Collaborative-Representation-Based Nearest Neighbor Classifier for Hyperspectral Image Classification Combined with Superpixel and Loopy Belief Propagation

Danning Lin, Zhijing Yang, Meilin Wang, Yongqiang Cheng, Qing Pan

Abstract: The k nearest neighbor (KNN) is one of the most popular classifiers for hyperspectral images (HSI). However, in hyperspectral imagery classification, since the pixel spectral signatures are usually mixed due to the relatively low spatial resolution, traditional KNN on pixel-level cannot handle it. To improve the performance of classification, a new KNN method based on superpixel and the collaborative-representation (KNNSCR) has been proposed. This proposed method can effectively overcome the intra-class variations and inter-class interference. Furthermore, we combine KNNSCR with loopy belief propagation (LBP) to catch more detailed spatial information. The proposed method can greatly improve the accuracy of HSI classification. The experiments demonstrate that the proposed method obtain very good results by comparing with some well-known methods.