Unsupervised and generalized orthogonal subspace projection and constrained energy minimization for target detection and classfication in remotely sensed imagery

Orthogonal subspace projection (OSP) and constrained energy minimization (CEM) approaches have been successfully applied to target detection and classification in hyperspectral imagery. However, in order for the OSP and the CEM to work effectively, the data dimensionality must be sufficiently large to accommodate all possible signal sources including target and background signatures present in an image scene. Two problems arise from this requirement. One is that the OSP and the CEM may not perform well when they are applied to multispectral imagery due to small data dimensionality. Another is that in many practical applications it is generally difficult to obtain signal knowledge a priori. This dissertation investigates the constraints of OSP and CEM, and further proposes several approaches to extension. It first develops an orthogonal projection-based method to extend the OSP to an unsupervised OSP (UOSP) which can detect and classify targets in an unknown image scene, particularly anomalies. The OSP is then further extended to a generalized OSP which can be applied to multispectral imagery. The CEM has been shown to be very effective in subpixel detection. However, it cannot classify multiple targets simultaneously. So, in this dissertation, the CEM is extended to a target classifier using a linearly constrained minimum variance (LCMV) approach which results in the LCMV classifier and target-constrained interference-minimized filter (TCIMF) can classify multiple targets in a single image as opposed to the CEM that can detect a single target at a time. Then using the unsupervised method proposed for UOSP an unsupervised version for CEM, LCMV and TCIMF can be also derived in a similar fashion. All the new approaches developed in this dissertation are evaluated by real hyperspectral and multispectral images for performance analysis.