Automatic target detection and classification for hyperspectral imagery

Automatic target detection and classification is one of primary tasks of hyperspectral imaging. Its detectability does not rely on prior knowledge. In many practical applications such as surveillance, this is a significant advantage over supervised target detection and classification methods which require some level of information. This dissertation designs and develops computer-automated algorithms to extract targets for detection and classification with no prior knowledge about the image data. Of particular interest are small targets, which are generally man-made objects and occur with low probabilities. Three approaches are investigated in this dissertation, projection pursuit (PP), linear spectral random mixture analysis (LSRMA), anomaly detection and classification. The proposed PP utilizes the criteria of skewness and kurtosis to design four projection indices to capture targets where an evolutionary algorithm (EA) is used to find optimization solutions. In order to segment targets from the background, a zero-detection thresholding technique is also introduced for target extraction. LSRMA models an image pixel as a random process resulting from a random composition of multiple spectra of distinct materials in the image where the commonly used independent component analysis (ICA) is modified and reformulated for hyperspectral image analysis. LSRMA does not require prior target knowledge as generally required for linear spectral mixture analysis (LSMA). Most importantly, LSRMA models each of materials of interest as an independent random signal source so that the spectral variability of materials can be captured more effectively in a stochastic manner. A third approach is anomaly detection and classification where RXD is modified to derive several variants. Among them is the causal RXD which can be implemented in real time. Since an anomaly detector does not necessarily classify the targets it detected, target discrimination measures are also proposed to accomplish target classification. Incorporating one of these target discrimination measures, anomaly detection can be extended to anomaly classification. The experimental results conducted in this dissertation demonstrate that all the three approaches developed in this dissertation perform effectively in extracting targets from image data with no required prior knowledge.