For the problem that the temperature change of the sample affects the prediction result of the model, firstly, the spectrum of the same sample at different temperatures is compared with the spectrum of the same sample at the same temperature. The results show that the spectral difference at different temperatures is large. Then the effect of sample temperature on the prediction of corn crude protein model was studied. Spectral collection of samples with a crude protein content of 6.04% was performed at different temperatures, and near-infrared spectra at different temperatures were pretreated in the same way as those used in modeling, so as to eliminate the influence of factors other than the temperature on the spectra. The pre-processed spectrum is substituted into the established model for prediction. The prediction results show that the difference between the predicted result and the measured value increases as the difference between the spectral temperature and the modeled temperature increases, and the maximum error is 1.12%. In order to solve the influence of temperature on the prediction results of the model, we further analyzed the relationship between temperature and spectral data at different temperatures, and found that after removing the areas with serious noise at both ends of the spectrum, there was a certain linear relationship between spectral data at the same wavelength point at different temperatures. According to this finding, a temperature correction theory is proposed. Taking the spectrum at the time of modeling as the reference spectrum, and then using the linear regression algorithm to perform linear regression on the spectra of different wavelength points according to the linear relationship between temperature and spectrum, the difference between the spectrum at different temperatures and the reference spectrum is obtained. Finally, the spectra at different temperatures are corrected to the reference spectrum. After the spectrum is corrected by the theory, the difference between the spectra has been greatly improved compared with before the correction. The corrected spectrum is brought into the model, and most of the prediction results are improved, which meets the requirements of ±0.5% of the national standard. Finally, the temperature correction theory was verified by using 34 different samples unrelated to the modeling. The model prediction values and standard physical and chemical value determination coefficients of the crude protein before and after the spectral correction were 0.910 and 0.982, respectively, and the root means square error was 0.558 and 0.172, and the average relative error was 6.05% and 1.75%, respectively. The temperature correction theory has been temperature-corrected from the nature of near-infrared spectroscopy, providing a reference for temperature correction of other samples, which is beneficial to the promotion of handheld near-infrared spectroscopy.