Identification of low-count, low-resolution gamma spectral radionuclide using 2DCNN-BiLSTM neural network

To address the challenges of feature extraction and low classification accuracy in low-count, low-resolution gamma spectra with emerging nuclide identification methods, we proposed a novel approach using a two-dimensional convolutional bidirectional long short-term memory neural network (2DCNN-BiLSTM). This method extracts spatial features from gamma spectrum gray images using two-dimensional convolution operations and analyzes temporal features with a bidirectional long short-term memory neural network, leveraging spatiotemporal dependencies for nuclide classification. In simulation experiments, we modeled a NaI(Tl) detector using Geant4 and simulated various types of gamma spectra. Results showed that the 2DCNN-BiLSTM model achieved an average identification accuracy of 96.58 %, surpassing the 95.70 %, 92.38 %, and 92.19 % accuracies of 2D-VGG16, 1D-CNN, and PCA-BPNN models, respectively. Additionally, 2DCNN-BiLSTM demonstrated superior performance in resolving overlapping peaks in multi-source gamma spectra, with parsing accuracies of 97.95 %, 96.38 %, and 93.17 %. The method also exhibited good generalization in low-count, background noise, and spectrum peak drift scenarios and showed usability on experimental data from a NaI(Tl) detector. These findings show that the proposed method can be applied to the task of nuclide identification in low-count, low-resolution gamma energy spectra obtained from short-term measurements, providing some insight into rapid nuclide identification.