Dual-domain mamba for seismic random noise suppression

Abstract

Noise suppression in seismic exploration is pivotal for recovering effective signals from noise-contaminated data. While mainstream deep-learning paradigms like convolutional neural networks (CNNs) and Transformers have demonstrated notable success in seismic denoising, their limitations remain pronounced. Specifically, CNNs prioritize local feature extraction at the expense of global context modeling, whereas Transformers suffer from quadratic computational complexity despite their superior global representation capacity. To address these limitations, we introduce Mamba, an emerging selective structured state space model (SSM), for seismic noise suppression. Mamba achieves efficient long-range dependency modeling with linear computational complexity, positioning it as a formidable competitor to Transformers. However, standard Mamba implementations face two key challenges in seismic applications: 1) requirement for substantial hidden states to memorize long-range dependency, inducing channel redundancy and impairing critical channel representation learning; and 2) neglect of frequency-domain characteristics essential for distinguishing noise from seismic signals. To address these challenges, we propose DDMamba, a dual-domain Mamba architecture synergistically unifying frequency-domain analysis with selective state space modeling. Specifically, we design a frequency-augmented state space module (FSSM) that harmonizes local-global perception via fast Fourier convolution (FFC) with Mamba's selective scanning mechanism, enabling joint frequency-spatial feature refinement. Additionally, we introduce a critical channel fusion module (CCFM) employing a multi-branch residual structure with channel attention and FFC to mitigate redundancy and enhance critical feature propagation. Synthetic and field experiments demonstrate DDMamba's superior denoising performance, with ablation studies validating the effectiveness of each proposed component.