A New Approach for Spatio-Spectral Feature Selection for Sensors with Noisy and Overlapping Spectral Bands

This paper extends a recently developed canonical-correlation feature-selection (CCFS) approach to a collective spatio-spectral feature-selection and classification framework for hyperspectral imagers. The work utilizes the concept of spectrally enhanced spatial features by integration of pixels' spatial and spectral information. In order to determine the most informative features, the proposed methodology employs a sequential spatio-spectral feature-selection approach that consists of two distinct stages: a spatially independent spectral feature extraction, based on the CCFS, followed by a spatially enhanced classification. The performance of the new methodology is tested on target detection and classification applications using remotely sensed imagery collected by the Air-borne Hyperspectral Imager (AHI). Sensitivity of the spatio-spectral feature-selection approach with respect to the initial set of sensor bands and with respect to the number and types of spatial features utilized during the classification stage is also studied.