Abstract

Seismic images play a crucial role in revealing subsurface geological structures and resource distribution for hydrocarbon exploration. However, seismic data often suffer from random noise, signal-to-noise ratio (SNR), and accuracy of geological interpretation. While traditional denoising methods exist, they typically require complex parameter selection, incur high computational costs, and are highly dependent on the operator's expertise. Deep learning (DL) techniques, on the other hand, offer promising alternatives with reduced reliance on user experience. Among DL methods, the Denoising Convolutional Neural Network (DnCNN) has demonstrated potential due to its local perception and weight sharing capabilities in signal processing. Despite this, DnCNN’s architecture has inherent limitations, often failing to adequately preserve signal structure and detail in denoised results. To address these shortcomings, we propose an Attention-based Denoising Convolutional Neural Network (ADnCNN), which enhances the DnCNN framework by integrating dual attention mechanisms. This novel model simultaneously captures spatial and channel-wise information, improving feature extraction and effectively preserving structural and textural details of seismic signals. Experimental results show that the proposed ADnCNN outperforms both conventional denoising approaches and standard deep convolutional neural network models in terms of quantitative evaluation and visual quality. The ADnCNN provides a robust solution for seismic data denoising, facilitating more accurate geological analysis and interpretation.