Semantic Correntropy Representation for Hyperspectral Image Classification

Abstract

Exploring semantic structure information contained in the hyperspectral image (HSI) is significant for accurate HSI classification (HSIC), especially when there are very few labeled samples. Besides, it is usually time-consuming to process HSI for many algorithms. In this paper, for a more accurate and efficient HSIC, a Semantic CorrEntropy Representation-based method (SCER) is presented. Specifically, superpixel segmentation is employed to capture structural information and improve computational efficiency. Each superpixel in HSI can be regarded as a semantic subregion. And for each semantic subregion, we define a semantic correntropy matrix to represent the intrinsic spectral similarity distribution, which can provide beneficial information for classification. However, the spatial sizes of diverse land covers in HSI have large variations, which means the rich semantic information cannot be fully captured by a single-scale segmentation map. Therefore, multiscale SCER (MSCER) with the multiscale segmentation strategy is proposed. Further, the features learned from multiple scales are integrated via multiple kernels to promote semantic information learning. The weights of multiple kernels can be determined adaptively. The semantic features of diverse land covers represented by MSCER have shown great intra-class compactness and inter-class separability. Compared with other state-of-the-art methods, our proposed MSCER has achieved competitive classification performance both in accuracy and efficiency with few labeled samples.