Aiming at the problem that the traditional feature representation ability is weak, we propose a polarization synthetic aperture radar image semantic segmentation method based on the multi-layer deep feature fusion. The pre-trained VGG-Net-16 model is used to extract multi-layer image features with strong representation ability, and then deep features of each layer are used to train the corresponding conditional random field model. The output results of multiple conditional random field models are finally merged to realize the final semantic segmentation of the images. The results show that compared with the methods based on classical features, the proposed method achieves the highest overall accuracy, indicating that the fusion features used by the proposed method have stronger representation ability than traditional features.