The Development of a Feature-Driven Analytical Approach for Gamma-Ray Spectral Analysis

Gamma-ray spectroscopy is an essential tool in nuclear science, nuclear security, and environmental monitoring. However, challenges arise in interpreting spectral data due to the presence of low counts, multiple sources, and dynamic backgrounds. To address these issues, a novel feature-driven analytical approach for gamma-ray spectral analysis using machine-learning techniques is developed. The method utilizes a series of random forest models for in-distribution (ID) multi-label classification, and the model-derived feature importance values to guide the out-of-distribution (OOD) detection task. The performance of this approach is quantitatively evaluated across various spectral parameters, including acquisition time, number of sources, energy of an OOD source, and background composition. Increasing the acquisition time from 1 s to 100 s leads to improved performance for multi-label classification, with 22 sources achieving F₁-scores ≥ 0.9 after 50 s acquisitions for a CLLBC handheld detector and a standoff distance of 30 cm. The feature-driven analytical approach also demonstrates robustness when handling complex source mixtures. Furthermore, it provides contextual energetic information for OOD detection. The results presented here highlight the interpretability of the approach, establishing clear links between the spectral features and underlying physics. Moreover, the approach effectively distinguishes overlapping spectral signatures of different ID gamma-ray sources, enhancing human reliability in machine learning-based gamma-ray spectral analysis. The feature-driven analytical approach offers a promising solution to automate gamma-ray spectral analysis by addressing existing limitations and providing insights into performance across diverse spectral parameters.