Mapping the spatial distribution and activity of 226Ra at legacy sites through Machine Learning interpretation of gamma-ray spectrometry data

Open Access

Abstract

Radium (226Ra) contamination derived from military, industrial, and pharmaceutical products can be found at a number of historical sites across the world posing a risk to human health. The analysis of spectral data derived using gamma-ray spectrometry can offer a powerful tool to rapidly estimate and map the activity, depth, and lateral distribution of 226Ra contamination covering an extensive area. Subsequently, reliable risk assessments can be developed for individual sites in a fraction of the timeframe compared to traditional labour-intensive sampling techniques: for example soil coring. However, local heterogeneity of the natural background, statistical counting uncertainty, and non-linear source response are confounding problems associated with gamma-ray spectral analysis. This is particularly challenging, when attempting to deal with enhanced concentrations of a naturally occurring radionuclide such as226Ra. As a result, conventional surveys tend to attribute the highest activities to the largest total signal received by a detector (Gross counts): an assumption that tends to neglect higher activities at depth. To overcome these limitations, a methodology was developed making use of Monte Carlo simulations, Principal Component Analysis and Machine Learning based algorithms to derive depth and activity estimates for 226Ra contamination. The approach was applied on spectra taken using two gamma-ray detectors (Lanthanum Bromide and Sodium Iodide), with the aim of identifying an optimised combination of detector and spectral processing routine. It was confirmed that, through a combination of Neural Networks and Lanthanum Bromide, the most accurate depth and activity estimates could be found. The advantage of the method was demonstrated by mapping depth and activity estimates at a case study site in Scotland. There the method identified significantly higher activity (0.4m), that conventional gross counting algorithms failed to identify. It was concluded that the method could ...