Deep Graph Cluster Based Unsupervised Representation Learning for PolSAR Image Classification

Accurate labeled samples for polarimetric synthetic aperture radar (PolSAR) images are usually difficult to obtain. So unsupervised learning is meaningful for PolSAR land cover classification tasks. In this paper, we proposed an unsupervised spontaneous clustering network named deep graph cluster. Spatial information and polarimetric coherency matrix are combined to represent and cluster the data. Firstly, an accurate and efficient clustering algorithm based on approximate nearest neighbor search is proposed. Then, we proposed the deep graph cluster based on spatial aggregation propensity of the same class and spatial dispersion of different classes. Two groups of experiments on PolSAR images shows that the accuracy of proposed method reaches 90-96%, 7-12% higher than classical unsupervised method and close to the some supervised models.