Enhancing seismic noise suppression using the Noise2Noise framework

ABSTRACT Efficient noise reduction is essential in seismic processing to improve signal clarity and subsurface imaging. Although supervised deep learning (DL) offers a solution for removing noise from seismic records, it faces challenges due to the scarcity of noise-free labels. To address this problem, we develop the Noise2Noise enhancement (N2NE) framework, which improves the noise reduction capabilities of conventional methods. By leveraging the Noise2Noise concept, this framework eliminates the requirement for clean reference signals, making it highly versatile for seismic processing. The framework is quantitatively examined using field noisy data with and without repeated shots. In the repeated-shot scenarios, the N2NE framework enhances the performance of conventional vertical stacking using deep neural networks trained on the repeated-shot pairs. Furthermore, using a substack strategy, which applies smaller substacks for preliminary noise suppression prior to DL training, boosts noise suppression. In scenarios without repeated shots, the cross-domain N2NE framework refines conventional denoising methods, such as f-x deconvolution, by using information from the common-shot and receiver domains. Quantitative assessments indicate significant improvements in the signal-to-noise ratio and structural similarity index, demonstrating the framework’s effectiveness. With the elimination of the need for noise-free labels, the N2NE framework improves data quality and also has the potential to enhance efficiency and reduce the costs of conventional reflection seismic surveys and long-term monitoring projects.