The quality of wavefield separation on vertical seismic profiling (VSP) data directly affects subsequent imaging and inversion processes. However, the traditional methods have various defects in separating upgoing and downgoing waves. The application of deep learning brings another train of thought to solve related problems. The traditional methods, such as f-k filtering and Radon transform, produce results with spatial aliasing and inaccurate amplitudes. The Median filtering method relies on accurate first-break picking and waveform consistency. Moreover, the results of traditional methods are subject to manual intervention. To overcome these problems, we proposed a deep learning method to separate the VSP wavefields automatically. First, the traditional Radon transform method was used to produce training datasets. Then, to ensure amplitude preservation and suppress spatial aliasing, we created the new input data by recombining the upgoing and downgoing waves extracted by Radon transform. The developed deep learning network has very high efficiency. Its accuracy exceeds that of the traditional methods just after one epoch training. We established a practical workflow of wavefield separation based on deep learning and applied it to synthetic data and field distributed acoustic sensing VSP (DAS-VSP) data. The results show that the proposed method is superior in terms of amplitude preservation and spatial aliasing suppression. The time consumption of the proposed method is well acceptable and can be further minimized by training the network using downsampling data.