Quality Assessment of Operational Sea Surface Temperature Product from FY-4B/AGRI with In Situ and OSTIA Data

Abstract

The Fengyun-4B (FY-4B) satellite is currently the primary operational geostationary meteorological satellite in China, replacing the previous FY-4A satellite. The advanced geostationary radiation imager (AGRI) aboard the FY-4B satellite provides an operational sea surface temperature (SST) product with a high observation frequency of 15 min. This paper conducts the first data quality assessment of operational SST products from the FY-4B/AGRI using quality-controlled measured SSTs from the in situ SST quality monitor dataset and foundation SSTs produced by the operational sea surface temperature and sea ice analysis (OSTIA) system from July 2023 to January 2024. The FY-4B/AGRI SST product provides a data quality level flag on a pixel-by-pixel basis. Accuracy evaluations are conducted on the FY-4B/AGRI SST product with different data quality levels. The results indicate that the FY-4B/AGRI operational SST generally has a negative mean bias compared to in situ SST and OSTIA SST, and that the accuracy of the FY-4B/AGRI SST, with an excellent quality level, can meet the needs of practical applications. The FY-4B/AGRI SST with an excellent quality level demonstrates a strong correlation with in situ SST and OSTIA SST, with a correlation coefficient R exceeding 0.99. Compared with in situ SST, the bias, root mean square error (RMSE), and unbiased RMSE (ubRMSE) of the FY-4B/AGRI SST with an excellent quality level are −0.19, 0.66, and 0.63 °C in daytime, and −0.15, 0.70, and 0.68 °C at night, respectively. Compared with OSTIA SST, the bias, RMSE, and ubRMSE of the FY-4B/AGRI SST with an excellent data quality level are −0.10, 0.64, and 0.63 °C in daytime, and −0.13, 0.68, and 0.67 °C at night. The FY-4B/AGRI SST tends to underestimate the sea water temperature in mid–low-latitude regions, while it tends to overestimate sea water temperature in high-latitude regions and near the edges of the full disk. The time-varying validation of FY-4B/AGRI SST accuracy shows weak fluctuations with a period of 3–4 months. Hourly accuracy verification shows that the difference between the FY-4B/AGRI SST and OSTIA SST reflects a diurnal effect. However, FY-4B/AGRI SST products need to be used with caution around midnight to avoid an abnormal accuracy. This paper also discusses the relationships between the FY-4B/AGRI SST and satellite zenith angle, water vapor content, wind speed, and in situ SST, which have an undeniable impact on the underestimation of the FY-4B/AGRI operational SST. The accuracy of the FY-4B/AGRI operational SST retrieval algorithm still needs to be further improved in the future.

1. Introduction

Sea surface temperature (SST) is a fundamental variable that can help us understand, monitor, and predict heat, momentum, and gas fluxes at various scales, which determine complex interactions between the atmosphere and ocean [1]. With the continuous warming of the ocean, many regions of the global ocean have experienced long-term extreme ocean warming events, known as marine heatwaves (MHWs), in recent decades [2]. MHWs are extremely warm SST periods that last for several days to months, and can extend for thousands of kilometers [3]. The warming of SST accelerates the rate of retreat in freeze onset and has an impact on the early freeze-up of sea water in early winter [4].

SST has been operationally retrieved from many satellites in recent decades, including polar orbit satellites, as well as geostationary satellites [5]. Compared with polar orbit satellites, the SST observed by geostationary satellites has a higher temporal resolution advantage [1], and can obtain large-scale diurnal variations (DVs) in SST [6,7]. DV is one of the dominant variations in SST due to solar radiation and the Earth’s rotation. Currently, geostationary satellites are the only practical method to obtain the SST with sufficient frequency across the vast ocean to resolve DV. SST with high temporal resolution is very important for studies of air–sea interaction. At present, commonly used geostationary satellites with the ability to retrieve the SST include the United States’ geostationary operational environmental satellite (GOES) series, the European Meteosat satellite series, Japan’s multi-functional transport satellite (MTSAT) and Himawari series meteorological satellites, and China’s Fengyun (FY) series meteorological satellites [1,5,8]. In addition, India, South Korea, and Russia have also successively launched their own geostationary satellites with the ability to retrieve the SST [1,5,9,10].

