Error Analysis and Correction of FENGYUN-4A GIIRS Temperature Profile Products in Summer over the Qinghai–Tibet Plateau

Abstract

To understand the applicability of the temperature profile product of the FENGYUN-4A (FY-4A) geostationary interferometric infrared detector (GIIRS) in summer over the Qinghai–Tibet Plateau and to improve product quality, the error of GIIRS temperature products was analyzed based on radiosonde data. A long short-term memory network model (LSTM) was used to correct the GIIRS temperature profile product in summer over the plateau at a high altitude (above 500 hPa), and further evaluation of the corrected product was conducted. The results show that summertime GIIRS temperature retrievals over the Qinghai–Tibet Plateau had a positive bias above 150 hPa and a negative bias below 150 hPa, resulting in an overall negative bias. The root mean square error was between 2 and 2.9 K, and the root mean square error was relatively large at 100 hPa and above. The LSTM established in this study could effectively correct the GIIRS temperature over the plateau. The correlation and root mean square error of the corrected GIIRS temperature and the radiosonde observation temperature were significantly improved. Using a trained LSTM correction model to correct the hourly GIIRS temperature can improve the accuracy and usability of the product. After correction, the average bias of the GIIRS temperature compared to the ERA5 temperature product was reduced from −0.26 K to 0.06 K, and the root mean square error was reduced from 2.25 K to 1.25 K. The correction model can be applied to different seasons, and it can also correct the GIIRS temperature in larger areas based on other high-precision and high-resolution data, achieving good results, thus indicating that the correction model has universal applicability.

1. Introduction

Atmospheric temperature is one of the most important parameters in atmospheric thermodynamics, and obtaining accurate atmospheric temperature profile information plays a crucial role in understanding the thermal structural characteristics and evolution laws of precipitation weather systems. Satellite remote sensing data have the advantages of wide coverage, long continuous observation times, and no influence from the terrain. The Cross-Track Infrared Sounder (CrIS) and Infrared Atmospheric Sounding Interferometers (IASIs), along with other instruments mounted on polar-orbiting satellites, can obtain global atmospheric temperature profiles [1,2,3,4]. Polar-orbiting satellites can only detect the same area twice a day and cannot observe continuously. However, severe weather warnings, nowcasting, and short-term forecasting require nearly continuous monitoring of the vertical temperature and moisture structure of the atmosphere [5]. The new-generation geostationary meteorological satellite Fengyun-4A (FY-4A), aimed at quantitative remote sensing applications in China [6], was successfully launched on 11 December 2016. It can provide high-spatial–temporal-resolution meteorological data for China and its surrounding areas, especially for plateau areas where it is difficult to establish high-density observation stations, and it is the best way to address the lack of observation data in plateau areas. The Qinghai–Tibet Plateau, as the third pole of the world, has a profound impact on the weather and climate in China and even around the world [7]. The peak precipitation of the plateau throughout the year often occurs in July and August, and heavy precipitation is mainly concentrated in the period from early July to mid-August [8]. Therefore, research on summer precipitation on the plateau has always been of interest.

