A spectra calibration method and a radionuclide identification method are developed for the improvement of the adaptability to the nuclear detector measurement in complex environment. In view of the low identification rate of multiple nuclides caused by γ-ray energy spectrum shifting with temperature change, we propose a radionuclide identification method based on sparse representation and multi-task learning. Firstly, a transfer matrix was constructed to represent the environment variation affecting currently measured spectra. Then, the model of the measurement spectra was established, which was used to describe the instantaneous superposition of scale copies of individual nuclide sub-spectra in standard spectra library. Thus, the problem of radionuclide identification was transformed into the problem of sparse decomposition of various radionuclides. In order to solve this non-convex optimization problem, the multi-task learning method based on alternating direction multiplier method (ADMM) was developed to optimize the transfer matrix and decompose the sparse matrix simultaneously. The feasibility and effectiveness of the developed method were verified by some experiments, in which the measurement environment of CsI (Tl) detector was simulated by using the programmable temperature and humidity chamber and the real radioactive spectrum data of 11 kinds of nuclides and typical mixed nuclides were measured, respectively. Meanwhile, the single and mixed nuclide data of 27 kinds of nuclides were used, which are specified by the simulation IAEA of Monte Carlo analysis software Geant4. The experiment results show that the developed method can accurately identify a variety of commonly used nuclides even at temperature range of -20~50℃.