Multispectral thermometry is based on Blackbody radiationlaw, and the temperature value can be calculated based on the radiation intensity and multiple sets of wavelengths. This method has become widely used in engineering practice, as it overcomes the constraints of the single spectrum and similar colorimetric spectrum requirements for colorimetric temperature measurement. In multispectral temperature inversion, the solution of spectral emissivity and multispectral data processing are the keys to accurate temperature measurement. At present, the solution of spectral emissivity is mostly based on the assumption model of spectral emissivity. When the hypothetical model is close to reality, the accuracy of the inverted temperature and spectral emissivity is very high; otherwise, the inversion result deviates significantly. For the temperature measurement of complex materials and the dynamic changes of material properties during the combustion process, the method of assuming the model of spectral emissivity is groundless; In recent years, the deep learning method based on the neural network has been applied to multispectral temperature measurement, which avoids the assumption model of spectral emissivity, and can establish the nonlinear statistical relationship between temperature and multi spectrum, but it requires massive data and supercomputing power support, and the modeling process is complicated.In order to solve the above problems, this paper proposes a multispectral temperature measurement method named multi-element extreme value optimization (MEVO) measurement method. This method utilizes the correlation between multispectral signals at different temperatures, and by analyses the relationship between the measured temperatures of each channel in the process of multispectral temperature inversion, based on the principle of multispectral radiation temperature measurement and the information correlation between the data of each channel in the process of temperature inversion, establish a multivariate temperature difference correlation function, and establish a high-precision temperature measurement model through the optimization of the correlation function. This method simplifies the modeling process to the optimization problem of multivariate temperature difference function, avoids the assumption of the relationship between spectral emissivity and other physical quantities, reduces the requirement of data sample size for deep learning methods, and simplifies the process of multispectral temperature measurement. A simple 8-channel temperature measuring device was used for experimental verification. In the experiment, we determined that the temperature emitted by the Blackbody furnace was the standard value. The spectral data of the 468~603 nm band in the 1 923.15~2 273.15 K temperature zone was calibrated, and the multispectral thermometry based on the optimization of multiple extreme values was realized. The temperature measurement accuracy is about 0.5%, and the temperature inversion time is within 2.5 s. Compared with the second measurement method (SMM) and the neural network method, the inversion accuracy is substantially improved. Moreover, the inversion speed is significantly faster than the SMM method.