Kernel-Based Machine Learning for Background Estimation of NaI Low-Count Gamma-Ray Spectra

Abstract

Virtually all gamma-ray spectrometry measurements contain background components due to the ubiquitous presence of primordial radionuclides in the Earth's crust and cosmic radiation interactions high in the Earth's atmosphere. In principle, spectral signatures due to radiation source(s) of actual interest can be extracted from the measured gamma-ray spectrum by background subtraction. However, if separate background measurements are unavailable or infeasible, and particularly for measurements exhibiting low signal-to-noise ratio (SNR), background subtraction is nontrivial, and it requires accurate background estimation . An example application of gamma-ray spectroscopy with low SNR is the “source search” scenario, where the position of a source is sought using measurements taken over very short time intervals by a detector in motion. We have developed an algorithm for background estimation in low-count gamma-ray spectra using kernel-based Gaussian processes (GP) taken from the field of machine learning. We have evaluated the performance of our algorithm using a group of three kernels tested against a dataset composed of background spectra measured in an urban environment using a mobile sodium iodide (NaI) detector. We have also simulated datasets containing nonbackground gamma-ray sources in an urban background measured with a NaI detector. The simulated scenarios employed a variety of source-detector distances and different types of source shielding. As a metric of algorithm performance, we calculated correlation coefficients, Theil inequality coefficients, and count difference statistics between estimated and actual backgrounds. We concluded that our method adequately estimates the gamma-ray background, but we also observed a strong dependence of the algorithm's performance on the selected kernel.