Unsupervised classification of polarimetric SAR images using deep embedding network

In this paper we propose an unsupervised polarimetric SAR image classification method using deep embedding network. In this method, we first use superpixel segmen-tation method to produce superpixel regions and use the density peaks clustering (DPC) method to generate the representation points of the superpixel regions. The simi-larity matrix of the representation point and the sample points is then constructed. The low-dimensional manifold features produced by singular value decomposition (SVD) of the similarity matrix are then input onto the deep embedding network which is composed of stacked auto-encoders (SAEs). The unsupervised classification results are finally obtained by clustering algorithm. By using super-pixel segmentation our method introduces spatial constraints. Comparing to other methods, the DPC method guarantees generating more robust representation points. Random mapping of the low-dimensional features in the deep embedding network guarantees the robustness of the method. The experimental results on real polarimetric SAR data demonstrate the effectiveness of the proposed method.