A seismic random noise suppression method based on CNN-Mamba

BackgroundSeismic random noise suppression is recognized as a key step to improve the quality of seismic data. Data-driven deep learning provides an intelligent solution for the noise suppression. However, mainstream random noise intelligent methods based on convolutional neural networks (CNNs) are constrained by their local receptive fields. This limitation results in insufficient collaborative optimization between local details and macroscopic structures during denoising, further reducing the noise suppression accuracy. Transformer models, which are widely applied to global feature extraction, can effectively capture long-distance dependencies through the self-attention mechanism, theoretically overcoming the limitations of CNNs in global modeling. However, these models face challenges such as slow computation, high resource consumption, and limited applications.Objective and Methods To address these issues, this study proposed a CMUNet seismic random noise suppression network that integrates CNN and Mamba. Based on the 2D-selective-scan (SS2D) mechanism, which can traverse the input data along horizontal and vertical directions, a global dynamic system was constructed using state-space equations. This enabled the trans-scale feature extraction of the spatiotemporal characteristics of seismic data. The hardware-aware parallel algorithm of Mamba was employed to reduce the computational resource consumption, thus ensuring the denoising performance while enhancing computational efficiency. Targeting the characteristics of seismic data, this study designed a CNN-Mamba hybrid block to construct hierarchical feature extraction pathways in the UNet encoder. Specifically, the CNN in a shallow layer focused on local noise pattern recognition, while Mamba in a deep layer was used to capture the correlations of large-scale geological structures. Additionally, residual channel attention gating was further introduced to enhance the feature separability between effective signals and noise. Results and Conclusions The results indicate that for synthetic data, the proposed CMUNet network increased the signal-to-noise ratio (RS/N), peak signal-to-noise ratio (RPSN), and structural similarity by 2.4 dB, 2.4 dB, and 0.005 6, respectively compared to UNet. These results suggest that the CMUNet network enables effective random noise suppression and preserves effective signals. This network was applied to field seismic data. An image-based local similarity analysis reveals that the network yielded low local similarity, further corroborating that it causes minimal damage to effective signals and exhibits superior amplitude preservation. Therefore, the CMUNet network proposed in this study holds great potential for application.