The ability to detect and identify gamma-ray sources by means of analyzing gamma-ray spectra is essential for nuclear safety and security, and accurately performing these tasks in environments with varying background radiation remains a challenge. One common approach to enhancing detection capabilities is directing research and development at novel detection materials and systems. Alternatively, detection sensitivity can be enhanced by making use of more sophisticated data processing methods on existing detection systems. Leveraging advances in data analysis methods, this dissertation introduces and characterizes novel data-driven approaches to spectral detection and identification. An emphasis is made on methods that can potentially be augmented with complementary non-radiological data (e.g., video streams), with the objective of enhancing performance by constraining models using information about the local environment.

Two general data analysis methods are examined for both detection and identification: non-negative matrix factorization (NMF) and neural networks. When applied to gamma-ray spectra, NMF yields accurate and interpretable models of background and sources using relatively few parameters. Neural networks are considered for their flexibility in design, the significant amount of active research in the area, and the ease with which models can be augmented with additional data sources. For both the NMF and neural network models, detection and identification methods are introduced, the performance of each is evaluated relative to benchmarks from the literature, and an assessment on tradeoffs, specifically as they relate to practical considerations, is discussed. The methods introduced in this work provide improvements over the examined benchmarks, and each method can be applied to existing systems. Additionally, discussion is provided on the potential to extend each method further using complementary non-radiological data.