Random noise attenuation using dual-module attention with multi-scale kernels

Improving the signal-to-noise ratio (SNR) of seismic data is a fundamental problem in exploration geophysics, as random and coherent noise severely degrade seismic record quality and hinder subsequent imaging and interpretation. Conventional methods often suppress weak but geologically informative signals along with noise, limiting their effectiveness. This research presents a novel Attention module, the Multi-Kernel Channel-Spatial Attention (MKCSA), that effectively addresses random seismic data noise. The network enriched its feature space to extract the complex geological features, effectively attenuating random noise. Through the integration of multi-kernel features extraction into Channel and Spatial attention, which yields more efficient extraction of local and global geological features. This Attention mechanism amplifies noise suppression even further, thereby reducing the signal in the local similarity graph. Experimental results show that the MKCSA-Net outperforms Denoising Convolutional Neural Network (DnCNN) and the DnCNN with Convolutional Block Attention (DnCNN-CBAM) model, improving Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) and Frequency Distance (FD) by ([Formula: see text], [Formula: see text], and [Formula: see text]) and ([Formula: see text], [Formula: see text], and [Formula: see text]) respectively.