The Geostationary Interferometric Infrared Sounder (GIIRS) carried on FY-4A is the first high-spectral-resolution advanced infrared sounder on board a geostationary weather satellite [6,9,10]. It has 1650 spectral channels, significantly improving its atmospheric vertical detection capability [11,12]. The GIIRS spectral range is divided into two bands. Band 1 (the longwave band) ranges from 700 to 1130 cm⁻¹ and is mainly used for atmospheric temperature and ozone detection. Band 2 (the middle-wave band) ranges from 1650 to 2250 cm⁻¹ and is used for atmospheric humidity detection [13,14]. The temperature and humidity profile products provided by GIIRS are of great significance for weather forecasting and improving the accuracy of numerical weather prediction [15,16]. They can also compensate for the shortcomings of insufficient coverage and the long time intervals of radiosonde stations. Researchers have conducted preliminary evaluations of the quality of FY-4A GIIRS temperature and humidity profile products. He et al. [17] used global radiosonde data and the cloud classification product of the Himawari-8 satellite to evaluate the inversion accuracy of FY-4A GIIRS temperature profile products (hereinafter referred to as GIIRS temperature) under cloud-free and different cloud-type conditions. The results showed that the GIIRS temperature had an uncertainty of 2.1 K under clear sky conditions and 3.7 K under cloudy sky conditions. The accuracy of the GIIRS temperature varied under different cloud types. Du Mingbin et al. [18] conducted a quality evaluation of GIIRS temperature products for the entire year of 2020 based on radiosonde data and found that the root mean square error of the entire layer of GIIRS temperature products under clear sky conditions was about 2 K. The quality of the GIIRS temperature products was better in summer and autumn than in winter and spring, and the overall quality for low-altitude areas was better than that for high-altitude areas. Ren et al. [19] evaluated the accuracy of GIIRS temperature products using radiosonde data in winter and found a root mean square error of approximately 2.5 K and used numerical model data to reconstruct the missing data caused by the cloud blocking effect. Huang Yiwei et al. [20] studied the applicability of GIIRS temperature in the East China Sea and South China Sea during typhoon-prone periods based on reanalysis data and found that the accuracy of GIIRS temperature was slightly different in the East China Sea and South China Sea regions and that typhoons have a significant impact on GIIRS temperature profiles. The above research indicates that GIIRS temperature products have usability, but there are still uncertainties regarding the accuracy of the products in different seasons, at different levels, and in different regions. There is also a lack of more detailed research on the usability of GIIRS temperature products under the special terrain conditions on the plateau. This article analyzed the errors of GIIRS temperature products over the Qinghai–Tibet Plateau at a high altitude (above 500 hPa) during the summers of 2020–2022 (June to August) based on radiosonde data and used a long short-term memory network model (LSTM) to correct GIIRS temperatures at different stations and levels in order to better leverage the advantages of Fengyun satellite data in the research and application of summer precipitation on the plateau, provide data for the study of summer precipitation weather processes on the plateau, and provide a basis for improving the inversion accuracy of GIIRS products in summer on the plateau. Furthermore, the same correction model was established for other seasons, and a correction model was established for a larger area based on an IASI Level2 temperature product, achieving correction for GIIRS temperature products in different seasons and over a larger range and proving the universality of the correction model.

2. Materials and Methods

2.1. Data Introduction

The data used in this article included the following: (1) Radiosonde temperature data from 15 upper-air meteorological stations over 3000 m on the Qinghai–Tibet Plateau, including daily observations at 08:00 and 20:00 Beijing Time Coordinated (BTC). The spatial distribution of the upper-air stations and the topography of the Qinghai–Tibet Plateau are shown in Figure 1. All radio sounding data underwent quality control, and stations with more than 20% of data missing were not considered. (2) GIIRS temperature products from the atmospheric vertical detection product (AVP) provided by the FengYun remote sensing data network (http://satellite.nsmc.org.cn/portalsite/default.aspx, accessed on 20 May 2024). AVP products are regional atmospheric vertical detection products inverted from GIIRS data, including geographic longitude and latitude, land and sea masks, elevation, solar zenith angle, solar azimuth, satellite zenith angle, satellite azimuth, and atmospheric temperature profile products for each field of view. These products have a time resolution of 1 h and a resolution of approximately 16 km for the understellar point. There are 101 layers in the vertical direction, providing air pressure, temperature, and quality control information on temperature for the 101 layers in the range of 0–1100 hPa. Based on the product descriptions, this article considered latitude and longitude, air pressure, temperature profile, and temperature quality control data. The quality control data were mainly related to two factors, which reflected the L1-level data quality and the difference between the inversion results and the reference numerical model field [21]. We selected temperature profile products with quality control labels of 0 (representing perfect) and 1 (representing good) for subsequent analysis. Data with label 2 (representing poor quality) or unavailable or invalid data were removed. (3) Fifth-generation reanalysis (ERA5) data from the European Centre for Medium Range Weather Forecasts [22], which had a time resolution of 1 h and a horizontal resolution of 0.25° × 0.25°, with a vertical resolution of 137 levels. (4) Infrared Atmospheric Sounding Interferometer (IASI) Level2 products from the Comprehensive Large Array-Data Stewardship System (CLASS, https://www.avl.class.noaa.gov, accessed on 20 May 2024). IASIs are the most innovative and some of the key instruments on Metop satellites and are Michelson interferometers that measure in the infrared region [23,24]. The main goal of the IASI mission was to provide atmospheric emission spectra to derive temperature and humidity profiles with high vertical resolution and accuracy. An IASI scans across track in 30 successive elementary fields of view, which provides global Earth coverage twice per day and vertically. The IASI Level2 temperature products provide 100 pressure layers. All data in summer were taken from June to August from 2020 to 2022.

