A Deep Similarity Clustering Network With Compound Regularization for Unsupervised PolSAR Image Classification

Polarimetric synthetic aperture radar (PolSAR) image classification is a critical application of remote sensing image interpretation. Most of the early algorithms that use hand-crafted features to divide the image into various scattering categories have a general classification performance. The convolutional neural networks (CNNs) show superior performance in image processing with powerful nonlinear representation capabilities. However, they also require a large amount of labeled training data, which limit the practical use of CNNs in PolSAR image classification scene where labeled data are rare and expensive. To address the previous issue, this article proposes a deep similarity clustering model with compound regularization for unsupervised PolSAR image classification. The proposed model combines an unsupervised feature extraction pipeline with Wishart distance metric and a deep clustering pipeline with feature similarity metric. The regularization combines two ingredients to preserve both the sharpness of edges and the semantic continuity of the image contents. The first is the differential constraint based on pixel-level features, and the second is the graph partition constraint based on superpixel-level features. Experiments prove the effectiveness of the proposed method on both spaceborne (RadarSat2) and airborne (E-SAR/UAVSAR) PolSAR images. The visual results and quantitative scores show that our method outperforms the traditional unsupervised methods and deep-learning-based unsupervised methods with a large margin.