Low-frequency information in seismic data can improve seismic resolution and imaging accuracy, enhance the quality of inversion, and play an essential role in imaging algorithms such as full-waveform inversion (FWI). Sufficiently low-frequency data can avoid the cycle skipping phenomenon during FWI. During seismic data processing, the protection and reconstruction for low-frequency information are therefore of great importance. In this article, we systematically investigate the extrapolation of prestack viscoacoustic seismic low-frequency data using a dense convolutional network (DenseNet) to effectively establish the nonlinear relationship between high- and low-frequency data and realize the extrapolation and reconstruction of viscoacoustic 0–5-Hz low-frequency data using 5–30-Hz high-frequency component. The generalizability of the method for different influencing factors such as wavelets, noise, and models is analyzed using the Marmousi2 velocity model forward data. It is demonstrated that the method has high robustness and can be applied to different situations, and the accuracy is higher than that of the traditional convolutional neural network (CNN) method. The feasibility of the low-frequency extrapolation method based on DenseNet is also verified by synthetic data, physical experiment simulation data, and field data testing and is superior to the traditional CNN method.