An optimization AI model based on MoE for retrieving land surface temperature and emissivity

The main drawbacks of contemporary MODIS land surface temperature (LST) and emissivity (LSE) outputs stem from cloud pollution, imprecision in atmospheric moisture vapor rectification, preliminary miscalculations of variable surface emissivity, and heterogeneity at the sub-pixel level, leading to significant potential for advancements in the reliability of these outputs. This study proposes an iterative optimization strategy based on an Optimization Artificial Intelligence Model Based on Mixture of Experts (Opt-MoE) model. Initially, a logical analysis based on physical principles was conducted using the radiative transfer equation to verify that the model’s input and output parameters align with the principles and determination criteria of parameter inversion. Subsequently, the radiative transfer framework was applied to model and progressively refine the MODIS thermal infrared channels, confirming the practicality of enhancement and determining the most suitable band configuration. Then, brightness temperature, LST, and LSE product data were collected in five MODIS thermal infrared bands. An iterative fine-tuning strategy was constructed based on the Adam optimizer to gradually optimize the Opt-MoE and obtain a product dataset that more closely approximates the real LST and LSE. Lastly, the Opt-MoE underwent retraining and was implemented for retrieving MODIS information. Outcomes demonstrate that, during the experiment with modeled data, in contrast to the four-band arrangement, the mean absolute error (MAE) in LST retrieval employing five-band inputs dropped from 0.6346 K to 0.5238 K, the Pearson correlation coefficient (PCC) rose from 0.9981 to 0.9993, and the precision of LSE retrieval likewise experienced substantial enhancement. Additionally, the discrepancy between the refined perturbed simulation data and the true simulation data was negligible. The optimization validation using observational data showed that, after iterative fine-tuning, the MAE of LST retrieval decreased from 1.7748 K to 0.7199 K, while the PCC increased from 0.9817 to 0.9973.