Unsupervised Learning With Waveform Multibranch Attention Mechanism for Erratic Noise Attenuation
Chenglin Fu, Yupeng Huo, Guodong Li, Yangkang Chen · IEEE Transactions on Geoscience and Remote Sensing · 2024
High signal-to-noise ratio (SNR) seismic data are crucial for oil and gas exploration, particularly for advanced high-precision migration processes and sophisticated reservoir inversion techniques. The erratic and random noise present in nature increases the difficulty of oil and gas exploration by reducing seismic signal quality. Deep learning (DL) networks facilitate rapid extraction of signal characteristics and effective noise reduction. The unsupervised DL U-Net network reconstructs signals through encoding and decoding. However, it overlooks the morphology and continuity of seismic data reflection waves, potentially missing detailed information in the reflected waveform. To address this, we introduce a waveform multibranch attention network (WMANet) for unsupervised learning (UL), designed to attenuate erratic and random noise using the U-Net architecture. WMANet uses the waveform attention mechanism to enhance the weight of waveform information in the network. In addition, the proposed loss function combining Welsch loss and local similarity enhances the erratic noise removal performance while effectively reducing the signal leakage phenomenon. Specifically, during the encoding, waveform information is captured using a fully connected layer capable of isolating critical data points. In the decoding stage, we integrate the waveform attention module with the fully connected module. This strategy gradually restores local information of the waveform while improving the efficiency of interbranch information utilization. We have evaluated our proposed method on both the synthetic and field datasets. The results demonstrate superior signal preservation and noise reduction compared with traditional denoising techniques.