Due to limitations in exploration costs and surface conditions, actual seismic data are often sparsely sampled and exhibit irregular coverage, which weakens the effects of migration imaging and wavefield inversion, thereby reducing the accuracy and reliability of geological interpretation. Deep learning, particularly generative adversarial networks (GANs), has shown considerable promise in seismic data reconstruction. However, existing GAN-based methods often suffer from limitations such as restricted feature extraction, blurring and artifacts. Additionally, 2D networks struggle to capture the spatiotemporal correlations of 3D seismic data, leading to inter-layer inconsistencies and the loss of critical tectonic features, which ultimately limits overall reconstruction quality. To address complex missing-data patterns in seismic datasets, this study proposes a 3D generative adversarial network integrating a Global Grouped Coordinate Attention (GGCA-GAN). The proposed network employs group-based feature modeling and, by jointly training on natural images and real seismic data, achieves high-resolution 3D seismic reconstruction with moderate robustness to noise. This method further reduces computational overhead by enhancing global spatial feature extraction in both the generator and discriminator via grouped processing. In addition, to address the scarcity of training data under limited observation conditions, natural image datasets are incorporated into the network training process. Their structural features are used to construct cross-domain priors, which, when combined with measured seismic data, enhance the network's ability to generalize to complex seismic signals. The effectiveness of the proposed method is validated on publicly available datasets.