Low-wavenumber noise attenuation of seismic reverse time migration with deep learning

Low-wavenumber noise attenuation in seismic reverse time migration (RTM) is crucial for the accurate identification of geological structures, seismic inversion, and reservoir characterization. However, the significant computational overhead associated with RTM limits its practical deployment. To address this, we propose a novel seismic low-wavenumber noise attenuation method based on residual blocks and an encoder–decoder architecture (U-Resnet). Firstly, an extensive seismic dataset is generated by employing RTM imaging combined with Hilbert transform wavefield decomposition. Subsequently, the U-Resnet architecture is developed and applied to attenuate noise from seismic RTM imaging results. The effectiveness of our method is validated through RTM datasets, including the SEG/EAGE salt model, the Marmousi model, and the Overthrust model. Results from the salt model show that, compared to the wavefield decomposition method, our approach achieves imaging accuracy of up to 99%, reduces computational costs by approximately 83.3%, and decreases computing time by about 88.9%. Additionally, tests on Marmousi and Overthrust model data demonstrate that our method performs better than Laplace filtering in terms of noise suppression and imaging reliability. This work offers a more efficient solution for suppressing low-wavenumber noise in seismic imaging and represents a meaningful advance over existing techniques.