
FY4A/GIIRS is the first geostationary orbit interferometric vertical sensor. It adopts Michelson interference spectroscopy. By measuring the change in target radiation intensity with the moving distance of the moving mirror, the atmospheric infrared radiation spectrum can be obtained by Fourier transform. Since different spectral channels reflect the atmospheric radiation contribution at different heights, a group of appropriate spectral channels are selected. According to the principle of atmospheric infrared radiation transmission, the vertical distribution of parameters such as temperature can be obtained through numerical calculation (Zhang et al., 2016).
FY4A/GIIRS has 1650 spectral channels belonging to the longwave infrared and mediumwave infrared bands, with a spectral resolution of 0.625 cm^{−1} wave number. The spatial resolution of the infrared channels is 16 km, and the observation region is 5°–55°N, 60°–140°E, covering China and its surrounding areas. The observation times are 0000, 0200, 0400, 0600, 0800, 1000, 1200, 1400, 2000, and 2200 UTC, for a total number of 10 times a day.
Because the temperature profile retrieved by FY4A/GIIRS is from the infrared band and the ability of infrared wave to penetrate the clouds is poor, the accuracy of the temperature is higher under clear sky, thin cloud conditions, or above the cloud top height. When deep cloud system activities occur, the accuracy of the temperature in and under the clouds is relatively lower. The FY4A/GIIRS temperature profile dataset also provides the data quality flag, according to the temperature. In this paper, we choose the data with high accuracy (quality flag is 0 or 1) for application research (Fig. 1), and the green (quality flag: 0) and blue (quality flag: 1) points marked in Fig. 1 are chosen in this paper.
Figure 1. FY4A/GIIRS temperature quality flag at 850 hPa (green: 00_perfect; blue: 01_good; red: bad) and FY4A/AGRI infrared channel cloud image at 0000 UTC 7 November 2021.
The data used in this paper include temperature and its quality flag. The temporal and spatial resolutions are the same as the observation of the instrument. There are 101 layers in the vertical direction. The pressure layers below 100 hPa are 103, 110, 118, 126, 134, 142, 151, 160, 170, 180, 190, 200, 212, 223, 235, 247, 260, 273, 286, 300, 314, 329, 344, 359, 375, 391, 407, 424, 442, 460, 478, 497, 516, 535, 555, 576, 596, 617, 639, 661, 685, 707, 730, 754, 778, 802, 827, 853, 879, 905, 932, 958, 986, 1014, 1042, 1071, and 1100 hPa. The time periods of data used in this paper are from October to December 2020, January to March 2021, and October to November 2021.

Because the research focus of this paper is the application of FY4A/GIIRS temperature in cold wave monitoring, the temperatures of meteorological radiosonde in China and high latitudes, which are covered by FY4A/GIIRS, are selected as the true values in the validation. The time periods of meteorological sounding data used are from October to December 2020, January to March 2021, and October to November 2021. There are two times each day, 0000 and 1200 UTC, and 11 layers in the vertical levels. Considering the influence of terrain height, a total of 10 layers are selected for validation in this paper, which are 100, 150, 200, 250, 300, 400, 500, 700, 850, and 925 hPa (Zhi and Xu, 2013; Chen et al., 2017).
The location, number (N01–N109), and terrain height of the radiosonde stations selected for validation are shown in Fig. 2, including 13 international meteorological radiosonde stations (9 in Russia, 3 in Mongolia, and 1 in Kazakhstan) and 96 Chinese meteorological radiosonde stations, for a total of 109. Most of the selected stations have low altitude and can effectively detect the temperature at 850 hPa.

The meteorological reanalysis data are used to evaluate the reconstruction effect of FY4A/GIIRS temperature in the cloudy area where the data are missing. Temperature of the fifth generation ECMWF reanalysis (ERA5) dataset (Hersbach et al., 2020) is selected. The reanalysis dataset is a fusion of model data and global observation data; the unit of the temperature data is K; the horizontal resolution is 0.25°, which is divided into 37 vertical layers; 500 and 850hPa pressure layers are selected; and the time resolution is 1 h.

