Accurate separation of P-and S-waves is crucial for multi-component seismic imaging in shallow subsurface studies. We proposed a transfer learning framework based on the U-FCN architecture that combines multiscale feature fusion from U-Net with a fully convolutional neural network. Synthetic datasets were generated using finite-difference simulation and Helmholtz decomposition to train the network in a data-driven manner. For field data adaptation, we employed a transfer learning strategy involving: (1) freezing early feature extraction layers, (2) fine-tuning the final some layers, and (3) incorporating dilated convolutions to enhance feature extraction. Numerical simulations and field experiments demonstrate that the proposed approach achieves accurate P-and S-wave separation. On the field seismic data acquired from the Loess Plateau, the separated P-waves achieved R2=0.952, SSIM=0.906, and PSNR=29.197 dB, while the S-waves reached R2=0.938, SSIM=0.885, and PSNR=28.846 dB, with an inference time of only 1.3 s. Compared with conventional methods, our approach achieves cleaner P-and S-wave separation with significantly higher computational efficiency and without relying on prior model parameters. The results confirm the robustness and scalability of the method for real-world seismic applications, effectively bridging the gap between synthetic-based training and field data interpretation.