Abstract:

To address the limitations of traditional retrieval methods under cloudy conditions and the inability of the GIIRS (Geostationary interferometric infrared sounder) operational products to provide full−layer atmospheric temperature and humidity profiles, this study proposed a deep learning based approach using convolutional neural networks (CNN). The method was applied to the GIIRS observations from December 2019 and July 2020 with  focusing on optimizing both the network architecture and its hyperparameters. Radiosonde measurements during the same periods were used as the truth to evaluate the performance of the proposed CNN retrieval algorithm alongside the GIIRS Level 2 (L2) operational products. The results showed that the CNN model was able to retrieve the full layer temperature and humidity profiles. When evaluated against 453 independent test samples, the CNN achieved a mean temperature RMSE of 3.152K across all vertical levels and a maximum correlation coefficient of 0.995, demonstrating strong generalization and predictive accuracy. With GIIRS observations, the CNN exhibited the highest temperature retrieval accuracy in the mid−to−lower troposphere under clear sky conditions, with RMSE values ranging from approximately 3K to 5K. Under cloudy conditions, while the GIIRS L2 products were limited to retrieving temperatures above the cloud tops, the CNN was capable of retrieving the full layer temperature profile. Moreover, the CNN significantly outperformed the L2 products in retrieval accuracy under cloudy conditions. Furthermore, through transfer learning, the CNN was adapted to retrieve humidity profiles for December 2019 and July 2020. In both clear−sky and cloudy conditions, the CNN achieved a full−layer humidity RMSE of less than 2g·kg⁻¹. 

Key words: Convolutional neural network, Temperature profiles, Infrared hyperspectral remote sensing