With the requirements of the continuity of weather process and the accuracy and beauty of images in meteorological services, the FY4A/GIIRS temperature field in cloudy areas or missing data areas must be reconstructed to form a spatially continuous and uniform temperature field. At present, there are many kinds of data interpolation algorithms (Zhang et al., 2009; Liu and Hu, 2010), such as Kriging interpolation method, inverse distance weighted method (IDW), spline interpolation method, and Cressman interpolation method. Cressman objective analysis method is a gradually revised interpolation method that interpolates discrete points into regular grid points, bringing less error. It is widely used in various diagnostic analyses in the field of meteorology. Based on the analysis of common interpolation algorithms, in this paper, the method of stepbystep interpolation analysis is adopted to optimize the Cressman interpolation algorithm and realize improvements in calculation speed and highprecision iteration. See Eqs. (1) and (2) for Cressman interpolation:
$$\hspace{42pt} \alpha ' = {\alpha _0} + \Delta {\alpha _{i,j}}, $$ (1) $$\hspace{42pt} \Delta {\alpha _{i,j}} = \dfrac{{ \displaystyle\sum\nolimits_{k = 1}^K \left( {W_{i,j,k}^2\Delta {\alpha _k}} \right)}}{{ \displaystyle\sum\nolimits_{k = 1}^K {W_{i,j,k}}}}, $$ (2) where
$\alpha $ is the meteorological variable (temperature T in this paper),${\alpha _0}$ is the first guess value of the variable$\alpha $ on the grid point$(i,j)$ , the first guess value selected in this paper is 0 (${\alpha _0} = 0$ ),$\alpha '$ is the revised value of the variable$\alpha $ on the grid point$(i,j)$ ,$ \Delta {\alpha _k}$ is the difference between the observed value of the observation point k and the first guess value (in this paper,${\alpha _0} = 0$ ;$ \Delta {\alpha _k}$ is the observed value of observation point k), and${W_{i,j,k}}$ is the weight function. The weight function is determined by Eq. (3):$$ {W_{i,j,k}} = \left\{ \begin{gathered} \frac{{{R^2}  d_{i,j,k}^2}}{{{R^2} + d_{i,j,k}^2}} \quad({d_{i,j,k}} < R) \\ 0\quad\quad\quad\quad\;\;({d_{i,j,k}} \geqslant R) \\ \end{gathered} \right., $$ (3) where R is the influence radius. In this paper, according to the maximum spatial resolution of FY4A/GIIRS temperature product of 16 km, R = 30 km is selected;
${d_{i,j,k}}$ is the distance from the grid point$(i,j)$ to the observation point k. The calculation method is Eq. (4):$$\begin{aligned}[b] {d_{i,j,k}} & = r \times {\text{arccos[cos(}}\frac{{{\text{lat}}{1_k} \times \pi }}{{180}}) \\ & \times {\text{cos(}}\frac{{{\text{lat}}{2_{i,j}} \times \pi }}{{180}}) \times \cos (\frac{{{\text{lon}}{1_k} \times \pi }}{{180}}  \frac{{{\text{lon}}{2_{i,j}} \times \pi }}{{180}}) \\ & + \sin (\frac{{{\text{lon}}{1_k} \times \pi }}{{180}}) \times \sin (\frac{{{\text{lon}}{2_{i,j}} \times \pi }}{{180}})], \end{aligned}$$ (4) where
${\rm{lon}}{1_k}$ and${\rm{lat}}{1_k}$ are the longitude and latitude of the observation point k, respectively;${\text{lon}}{{\text{2}}_{i,j}}$ and${\text{lat}}{{\text{2}}_{i,j}}$ are the longitude and latitude of the grid point$(i,j)$ , respectively; and$r = 6371{\text{ km}}$ , which is Earth’s radius, and$ \pi = 3.1415 $ .In this paper, the spatial resolution of the grid interpolation is 0.1° by latitude and longitude. FY4A/GIIRS temperature data include about 1628 × 32 observation points for one single level at each time (101 levels in total). In order to improve the calculation speed, except for the first interpolation, the observed values of subsequent iterative interpolation algorithms are the results of the previous grid interpolation. We searched and calculated the value less than or equal to the influence radius in the range of 10 grid points for multiple interpolation iterations, greatly improving the calculation efficiency.

