CNN model edges NMF on RADAI benchmark for urban gamma-ray source detection at low false-alarm rates

The arXiv preprint converts sequential gamma-ray spectra into two-dimensional waterfall representations that treat consecutive time bins as input channels analogous to RGB. Three architectures were trained on the RADAI benchmark: an MLP, a CNN, and a ViT. At the one-false-alarm-per-hour operating point the CNN delivered detection, classification, and identification rates of 0.4334, 0.3965 and 0.2950, respectively, versus 0.4151, 0.3611 and 0.2625 for non-negative matrix factorization.

Urban radiological search faces non-stationary backgrounds, brief source transits and extreme class imbalance. By encoding both spectral shape and temporal persistence inside the same tensor, the CNN exploits spatial features that NMF treats as independent. Performance nevertheless drops below NMF when false-positive budgets tighten further, indicating that learned features remain sensitive to rare background fluctuations not captured in training.

If deployed on street cameras or phone networks, the same pipeline could trigger public alerts for concealed medical or industrial isotopes. Historical false alarms from NORM cargo and medical patients already strain response resources; an order-of-magnitude increase in candidate events would require new adjudication protocols before any city-scale rollout.

Next steps include controlled injections of realistic urban clutter and live streaming tests on mobile platforms to measure end-to-end latency and calibration drift.