A high-precision interpretation method for airborne gamma-ray spectra multi-nuclide analysis based on spatio-temporal neural network

Aiming at the problems of peak overlap, noise interference and statistical error faced by traditional methods in high background, low activity and multi-nuclide scenarios, as well as the limitations of deep learning models in the modeling of spatio-temporal features of energy spectra, this paper proposes a high-precision multi-nuclide interpretation method for airborne gamma spectrometry based on ConvLSTM and CNN-LSTM. A hybrid dataset was constructed through Monte Carlo simulation and the measured data from the CeBr 3 detector, covering seven nuclides: 60 Co, 137 Cs, 232 Th, 226 Ra, 238 U, 40 K, and 241 Am. ±5% channel address dynamic energy drift and natural background noise were incorporated to simulate the real measurement environment. The experimental results show that the proposed method has excellent performance: the multi-nuclide recognition accuracy of the ConvLSTM model reaches 99.65%, and that of the CNN-LSTM model reaches 92.10%, which is 35.00% higher than that of the traditional neural network. The mean absolute error (MAE) of activity quantification based on pseudo-energy spectrum labels is only 1.28%, and the sensitivity to low-activity 241 Am is increased by 37.00%. This method integrates the characteristics of energy spectra with the advantages of deep learning, enhancing the ability to analyze nuclides in complex scenarios and providing high-value technical support for nuclear radiation monitoring and nuclear emergency response.