MGF-UNet: Mask-Guided Gated Skip Fusion for Seismic Interpolation with Randomly Missing Traces

Seismic trace interpolation reconstructs missing traces caused by incomplete spatial sampling while preserving reflection-event continuity, amplitude fidelity, and waveform character. Randomly missing traces introduce an inherent reliability asymmetry: observed positions contain measured amplitudes, whereas missing positions must be inferred from neighboring seismic events. In U-Net-based encoder–decoder architectures, shallow features extracted around zero-filled missing traces may carry sampling artifacts and unreliable high-frequency details through direct skip connections. To address this limitation, this paper proposes MGF-UNet, a mask-guided gated skip fusion network for 2D seismic interpolation with randomly missing traces. The incomplete seismic patch and binary trace mask are used as a dual-channel input, and mask-guided gates regulate encoder features before decoder fusion. Known-trace preservation is then applied to retain measured traces in the final output, while a hybrid objective supervises missing-trace recovery, full-patch fidelity, and structural coherence. In five-run experiments on Marmousi synthetic data, MGF-UNet achieves the lowest mean MissingRMSE at the 30% and 50% missing ratios and remains competitive in global RMSE, SSIM, and SNR; however, the seed-matched paired comparisons do not reach statistical significance. Field-data comparisons and local waveform analyses further show coherent event reconstruction and reduced residual artifacts. These results suggest that mask-guided skip regulation is a promising strategy for moderate-to-severe random missing-trace interpolation.