Air temperature prediction models using satellite data are based on two variables of land surface temperature and vegetation cover index. These variables are obtained by atmospheric corrections in the values for the above data. Water vapor, ozone, and atmospheric aerosol optical depth are required for the atmospheric correction of visible bands. However, no measurements are available for these parameters in most locations of Iran. Using the common methods, land surface temperature can be measured accurately at 2 ° C. Given these limitations, efforts are made in this study to evaluate the accuracy of predicting maximum air temperature when uncorrected atmospheric data from the NOAA Satellite are used by a neural network. For this purpose, various neural network models were constructed from different combinations of data from 4 bands of NOAA satellite and 3 different geographical variables as inputs to the model in order to select the best model. The results showed that the best neural network was the one consisting of 6 neurons as the input layer (including 4 bands of NOAA satellite, day of the year, and altitude) and 19 neurons in the hidden layer. In this structure, about 91.4% of the results were found to be accurate at 3 ° C and the statistical criteria of R², RMSE, and MBE were found to be 0.62, 1.7 ° C, and -0.01 ° C, respectively.