An effective self-supervised learning method for attenuating various types of seismic noise

In real-world scenarios, the effectiveness of seismic denoising methods based on supervised learning (SL) is hindered by the scarcity of clean labeled data. A network trained on synthetic data often struggles to adapt to the distinct feature distributions of field data. To address this challenge, we develop an self-SL strategy that effectively attenuates various types of seismic noise using a single denoising network model. Our approach starts with a warm-up phase that pretrains the network on artificially noisier-noisy data pairs, enhancing its stability and ability to recognize specific noise characteristics. This is followed by an iterative data refinement phase, in which the model iteratively refines its predictions, narrowing the gap between noisy inputs and clean data. Field data tests demonstrate that our method outperforms traditional SL techniques on random, backscattered, and blending noise.