Abstract: The reconstruction of randomly missing seismic data plays a critical role in subsurface imaging. However, the performance of existing self-supervised learning methods heavily depends on the data missing rate, and most approaches ignore feature learning in the transform domain, especially under high missing rates. To address this issue, this paper proposes a self-supervised cross-domain feature fusion network for seismic data reconstruction. A novel cross-domain feature fusion module is designed within the network, where the fast Fourier transform captures global frequency components, enhances effective frequency components while removing irrelevant information. The output after the inverse fast Fourier transform is fused with the module’s input to compensate for potential losses caused by using only frequency-domain information. This cross-domain interaction architecture overcomes the limitations of traditional single-domain representations, achieving a balance between global structural preservation and local detail recovery of seismic signals under self-supervised conditions. Compared with the weighted nuclear norm minimization method, the deep seismic prior method, the blind-trace network method, and the blind-trace network method with regenerative constraints, the reconstruction results of the proposed method in this paper show significant improvements: On the synthetic dataset with 75% and 95% random missing traces, the signal-to-noise ratios of the reconstructed results are 10.451 5 and 5.151 4 dB, respectively; On the real seismic dataset, the signal-to-noise ratios of the reconstructed results are 8.383 0 and 4.376 6 dB, respectively.

Key words: cross-domain feature fusion module, fast Fourier transform, self-supervised learning, seismic data reconstruction