Dip-aligned structural constraints for suppression of partially coherent seismic noise using self-supervised learning

Summary

Partially coherent noise, including swell and scattered noise, remains difficult to suppress because its spatial correlation and structural similarity to reflection events reduce the effectiveness of methods that assume random, incoherent, or stationary noise. This problem is particularly challenging for self-supervised blind-spot denoising, where correlated noise can violate the independence assumptions used to separate signal from noise. We propose a structure-aware self-supervised denoising method based on the hypothesis that reflection signal and partially coherent noise can be better separated by enforcing directional consistency with local geological structure. The key component is a dip-aligned second-order directional regularization term that penalizes curvature transverse to locally estimated reflector dips while preserving continuity along structural trends. This constraint acts as a seismic structural prior and is combined with a self-attention-enhanced struct blindspot network, which captures long-range dependencies across traces and improves the representation of spatially extended noise. To reduce dependence on a fixed dip estimate, the dip field is updated from intermediate denoised results during training, allowing the structural constraint to become progressively more reliable as the signal-to-noise ratio improves. Synthetic and field examples show that the proposed method attenuates partially coherent noise while preserving reflector continuity and reducing signal leakage. Comparisons with conventional filtering methods and existing self-supervised approaches demonstrate improved noise suppression and structural preservation, particularly when noise exhibits spatial correlation. The results suggest that dip-aligned directional regularization provides a physically interpretable and effective constraint for self-supervised seismic denoising.