Abstract: Due to the limitations of complex field environments and acquisition equipment conditions, seismic data inevitably contain noise during the recording process, which adversely affects subsequent data processing and interpretation. The U-Net architecture is effective in capturing local feature details, while the Transformer excels at modeling global contextual information. Both types of networks have shown significant potential in seismic data denoising, but there exists a semantic gap when fusing features from these two architectures, which reduces the accuracy of signal reconstruction and hinders the preservation of weak but valid signals. To address this issue, a GLT-Unet based on the Transformer architecture was proposed. Built upon U-Net and Transformer, a G-LCF module was designed, which effectively fused local and global features of seismic data, thereby suppressing noise while preserving valid signals. Specifically, GLT-Unet first employed a CNN-Transformer hybrid encoder to extract local and global features from the seismic data. Then, the G-LCF module integrated these features to enhance the fusion effect. Finally, the decoder adopted a cascaded up-sampling structure to progressively reconstruct signal details, improving the preservation of weak signals. Denoising experiments on both synthetic and field data demonstrated that the proposed method achieved superior noise suppression compared to the K-SVD dictionary learning method and the TransUNet network, while also providing better protection for weak but valid signals.