2.2. Data Processing

In order to facilitate the comparative analysis of the GIIRS temperature and radiosonde temperature data, it was necessary to match the two datasets spatiotemporally based on radiosonde data. Considering that the average drift of radiosonde balloons in summer is about 40 km, it takes about 50 min to rise to 100 hPa [25,26]. In terms of time matching, GIIRS temperatures within 1 h from 08:00 and 20:00 BTC were selected for each day. In terms of space matching, the nearest-neighbor matching method was used horizontally, and satellite scanning point data closest to the radiosonde station and not exceeding 0.5 degrees were selected to match the radiosonde data. Vertically, considering the high altitude of the Qinghai–Tibet Plateau, where the surface pressure in some areas is lower than 600 hPa, 10 commonly used levels above 500 hPa were selected, namely, 50, 100, 150, 200, 250, 300, 350, 400, 450, and 500 hPa. The GIIRS temperature was interpolated onto these 10 levels using the spline interpolation method, resulting in GIIRS temperature and radiosonde temperature data for 15 radiosonde stations twice a day and at 10 vertical levels. In addition, in order to avoid losing a large amount of effective information and to minimize the impact of outliers [20], data with an absolute difference between the processed GIIRS temperature and the radiosonde temperature exceeding three times the standard deviation were defined as outliers and excluded from consideration. Similarly, when comparing GIIRS temperature with ERA5 temperature products, the ERA5 data were bilinearly interpolated onto the stations to obtain ERA5 temperature products for 15 stations hourly and 10 vertical levels, where these values were interpolated as a function of log(pressure) to reduce the impact of the log relationship between pressure and height. When comparing GIIRS temperature with IASI temperature products, linear interpolation was used in the horizontal direction to interpolate the two to grid points with a resolution of 0.25° and vertically to interpolate them to the 10 commonly used levels.

2.3. Error Analysis Methods

This article selected the sample mean bias (BIAS), the root mean square error (RMSE), and the correlation coefficient (CORR) to evaluate the quality of the GIIRS temperature products. The calculation formulae were as follows:

$$\mathrm{B}\mathrm{I}\mathrm{A}\mathrm{S}=\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{X}}_{\mathrm{i}}-{\mathrm{X}}_{\mathrm{o}})}{\mathrm{n}}$$

(1)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{X}}_{\mathrm{i}}-{\mathrm{X}}_{\mathrm{o}})}^2}{\mathrm{n}}}$$

(2)

$$\mathrm{C}\mathrm{O}\mathrm{R}\mathrm{R}=\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{X}}_{\mathrm{i}}-{\overline{\mathrm{X}}}_{\mathrm{i}}){(\mathrm{X}}_{\mathrm{o}}-{\overline{\mathrm{X}}}_{\mathrm{o}})}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{X}}_{\mathrm{i}}-{\overline{\mathrm{X}}}_{\mathrm{i}})}^2}\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{X}}_{\mathrm{o}}-{\overline{\mathrm{X}}}_{\mathrm{o}})}^2}}$$

(3)

where X_i represents the GIIRS temperature data, Xo represents the radiosonde or ERA5 temperature data, i (i = 1, 2, …, n) represents the i-th matching sample, and n represents the total number of samples.

2.4. LSTM Correction Model

The long short-term memory network (LSTM) is a commonly used time-series prediction model that introduces concepts such as forget gates, input gates, and output gates [27] to achieve selective memory of information. It can effectively solve the problems of gradient explosion and vanishing in the process of processing long sequence data [28]. A schematic diagram of the neural structure of the LSTM is shown in Figure 2a, where ⊕ represents the addition operation, $\times$ represents multiplication operations, σ represents the sigmoid activation function, tanh represents the tanh activation function, x^<t> represents the data input at time t, h^<t−1> represents the historical data information calculated by the previous neuron, c^<t−1> represents the memory data formed by the previous neuron’s processing, and h^<t> represents the data information calculated by the current neuron. h^<t> is a combination of historical data and new input data and serves as the historical data input for the next neuron, and c^<t> represents the current memory data formed by the current neuron processing. Due to the independent calculation of c^<t> from h^<t>, it can form independent storage of memory data, forming the long-term memory of the LSTM network.