According to the horizontal maximum spatial resolution of FY4A/GIIRS sounding of 16 km, and considering that the spatial resolution is lower than 16 km in high latitude, we take the selected meteorological sounding station location as the center and search the nearest FY4A/GIIRS sounding point within 50 km for horizontal spatial matching. The distribution of the nearest distance between 72 meteorological sounding stations and FY4A/GIIRS sounding matching is shown in Fig. 3; the matching distance affects the accuracy of data verification to a certain extent. The shorter the matching distance is, the more objective the data accuracy verification.
Figure 3. The average matching distance between the 109 meteorological sounding stations and FY4A/GIIRS temperature sounding points of the validation samples. The horizontal coordinate is the number of meteorological sounding stations in Fig. 2.
As can be seen from Fig. 3, except for the matching distances of N03 (29862, 28.42 km in Russia), N14 (51076, 23.86 km in Xinjiang), N18 (51709, 21.24 km in Xinjiang), and N60 (50953, 23.52 km in Heilongjiang), which are more than 20 km, the matching distances of the other 105 stations are close to or less than 16 km, which is the spatial resolution of FY4A/GIIRS temperature data. Among them, the number of the Beijing meteorological station (54511) is N71, and the average matching distance is 16.00 km. The number of the Zhangjiakou meteorological station (54401) in Hebei Province is N68, and the average matching distance is 7.49 km, which is equal to or less than the horizontal spatial resolution of FY4A/GIIRS sounding.

According to the vertical layer distribution of meteorological sounding stations, 100, 150, 200, 250, 300, 400, 500, 700, 850, and 925 hPa are selected for vertical layer matching. The vertical layers of FY4A/GIIRS temperature data are 103, 151, 200, 247, 300, 407, 497, 707, 853, and 932 hPa, respectively. The difference between two data points in matching vertical layers will also give objectivity in data verification.

The FY4A/GIIRS temperature at 0000 and 1200 UTC is matched with that of the meteorological sounding station temperature at 0000 and 1200 UTC.

In FY4A/GIIRS temperature error analysis, the bias (
$ {B_i} $ ), mean bias (MB), error (${\rm{Er}}$ ), mean absolute error (MAE), rootmeansquare error (RMSE), and correlation coefficient (RR) are mainly considered. The calculation methods are Eqs. (5)–(8):$$ {\text{MB}} = \frac{1}{N}\sum\nolimits_{i = 1}^N {{B_i}},\; {B_i} = {E_i}  {O_i} , $$ (5) $$ {\text{MAE}} = \frac{1}{N}\sum\nolimits_{i = 1}^N {{{\rm{Er}}_i}},\; {{\rm{Er}}_i} = {E_i}  {O_i} , $$ (6) $$ {\text{RMSE}} = \sqrt {\frac{1}{N}\sum\nolimits_{i = 1}^N {{{(E{}_i  O{}_i)}^2}} } , $$ (7) $$ {\text{RR}} = \dfrac{{\displaystyle\sum\nolimits_{i = 1}^N {[(E{}_i  \bar E)(O{}_i  \bar O)]} }}{{\sqrt {\displaystyle\sum\nolimits_{i = 1}^N {{{(E{}_i  \bar E)}^2}} } \sqrt {\displaystyle\sum\nolimits_{i = 1}^N {{{(O{}_i  \bar O)}^2}} } }} , $$ (8) where
$ N $ is the number of samples,$ E{}_i $ is the FY4A/GIIRS temperature of sample i,$ O{}_i $ is the sounding temperature of sample i,$ \bar E $ is the average FY4A/GIIRS temperature of the$ N $ samples, and$ \bar O $ is the average sounding temperature of the$ N $ samples. 
