Anomaly detection in hyperspectral signatures using automated derivative spectroscopy methods

The goal of this research was to detect anomalies in remotely sensed Hyperspectral images using automated derivative based methods. A database of Hyperspectral signatures was used that had simulated additive Gaussian anomalies that modeled a weakly concentrated aerosol in several spectral bands. The automated pattern detection system was carried out in four steps. They were: (1) feature extraction, (2) feature reduction through linear discriminant analysis, (3) performance characterization through receiver operating characteristic curves, and (4) signature classification using nearest mean and maximum likelihood classifiers. The Hyperspectral database contained signatures with various anomaly concentrations ranging from weakly present to moderately present and also anomalies in various spectral reflective and absorptive bands. It was found that the automated derivative based detection system gave classification accuracies of 97 percent for a Gaussian anomaly of SNR -45 dB and 70 percent for Gaussian anomaly of SNR -85 dB. This demonstrates the applicability of using derivative analysis methods for pattern detection and classification with remotely sensed Hyperspectral images.