A deep learning identification method for the core of fracture-cavity bodies based on Angle domain data-driven

Deep carbonate fracture-cavity reservoirs represent a class of hydrocarbon-rich reservoirs and are currently a focal area in exploration. Due to their considerable burial depth, seismic reflection structures are often coupled with sedimentary strata, making accurate identification challenging using conventional methods. Furthermore, fracture-cavity reservoirs differ from horizontally layered sedimentary reservoirs, being primarily characterized by vertical strike-slip faults and dissolution features. In seismic exploration, signals acquired at different angles contain distinct reservoir information, providing an advantage for the detailed characterization of fracture-cavity systems. To address the challenge of identifying fracture-cavity core zones, this study proposes a data-driven, angle-domain deep learning method. Unlike previous studies, our approach begins by analyzing variations in angle-domain seismic data: small-angle data (0-6°, not fixed) exhibit strong responses to horizontal strata, whereas large-angle data (27-33°, not fixed) show weaker horizontal stratigraphic influence, thereby mitigating its interference. Additionally, we introduce a neural network model incorporating multi-scale convolution, channel attention mechanisms, and spatial attention mechanisms. Comparative testing demonstrates that our method effectively resolves the mismatch between forward-modeled labels and actual samples—an issue unresolved by conventional models. Validation on field data confirms that our approach significantly outperforms second-generation coherence (C2) and the latest coherence (C4) methods in identification accuracy while also enhancing noise suppression and prediction reliability compared to small-angle and conventional migration data. In summary, our method improves the precision of fracture-cavity core prediction by optimizing both data sources and the predictive model, demonstrating broad applicability in analogous reservoirs.