Unsupervised classification of polarimetric SAR images using deep embedding network

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

“…In 2016, Xie [44] proposed the Deep Embedded Clustering (DEC) model by combining the K-means clustering algorithm with the autoencoder network while introducing KL divergence as the clustering loss. In 2017, Yan et al [20] applied the DEC network to PolSAR image classification, and used Density Peaks Clustering and SVD for data processing, achieving certain classification effects. In 2017, Liu et al [45] proposed a neighborhood-preserving network combined with superpixels for joint spatial constraints to implement unsupervised classification.…”

“…Finally, to verify the performance of the proposed classification model, we compared it with classic unsupervised classification algorithms H/alpha-Wishart (HA-W) [7], Freeman-Durden-Wishart (FD-W) [8], and GD-Wishart (GD-W) [9]. In addition, we also compared it with unsupervised classification algorithms based on deep learning, namely Deep Embedded Clustering (DEC) [20] and Vector Quantized Convolutional Autoencoder (VQ-CAE) [49]. In addition, the state-of-the-art supervised models: FCN [52] and Unet [53] are used further to compare the classification performance of the proposed method.…”

“…Therefore, unsupervised PolSAR image classification based on deep clustering is an important research direction. However, there is relatively little research on deep unsupervised learning and clustering methods [20]- [22] for PolSAR image. And the semantic discontinuity problem is serious during classification due to the lack of certain classification constraints, which resulting in poor classification performance.…”

A Comprehensive Survey of Machine Learning Applied to Radar Signal Processing

“…An unsupervised deep generative networkpoisson gamma belief network (PGBN) was proposed to extract multi-layer feature from SAR images data for targets classification tasks in [352]. An unsupervised PolSAR image classification method using deep embedding network-SAEs was built in [353], which used SVD method to obtain lowdimensional manifold features as the inputs of SAEs, and the clustering algorithm determined the final unsupervised classification results. As for In-SAR data, a DBN was used to model data in [366] for classification, which could fully explore the correlation between intensity and the coherence map in space and time domain, and extract its effective features.…”