However, due to differences in the performance of satellite instruments and retrieval algorithms, the accuracy of SST products for various geostationary meteorological satellites also varies. Petrenko et al. [11] developed four SST algorithms using GOES-R data, among which the incremental regression algorithm had the highest accuracy, with a bias and standard deviation (SD) of −0.09 and 0.52 K. Azevedo et al. [12] evaluated the accuracy of GOES-16 SST in the tropical and southwestern Atlantic, with a mean bias and root mean square error (RMSE) of 0.1 and 0.5 °C, respectively. Woźniak and Krężel [13] retrieved the SST from the spinning enhanced visible and infrared imager (SEVIRI) onboard Meteosat second generation (MSG) in the Baltic Sea, and compared it with advanced very-high-resolution radiometer (AVHRR) SST; the SD was less than 1 °C. Le Borgne et al. [14] estimated the operational SST products derived from SEVIRI; the bias and SD were −0.31 and 0.59 K. Kawamura et al. [15] added satellite zenith angle and solar zenith angle correction terms to the nonlinear SST (NLSST) algorithm, and estimated the SST of MTSAT with a bias of approximately zero and an RMSE of 0.8 K. Park et al. [16] developed SST retrieval algorithms for Himawari-8, and revealed that the multi-band algorithm performed markedly well, with the smallest RMSE of 0.4 °C. Woo et al. [9] retrieved SST from the communication, ocean, and meteorological satellite; the bias and RMSE of SST during the day and night were −0.01 and 0.58 °C, and −0.08 and 071 °C, respectively. Wang et al. [17] developed SST algorithms for the FY-2 satellite, which was China’s first-generation geostationary meteorological satellite series, and the optimal result showed a mean bias of −0.50 °C and an RMSE of 1.47 °C relative to the in situ SST. Jiang et al. [18] also developed an SST retrieval algorithm based on the radiative transfer model experiment for FY-2C, FY-2D, and FY-2E; the total errors were −0.6 ± 1.3, −0.2 ± 1.5, and 0.4 ± 1.8 K, respectively. Cui et al. [19] developed an operational SST product for FY-4A and evaluated the data quality in different time periods. The results showed a bias of −0.45 to −0.42 °C and an SD of 0.81 to 0.88 °C for FY-4A/AGRI SST, with an excellent quality level. Luo and Minnett [20] compared the skin SST derived from the advanced baseline imager onboard the GOES-16 with the data from marine–atmospheric emitted radiance interferometers deployed on ships. The results showed a mean bias of 0.086 K and a robust SD of 0.220 K.

FY-4 is China’s second-generation geostationary meteorological satellite series [21]. FY-4B is the second satellite of this series after FY-4A, and also the first operational satellite of this series [22]. FY-4B was successfully launched on 3 June 2021 and was located over the equator at 123.5°E on 10 June 2021. It was repositioned at 133°E above the equator on 12 April 2022. From 1 February to 5 March 2024, FY-4B drifted from 133°E to 105°E, replacing the FY-4A satellite and inheriting the main operational position observation system. Starting at 0:00 Co-ordinated Universal Time (UTC) on 5 March, FY-4B began operational observation services at 105°E. The advanced geostationary radiation imager (AGRI) is one of the primary payloads aboard the FY-4B satellite. The coverage range, spectral range, spatial resolution, and temporal resolution of FY-4B/AGRI are consistent with those of FY-4A/AGRI, but the spectral bandwidth of some channels is optimized and adjusted. The number of channels in FY-4B/AGRI increases from 14 to 15, with the addition of a low-level water vapor channel in the 7.24–7.60 µm band. The calibration accuracy of the infrared band is improved from better than 1.0 K to better than 0.7 K. Channels with central wavelengths of 3.8, 10.8, and 12.0 µm can be used for SST retrieval. Starting from 17 July 2023, FY-4B/AGRI began providing operational SST products, with a spatial resolution of 4 km at nadir and a temporal resolution of 15 min for full-disk earth-view images. There are approximately 95 full-disk observations per day.

