Feature Extraction for Hyperspectral Images Using Local Contain Profile

Abstract

Spectral-spatial information extraction is always important for hyperspectral image analysis, such as classification and detection. Extinction profile (EP), based on component tree (max-tree/min-tree), has been recently-proposed as one of the best morphological feature extraction methods. As an alternative, a new local contain profile (LCP), employing topology tree in the tree generation process, has been proposed. Topology tree, including the tree of shapes and the inclusion tree, is constructed by the inclusion relationship between the connected components belonging to the same level in the image. Furthermore, several new morphological properties, such as compactness, and elongation, are designed to accurately exploit specific shape information. The proposed LCP is expected to discard some irrelevant spatial information while preserving useful spatial characteristics. Experimental results validated on several real hyperspectral data demonstrate that the proposed LCP can significantly improve accuracy and decrease the half of feature dimension when compared to the state-of-the-art EP.