Abstract: The land surface temperature (LST) and land surface emissivity (LSE) derived from the moderate-resolution imaging spectroradiometer (MODIS) data have been widely used for climate monitoring, environmental assessment, and prevention of agricultural disasters. However, their accuracy is insufficient due to the influence of factors such as cloud, aerosol, precipitable water vapor, and mixed pixels. To enhance the inversion accuracy of both LST and LSE, this study proposed a deep learning-based iterative optimization strategy. First, the radiative transfer equation was used for physical logic reasoning to ensure the input and output variables for deep learning meet the parameter inversion theory and judgment conditions. Second, MODTRAN4 was employed to simulate and iteratively optimize the MODIS thermal infrared bands. The feasibility of the optimization was verified, and the optimal band combination was selected. Third, the brightness temperature, LST, and LSE data of five MODIS thermal infrared bands were collected. An iterative fine-tuning strategy was constructed based on the Adam optimizer to gradually optimize the deep learning neural network, thereby obtaining more accurate LST and LSE products. Finally, the optimized network was retrained and applied to the inversion of the MODIS remote sensing data of North America. In the experiments using simulated data, compared to the 4-band combination, the 5-band input for LST inversion led to a mean absolute error (MAE) decreasing from 0.747 5 K to 0.583 5 K and a Pearson correlation coefficient (PCC) increasing from 0.997 7 to 0.998 6. Meanwhile, the LSE inversion accuracy was also significantly enhanced. The optimized disturbance simulation data exhibited a minimal error compared to the actual raw data. The optimization of the actual data shows that through iterative fine-tuning, the LST inversion yielded a MAE decreasing from 1.823 7 K to 1.154 3 K and a PCC increasing from 0.980 3 to 0.991 8. Measured data were then introduced to validate the simulation data, with MAE decreasing from 2.180 4 K to 1.828 0 K and PCC increasing from 0.913 5 to 0.941 8. These results suggest that the simulation data are more close to actual values, further confirming the optimization effects of the strategy. Overall, this study provides reliable data support for fields such as climate and environment, holding broad application prospects.