A Novel Approach to Train Self-Supervised Seismic Denoising Dnn Architectures

Abstract

Summary Removing noise present in seismic data is of prime importance for seismic processing workflows and a matter of continuous research in the academic community. The challenging part of seismic noise suppression is the diverse nature of seismic noise: it is found as a combination of random and coloured noise, which can be both structured and unstructured. Algorithms based on signal decomposition, domain transformation, and filtering, among others, have been traditionally applied to denoise seismic data and have been successful for specific imaging targets, hence mostly identifying a specific seismic noise component. Recently, convolutional neural networks-based (CNN) denoisers have greatly outperformed standard denoising techniques mostly in natural and medical imaging applications, and furthermore, self-supervised frameworks have been proposed as a clever alternative to denoising when no ground truth exists. This work leverages four state-of-the-art U-Net type architectures in a novel self-supervised fashion to remove seismic noise. The training seismic data corresponds to a generous number of real seismic surveys. For the labelling, trace-wise corruption is applied to patches of the input data, so the CNN learns to predict the corrupted traces based on the receptive field. Our findings indicate that self-supervised learning using U-Net type architecture trained on real data is able to considerably remove both structured and unstructured seismic noise.