Separation of Ocean Blended Data Based on Deep Learning Residual U-Net network

With the gradual increase of spatial sampling density of seismic data, blended source acquisition has gradually become one of the effective means to improve the acquisition efficiency, and the effective deblending of blended seismic data is an important part of seismic data processing. We proposed a smart deblending technology for marine dual-source alternating excitation blended data based on the residual U-Net network. The method first converts the blended data from the common shot gather to the common receiver gather to reduce the correlation of the non-primary source excitation signals, and then achieves the intelligent deblending of dual-source blended data based on the residual U-Net network. Compared to traditional U-Net network, our new network model increased the network depth and introduced convolutional residual modules during the downsampling process, which effectively avoided the problems of gradient disappearance and gradient explosion, enhanced the feature extraction capabilities especially in the processing of detail issues, and better protected the valid information. Through model calculations and actual data processing, the good performance of the network in marine data deblending was verified. The experimental results show that the residual U-Net network could effectively deblend data without losing valid signals and significantly improve the signal-to-noise ratio of the deblending results. The research results provide a new idea for high-precision deblending of marine seismic data and lay the foundation for subsequent seismic data processing.