Feature Extraction And Design For Gamma-ray Spectra For Radionuclide Identification

In this study we compare the performance of various strategies in extracting features from gammaray spectra for radionuclide identification. The primary objective of feature design is to reduce the number of dimensions for the classifier, therefore improving performance while avoiding overfitting. We used two feature extraction methods, principal component analysis (PCA) and peak integration, and also used the raw spectra. Multilayer Perceptron (MLP) classifier was used to compare the performance between the different feature extraction methods. Training and testing samples were generated with a 3”x3” NaI detector model with a source library of 33 radionuclides with a spanning set of shielding configurations. The drawn samples included variable background and mixtures of SNM (Special Nuclear Material) with masking sources. The overall performance of each feature design was assessed using the F1-score. Individual radionuclides that performed best and worst in each feature design were compared as well.