A semi-supervised classification method of polarimetric SAR images based on DSFNN and non-local decision is proposed. The method comprises steps: the data of polarimetric SAR images being input; superpixel segmentation of polarimetric SAR image; extracting original features and super-pixel features of each pixel in polarimetric SAR image; training sample set and test sample set being selected; using the training sample set to train the depth super-pixel filter network; the depth superpixel filter network being used to predict the test samples; based on non-local decision, the training set being expanded by selecting samples from the test set; updating the depth superpixel filter network; the trained network being used to classify the test samples, so that a classification result diagram isobtained. The depth super pixel filter network of the invention extracts super pixel features to overcome coherent speckle noise, and utilizes semi-supervised classification algorithm of non-local decision to reduce the number of training samples and effectively improve classification accuracy, and can be used in the technical fields of polarimetric SAR image ground object classification and target recognition and the like.