Gamma-ray spectrum unfolding and air-absorbed dose-rate estimation using a two-stage spectral learning framework

Abstract

Gamma-ray spectrum unfolding and air-absorbed dose-rate estimation are important for environmental radiation monitoring. This study presents a sequential spectral analysis strategy linking incident-spectrum reconstruction with model-based dose-rate estimation. A LaBr$_{3}$:Ce response matrix was constructed from 591 monoenergetic Geant4 spectra, and 4000 mixed spectra were generated for supervised learning. Across three repeated runs, the unfolding model achieved a mean absolute error of $0.021732 \pm 0.001035$ and an $R^{2}$ of $0.932326 \pm 0.004621$. Using the reconstructed incident-spectrum weights as inputs, the direct tri-branch dose-regression model achieved an $R^{2}$ of $0.99335 \pm 0.00456$ and a mean absolute percentage error of $1.780\% \pm 0.620\%$. Application to a measured soil gamma-ray spectrum demonstrated the applicability of the unfolding model under a complex environmental condition. The results support a sequential pathway from detector spectra to reconstructed incident spectra and subsequent dose-rate estimates under the investigated conditions.