A Machine Learning Approach for Deriving Atmospheric Temperatures and Typhoon Warm Cores From FY‐3E MWTS‐3 Observations

Abstract Machine learning has gained an increasing popularity in the fields of satellite retrieval and numerical weather modeling. In this study, machine‐learning (ML) neural‐network (NN) models are utilized to retrieve atmospheric temperatures from observations of brightness temperature from the Microwave Temperature Sounder‐3 (MWTS‐3) onboard the China's first dawn‐dusk polar‐orbiting satellite Fengyun (FY)‐3E, with the ERA5 reanalysis serving as training data sets. The root mean square errors of the ML‐retrieved temperatures at all pressure levels are smaller than those obtained by a previously used traditional linear regression method compared to the ERA5 reanalysis over global oceans as well as radiosonde observations over land. Less than a 1‐week period of training data is usually sufficient for an ML NN model to converge in less than 50–100 iterations. The shortest time period of training data is 3‐days right before the testing data period. While the horizontal patterns and temporal evolutions of the ML‐retrieved warm cores of Typhoon Malakas (2022) and Typhoon Haikui (2023) in the upper troposphere compared favorably with those obtained by traditional regression methods as well as the ERA5 reanalysis in the testing periods. The vertical structures of ML‐retrieved warm cores extend further down to the middle and lower troposphere while those from the traditional regression method are confined in the upper troposphere. A comparison among ML results with additional training data sets across different seasons confirms the above conclusion.