Application of the adaptive subspace detector to Raman spectra for biological threat detection

Effective application of point detectors in the field to monitor the air for biological attack imposes a challenging set of requirements on threat detection algorithms. Raman spectra exhibit features that discriminate between threats and non-threats, and such spectra can be collected quickly, offering a potential solution given the appropriate algorithm. The algorithm must attempt to match to known threat signatures, while suppressing the background clutter in order to produce acceptable Receiver Operating Characteristic (ROC) curves. The radar space-time adaptive processing (STAP) community offers a set of tools appropriate to this problem, and these have recently crossed over into hyperspectral imaging (HSI) applications. The Adaptive Subspace Detector (ASD) is the Generalized Likelihood Ratio Test (GLRT) detector for structured backgrounds (which we expect for Raman background spectra) and mixed pixels, and supports the necessary adaptation to varying background environments. The structured background model reduces the training required for that adaptation, and the number of statistical assumptions required. We applied the ASD to large Raman spectral databases collected by ChemImage, developed spectral libraries of threat signatures and several backgrounds, and tested the algorithm against individual and mixture spectra, including in blind tests. The algorithm was successful in detecting threats, however, in order to maintain the desired false alarm rate, it was necessary to shift the decision threshold so as to give up some detection sensitivity. This was due to excess spread of the detector histograms, apparently related to variability in the signatures not captured by the subspaces, and evidenced by non-Gaussian residuals. We present here performance modeling, test data, algorithm and sensor performance results, and model validation conclusions.