In the period studied in this paper, through time and space matching methods for 109 sounding stations at 10 vertical levels for two kinds of datasets, the number of matching samples is
$N = 357,097$ . On average (Fig. 4), FY4A/GIIRS temperature MB = 0.07°C, MAE = 1.80°C, RMSE = 2.46°C, and RR = 0.95. From −50 to −20°C (at middle troposphere), the scatter distribution is closer to the linear regression line, and the FY4A/GIIRS temperature accuracy is higher and more stable.Figure 4. Scatter distribution of FY4A/GIIRS temperature and meteorological sounding station temperature.
The number of matching samples N of 10 vertical layers is shown in Fig. 5a. It is the most at 150 hPa (43,692) and the least at 925 hPa (21,575), and below 850 hPa, the number significantly decreases. The vertical distribution characteristics of the number of matching samples are related to the influence of clouds. When there are clouds with low cloud top height in the observation area, the temperature accuracy is less affected at the level above the cloud top height, and the temperature accuracy is low or the data are missing at the level with clouds or under the clouds.
Figure 5. The average accuracy of FY4A/GIIRS temperature in 10 vertical layers: (a) total number of matching samples; (b) RR; (c) MB (°C; red line), MAE (°C; green line), and RMSE (°C; blue line); and (d) MB (purple for 0000 UTC and green for 1200 UTC) and MAE (blue for 0000 UTC and red for 1200 UTC).
As can be seen from the distribution of RR at different levels (Fig. 5b), the maximum coefficient is 0.98 at 300, 400, and 500 hPa in the middle troposphere, and the correlation coefficient at 200 hPa is relatively low (0.86). The layers with large MB are 100 hPa (1.36°C), 925 hPa (−0.91°C), 400 hPa (0.76°C), and 850 hPa (−0.62°C) (Fig. 5c), which are generally positive in the upper troposphere and negative in the lower troposphere. The distributions of MAE and RMSE are relatively larger in the upper and lower troposphere, and relatively smaller in the middle troposphere. The minimum MAE is 1.33°C, and the minimum RMSE is 1.85°C, which are located in the middle troposphere. The maximum MAE is 2.44°C, and the maximum RMSE is 3.36°C, which are located at 925 and 100 hPa, respectively. The MAE and RMSE at 850 hPa are 2.27 and 2.99°C, respectively. The relatively large error at 100 hPa in the upper troposphere may be related to the spatial matching distance of the meteorological sounding stations. When the sounding balloon rises from the ground to the upper air, the higher the height from the ground, the greater the distance from the sounding station due to the influence of the wind field. In the data verification, this may cause a greater matching horizontal distance between FY4A/GIIRS temperature and radiosonde observation temperature. The gradually increasing error below 700 hPa may be related to the influence of clouds. It can also be seen from Fig. 5a that the number of matching samples below 700 hPa is also significantly less, and the reduction in matching samples is mainly affected by clouds.
Yin et al. (2020) showed that FY4A/GIIRS has temporal variation, where the diurnal variation in biases is obvious only for the upper tropospheric channels, and the biases for high tropospheric channels are smaller than the biases for low tropospheric channels. The FY4A/GIIRS temperature mean bias and mean absolute error show the same characteristics in vertical distribution by comparing the accuracy at 0000 and 1200 UTC (Fig. 5d), showing relatively poor data quality at 100 hPa and at lower than 800 hPa at 1200 UTC. In the upper and midtroposphere, the temporal variation is small.

