Seismic random noise attenuation via enhanced similarity self-supervised learning

Attenuation of seismic random noise is an important step of seismic data processing, and it is crucial to improve the signal-to-noise ratio (SNR) of seismic data. With the development of deep learning, various of DL based denoising models are proposed to solve the problems about the low effectiveness and the low efficiency of the traditional denoising methods. However, supervised learning models require a large amount of noise-free data to train the denoising network. In practical seismic data denoising, noise-free data are not available. To release the requirement of noise-free labels, we propose an enhanced similarity self-supervised learning (ESSL) model by effectively utilizing the self-similarity of seismic data. We first generate the training pairs by the proposed enhanced self-similarity based sampler and then train the denoising network with the generated training pairs. After the model training, we test the validity and effectiveness of our proposed model on both synthetic and field data.