Abstract: Velocity model construction is a critical step in seismic data processing, and its accuracy directly affects the reliability of subsequent seismic imaging, inversion interpretation, and subsurface structural characterization. In recent years, deep learning methods have made notable progress in velocity modeling using surface seismic data, demonstrating strong capability in nonlinear feature representation and end-to-end prediction. Compared with surface seismic data, vertical seismic profiling (VSP) data possess several advantages, including richer wavefield information, stronger signal energy, higher resolution, and reduced influence from complex near-surface conditions. Despite these benefits, research on end-to-end velocity modeling based on deep learning for VSP data remains relatively limited, and its potential has not yet been fully explored. To address this issue, this study proposes a VSP seismic velocity modeling method based on a spatial–channel dual-attention residual-enhanced U-Net architecture. The proposed approach takes VSP gathers as the network input and enables direct prediction of subsurface velocity models without requiring complex preprocessing or manual feature extraction. Building upon the classical U-Net encoder–decoder framework, the network is improved by enhancing its capability to model spatial positional information and channel-wise feature responses, thereby improving the identification and reconstruction of subsurface velocity structures. In addition, a residual-enhanced feature fusion strategy is introduced to facilitate deep feature propagation, improve training stability, and enhance prediction accuracy. To evaluate the effectiveness of the proposed method, numerical experiments are conducted on layered aquifer models, undulating interface aquifer models, and the classical SEG salt dome model. The proposed approach is compared with a U-Net-based velocity modeling method using surface seismic data and a U-Net-based velocity modeling method using VSP data. Experimental results demonstrate that the proposed method can accurately reconstruct the layered distribution of subsurface velocity fields and the morphology of undulating interfaces, outperforming the comparison methods in terms of structural boundary delineation, local detail recovery, and quantitative evaluation metrics. Furthermore, generalization and transfer experiments indicate that the proposed method maintains a certain level of adaptability across different geological scenarios and can further improve prediction performance with a limited number of target-domain samples. These findings suggest that VSP data retain significant advantages within deep learning-based velocity modeling frameworks, and the proposed method provides an effective approach for achieving high-precision subsurface velocity model prediction.