucial for high-precision processing. The self-supervised denoising methods based on correlation differences are gaining attention due to their low cost of training data construction and the ability to build training data directly on test data. However, as the volume of data processed increases, these methods must maintain a denoising effect by adding extra processings. This increase in processing times can lead to a rise in total cost, making these methods less cost-effective compared to other deep learning methods for processing large data. Therefore, we improved these self-supervised methods by altering the training data construction process, thereby retaining its cost advantage. Specifically, we modified the selection process of zones used for training data construction. Through correlation analysis, these methods can obtain higher-quality zones from the original zone for training, indicating that the network needs to be trained only once to process any data in the original zone. Without the requirement for additional processing to ensure noise attenuation, these selfsupervised methods can maintain their cost advantage when processing large data. We applied one of these self-supervised methods to synthetic and field examples to demonstrate the enhancement's effectiveness. Experimental results show that the improved method performs as well as the conventional self-supervised method in suppressing random noise and constructing reflection events but with a significant cost advantage.