Enhancing Bayesian methods for radioactive source localization: a parameter study, prior construction and signal smoothing

This study clarifies when Bayesian analysis provides a practical advantage, particularly in large-area searches, hazardous or inaccessible environments, and low signal-to-noise ratio (SNR) scenarios where direct source approach is not feasible. We present a PyMC-based Bayesian framework for localizing a single unshielded gamma-emitting source and estimating its activity, with the ability to switch between generic priors and measurement-derived informed priors. Model performance was evaluated using 1240 synthetic datasets spanning varying source activities, detector-to-source distances, and background levels, and further tested using controlled field experiments. Two workflows were assessed: a rapid single-step analysis using generic priors and a two-step approach in which preliminary estimates of source distance and activity define an informed prior and constrain the effective search area. Informed priors improved parameter stability near detection limits and reduced computation compared with generic priors. Savitzky-Golay smoothing enhanced SNR and improved robustness in marginal cases but could not compensate for insufficient signal strength. These results define practical conditions under which Bayesian localization is operationally beneficial.