Meteorological data are typical time-series data, and relevant work has introduced LSTM models or their variants or other artificial intelligence methods into the field of weather forecasting to improve accuracy [29,30,31]. However, there are currently limited research on and applications of satellite product correction. This article attempted to use the LSTM model to extract the temporal characteristics of the errors between the GIIRS temperature and radiosonde temperature observations and utilize its ability to recall historical information to predict current results based on historical error information, achieving error correction of the GIIRS temperature. This article established LSTM correction models for each station and level and used radiosonde temperature data as the true values to correct the GIIRS temperature. The correction model consisted of an LSTM layer, an activation layer, and a fully connected layer. The LSTM layer contained 64 hidden neural units, and the activation layer used the tanh activation function. Finally, it was output as a neuron through the fully connected layer. A schematic diagram of the model is shown in Figure 2b. Taking the current time as t₀, the model inputs were GIIRS temperature data (including t₀) and sounding temperature data (excluding t₀) for the ten time periods prior to t₀, and output Y was the radiosonde temperature data at t₀. For each station and level of the GIIRS temperature and radiosonde temperature data, 75% was selected as the training set to train the model and 25% as the test set for performance evaluation. To maintain the consistency of the data distribution between the test set and the training set as much as possible, a random segmentation method was used to segment the data. The trained model used the temperature information from the first ten hours of GIIRS and radiosonde observations to predict the current temperature and obtain the corrected GIIRS temperature.

3. Results

3.1. Error Analysis of the GIIRS Temperature and Radiosonde Temperature

Figure 3 shows the distribution of BIAS and RMSE between the GIIRS temperature and radiosonde temperature at different levels over the Qinghai–Tibet Plateau in summer. From the perspective of average bias, all stations at the upper level (100 hPa) had a positive bias, while stations at 250, 350, and 450 hPa had a negative bias. The bias between stations at each level did not change much. On average, at each level, all stations had a negative bias, with the bias gradually decreasing from the east and west sides of the Qinghai–Tibet Plateau towards the middle. The Lhasa station, located in the westernmost part, and the Hezuo station, in the easternmost part, had the highest bias, with an absolute negative bias of over 0.6 K. The bias distribution at other stations was relatively uniform, ranging from −0.6 K to 0 K, with an average value of −0.259 K. The RMSE at the 100 hPa and 350 hPa levels was the highest, with the RMSE exceeding 2.4 K at each station. The spatial distribution of the RMSE at these two levels was also relatively uniform. The RMSE at 250 hPa and 450 hPa was relatively small, ranging from 1.8 to 2.8 K, and the RMSE at the stations on the east and west sides of the Qinghai–Tibet Plateau was slightly higher than at the middle station. The average RMSE of each layer was between 2.2 and 2.6 K, and the average RMSE of the southwest station was about 0.2 K higher than that of the northeast station.

Figure 4 shows the time series of the mean bias and root mean square error between the GIIRS temperature and radiosonde temperature at 08:00 BTC (a) and 20:00 BTC (b). The mean bias was obtained by calculating the average deviation between the GIIRS temperature and the radiosonde temperature at each level for all stations for a specific day at 08:00 and 20:00. A similar process was used for the RMES. The time series of the error of the GIIRS temperature in summer over the Qinghai–Tibet Plateau was relatively stable, with little difference between different years and months. The RMSE was stable between 2 and 3 K, while the BIAS was mostly negative, concentrated between −1 K and 0 K. The error between daytime (08:00) and nighttime (20:00) was slightly different. The average RMSE at nighttime (20:00) was 2.34 K, which was slightly higher than the 2.24 K value during the daytime (08:00). The primary reason for this was an undetected cloud blocking effect that contributed to the differences. Due to the weak penetration ability of infrared channel radiation into clouds and infrared channel radiation being greatly affected by cloud emissions, attenuation, and other factors, under cloudy sky conditions, the error of temperature retrievals is often higher than that under clear sky conditions [32]. Affected by thermal conditions, convection often develops from afternoon to evening [33], with more convection at night and more cloudy conditions, which may lead to more undetected cloud pollution.

