There are some important factors that have an impact on the measurement accuracy of the temperature measurement in multi-spectral radiation, including surface emissivity of measured target, variability emissivity models and effects of high temperature thermal radiation. In this paper, these factors were analyzed. And the BP neural network improved model is applied to multi-spectral temperature measurement data processing. With a variety of launch training sample models, automatic recognition of the emissivity of the measured object model is realized. So the real temperature and spectral emissivity can be obtained. The simulation results show that the method studied in this paper can more accurately obtain the true temperature and emissivity.