<p>The low-frequency components in seismic data serve as critical constraints for constructing macro-scale velocity models of subsurface media. However, due to data acquisition limitations and noise contamination, seismic data often lack low-frequency information, resulting in compromised accuracy in subsequent imaging outcomes. To tackle this issue, a Swin-UNet–based extrapolation framework with Squeeze-and-Excitation modules is proposed, combining the Swin Transformer’s global representation capability with the local feature extraction strength of convolutional neural networks. The Squeeze-and-Excitation modules improve the channel information of skip connections through an attention mechanism, while the structural similarity index is integrated into the loss function to preserve the structural integrity of seismic data. Experiments on synthetic data show that the network effectively extrapolates low-frequency data while demonstrating noise resilience and generalization ability. Full waveform inversion using the reconstructed seismic data yields improved delineation of stratigraphic boundaries and complex geological features. Finally, field data experiments confirm the practical utility of the proposed method.</p>