On combining spectral and spatial information of hyperspectral image for camouflaged target detecting

Abstract

Detecting enemy’s targets and being undetectable play increasingly important roles in modern warfare. Hyperspectral images can provide large spectral range and high spectral resolution, which are invaluable in discriminating between camouflaged targets and backgrounds. As supervised classification requires prior knowledge which cannot be acquired easily, unsupervised classification usually is adopted to process hyperspectral images to detect camouflaged target. But one of its drawbacks—low detecting accuracy confines its application for camouflaged target detecting. Most research on the processing of hyperspectral image tends to focus exclusively on spectral domain and ignores spatial domain. However current hyperspectral image provides high spatial resolution which contains useful information for camouflaged target detecting. A new method combining spectral and spatial information is proposed to increase the detecting accuracy using unsupervised classification. The method has two steps. In the first step, a traditional unsupervised classifier (i.e. K-MEANS, ISODATA) is adopted to classify the hyperspectral image to acquire basic classifications or clusters. During the second step, a 3×3 model and spectral angle mapping are utilized to test the spatial character of the hyperspectral image. The spatial character is defined as spatial homogeneity and calculated by spectral angle mapping. Theory analysis and experiment shows the method is reasonable and efficient. Camouflaged targets are extracted from the background and different camouflaged targets are also recognized. And the proposed algorithm outperforms K-MEANS in terms of detecting accuracy, robustness and edge’s distinction. This paper demonstrates the new method is meaningful to camouflaged targets detecting.