Polarized synthetic aperture radar (SAR) high resolution remote sensing image technology has been widely used in commerce, agriculture, ocean transportation and other fields. The results show that the Deep Learning method can better mine the image information of high-resolution polarimetric SAR image than the traditional method. In this paper, a new method is proposed, which performs image preprocessing before substituting deep learning framework to improve classification accuracy. Firstly, Pauli, H/A/Alpha, Freeman, Yamaguchi and other characteristic components obtained by polarization decomposition were processed by principal component analysis (PCA) dimensionality reduction algorithm and three images were obtained. Wavelet decomposition and reconstruction, anisotropic filtering and grayscale of each image were performed respectively, and the three images were fused. Finally, the result graph is input into DeepLabV3+ network architecture for training. The data adopted in this paper is Flevoland regional image. By comparing 13 features and 17 features, the training results are brought into DeeplabV3+ network architecture with or without using the proposed method. The results show that the Kappa coefficient of deep learning classification is significantly increased by 2.2-2.3% after using the descending and anisotropic filtering, which proves that the proposed method can improve the classification accuracy of polarimetric SAR from the perspective of image preprocessing without increasing stacking.