Multi-feature based label propagation for semi-supervised classification of hyperspectral data

摘要

For hyperspectral remote sensing, some spatial features e.g., texture and morphological feature, have been successfully employed for classification task. In this paper, we investigate the utilization of multiple features in label propagation (LP) classifier, attempting to improve the classification accuracy with small size samples. The graph of LP classifier is constructed by using multiple features, where each feature calculates a similarity matrix and a composite similarity matrix is obtained via linear combination. To weight the importance of each feature in the combination, leave-one-out strategy is applied to obtain a suitable weight for each feature. Experimental results on two high resolution hyperspectral data show that the proposed approach obtained better classification performance than LP classifier with single spectral feature and LP with equal weighted multi-feature.