Abstract: Wavefield separation serves as a key step in processing the data of vertical seismic profiles (VSPs). Its accuracy directly influences seismic imaging, inversion of elastic parameters, lithology identification, and interpretation of hydrocarbon-bearing properties. Traditional methods face challenges in wavefield separation. For example, the median filtering requires manual intervention, often introducing errors and thus compromising separation accuracy; the FK filtering yields high accuracy but low efficiency. In contrast, deep learning techniques offer high automation, enabling both high accuracy and efficiency in wavefield separation. Hence, this study proposed a deep learning-based method for separating up- and down-going waves in zero-offset VSPs. First, the up- and down-going waves were separated through FK transform, generating a dataset. Second, a deep learning-based model, Unet++, was constructed for separating these waves in VSPs. Third, the relative down-going wavefield (obtained by subtracting the predicted up-going wavefield from the full wavefield) was incorporated into the loss function to mitigate the impacts of amplitude differences between up- and down-going waves on network updates. Moreover, the structural similarity index measure (SSIM) was employed as a regularization constraint to assist the network in learning the structural characteristics of the wavefield. The test results of actual VSP data demonstrate that the trained network can effectively learn the characteristics of the up- and down-going waves, achieving high accuracy and efficiency in wavefield separation.