Hyperspectral Image Classification Based on Extended Morphological Attribute Profiles and Abundance Information

Due to the high dimensionality of hyperspectral data and the present of mixed pixels in the image, classification is a challenging task in hyperspectral image applications. In this paper, a strategy for hyperspectral image classification was presented, where the spectral-spatial features and abundance information of hyperspectral image were extracted and provided as input to the sparse multinomial logistic regression classifier. The mathematical morphology method was used to extract the extended multi-attribute profiles (EMAPs), and the class-based end member extraction and sparse unmixing method was used to obtain the abundance information for each class. The experiments were conducted on the well-known Indian Pines dataset. Results indicated that the extended multi-attribute profiles and abundance information significantly improved the classification performance. Compared with the overall accuracy 80.46% of classification using original spectral bands, the overall accuracy using extended multi-attribute profiles and abundance information reached 94.27% and 88.75%, respectively.