*Article* **Comparison of Three Convolution Neural Network Schemes to Retrieve Temperature and Humidity Profiles from the FY4A GIIRS Observations**

**Shuhan Yao and Li Guan \***

Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Key Laboratory for Aerosol-Cloud-Precipitation of China Meteorological Administration, Nanjing University of Information Science and Technology, Nanjing 210044, China

**\*** Correspondence: liguan@nuist.edu.cn

**Abstract:** FY4A/GIIRS (Geostationary Interferometric Infrared Sounder) is the first infrared hyperspectral atmospheric vertical sounder onboard a geostationary satellite. It can achieve observations of atmospheric temperature and humidity profiles with high vertical and temporal resolutions. Presently, convolutional neural network algorithms are relatively less used in the field of atmospheric profile retrieval, and different convolutional neural network approaches have different characteristics. The one-dimensional convolutional neural network scheme 1D-Net and two three-dimensional retrieval schemes U-Net 1 and U-Net 2 are used to achieve atmospheric temperature and humidity profiles under all skies based on GIIRS-observed brightness temperatures in this paper. After validation with test training data, the retrievals of different schemes derived from actual GIIRS observations and level 2 operational products were verified with ERA5 reanalysis data and radiosonde measurements in summer and winter respectively. The retrieved three-dimensional temperature and humidity fields from U-Net 1 and U-Net 2 are closer to the ERA5 reanalysis field in both distribution and value than the retrievals from the 1D-Net scheme and level 2 operational products. In particular, the inversion field of the U-Net 2 scheme is more continuous in space. Compared with radiosonde observations, the accuracy of the level 2 temperature product is the highest when the field of view is completely clear both in winter and summer month. The root mean square error (RMSE) of temperature retrieval of the two U-Net schemes is the second highest, and the RMSE and bias of the 1D-Net scheme are both large. Two U-Net schemes overestimate the temperature and humidity slightly in winter and underestimate it in summer in both clear and all sky cases. Under all sky conditions, the temperature retrieval RMSE and bias of the two U-Net schemes above 800 hPa are lower than those of the level 2 products, especially the U-Net 2 scheme with an RMSE of approximately 2.5 K. The U-Net 2 scheme bias is the smallest, with a value of approximately 0.5 K in winter. Since the level 2 product only provides the atmospheric temperature above the cloud top, it indicates that its temperature product accuracy is very low when the field of view is influenced by clouds. The humidity retrieval RMSEs of the two U-Net schemes is within 2 g/kg, better than that of the 1D-Net scheme. The retrieval accuracy of the U-Net 2 scheme is approximately 0.3 g/kg better than that of the U-Net 1 scheme below 600 hPa in winter. Level 2 does not provide humidity products. The summer humidity retrieval is worse than in winter. In general, among the three deep machine learning algorithms, 1D-Net has a large retrieval error, and the temperature and humidity from U-Net 2 have the highest accuracy. The retrieval speeds of the two U-Net schemes are nearly the same, and both are faster than that of scheme 1D-Net.

**Keywords:** FY4A/GIIRS; temperature and humidity profiles; convolutional neural network (CNN); U-Net

**Citation:** Yao, S.; Guan, L. Comparison of Three Convolution Neural Network Schemes to Retrieve Temperature and Humidity Profiles from the FY4A GIIRS Observations. *Remote Sens.* **2022**, *14*, 5112. https:// doi.org/10.3390/rs14205112

Academic Editor: Ji Zhou

Received: 26 August 2022 Accepted: 10 October 2022 Published: 13 October 2022

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

#### **1. Introduction**

Atmospheric temperature and humidity profiles are two important parameters to study atmospheric state and play an important role in the research of atmospheric science. Accurately obtaining these parameters is of great significance to improve the accuracy of numerical weather forecast and short-term weather warning and forecast [1]. The traditional way of obtaining atmospheric temperature and humidity profiles through radiosonde observations has limited spatial representativeness due to the influence of geographical conditions and other factors and has failed to meet the needs of operational and modern meteorological development. However, the development of satellite remote sensing technology has made up for the shortage of radiosondes and provided technical support for the acquisition of global atmospheric temperature and humidity profile distributions with high spatial-temporal resolution [2].

The horizontal and vertical resolutions and accuracies of atmospheric temperature and humidity profiles must meet the requirements of numerical weather prediction systems, so it is necessary to study the remote sensing accuracy that can be achieved by using satellite technology and to study which retrieval method can obtain the optimal retrieval results. The number of setting channels for atmospheric sounding radiometers uploaded on meteorological satellite platforms determines their vertical observation resolution. Due to the small number of channels, the spectral resolution of microwave and imager instruments is low, and the weight function is too wide, so it is impossible to retrieve the fine atmospheric profile in the vertical direction. In contrast, the spaceborne infrared hyperspectral atmospheric vertical sounders have thousands of channels in the thermal infrared band with high spectral resolution and narrow weighting function and can obtain high vertical resolution three-dimensional finer observations of atmospheric temperature and humidity parameters.

