A Lightweight Multi-Scale Network with High Frequency Learning for Seismic Data Denoising

Facing complex geological conditions, the acquired seismic data are often seriously contaminated by noise. How to recover seismic signals in a strong noise background remains a challenging issue. In this paper, we propose a Lightweight Cascaded Multi-Scale (LCMS) network for seismic data denoising, which efficiently and effectively mines seismic signals from low Signal-to-Noise Ratio (SNR) data while reducing computational costs. The network comprises a series of LCMS residual blocks to fully explore the multi-scale dense features of seismic data, in which each layer directly accesses the features of all previous layers to implement a continuous memory mechanism, not only enhancing feature utilization but also reducing the redundant computations. The architecture of network is optimized to achieve a compact structure, ensuring that it requires fewer parameters and computational resources compared to traditional deep networks, thus enabling faster processing without sacrificing performance. Furthermore, high-frequency information mainly presents seismic reflection details, and the lack of high frequencies affects the bandwidth of seismic data. To this end, we employ the non-subsampled shearlet transform (NSST) to acquire high-frequency prior information and integrate it into the network, which guides the network to focus more on the learning of high-frequency features. The qualitative and quantitative performance evaluations are conducted on synthetic data and real seismic data. Compared with the SOTA methods, our method achieves higher Peak Signal-to-Noise Ratio (PSNR) values and fewer parameters. When applied to real seismic data, our method efficiently recovers more effective signals of seismic data.