3.2. Analysis of LSTM Correction Model Effectiveness

The analysis in the previous section showed that there was a certain degree of error in the GIIRS temperature during summer on the Qinghai–Tibet Plateau. It was necessary to correct the error and improve product quality. Considering that different stations and levels have different error distribution characteristics, LSTM correction models were established for each station and level. Taking the 50, 150, 250, and 350 hPa of Lhasa Station (station number: 55591) with larger errors as an example, the correction effect of the model on the test set was analyzed. Figure 5 shows the predictive performance of the correction models established at the four levels of 50, 150, 250, and 350 hPa at Lhasa Station on the test set. The horizontal axis represents the sample number in the test set. For each model, which included 3 years × 92 days in summer × 2 h per day, a total of 552 original samples were obtained. After removing missing or abnormal values, about 400 samples were left, and the test set accounted for 25%, or about 100 samples. Figure 5 shows that the models at each level had good correction effects. After model correction, extreme values of GIIRS temperature were significantly reduced, and the fluctuation range was smaller, being closer to that of the sounding temperature. The results before and after correction for the test set are shown in

Table 1. The improvement of LSTM correction models for the four levels of 50, 150, 250, and 350 hPa at Lhasa Station on the test set.

	Before Correction			After Correction			Proportion of Improvement
Levels	BIAS (K)	RMSE (K)	CORR	BIAS (K)	RMSE (K)	CORR	BIAS (%)	RMSE (%)	CORR (%)
50 hPa	−0.65	3.16	−0.09	−0.32	2.20	0.19	51	30	111
150 hPa	1.25	2.42	0.06	0.30	0.99	0.57	76	59	850
250 hPa	−1.49	2.48	0.34	0.11	1.02	0.69	93	59	103
350 hPa	−1.94	3.25	0.07	−0.12	1.63	0.56	94	50	700

: the calculated BIAS between the GIIRS temperature and the radiosonde temperature of Lhasa Station at 50, 150, 250, and 350 hPa was reduced from −0.65 K, 1.25 K, −1.49 K, and −1.94 K to −0.32 K, 0.30 K, 0.11 K, and −0.12 K, and the RMSE was reduced from 3.16 K, 2.42 K, 2.48 K, and 3.25 K to 2.20 K, 0.99 K, 1.02 K, and 1.63 K, respectively. At the same time, the correlation between GIIRS temperature and radiosonde temperature was significantly improved. The correlation coefficients between the two before correction at the four levels of 50, 150, 250, and 350 hPa were −0.09, 0.06, 0.34, and 0.07, respectively. After correction, they increased to 0.19, 0.57, 0.69, and 0.56, respectively. However, due to the reduced extreme values of GIIRS temperature after model correction, the model’s ability to correct more extreme values was slightly lacking. The models for other stations and levels on the test set were able to achieve similarly good correction results (figure omitted).

3.3. Comparative Analysis of Errors between GIIRS Temperature and Radiosonde Temperature before and after Product Correction

Figure 6 shows the vertical distribution of the BIAS, RMSE, and CORR between the GIIRS temperature and the radiosonde temperature before and after correction at 15 stations on the Qinghai–Tibet Plateau, using the test set, in summer. Before correction, there were differences in the BIAS of the GIIRS temperature relative to the radiosonde temperature observations at various levels, ranging from −1.6 K to 1.5 K. Among them, the BIAS at 100, 250, and 350 hPa was the highest, and the corrected BIAS at each level remained stable at around 0 K, with a large negative BIAS of −0.3 K only in the upper layers above 100 hPa. The RMSE before correction was between 2.0 K and 2.9 K, while the RMSE at 450 hPa, 350 hPa, 250 hPa, and 100 hPa was relatively large. The RMSE after correction decreased to 1.1–1.9 K, and the vertical distribution of the RMSE was similar to the results before correction. The reduction in the RMSE at each level was equivalent, indicating that the improvement in data quality by the correction models at each level was relatively consistent. Before correction, the CORR was relatively consistent below 250 hPa, at around 0.8, and lower above 250 hPa, with a minimum of only 0.4. After correction, the CORR increased on average by about 0.05 below 250 hPa and increased more above 250 hPa. The CORR mostly exceeded 0.7, and the average of the entire layer increased from 0.69 to 0.82.

