Adaptive seismic diffraction separation via a hybrid attention-based CASCU network

Abstract

Diffractions carry critical information about small-scale subsurface heterogeneities, such as faults, pinch-outs, salt domes, and fractured-vuggy reservoirs; therefore, identifying and characterizing these diffractive bodies has been a central focus in geophysical research. However, existing deep learning methods still suffer from insufficient separation accuracy when dealing with complex seismic data. In this study, we propose a hybrid model named CASCU, which integrates a U-shaped convolutional architecture with the Swin Transformer. The model overcomes the inherent limitations of conventional convolutional neural networks (CNNs), including restricted local receptive fields and fixed-size kernels. By enhancing the capture of long-range dependencies and global contextual information, CASCU more thoroughly exploits the spatiotemporal continuity features within seismic data. We also efficiently constructed a training dataset based on convolutional modeling and used CASCU to achieve high-fidelity diffraction separation from the total wavefield. Numerical results show that CASCU exhibits excellent robustness and generalization performance in diffraction wave separation tasks. These findings indicate that the proposed hybrid architecture provides an effective and reliable deep learning strategy for diffraction separation in complex seismic data, thereby supporting improved identification and characterization of small-scale subsurface heterogeneities.