Semisupervised Complex Network With Spatial Statistics Fusion for PolSAR Image Classification

Deep learning has achieved satisfactory results in polarimetric synthetic aperture radar (PolSAR) image classification, which requires a large number of labeled samples for training. However, in practice, labeling work is time-consuming and laborious. As a result, an insufficient number of labeled samples will lead to a limited ability of the network to recognize different terrains. To alleviate this problem, we take advantage of labeled and unlabeled samples simultaneously to train the deep learning model and, thus, propose a semisupervised complex network with spatial statistics fusion (SCN-SSF) for PolSAR image classification. First, the semisupervised complex network (SCN) continuously updates the pseudolabels of unlabeled samples during the training of complex-valued CNN, and their errors constitute the regularization term of the objective function, which improves the generalization of the network. As a result, SCN can recognize different terrains more accurately, and the classification has a higher belief. Then, a parameter-free spatial statistics module is built to model neighborhood label interactions based on the product of experts (POEs), thus enhancing contextual smoothness and correcting some misclassifications. Finally, based on the Dempster–Shafer theory, the contextual label information of POE and pixel-level information obtained by SCN are integrated to preserve image structure. Overall, with only a small number of labeled samples, SCN-SSF can accurately identify each terrain and obtain smooth classification while preserving edge information. The effectiveness of SCN-SSF is demonstrated by classifying PolSAR images with a small number of labeled samples.