Energy-Sensitive Learnable Directional Prior-Fused Residual Clique Network for Seismic Data Denoising

Accurate reservoir prediction is essential for oil and gas exploration, yet strong random noise under complex geological conditions often obscures weak seismic reflections and limits the reliable characterization of dolomite reservoir sweet spots. To address this issue, we propose an Energy-Sensitive Learnable Directional Prior-Fused Residual Clique Network (ELDP-RCNet) for seismic denoising. The method combines a learnable directional high-frequency prior, residual Clique feature learning, and energy-sensitive amplification (ESA) to jointly enhance structural representation and weak-signal preservation. A non-subsampled multi-scale decomposition with learnable directional filtering is employed to capture anisotropic structural information, while ESA preferentially strengthens low-energy directional responses before prior aggregation. The resulting prior is progressively injected into cascaded residual Clique blocks, which promote hierarchical feature interaction and reuse. ESA is further incorporated into local residual learning to enhance weak seismic responses during deep feature extraction. Multi-level features are finally fused through global residual reconstruction. Experiments on synthetic seismic data demonstrate that ELDP-RCNet effectively suppresses random noise while improving signal-to-noise ratio, reflection-event continuity, spectral fidelity, and weak-signal preservation.