In this paper, we focus on denoising seismic signals effectively to obtain high-quality data, which is crucially important for oil-gas reservoir prediction and seismic interpretation tasks. The rapid progress of deep learning has brought new development opportunities to seismic oil and gas exploration technologies. However, current deep learning-based seismic denoising models have limited learning ability due to insufficient extraction features in strong noise backgrounds. For this reason, we develop a new multi-scale generative adversarial networks (GAN) with transform prediction for seismic signal denoising. First, we develop a deep multi-scale diversion fusion (MSDF) network in a GAN generator considering the advantages of combining GAN and convolutional neural networks. MSDF network primarily contains several MSDF blocks that mine abundant long-short path features in multiple receptive fields to restore seismic signals in a rough to detailed manner. Additionally, current deep learning-based seismic denoising models only exploit features in the spatial domain, not considering high-frequency characteristics, which results in insufficient high-frequency details when reconstructing seismic data. So, we propose to predict high-frequency components during learning by employing the superior non-subsampled contourlet transform (NSCT), which further preserves the better global topological structure and local texture characteristics of MSDF in GAN generator than the spatial domain, promoting the discrimination ability in GAN discriminator. The qualitative and quantitative results on our constructed synthetic dataset and actual seismic data demonstrate that the proposed method surpasses other deep learning-based approaches in realizing higher signal-to-noise ratio, as well as mining more effective high-frequency signals.

We propose a robust seismic noise attenuation method by an alternative denoising procedure using variational mode decomposition (VMD) combined with the Schrodinger equation thresholding, referred as VMD-Schrodinger interval-thresholding method. In the proposed method, VMD is used as a subband-like filter to get a series of uncorrelated intrinsic mode functions (IMFs) from low frequency to high frequency. For IMFs having strong correlations with the original signal, the Schrodinger equation thresholding-based denoising is used to obtain the denoised IMFs. For IMFs existing weak correlations with the original signal, the noise extracted from these IMFs is used to generate several different versions of noise by altering the sample positions. These different versions of noise combined with the information extracted from these IMFs are used to reconstruct different noisy versions of the original IMFs existing weak correlations. Then the Schrodinger equation thresholding-based denoising is applied to these different noisy data. The denoised IMFs existing weak correlations are the average of these denoised noisy versions signals. Finally, the denoised seismic data is the sum of these denoised IMFs. Compared with the wavelet thresholding method, iterative EMD interval-thresholding method and quantum mechanics-based signal denoising method, the proposed method can more effectively extract weak signals from the noise-dominant components of seismic data, while protecting both the low-frequency and high-frequency components of the seismic trace, especially exhibiting good performance for complex dipping events. The synthetic and field data applications illustrate the robustness and the superiority of the proposed method.