Improved Feature Extraction, Feature Selection, and Identification Techniques That Create a Fast Unsupervised Hyperspectral Target Detection Algorithm

This research extends the emerging field of hyperspectral image (HSI) target detectors that assume a global linear mixture model (LMM) of HSI and employ independent component analysis (ICA) to unmix HSI images. Via new techniques to fully automate feature extraction, feature selection, and target pixel identification, an autonomous global anomaly detector, AutoGAD, has been developed for potential employment in an operational environment for real-time processing of HSI targets. For dimensionality reduction (initial feature extraction prior to ICA), a geometric solution that effectively approximates the number of distinct spectral signals is presented. The solution is based on the theory of the shape of the eigenvalue curve of the covariance matrix of spectral data containing noise. For feature selection, previously a subjective definition called significant kurtosis change was used to denote the separation between targets classes and non-target classes. This research presents two new measures, potential target signal to noise ratio (PT SNR) and max pixel score which computed for each of the ICA features to create a new two dimensional feature space where the overlap between target and non-target classes is reduced compared to the one dimensional kurtosis value feature space. Finally, after target feature selection, adaptive noise filtering, but with an iterative approach, is applied to the signals. The effect is a reduction in the power of the noise while preserving the power of the target signal prior to target identification to reduce false positive detections. A zero-detection histogram method is applied to the smoothed signals to identify target locations to the user. MATLAB code for the AutoGAD algorithm is provided