Aliaksandr Dvornik - Institutionen för translationell medicin

Title: Localization of radioactive sources out of regulatory control using moving gamma spectrometry and recursive frequentist and Bayesian inference

Main supervisor: Prof. Christopher Rääf

Reviewers: Magnus Dustler, Rani Basna, Francisco Piñero García

Abstract

Background

The objective of the project is to investigate how existing Bayesian methods for source localization can be improved by combining traditional frequentist statistics with Bayesian approaches. It also examines how ambient conditions, such as fluctuating natural background radiation, can be accounted for to yield reliable predictions based on representative scenarios, including a single point source, two sources, and two-dimensional contamination. Furthermore, the project aims to explore whether the blended algorithms can be implemented to enable real-time computational projections of the radiation field, thereby supporting more efficient in-situ radiation search strategies.

Research questions

1. Enhancement of the Bayesian localization framework through parameter analysis and informed prior construction. Investigate how model parameterization and the incorporation of informed priors, derived from preliminary signal characteristics from a gamma detector, improve the stability, accuracy, and computational efficiency of Bayesian radiation source localization. 
2. Influence of background radiation on localization performance. Assess how varying background radiation levels affect hybrid source localization methods and develop strategies to robustly account for background-induced noise in portable gamma spectrometry. 
3. Resolution limits in multi-source localization scenarios Investigate the capability of the model to localize two spatially proximate gamma-emitting sources and quantify its spatial resolution limits, including conditions under which source signals become indistinguishable. 
4. Computational performance and real-time applicability. Optimize the computational performance of hybrid localization algorithms and identify the requirements for achieving reliable real-time source localization under low signal-to-noise and operationally constrained conditions

Preliminary results

The results demonstrate that the performance of Bayesian radiation source localization strongly depends on the use of informed priors and appropriate parameterization. The two-step workflow, in which preliminary estimates of source distance and activity are used to construct informed priors, consistently improved the stability and accuracy of posterior estimates compared to the use of generic priors. This approach also reduced computational cost by effectively constraining the parameter space. These improvements were most pronounced in low signal-to-noise ratio (SNR) conditions and near the detection limit, where the model with generic priors often exhibited unstable or ambiguous solutions (Paper I).

At the same time, the study shows that the fundamental limitation of the method is governed by signal quality rather than model structure. While signal smoothing techniques, such as Savitzky–Golay filtering, improved robustness and reduced noise in marginal cases, they could not compensate for insufficient signal signal-to-background ratio (SNR). The results therefore define a clear operational domain in which Bayesian methods provide practical benefit – namely, in scenarios with constrained accessibility, large search areas, or weak signals – while highlighting that reliable localization ultimately requires a sufficient SNR and careful integration of prior information 

Significance

This work clarifies the practical conditions under which Bayesian methods provide a meaningful advantage for radiation source localization. By systematically analysing parameter sensitivity, prior construction, and signal quality, the study demonstrates that model performance is primarily driven by the integration of informed priors and the effective reduction of the parameter space. This shifts the focus from purely methodological development of mobile radiometry toward data-informed modelling strategies that improve both robustness and computational efficiency in realistic scenarios.

The results are particularly relevant for operational applications of mobile gamma spectrometry, where measurements are often limited by low SNR, restricted accessibility, and time constraints. By defining the limits of applicability and identifying conditions for reliable performance, this work provides a foundation for deploying Bayesian localization methods in real-world search missions, including complex environments where traditional approaches are insufficient.

Published studies

Dvornik, A., Finck, R. and Rääf, C. (2026). Enhancing Bayesian methods for radioactive source localization: a parameter study, prior construction and signal smoothing. Journal of Environmental Radioactivity, 295, 107960. https://doi.org/10.1016/j.jenvrad.2026.107960