Enhancements to cSAM spectral comparison

Abstract

Growing use of infrared spectral signature data in scientific and forensic analysis requires collecting large quantities of data from a variety of sensors using a variety of techniques in a variety of environmental conditions. While individual collections are internally valid, variations and clutter in observed spectral features make signature classification and comparison both challenging for spectral analysts and often impossible for automated systems. As the quantity of collected data continues to grow rapidly, automated solutions are increasingly critical. For example, the Joint Improvised Explosive Device (IED) Defeat Organization (JIEDDO) Integrated Signatures Program (ISP) has collected approximately one million infrared (IR) spectra and relies on manually produced metadata to identify and classify signatures. The manual production of metadata, at this scale, requires significant and unsustainable cost and schedule investment to reduce errors and inconsistencies. The primary methods used to autonomously identify and classify IR spectral data include spectral angle mapping and key feature detection. Spectral angle mapping cannot compare spectra with differing domains (e.g., due to removed bands). Additionally, spectral mapping is a computationally slow process, running in linear time against an entire reference library to identify a single new signature. Key feature detection improves computation time by pre-calculating feature locations in the reference spectra, but this method is highly susceptible to noise and clutter features.