Multiple attenuation algorithm based on improved pix2pix GAN network

The effective attenuation of seismic multiples plays a crucial role in the seismic data processing work-flow. Despite the existence of numerous multiple attenuation methods,traditional approaches heavily rely on prior geological structure information and require extensive calculations,resulting in slow attenuation speed. This poses an even greater challenge for multiple attenuation under complex geological conditions. To overcome the limitations of traditional methods and improve efficiency,this paper applies the pix2pix GAN network to the problem of multiple attenuation and utilizes the feature learning capability of neural networks to improve the processing speed. It proposes an enhanced multiple attenuation method for the pix2pix GAN network, which integrates ResNet and U-Net as the network generator to avoid gradient vanishing or exploding phenomena used by deep netwoorks,while incorporating the SE attention mechanism. The improved generator can better perceive the characteristics of both first - order and multiples,thereby enhancing its performance. Additionally,a multi-scale discriminator is employed to discern detailed features and texture information on finer seismic images for accurate identification of authenticity. The input data for the network consists of full wave field data labeled as primary wave data,with training conducted using a dataset synthesized from two simple formation models and a public Sigbee2B model. Experimental results demonstrate that the improved GAN network exhibits superior accuracy in multiple attenuation compared to pix2pix GAN,effectively improving attenuation speed. © 2024 Editorial office of Oil Geophysical Prospecting. All rights reserved.