| Summary: | In this paper, we propose to extract spectral absorptions as the discriminative features to classify hyperspectral imagery. Different from previous researches that mainly take hyperspectral curves as high-dimensional inputs, we analyze hyperspectral data more from its physical and chemical origins. In the proposed approach, the discriminatory information, which is characterized by the observed materials' constituents, is extracted as a group of absorption features. First, the original hyperspectral spectra are transformed to a normalized spectra, in which a modified continuum removal algorithm is adopted to highlight all spectral valleys. Next, a standard peak detection method is applied to the continuum-removed spectra, and all effective absorptions are found as the candidate features. Then, to obtain the most informative absorptions to classification, a novel mutual-information based feature selection method is used to search for the key absorption spectra. Finally, we put forward a matching algorithm to classify the absorption features using the multi-label learning. To testify the proposed method, both laboratory and remotely sensed hyperspectral data are used to evaluate the classification performance. Experimental results show that the proposed method achieves competitive classification accuracy against the state-of-the-art methods, but with an advantage of more compact feature representation.
|
|---|
