Local Approach to Orthogonal Subspace-Based Target Detection in Hyperspectral Images

Airborne or satellite hyperspectral sensing has proven valuable in many target detection applications, thanks to the dense spectral sampling of the sensed data, which provides a high material discriminability. Within this framework, this paper focuses on detection algorithms that rely upon sub-space-based characterization of background. Whereas background subspace estimation has been typically accomplished through a global approach, which employs the whole image, a local methodology is here adopted. In fact, most of the interference affecting targets derives from the background materials in which they are inserted. Such a background interference lies in a subspace that is more likely spanned by the spectra of the pixels in the target neighborhood, rather than by endmembers/eigenvectors extracted from the whole image. Real hyperspectral imagery from the HyMap sensor is used to experimentally compare both global and local approaches to background subspace estimation. On this data, which exemplifies a mixed-pixel cluttered detection problem, detection results were strongly in favor of the local approach.