Intelligent Deblending via an Automatic Multistage Training Framework

Simultaneous‐source acquisition greatly improves seismic data acquisition efficiency but introduces blending noise, which adversely affects subsequent seismic data processing. Deep learning‐based deblending methods gain considerable attention in both academia and industry due to their ability to characterize seismic data nonlinearly, thereby enhancing deblending accuracy. The recently proposed multistep deblending method combines U‐Net with blending noise simulation–subtraction (BNSS) operations, incorporating the workflows of traditional iterative deblending into a deep learning framework. However, the multistep strategy requires manual, sequential intervention of multiple individual network training steps and BNSS operations, accompanied by frequent data loading and storage, resulting in high operational complexity and limited applicability. To better utilize the computational resources and enhance algorithmic automation, we propose an automatic multistage deblending framework that incorporates multiple networks and BNSS operations. The proposed framework minimizes manual intervention during training, with a weighted loss function and a stage‐wise activation training strategy to ensure its effectiveness. The proposed deblending method shows improved automation and slightly higher deblending accuracy compared to the original multistep method, as demonstrated by synthetic and artificially blended field data examples.