False Alarm Mitigation in Hyperspectral Detection of Gaseous Chemicals using Knowledge of Chemical Library and Residuals

Remote sensing of chemical vapors with hyperspectral imaging devices plays an important part in many civilian and military applications. The EPA uses hyperspectral imaging to monitor the chemical vapor output from smoke stacks while the military uses it for detection of chemical warfare agents (CWAs) in the battlefield. Detection algorithms have been studied for decades now but a major limitation thus far has been the library of signatures available. In previous years, the signature libraries have been of very good quality but because of the effort needed per signature, only a handful of chemical plume signatures were available. On the other hand, reducing the resolution allowed for having a much more expansive set of signatures, but the quality of each signature becomes significantly worse. In recent years, the technology to create high resolution chemical vapor signatures have become available, increasing the size and quality of the hyperspectral signature libraries.;With the advent of a comprehensive, high quality hyperspectral chemical signature library, we can now study the chemicals themselves. This thesis presents a study of the chemical vapor signature library. The main objective of studying the library is to determine the discriminability of the signatures. We attempt to form clusters of signatures in order to see if there are inherent similarities between signatures. If clusters form naturally, then we can expect signatures within a cluster to be easily mistaken for one another. We use two methods to cluster the library: K-means clustering and hierarchical tree-based clustering. With K-means clustering, we utilize a binomial search in order to determine the number of cluster centers given a spectral angle threshold. For tree-based clustering, we analyze the advantages and disadvantages of using the single, average and complete linkage methods for our clustering study.;We also utilize the knowledge of the library to study False Alarm Mitigation (FAM) by comparing detection results with each chemical's nearest confusers in terms of Spectral Angle. We embed both a plume along with it's confusers and compare detection results. The Adaptive Coherent/Cosine Estimator (ACE) detector, which is the basis detector for this study, can give both plume and confuser very similar scores. Therefore, we utilize information found in the residuals in order to separate false alarms from true positives. We also use images with real data and real background based false positives. We later conclude that, due to the nature and origin of these false positives as compared to embedded false positives, metrics which work for one case in general does not work for the other.