Jonathan Oliveira dos Santos

The limited availability of experimental gamma-ray spectrometry data remains a major obstacle to the development of reliable machine learning models for predicting radiological hazard indices in building materials. Although oversampling techniques have been widely adopted to alleviate data scarcity, their relative effectiveness for radiological prediction tasks has not been systematically established. This study addresses this gap by benchmarking three oversampling strategies, Random Oversampling (ROS), Synthetic Minority Oversampling with Gaussian Noise (SMOGN), and MixUp, using a dataset of 248 HPGe gamma-ray spectra acquired from 23 categories of Brazilian building materials. Each strategy was evaluated using a Dense Rectifier Neural Network (DRNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The DRNN significantly outperformed XGBoost for D𝛾 and I𝛾 (p < 0.05), while all three models exhibited comparable performance for IBra. Under LeaveOne-Group-Out (LOGO) validation, the DRNN maintained RMSE values below 0.04 for previously unseen materials categories, demonstrating strong generalization capability. SHAP analyses confirmed the physical consistency of the learned relationships, identifying 232Th as the dominant contributor across D𝛾, I𝛾, and IBra, followed by 40K and 226Ra. Monte Carlo uncertainty propagation revealed substantially wider confidence intervals for low-activity samples than for samples with higher activity concentrations. MixUp consistently outperformed SMOGN across all models and targets, reducing RMSE by up to 63% for I𝛾 with the DRNN. Despite its simplicity, ROS matched MixUp performance for IBra and outperformed both MixUp and SMOGN for D𝛾 and I𝛾 across all evaluated models, emerging as the most effective overall oversampling strategy. These findings demonstrate that the interaction between oversampling strategy and learning architecture is a key determinant of predictive performance under limited-data conditions, while providing a validated, physically interpretable, and uncertainty-aware framework for machine learning-assisted radiological assessment of building materials.