Figure 7 shows the probability density distribution between the GIIRS temperature and the radiosonde temperature at all 15 stations and 10 levels on the Qinghai–Tibet Plateau before and after correction in summer, with a total sample size of 14,041. From the figure, it can be seen that, before correction, there were many points below the diagonal in the range of 190 to 230 K, which meant that the average GIIRS temperature in the range of 190 to 230 K was slightly lower than that of the radiosonde observations, while the average GIIRS temperature in the range of 230 to 280 K was slightly higher than that of the radiosonde observations. The average BIAS of all samples was −0.295 K, and the RMSE was 2.344 K. The corrected data were more symmetrically and centrally distributed on both sides of the diagonal, indicating that the corrected GIIRS temperature was more consistent with the temperature observed by radiosonde. The BIAS after correction decreased to −0.057 K, the RMSE decreased to 1.546 K, and the relative error before correction decreased by 34%.

Although we mainly considered the errors and correction methods for GIIRS temperature in summer over the Qinghai–Tibet Plateau, the same correction method can be easily extended to other seasons. In order to understand the applicability of the model in other seasons, we used the same method and established correction models for each station and level for other seasons. The vertical distribution of the BIAS, RMSE, and CORR before and after correction is shown in Figure 8a–c, representing the error distribution in the spring, autumn, and winter, respectively. The vertical distribution of the BIAS, RMSE, and CORR in each season was relatively consistent, with only slight differences. The GIIRS temperature had a significant positive bias at 400 hPa and 100 hPa, and a significant negative bias at 500 hPa, 450 hPa, and 350 hPa. The RMSE was relatively large at 450 hPa, 350 hPa, 250 hPa, 100 hPa, and 50 hPa, and relatively small at other levels. The CORR was consistent for all layers below 250 hPa, exceeding 0.8, and was relatively small above 250 hPa. After correction, all other seasons and levels showed positive correction effects. After correction, the BIAS of each layer was almost 0 K, with an average decrease of 0.14 K, 0.22 K, and 0.20 K in the spring, autumn, and winter, respectively. The average RMSE of each layer after correction in the spring, autumn, and winter decreased from 2.52 K, 2.18 K, and 2.52 K to 1.82 K, 1.52 K, and 1.79K, respectively, and the reduction was 0.70 K, 0.66 K, and 0.73 K, respectively. There were differences in the reduction in the RMSE at different levels and seasons. The CORR increased slightly above 200 hPa in all seasons, and there was no significant change at other levels. The correction models established for spring, autumn, and winter all produced a correction effect on the GIIRS temperature.

3.4. Comparative Analysis of Errors between the GIIRS Temperature and the ERA5 Temperature before and after Product Correction

To further verify the correction effect of the model, third-party ERA5 data were used to evaluate the model’s performance. Using the trained models for each individual station point in summer, the GIIRS temperature was corrected hourly, with the input variables still being GIIRS temperature data (including t0) and radiosonde temperature data (excluding t0) from ten hours prior to t0. At this time, t0 could be hourly and was not limited to 08:00 and 20:00 BTC to obtain hourly correction products. The correction effect was further verified using hourly ERA5 temperature products. Figure 9a shows the vertical distribution of the BIAS, RMSE, and CORR between the GIIRS temperature and ERA5 temperature products before and after correction at 15 stations on the Qinghai–Tibet Plateau in summer. It can be seen from the figure that the BIAS, RMSE, and CORR between the two before correction were approximately between −1.60 and 1.88 K, 1.90 and 2.92 K, and 0.5 and 0.8, respectively. The vertical distribution characteristics of the BIAS, RMSE, and CORR were consistent with those between the GIIRS temperature and radiosonde temperature in Figure 6. Since the radiosonde data of good quality were assimilated into the model to adjust it towards the radiosonde, the temperature products of ERA5 were consistent with the radiosonde temperature. Similarly, the BIAS was highest at 100, 250, and 350 hPa; the RMSE was relatively high at 450, 350, 250, and 100 hPa; and the CORR above 250 hPa was relatively small. The corrected BIAS was −0.36 to 0.39 K, with some differences between each layer. The RMSE was 0.99 to 1.47 K, and the distribution of layers below 250 hPa was relatively consistent, while there were slight fluctuations above 250 hPa. The CORR was 0.76 to 0.92, and each layer was relatively consistent. Figure 9b,c show the vertical distribution of the RMSE reduction of the product after correction at different times. Among them, Figure 9b shows the results of 6 times when the correction effects were more similar at 16:00, 17:00, 18:00, 19:00, 22:00, and 23:00 UTC, while Figure 9c shows the results of the remaining 18 times. The overall correction effect of the six times represented in Figure 8b was lower than that of other times. At 22:00 and 23:00 UTC, the low-level correction effect at 400 hPa and below was relatively average, and the RMSE reduction was less than 0.4 K, while there was a certain correction effect above 400 hPa. At 16:00, 17:00, 18:00, and 19:00 UTC, the RMSE reduction at 300 hPa was almost zero, while at other levels, the RMSE reduction exceeded 0.5 K, indicating that there was no correction effect at 300 hPa, while the correction effect at other levels was better. The correction effects of the remaining 18 times represented in Figure 8c had good results in each layer, with RMSE reductions mostly exceeding 0.4 K. Among them, the correction effects of the four levels of 450 hPa, 350 hPa, 250 hPa, and 100 hPa were relatively good.

