This paper presents a metrological approach of spectral unmixing for automatic identification and quantitative analysis of γ-emitting radionuclides in natural background radiation at low statistics. Based on full-spectrum analysis, the proposed method relies on the maximum likelihood estimation based on Poisson statistics that accounts for the spectral signatures of the γ-emitters to be identified and natural background. In order to obtain robust decision-making at low statistics, a sparsity constraint is implemented along with counting estimation given by spectral unmixing. In contrast with the standard approach, this technique relies on a single decision threshold applied for a likelihood ratio test. Standard deviations on estimated counting are determined using the Fisher information matrix. The robustness of decision-making and counting estimation was investigated by means of Monte Carlo calculations based on experimental spectral signatures of two types of scintillation detectors [NaI(Tl), plastic]. This study demonstrates that sparse spectral unmixing is a reliable method for γ-spectra analysis based on low-level measurements. The sparsity constraint acts as an efficient technique for decision-making in the case of complex mixtures of γ-emitters with significant contribution of natural background. This method also yields unbiased counting estimation related to the identified radionuclides. Reliable assessment of standard deviations are obtained and the Gaussian approximation of the coverage intervals is validated. The proposed method can be applied either by non-expert users for automatic analysis of γ-spectra or to help experts in decision-making in the case of complex mixtures of γ-emitters at low statistics.
Read full abstract