Application of neural networks in the analyses of gamma spectra collected during UAV flights

Abstract

In the characterization of nuclear contamination, UAVs with radiological equipment are an interesting approach for radiation mapping because of their autonomy and flexibility. Yet, UAVs have limitations and combined with the dynamic character of a flight this will present gamma spectra with poor statistics. To tackle the difficulties in quantification and identification, a new approach with neural networks is proposed, where a model is trained on similar data as collected by the UAV system. This thesis aimed to investigate if it was possible to implement an ANN that could identify 137Cs in various background cases. To achieve this, a DenseNet architecture was implemented with TensorFlow and the model was trained with 6500 spectra of a weak 137Cs source at different distances. This model was tested on a validation dataset which consisted of 2090 spectra collected with differences in time and distance, these results were then compared with a standard radionuclide identification method (i.e. MultiSpect). The model reached an accuracy of 85% on the validation set, MultiSpect only achieved 44%. At the base of misclassified spectra by the model was a low amount of counts in the photopeak, the minimum amount needed for good results was 34±3 counts. Reducing the number of channels within the spectrum lead to little improvements in accuracy. In conclusion, with the use of a strong variety of background situations, this method is an effective means for automated radionuclide identification and could be expanded with more radionuclides.