Abstract: In complex seismic exploration, denoising and reconstruction of seismic data serve as critical prerequisites for subsequent processing workflows. However, staged processing of seismic data for denoising and reconstruction is prone to introducing errors. Meanwhile, slice operations on 2D seismic data are liable to lose spatial information and cause error accumulation. To address these issues, this paper proposes a seismic data synchronous denoising and reconstruction method based on a three-dimensional generative adversarial U-Net (3D GA-UNet) network.First, the model constructs a generator based on U-Net, integrating residual skip connections and attention module to address the problem of gradient vanishing and enhance the model's feature extraction capabilities. In the decoder of the U-Net model, trilinear interpolation upsampling is employed to restore the spatial dimensions of the data, preserving the continuity of seismic events (reflection axes).Second, conditional constraints and instance normalization are added to the discriminator to enhance the model's reconstruction performance.Finally, a hybrid loss function is used to design differentiated loss weights for noisy regions, missing regions, and normal regions in seismic data. This approach mitigates the impact of feature imbalance across different regions on model training and improves the model's ability to recover seismic event continuity and texture information.Experimental results demonstrate that the proposed method enhances the signal-to-noise ratio of the data and achieves superior seismic data reconstruction performance. Notably, it exhibits stronger robustness in scenarios involving large-scale trace missing and complex strong noise.