The Geostationary Interferometric Infrared Sounder (GIIRS) onboard FY-4A launched on 11 December 2016 (China's new generation of quantitative remote sensing meteorological satellites with geostationary orbits) is the first infrared hyperspectral atmospheric sounder mounted on a geostationary meteorological satellite. The GIIRS can provide time-continuous, high-spectral-resolution atmospheric sounding information [3], which is used to retrieve the vertical structure of atmospheric temperature and humidity parameters with high vertical resolution. The improved vertical resolution provides higher accuracy services for numerical weather forecasting, weather monitoring, and warning [4].

To date, the traditional methods commonly used to retrieve atmospheric temperature and humidity profiles based on satellite-based infrared hyperspectral observations are statistical regression methods and physical retrieval methods. The statistical regression approaches include the eigenvector method [5–7], empirical orthogonal function expansion method [8], and least squares method. Although the statistical regression algorithm is simple to calculate and retrieval results are more stable, it does not consider the physical nature of the atmospheric radiation transmission process, and the retrieval accuracy needs to be improved. Moreover, the method relies on training samples and cannot be applied to areas where the sample information is not sufficiently representative. Although the retrieval accuracy is high, the physical retrieval method, such as the one-dimensional variational method, requires an initial field, complex physical processes to be considered, and a long computation time [9,10].

With the continuous development of artificial intelligence, machine learning algorithms have been applied to various research fields and have demonstrated a powerful ability to handle big data. Machine learning algorithms with features such as adaptive, self-organizing, and real-time learning have gradually been introduced into the field of meteorology, providing new ideas for atmospheric remote sensing. Singh et al. [11] retrieved atmospheric temperature and humidity profiles from microwave and infrared hyperspectral data, respectively, with a shallow learning neural network approach, and the results showed good agreement for all atmospheric pressure levels except for below 850 hPa. The deep learning algorithm is now also gradually applied to satellite remote

sensing retrieval with the continuous development of machine learning. For example, Malmgren-Hansen et al. retrieved atmospheric temperature profiles with a convolutional neural network (CNN) algorithm based on infrared atmospheric sounding interferometer (IASI) observations and found that the CNN algorithm has higher retrieval accuracy than the linear regression method [12].

As one of the representative methods of deep machine learning (namely, hierarchical machine learning methods including multilevel nonlinear transformations), the convolutional neural network algorithm mainly integrates feature extraction into multilayer perceptions through structural reorganization and weight reduction. It unifies feature representation and regression prediction with obvious advantages compared to shallow models in feature extraction and modelling and has achieved good performance in image recognition and classification [13–15]. However, its application in the field of atmospheric parameter profile retrieval has rarely been reported in the literature. Therefore, three convolutional neural network schemes are applied to retrieve atmospheric temperature and humidity profiles based on the infrared hyperspectral GIIRS observations in this paper, namely, the traditional convolutional neural network scheme 1D-Net to retrieve the onedimensional atmospheric temperature and humidity profiles and the U-Net 1 and U-Net 2 schemes to retrieve the three-dimensional profiles. Meanwhile, the retrieval accuracy and efficiency of each scheme are evaluated by comparison and verification.

#### **2. Data**

#### *2.1. GIIRS Data*

GIIRS is the first hyperspectral infrared instrument onboard the geostationary meteorological satellite, and two infrared spectral bands (longwavelength IR (LWIR) band of 700–1130 cm−<sup>1</sup> and the medium-wavelength IR (MWIR) band of 1650–2250 cm<sup>−</sup>1) with a spectral resolution of 0.625 cm−<sup>1</sup> are designed to detect the atmosphere. There are 1650 channels in total, and the nadir spatial resolution is 16 km. The LWIR band is 689 channels containing the CO2 absorption band near 15 μm, the thermal infrared window region in the 8–12 μm and the 9.6 μm O3 absorption band. The MWIR band includes the strong water vapor absorption band (5–8 μm) centered at 6.3 μm, which can be used to retrieve atmospheric water vapor profiles, as well as the CO2 absorption band near 4.3 μm, with a total of 961 channels. The main GIIRS instrument performance parameters are given in Table 1. GIIRS observes China and its surrounding area (3◦~55◦N, 66◦~144◦E), including 7 latitude belts from north to south, with each belt taking 15 min. Each scan belt has 59 fields of regard (FORs), and each FOR contains 128 fields of view (FOVs). Each GIIRS FOR is composed of 128 FOVs arranged in a 32 × 4 array.

**Table 1.** Specifications of FY4A/GIIRS.


GIIRS Level 1 (L1) radiation observed data for the whole month of February 2020 and February 2021 were used in this study, including radiance measurement, latitude, longitude, and satellite zenith angle information. At the same time, cloud mask (CLM)

and atmospheric temperature profile products from GIIRS Level 2 (L2) operational products were used. The GIIRS L1 and L2 datasets are from the Chinese National Satellite Meteorological Center (http://satellite.nsmc.org.cn, accessed on 12 January 2022).