Figure 6 shows the accuracy in 10 layers of 109 meteorological sounding stations. The accuracy difference of 109 stations at 250, 300, 400, 500, and 700hPa layers is small. The temperature correlation coefficient for most stations is more than 0.9, the mean bias is −1 to 1°C, the mean absolute error is 1–2°C, and the rootmeansquare error is about 1–2.5°C. At these barometric layers, the stations with relatively low correlation coefficients and large errors are N02, N23, N63, and N100–N109 (Fig. 1). The temperature accuracy of FY4A/GIIRS varies greatly at different radiosonde stations at 100, 850, and 925hPa pressure layers. Stations with large errors are N01–N26 (mainly in Russia and Xinjiang) and N56–N64 (in Northeast China). The FY4A/GIIRS temperatures in the regions where N42–N53 and N65–N109 stations are located (in Inner Mongolia, North China, the east of Northwest China, and South China) have high accuracy at most vertical levels.

In order to obtain more detailed information on FY4A/GIIRS temperature accuracy, according to the meteorological operational service for winter cold waves in China, Beijing (station No. 54511, N71) and Zhangjiakou (station No. 54401, N68) meteorological sounding stations are selected for data accuracy verification. Tables 1 and 2 show the RR, MB, MAE, RMSE, and the number of samples
$ N $ of the 10 pressure layers for the two radiosonde stations.Level (hPa) RR MB (°C) MAE (°C) RMSE (°C) N 100 0.62 0.54 1.83 3.65 460 150 0.61 −0.18 1.44 3.34 466 200 0.74 −0.19 1.59 3.53 476 250 0.76 0.13 1.56 3.50 440 300 0.80 0.04 1.26 3.13 414 400 0.89 0.92 1.44 2.55 397 500 0.93 −0.54 1.30 2.16 411 700 0.95 0.80 1.61 2.24 427 850 0.97 −0.43 1.56 2.07 396 925 0.96 −1.06 1.93 2.36 363 Average 0.02 1.55 2.95 Table 1. FY4A/GIIRS temperature accuracy at Beijing meteorological sounding station
Level (hPa) RR MB (°C) MAE (°C) RMSE (°C) N 100 0.79 −0.14 1.75 2.22 484 150 0.85 0.22 1.20 1.56 493 200 0.89 0.14 1.40 1.86 499 250 0.93 −0.10 1.25 1.60 487 300 0.95 0.13 0.97 1.29 442 400 0.97 1.11 1.30 1.59 437 500 0.97 −0.38 1.04 1.38 448 700 0.96 0.90 1.61 2.19 473 850 0.97 0.01 1.77 2.33 444 925 0.92 0.08 2.60 3.28 363 Average 0.19 1.47 1.97 Table 2. FY4A/GIIRS temperature accuracy at Zhangjiakou meteorological sounding station
The horizontal spatial matching distance of the Beijing meteorological sounding station is about 16.07 km (Fig. 3), and the number of matching samples is about 350–450 (Table 1). The correlation coefficient of the two datasets is between 0.61 and 0.97; the correlation coefficient is slightly smaller at the upper troposphere and increases gradually from the middle to lower troposphere. The 10 layers’ average MB = 0.02°C, MAE = 1.55°C, and RMSE = 2.95°C. At 850 hPa in the lower troposphere, which is very impotent in cold wave monitoring in winter, RR = 0.97, MB = −0.43°C, MAE = 1.56°C, and RMSE = 2.07°C. At 500 hPa in the middle troposphere, RR = 0.93, MB = −0.54°C, MAE = 1.30°C, and RMSE = 2.16°C.
