Seismic data quality frequently deteriorates due to random noise contamination, substantially impeding subsequent processing and geological interpretation. While deep learning approaches have emerged as powerful tools for noise suppression, conventional single-stage architectures exhibit inherent limitations in handling complex seismic features while preserving subtle geological details. These challenges motivate the development of advanced multi-stage neural networks for seismic data enhancement. The proposed multi-stage progressive U-shaped convolutional network (MPU-Net) architecture addresses these limitations through supervised cross-stage attention mechanisms that maintain feature connectivity throughout the network. Building upon this foundation, group enhanced convolutional blocks (GEB)-MPU-Net introduces GEB to specifically counteract the progressive attenuation of shallow features in deep networks. This dual-stage enhancement strategy combines hierarchical feature preservation, adaptive information fusion, and stable gradient propagation. Comprehensive evaluation using both synthetic and field datasets demonstrates GEB-MPU-Net’s superior performance compared to conventional time-frequency analysis methods and established networks, such as U-Net, residual dense network, residual dense block U-Net, and MPU-Net. The architecture consistently achieves enhanced reflection continuity, improved geological feature resolution, and robust noise suppression. These advancements provide more reliable input for seismic interpretation, better preservation of subtle stratigraphic features, and increased applicability to challenging field conditions.