Figure 10 shows the probability density distribution between the hourly temperature profile products of FY4A-GIIRS and the ERA5 temperature products at 15 stations over the Qinghai–Tibet Plateau before and after correction at 10 levels in summer. The total number of samples was 619,802. The figure shows that the GIIRS temperatures in the range of 190 to 220 K were mostly slightly lower than the ERA5 temperature products, while the GIIRS temperatures in the range of 220 to 250 K were mostly slightly higher than the ERA5 temperature products. The two were equivalent in the range of 250 K and above, and the corrected data were more uniformly and consistently distributed near the diagonal. The BIAS and RMSE decreased from −0.264 K and 2.247 K before correction to 0.062 K and 1.248 K, respectively, with an accuracy improvement of 44%.

3.5. Correction Model Based on the IASI Level2 Temperature Product

In the previous section, based on radiosonde data, the LSTM correction model was used to correct the GIIRS temperature near the radiosonde stations and achieved good results. However, the distribution of radiosonde stations is sparse. In order to achieve better correction results, it was necessary to extend the application of the correction model to the entire plateau and achieve correction of summertime GIIRS temperatures over the entire plateau. Due to the orbit height of polar-orbiting satellites being lower than that of geostationary satellites, the observation accuracy of polar-orbiting satellites is often higher [34,35]. Research on the accuracy of IASI temperature profile products showed that IASI temperature products have good reliability and accuracy [36,37,38]. Therefore, we used IASI temperature as a reference to correct a larger range of GIIRS temperatures. The correction model based on IASI temperature was similar to the correction model based on radiosonde, simply analogizing each grid point to radiosonde stations. Here, for the convenience of calculation and comparison with the correction effect of the correction models based on each station and level, a correction model was established for all grid points with a resolution of 0.25° in the range of 75°E–105°E and 25°N–40°N, and the effectiveness of the model was evaluated. Figure 11 shows the spatial distribution of the RMSE between the GIIRS temperature and the IASI temperature before and after the correction and the reduction in the RMSE after the correction. Before the correction, the RMSE between the GIIRS temperature and the IASI temperature was larger in the central and southern parts of the plateau. The region from the west of Qinghai Province to the middle of Xizang Province and the southwest of Xizang were two regions with a large RMSE exceeding 3.6 K. The RMSEs in the east of Qinghai Province and the west of Sichuan Province were relatively small, mostly below 2.4 K. After correction, the RMSEs in most areas of the plateau showed a degree of decline. The positions of the two regions with large RMSEs remained unchanged, but the RMSE values were significantly decreased. After correction, the RMSE in most areas was between 2 and 2.2 K, with only a small portion exceeding 3.6 K. From Figure 11c, it can be seen that the RMSE of most regions decreased. In general, the model had a good correction effect for the entire region. However, in some regions, such as the northwest and northeast corners in the figure, the correction effect actually deteriorated. This was mainly because the correction model was established based on the error characteristics of all grid points, representing the overall error characteristics of the entire region, which could have reduced the overall error. There may be insufficient representativeness for some grid points. In the future, smaller grid areas could be divided for correction, for example, an error correction model could be created for each grid point within the range of 5° × 5°, which may have better representativeness and achieve better correction results.

