Multi-label learning embedding approach based on multi-temporal spectral signature for hyperspectral images classification

Currently, Hyperspectral signal processing is a crucial area of research. Respectively, various techniques have been investigated to apprehend features combination and multi-label classification issues. Indeed, significant consideration has been given to approaches supporting the use of a single type of features. Moreover, few efforts have been dedicated to model the multi-label aspect of hyperspectral pixels and to integrate simultaneously divergent kinds of interdependent features. In this paper, we propose a novel embedding multi-label learning approach integrating complementary weighted features. The proposed framework combines the singular statistical characteristics of each feature to accomplish a physically meaningful cooperative low-dimensional representation of extracted features. This will grant, in one hand, the refinement of classification process and the propagation of narrow class information to unlabeled sample, in the other hand, when only partial labeling knowledge is available. This paper makes the following contributions: (i) the extraction of multi-view features based on the 3D model of the spectral signature and (ii) an embedding multi-label based approach by better tackling unbalanced and dimensionality issues. A set of complementary spatial/spectral features is extracted in the experimental section from a series of hyperspectral images. The obtained results reflect the efficiency of the proposed classification schema while maintaining a reasonable computational complexity.