Abstract: A sparse representation based method for nuclide spectrum feature extraction is proposed. The essence of this method is to decompose the energy spectrum on the best distinguishable sparse atom. The sparse decomposition method is used to decompose the nuclide energy spectrum, and the decomposition coefficient vector is taken as the feature to represent the energy spectrum. The classification model is established by the pattern recognition algorithm to realize the nuclide identification. The main difference from the traditional sparse decomposition method is that we decompose the energy spectrum in accordance with the sparse atoms in the sequential order in sparse dictionary. In the experiments, 6 kinds of radionuclide including ²⁴¹Am, ¹³³Ba, ⁶⁰Co, ¹³⁷Cs, ¹³¹I and ¹⁵²Eu, 1200 energy spectra are used and the average nuclide identification accuracy on 7 different pattern recognition algorithms is 91.71%. The results of statistical tests show that the proposed algorithm performs significantly better than two traditional nuclide spectrum feature extraction methods.