Machine Learning‐Driven MLT Temperature Retrieval Based on O2b1Σg+ ${\text{O}}_{2}\left({\text{b}}^{1}{\Sigma }_{g}^{+}\right)$ Airglow Observation From ICON/MIGHTI

Abstract The ICON/MIGHTI designed by NASA is an advanced satellite payload for detecting temperature characteristics in the mesosphere and lower thermosphere (MLT), with an observation range of 90–140 km. However, the maximum height of its temperature profile product is only 125 km. To improve the spatial coverage, this study proposes a machine learning (ML) method called Bayesian genetic backpropagation neural network (BGBN) to achieve high‐precision and large spatial‐coverage of the temperature profile retrieval through combing airglow data observed by the instrument with its existing temperature products. By comparing and analyzing with the official temperature product (MIGHTI v05) and other ML/deep learning models, the reliability and scientificity of the BGBN model were verified. The accuracy of the BGBN was analyzed by comparing independent results of two vertical fields of view. With the help of the interpretability analysis method, the relative variation characteristics consistent with the theory of molecular energy level in the model results were further revealed. On this basis, the model was optimized to improve the spatial coverage of its temperature retrieval results. The results indicate that the average deviation between BGBN retrieval results and MIGHTI‐v05 temperature is less than 5 K, and the self‐consistency is improved by 6.31%. The optimized model extends temperature profiles to 140 km, as validated against MIGHTI v06 data, and provides effective temperature data to some extent for nighttime observations and in extreme regions (like South Atlantic Anomaly). This methodology is expected to establish an effective paradigm for synergistic fusion of machine learning with MLT retrieval.