Figure 12 shows the vertical distribution of the BIAS, RMSE, and CORR between the GIIRS temperature and IASI temperature before and after correction within the range of 75°E–105°E and 25°N–40°N. Before correction, the BIAS between the GIIRS temperature and the IASI temperature ranged from −2.0 K to 1.5 K. The BIAS at 100, 450, and 500 hPa exhibited significant positive deviations, and the BIAS at 200, 250, 300, and 400 hPa exhibited significant negative deviations. The corrected BIAS at each level remained stable at around 0 K. The RMSE before correction was between 2.0 K and 2.8 K, while the RMSE at 200 hPa and 100 hPa was relatively large. The RMSE after correction decreased to 1.3–2.4 K, with no correction effect at 350 hPa. Other levels of the RMSE showed varying degrees of reduction, with an average RMSE reduction of 0.5 K. The change in CORR after correction was not significant, with only a slight improvement at 150 hPa and above.

4. Discussion

The above analysis indicated that the GIIRS temperature error correction model established based on radiosonde observation data in this article can effectively achieve correction of the product at a single location, which is of great significance for improving product quality. The corrected product showed significant improvements in the BIAS, RMSE, and CORR compared to both radiosonde temperature observations and ERA5 hourly temperature products, with product accuracy increasing by 34% and 44%, respectively. In future satellite product business applications, this error correction model has good application prospects.

In this paper, the LSTM model was used to correct the summer GIIRS temperature at 15 radiosonde station points on the Tibetan Plateau based on radiosonde data, and good results were obtained. Furthermore, the same correction method could be used for different seasons, and it could also be used to correct GIIRS temperature based on other high-precision data. Taking the IASI Level2 temperature product as a reference, a correction model for GIIRS temperature representing all grid points at a larger scale (75°E–105°E, 25°N–40°N) was established, which also achieved good results overall. Therefore, we could also infer that if the correction method based on all radiosonde stations was applied to the whole plateau, although the correction effect in some areas may be poor, the overall effect will be positive. This indicates that the LSTM correction model has a good correction ability for GIIRS temperature in different seasons and over a larger range. In the future, multiple artificial intelligence methods could be explored to conduct more extensive research on GIIRS temperature error correction. It is also possible to divide the plateau into finer grid regions and, based on IASI products or other high-precision and high-resolution data, establish a correction model for each small grid area, which may be more representative and produce a better correction effect. In addition, this paper mainly focused on high-altitude corrections of GIIRS temperature profiles over the Qinghai–Tibet Plateau in summer. In future, research could be further extended to lower altitudes or other regions and other seasons to achieve GIIRS correction at a larger space–time scale.

5. Conclusions

This article analyzed the error of summertime GIIRS temperatures over the Qinghai–Tibet Plateau using radiosonde temperature data from June, July, and August 2020 to 2022. An LSTM model was used to correct the GIIRS temperature at different stations and levels. The main conclusions are as follows:

• In the summer on the Qinghai–Tibet Plateau, GIIRS temperature compared to radiosonde temperature observations above 150 hPa mostly had a positive bias, while those below 150 hPa mostly had a negative bias. Lhasa and Hezuo Stations had the highest BIAS, exceeding −0.6 K. The distribution of other stations was relatively uniform, with an average BIAS of −0.259 K. The vertical distribution of the RMSE was relatively uniform, between 2 and 2.9 K, with slightly higher errors at higher levels. On average, the RMSE at the southwest station was about 0.2 K higher than at the northeast station. The time-series changes in the BIAS and RMSE were relatively stable, and the error of the GIIRS temperature at 20:00 was slightly higher than at 08:00.
• The LSTM model could effectively correct GIIRS temperature products. On the test set, the correlation between the corrected product and the radiosonde observation temperature significantly improved, but the ability to correct extreme values was slightly lacking. The BIAS between the corrected GIIRS temperature and the radiosonde observation temperature decreased from −0.295 K to −0.057 K, the RMSE decreased from 2.344 K to 1.546 K, and the CORR increased from 0.69 to 0.82.
• By using the trained LSTM correction model to correct the hourly GIIRS temperature, higher-quality products could be obtained for 15 radiosonde station points. After correction, the BIAS between the GIIRS temperature and the ERA5 temperature product was reduced from −0.264 K to 0.062 K, and the RMSE was reduced from 2.247 K to 1.248 K, with an accuracy improvement of 44%.
• The LSTM correction model can be applied to different seasons and can also produce good correction effects. The same correction model can also be used to correct GIIRS temperature based on other high-precision data. The correction model based on IASI Level2 temperature products within the range of 75°E–105°E, 25°N–40°N achieved good results, reducing the RMSE of the entire region by 0.5 K. This indicates that the correction model has a certain degree of universal applicability.