Modeling Shielded Gamma-ray Source Spectra using Non-negative Matrix Factorization

Accurately identifying shielded gamma-ray sources remains a challenge to spectroscopic identification algorithms. In this paper, a data-driven approach for modeling spectral contributions from shielded sources using Non-negative Matrix Factorization (NMF) is examined. When performed on gamma-ray spectra, NMF yields spectral basis vectors that can be used to approximate various gamma-ray sources. This approach for generating source representations using NMF allows for more accurate estimates of source contributions in gamma-ray spectra in the presence of source shielding, ultimately resulting in enhanced identification capabilities for shielded sources over methods that do not account for the effects of shielding. Likelihood ratio tests are performed using NMF-based spectral models for performing simultaneous detection and identification, and this paper introduces a method for analytically computing thresholds used in hypothesis testing. Performance of this method is characterized by injecting simulated source gamma-ray spectra into background spectra measured from a mobile detection system with six 2" × 4" × 16" NaI(Tl) scintillation detectors. Specifically, background spectra injected with 60Co and 137Cs shielded by 50 mm of steel show an improvement in detection performance over an NMF-based template matching identification approach that does not account for spectral variation due to source attenuation. The methods discussed here are general enough to be broadly applicable to gamma-ray spectroscopic analysis, and can be used in applications relying on template matching.