Unsupervised Multi-Stage Deep Learning Network for Seismic Data Denoising

Seismic data denoising plays an essential role at various stages of the seismic processing workflow. However, it is always a challenge to find the right balance between preserving the seismic signals and attenuating the seismic noise. So, we propose a multi-stage deep learning model designed to suppress seismic noise with minimal signal leakage. Operating as a patch-based method, the model extracts overlapped patches from noisy data, flattening them into a 1D vector for input. The proposed model comprises two identical sub-networks with different configurations. Inspired by the transformer architecture, each sub-network uses an embedding layer, encompassing two fully connected layers, to extract features from the patches. Afterward, a multi-head attention module assigns a high attention weight to the important features. The primary difference between the first and second sub-networks lies in the number of neurons in their fully connected layers. The first sub-network acts as a strong denoiser with fewer neurons to attenuate the seismic noise, while the second sub-network serves as a weak denoiser with more neurons to retrieve the signal leakage from the output of the first sub-network. The proposed model has two outputs, each corresponding to a sub-network, and both sub-networks are optimized simultaneously. Testing on synthetic and field data demonstrates the model's capacity to suppress seismic noise with minimal signal leakage compared to benchmark methods.