The horizontal spatial matching distance of the Zhangjiakou meteorological sounding station in Hebei Province is about 7.49 km (Fig. 3), and the number of matching samples is slightly larger than that of the Beijing meteorological station, which is about 350–500 (Table 2). The correlation coefficient of the two data is between 0.79 and 0.97, which also shows that the highlevel correlation coefficient is slightly smaller, and the correlation coefficient below 250 hPa is greater than 0.92. On the whole, the data correlation coefficient is slightly larger than that of the Beijing meteorological station. The 10 layers’ average MB = 0.19°C, MAE = 1.47°C, and RMSE = 1.97°C. At 850 hPa in the lower troposphere, RR = 0.97, MB = 0.01°C, MAE = 1.77°C, and RMSE = 2.33°C. At 500 hPa in the midtroposphere, RR = 0.93, MB = −0.38°C, MAE = 1.04°C, and RMSE = 1.38°C. Overall, the FY4A/GIIRS temperature accuracy at the Zhangjiakou meteorological sounding station in Hebei Province is slightly better than that of the Beijing meteorological sounding station.
It can be seen from the comparison of temperature and bias of all matching samples of the two datasets one by one (Figs. 7, 8) that the trend of FY4A/GIIRS temperature at 850 hPa (Fig. 7a) at the Beijing meteorological sounding station is consistent with sounding temperature. Except for a few cases, the warming and cooling processes of the 396 matching samples agree with each other. There are more negative bias samples (Fig. 7b). The total number of samples with absolute error greater than 4°C (bias greater than 4°C and less than −4°C) is 20, the number of samples with absolute error of 3–4°C is 26, and the number of samples with absolute error greater than 3°C accounts for about 11.6% of the total number of samples (369). The FY4A/GIIRS temperature mean absolute error at 500 hPa (1.29°C) is smaller than that at 850 hPa (1.56°C) (Table 1; Figs. 7c, d), and the consistency of data accuracy is better. The number of samples with absolute error greater than 4°C is 12, the number of samples with absolute error of 3–4°C is 13, and the number of samples with absolute error greater than 3°C accounts for about 6.8% of the total number of samples (411).
Figure 7. FY4A/GIIRS temperature accuracy verification at Beijing meteorological sounding station. Temperature at (a) 850 hPa and (c) 500 hPa, and bias at (b) 850 hPa and (d) 500 hPa; sample number N = 396 at 850 hPa, and sample number N = 411 at 500 hPa.
Figure 8. As in Fig. 7, but for Zhangjiakou meteorological sounding station (sample number N = 444 at 850 hPa, and sample number N = 448 at 500 hPa).
The trend of FY4A/GIIRS temperature at 850 hPa at the Zhangjiakou meteorological sounding station (Fig. 8a) is also in agreement with the sounding temperature. The total number of samples with absolute error greater than 4°C (Fig. 8b) is 23, the number of samples with absolute error of 3–4°C is 36, and the number of samples with absolute error greater than 3°C accounts for about 13.3% of the total number of samples (444). The proportion of samples with error greater than 3°C is slightly higher than that of the Beijing meteorological sounding station. The FY4A/GIIRS temperature mean average error (1.04°C) at 500 hPa is better than that at 850 hPa (1.77°C) (Table 1; Figs. 8c, d). Similarly, the consistency of data accuracy is also better. The number of samples with absolute error greater than 4°C is 4, the number of samples with absolute error of 3–4°C is 10, and the number of samples with absolute error greater than 3°C accounts for about 3.1% of the total number of samples (448), which is better than that of the Beijing meteorological sounding station at 500 hPa.

FY4A/GIIRS has less effective sounding in cloudy areas, and the quality of most data is identified as poor (red points in Fig. 1). In order to achieve a better application effect, it is necessary to remove data with poor accuracy and only select the data with high accuracy (green and blue points in Fig. 1). Due to the influence of cloud and quality identification control, there sometimes will be a large range of missing data in some areas (Fig. 9a). The missing data must be reconstructed to meet the operational service requirements for the continuity of weather system monitoring. In this paper, the data reconstruction method described in Section 2.4 is adopted.
