Polarimetrie SAR image classification based on deep belief network and superpixel segmentation

Inspired by recent successful deep learning methods, this paper presents a new approach for polarimetric synthetic aperture radar (PolSAR) image classification. It combines both advantages of pixel-based and object-based methods. An improved simple linear iterative clustering (SLIC) superpixel segmentation algorithm is used to obtain spatial information in the PolSAR image. Then, a Deep Belief Network (DBN) is introduced to make full use of the limited training data sets, which is trained in an unsupervised manner to extract high-level features from the unlabeled pixels. The DBN's preliminary classification results are finally refined according to the spatial information contained in superpixels. Experimental results over real PolSAR data show that the proposed approach is more efficient with less training data and higher classification accuracy compared with the conventional manners.