Seismic Deblending Based on Iterative Self-supervised Learning in the Common-Receiver Domain

The low-cost potential of multi-source simultaneous seismic acquisition critically depends on the accuracy of deblending methodologies. Effective suppression of crosstalk noise in blended seismic data represents a fundamental prerequisite for subsequent imaging and analysis workflows, requiring reliable separation of interference into distinct source responses. In recent years, deep learning-based methods have provided promising separation results. However, classical supervised algorithms require unblended clean data for labels, which is rare in field cases. This study presents an iterative self-supervised deblending method operating in the common receiver domain, which eliminates the need for clean data labels while incorporating the physical constraints inherent to iterative inversion algorithms. The proposed workflow addresses the absence of clean common receiver domain data through a progressive noise augmentation strategy inspired by self-supervised techniques developed for common shot domain processing. Each iteration constructs incrementally cleaner datasets by training the model for a single epoch on progressively refined data approximations, thereby enabling gradual noise suppression through successive refinement cycles. This methodology facilitates the creation of near-clean data representations while maintaining compatibility with field acquisition constraints. Field data experiments demonstrate superior performance compared to self-supervised training strategies implemented in the common shot domain, particularly in preserving coherent signal components and attenuating residual interference. The results highlight the enhanced capability of our method to recover individual source responses without requiring clean training labels, suggesting practical advantages for real-world seismic acquisition scenarios where conventional supervised learning approaches prove infeasible.