Figure 9. (a) FY4A/GIIRS highprecision temperature, (b) temperature after data reconstruction, and (c) ERA5 reanalysis temperature at 850 hPa at 0000 UTC 7 November 2021, and (d) the correlation coefficient spatial distribution of the two datasets in November 2021.
As shown in Figs. 9a, b, the FY4A/GIIRS temperature distribution at 850 hPa is compared before and after the data reconstruction. Following the optimized interpolation algorithm, the missing data are filled in the cloudaffected regions of Huanghuai, North China, Northeast China, central and eastern Inner Mongolia, and northwestern Mongolia, which were the key regions of cold air activity during the cold wave event on 7 November 2021. The comparison with the ERA5 reanalysis data at the same time (Fig. 9c) shows that the reconstructed FY4A/GIIRS temperature in the regions of missing data is consistent with that of the reanalysis data. It can be seen from the contour line of −4°C in Figs. 9b, c that they are all located in central Jilin, west of Liaoning, southeast of Hebei, west of Henan, west of Hubei, south of Shanxi, and south of Gansu. Research has shown that the −4°C temperature contour line at 850 hPa is the phase transition temperature of rain and snow in Beijing (Zhang et al., 2013). At the same time, FY4A/GIIRS temperature reconstruction completely retains the cold center of the highprecision and effective sounding area (Figs. 9a, b). The intensity and location of the cold center in central Inner Mongolia and central and eastern Mongolia have more fine features than that of the ERA5 reanalysis data (Fig. 9c).
Figure 9d shows the spatial distribution of FY4A/GIIRS reconstructed temperature and ERA5 temperature correlation coefficients at 850 hPa. At 0000 UTC 1–30 November 2021, the data show that except for the Tibetan Plateau, the correlation coefficient in most regions of China is greater than 0.8, and the correlation coefficient in Mongolia, the east of Northwest China, and North China is greater than 0.9 in the key areas of cold air activity. It shows that the reconstructed temperature has efficient performance in cold wave monitoring.

From 4 to 8 November 2021, a strong cold wave weather event occurred in China, with strong cooling and wide influence. The cold air activity caused a sharp drop in temperature. The daily minimum temperature in most areas of China decreased by 10–14°C, and even up to 16°C in some areas. The daily minimum temperature at many national meteorological observation stations reached or exceeded the extreme historical records in early November. It also brought a wide range of rain and snow. The daily precipitation in many places exceeded the extreme historical values. The daily snowfall in eastern and northeastern Inner Mongolia was characterized by heavy snow and extra heavy snow. FY4A/GIIRS temperature after data reconstruction is used to monitor this strong cold wave process.
It can be seen from the 24h temperature difference at 0000 UTC (Fig. 10) from 4 to 5 November that the cold air affected most of Mongolia, Xinjiang, and Inner Mongolia, and the maximum cooling center exceeded 14°C. From 5 to 6 November, the cold air pushed southeastward, and it moved faster in the western part (Gansu Province). The strong temperature decrease appeared in Gansu, Ningxia, Inner Mongolia, and west of Heilongjiang. From 6 to 7 November, the cold air reached the east of Southwest China, the east of Northwest China, North China, and Huanghuai and Jianghan regions. From 7 to 8 November, the cold air continued to move southeastward, influencing the Huanghuai, Jianghuai, and Jiangnan regions. The maximum 24h cooling was more than 14°C, even up to 16°C in some areas. Figure 11 shows the vertical profile of the 24h temperature difference from 0000 UTC 6 to 0000 UTC 7 November along the east longitudes 105°, 110°, and 115°E. It can be seen that during this cold wave process, the cold air mass was deep, reaching up to 350 hPa. The cold air affected the lower troposphere earlier than the middle layer, and the temperature drop intensity in the lower troposphere was stronger.
Figure 10. The 24h temperature difference (°C) at 850 hPa at 0000 UTC from 4 to 8 November 2021. (a) 4–5 November, (b) 5–6 November, (c) 6–7 November, and (d) 7–8 November.
