Hyperspectral image anomaly detection based on local orthogonal subspace projection

The Orthogonal Subspace Projection (OSP) algorithm is a supervised classifier that needs the information of the classified objects. To expand its application, a local OSP (LOSP) is design to apply to detect the hyperspectral image. The anomaly detection algorithms are usually used to extract the isolated man-made objects in the nature background, where the substances in the small local region are usually uniform. Based on the principle, the LOSP is constructed by choosing the detected pixel as the interested object and the mean of its nearby pixels as the suppressed object. The experiments show that the LOSP can detect the sub-pixel targets with a content greater than 30%, and can also detect the targets occupying more pixels by enlarging the window size. In addition, LOSP is proved not to be affected by the Hughes phenomenon, and the computing time is less than 1/10 that by RX detector when the number of wavelengths is 80. LOSP is effective both in precision and in efficiency, and is applicable to the real-time detection of the hyperspectral image.