Seismic multiple attenuation based on improved U-Net

Abstract

Effective attenuation of seismic multiples is a crucial step in the seismic data processing workflow. Despite the existence of various methods for multiple attenuation, challenges persist, such as incomplete attenuation and high computational requirements, particularly in complex geological conditions. Conventional multiple attenuation methods rely on prior geological information and involve extensive computations. Using deep neural networks for multiple attenuation can effectively reduce manual labor costs while improving the efficiency of multiple suppression. This study proposes an improved U-net-based method for multiple attenuation. The conventional U-net serves as the primary network, incorporating an attentional local contrast module to effectively process detailed information in seismic data. Emphasis is placed on distinguishing between seismic multiples and primaries. The improved network is trained using seismic data containing both multiples and primaries as input and seismic data containing only primaries as output. The effectiveness and stability of the proposed method in multiple attenuation are validated using two horizontal layered velocity models and the Sigsbee2B velocity model. Transfer learning is employed to endow the trained model with the capability to suppress multiples across seismic exploration areas, effectively improving multiple attenuation efficiency.