Figure 11. Vertical distribution of 24h temperature difference (°C) at 0000 UTC from 6 to 7 November 2021. (a) 105°E, (b) 110°E, and (c) 115°E.
Under the influence of cold air and warm air with high humidity, rainy and snowy weather occurred in Inner Mongolia, Northeast China, North China, the Huanghuai region, and east of Northwest China from 6 to 7 November, and snowstorm or heavy snowstorm occurred in some areas. As can be seen from the evolution of the −4°C FY4A/GIIRS temperature contour line at 850 hPa (Fig. 12a), at 0000 UTC 6 November, the −4°C temperature contour line was located in the west of Northeast China, the southeast and middle of Inner Mongolia, the north of Shanxi Province, and the south of Gansu Province, which is almost consistent with the snow line monitored by the ground meteorological observation stations (Fig. 12b). At 0600 UTC 6 November, the −4°C temperature contour line was slightly extended southward, and it began to snow in the north and northeast of Hebei Province. At 2000 UTC 6 November, snowfall occurred successively in Beijing, northcentral Hebei Province, and the north of Shanxi Province, and the −4°C temperature contour line was consistent with the snow line observed on the ground. At 0000 UTC 7 November (Figs. 9b, 12), the −4°C temperature contour line began to push southward to the south of Hebei Province and west of Henan Province, and there was sleet in the areas with −4 to 0°C temperature (blue and yellow contour lines in Fig. 12). FY4A/GIIRS temperature effectively monitored the transformation of rain and snow phases during this cold wave. Generally speaking, at 850 hPa, the −4°C temperature contour line can be used as the key indicator of the snow line, −4 to 0°C is for sleet, and above 0°C is for rain.

Another application of FY4A/GIIRS temperature in cold wave monitoring is the accuracy verification of model prediction. The verification of the GRAPESGFS (Global/Regional Assimilation and Prediction System, Global Forecast System) model of 6 and 12h predicted temperature at 850 hPa (Fig. 13) shows that, compared with the FY4A/GIIRS data (Fig. 9b), the 6 and 12h predicted temperatures in Mongolia, the east of Northwest China, and Inner Mongolia are 1–4°C higher, especially in Mongolia, where the temperature difference is more than 5°C. At the beginning of the cold air advancing, the difference between the predicted temperature and FY4A observed temperature is relatively small in North China, Huanghuai, and the east of Southwest China, and the distributions of −4 and 0°C contour lines are very similar, indicating that the model has a good prediction for cold air advancing, while the predicted temperature deviated near the cold air center. From the predicted 24h temperature change, the maximum cooling center is located in the central and western part of Inner Mongolia, north and west of the cooling center monitored by FY4A (Fig. 10c), and the range of temperature decrease at the beginning of the cold air advancing is lower.
Figure 13. The 6 and 12h prediction of GRAPESGFS model at 0000 UTC 7 November 2021 at 850 hPa: (a) 6h and (b) 12h forecasting temperature, (c) 6h and (d) 12h forecasting temperature difference between GRAPESGFS model and FY4A/GIIRS, and 24h temperature change in (e) 6h and (f) 12h prediction.
Level (hPa)  RR  MB (°C)  MAE (°C)  RMSE (°C)  N 
100  0.62  0.54  1.83  3.65  460 
150  0.61  −0.18  1.44  3.34  466 
200  0.74  −0.19  1.59  3.53  476 
250  0.76  0.13  1.56  3.50  440 
300  0.80  0.04  1.26  3.13  414 
400  0.89  0.92  1.44  2.55  397 
500  0.93  −0.54  1.30  2.16  411 
700  0.95  0.80  1.61  2.24  427 
850  0.97  −0.43  1.56  2.07  396 
925  0.96  −1.06  1.93  2.36  363 
Average  0.02  1.55  2.95 