ObjectiveAtmospheric temperature and humidity profiles are two important parameters for studying the state of the atmosphere, which have important applications in the research of atmospheric science. FY-4A/GIIRS (Geostationary Interferometric Infrared Sounder) has achieved the first geostationary orbit infrared hyperspectral detection, which can continuously obtain high vertical resolution atmospheric temperature and humidity profile information. Currently, when clouds exist in the GIIRS field of view (FOV), the Level 2 operational products only provide temperature profiles above cloud top within the observed field of view, and do not retrieve humidity profiles for the entire field of view. In addition, the commonly used atmospheric temperature and humidity profiles retrieval algorithms, including statistical methods, physical methods, and machine learning algorithms, are only based on a single observation field of view and do not consider the continuity of spatial information (especially horizontal dimensions) and feature transformations between fields of view. Furthermore, there was a lack of methods to retrieve atmospheric temperature and humidity profiles when the observational field of view was affected by clouds. The U-Net convolutional neural network algorithm is used to achieve the GIIRS all-sky atmospheric temperature and humidity profiles retrieval, which can obtain high retrieval accuracy under cloudy field of view.
MethodsAll-sky retrieval of atmospheric temperature and humidity profiles, including clear sky and full cloud coverage fields of view, is realized with the U-Net convolutional neural network algorithm based on the GIIRS radiance observations. The algorithm converts the atmospheric temperature and humidity profiles retrieval problem into an image processing problem from the perspective of an image, and considers the image features of multiple neighboring fields of view with different weather conditions to obtain the atmospheric parameter information. This article based on radio sounding observations focuses on the accuracy assessment of the all-sky atmospheric temperature and humidity profiles retrieved by the U-Net machine learning algorithm, especially in cloudy fields of view, and analyzes the effects of different cloud amounts and different cloud optical thicknesses on the retrieval accuracy of the temperature and humidity profiles.
Results and DiscussionsFrom Fig.6, it can be shown that the
ME (Mean Error) and
RMSE (Root Mean Square Error) of clear and all-sky retrieved temperatures by the U-Net algorithm are similar, but the
RMSE of retrieval is larger below 800 hPa, especially for the clear sky in winter. From Fig.8, it can be shown that the
ME is within ±0.5 g/kg for both winter and summer, all-sky and clear, with negative
ME above 600 hPa, and the
RMSE of clear sky is slightly smaller than all-sky. In general, the U-Net algorithm has comparable retrieval capabilities for atmospheric temperature and humidity profiles retrieval in clear sky and cloudy fields of view. From Fig.9, it can be shown that the temperature retrieval error increased with increase in cloud amount in winter time. In summer, on the contrary, the retrieval error decreased with the field of view gradually filled with clouds. This indicates that the algorithm has a slightly higher retrieval accuracy in summer compared to winter, and the retrieval accuracy is higher when the field of view has more clouds, and the algorithm is very suitable for the retrieval of atmospheric temperature profiles under cloudy conditions. Figure 10 shows that the
RMSE of humidity retrieval increases with increasing cloud amount both in winter and summer. Figure 11 and 12 show that the difference of temperature retrieval error at different cloud optical thicknesses is small, while the humidity retrieval error increases with the increase of cloud optical thickness.
ConclusionsThe U-Net retrieval ability of temperature and humidity profiles with cloudy field of view is equivalent to that of clear sky, and the accuracy of temperature retrieval in summer is better than in winter, which is beneficial to the monitoring of disastrous weather in the season of frequent occurrence. In the summer when the cloud system is more active, the retrieval accuracy of the temperature profile becomes gradually higher with the increase of the clouds in the field of view, indicating that the algorithm was applied to retrieve the atmospheric temperature profile under cloudy conditions. And the GIIRS can obtain a good retrieval accuracy for thin clouds. Although the physical significance of the U-Net algorithm is not clear, it can quickly retrieve the all-sky atmospheric temperature and humidity profiles, especially in cloudy conditions, and can obtain higher retrieval accuracy.