Multi-label Classification of Hyperspectral Images Based on Label-Specific Feature Fusion

For hyperspectral classification, the existence of mixed pixels reduces the classification accuracy. To solve the problem, we apply the multi-label classification technique to hyperspectral classification. The focus of multi-label classification is to construct label-specific features. However, some algorithms do not consider the construction of label-specific features from multiple perspectives, resulting in that useful information is not selected. In this paper, we propose a new hyperspectral image multi-label classification algorithm based on the fusion of label-specific features. The algorithm constructs label-specific features from the three perspectives: distance information and linear representation information between instances, clustering information between bands, and then merges three feature subsets to obtain a new label feature space, making each label has highly discriminative features. Comprehensive experiments are conducted on three hyperspectral multi-label data sets. Comparison results with state-of-the-art algorithms validate the superiority of our proposed algorithm.