Quality assessment is crucial for the quantitative application of remote sensing retrieval products. The national satellite meteorological center (NSMC) has already carried out the accuracy evaluation of the preliminary algorithm when releasing the FY-4B/AGRI operational SST products. However, the data quality of satellite-derived SST is influenced by various factors, such as anomalous atmospheric conditions, instrument working state, in situ observation errors, and cloud detection failures, which may generate lots of biases and uncertainties in the SST. Therefore, before putting the FY-4B/AGRI SST product into practical operational applications, its accuracy still needs further assessment. Assessing the data quality of the FY-4B/AGRI operational SST product will promote its application in analyzing and quantifying SST variations. This study aims to evaluate the quality of the first batch of operational SST products from FY-4B/AGRI by using in situ observations and reanalysis data. This study is carried out based on the FY-4B/AGRI data, the in situ data from the in situ SST quality monitor (iQuam) [23] dataset, the foundation SST data produced by the operational sea surface temperature and sea ice analysis (OSTIA) system [24], and the fifth-generation global climate monitoring reanalysis data (ERA5) [25] collected from 17 July 2023 to 31 January 2024. This paper is structured as follows: The employed data and methodology are introduced in Section 2. The results and discussions are presented in Section 3 and Section 4, respectively. Finally, the conclusions are provided in Section 5. The flowchart of the step-by-step research is shown in Figure 1.

2. Materials and Methods

2.1. Materials

2.1.1. FY-4B/AGRI Data

The operational SST product of FY-4B/AGRI was developed with the NLSST algorithm in real time by the NSMC of the China meteorological administration (CMA). NLSST is an operational SST retrieval algorithm developed by Walton et al. [26] for the AVHRR aboard the national oceanic and atmospheric administration (NOAA) satellite. It improves the accuracy of satellite-derived SST by introducing the first-guess temperature. The coefficients of the NLSST algorithm are obtained through regression using the least squares method between satellite data and in situ data, which is similar to the FY-4A/AGRI SST retrieval algorithm [19]. The first-guess SST used in the FY-4B/AGRI NLSST algorithm is from the OSTIA SST, and the in situ data are from the iQuam dataset. Different SST retrieval coefficients are used for daytime and nighttime data, respectively, but the same set of coefficients is used for full-disk data. To calculate the regression coefficients, it is necessary to accumulate a matchup dataset for 1–3 months. When the instrument status changes or the product accuracy deteriorates, the regression coefficients are usually replaced [19]. The NLSST algorithm can be described as follows:

$$T_s=A_0+A_1\times T_{11}+A_2\times T_{fg}\times \left(T_{11}-T_{12}\right)+A_3\times \left(T_{11}-T_{12}\right)\left(\sec \theta -1\right)$$

(1)

where T_s is the satellite-derived SST (in °C), T₁₁ and T₁₂ are the brightness temperature (BT) of 10.7 and 12.0 µm bands (in K), A_i (where i = 0, 1, 2, and 3) is the coefficient regressed by satellite BT and in situ SST, T_fg is the first-guess SST acquired from the daily OSTIA (in °C), and θ is the sensor zenith in angular.

Table 1. Coefficients for the operational SST product from FY-4B/AGRI.

	A0	A1	A2	A3
Daytime	−245.276	0.908994	0.0722074	0.793245
Nighttime	−243.704	0.904217	0.0715105	0.807235

presents the coefficients A₀–A₃ for the daytime and nighttime algorithms of the FY-4B/AGRI operational SST product. The FY-4B/AGRI operational SST algorithm is described in detail in the instruction for the SST product of FY-4B satellite, which is available online (https://satellite.nsmc.org.cn/PortalSite/StaticContent/FileDownload.aspx?CategoryID=1&LinkID=1083, accessed on 1 April 2024).

The FY-4B/AGRI level 1 GEO data and level 2 SST products are used in this study, both of which are all full-disk normalized geostationary projection data with a spatial resolution of 4 km and a temporal resolution of 15 min. The GEO data are stored in a network common data form (NetCDF) file, which includes satellite zenith angle and azimuth angle, solar zenith angle and azimuth angle, and other datasets. This study uses both the satellite zenith angle and the solar zenith angle. The SST product is stored in a hierarchical data format version 5 (HDF5) file, which contains datasets including SST, data quality, and others. There are two SST datasets in the FY-4B/AGRI SST product. One dataset named “SST” only includes the best-quality SST data, while the other dataset named “SST_ALL” includes SST data with various quality levels. The data quality flags for FY-4B/AGRI SST contain four values: 0 (excellent), 1 (good), 2 (bad), and 3 (invalid), which correspond to SST data pixel-by-pixel. The quality level is determined by using quality control methods including angle control, spatial consistency testing, and climate threshold testing. To comprehensively evaluate the accuracy of the FY-4B/AGRI SST product, this study selected the dataset named “SST_ALL” for sample data extraction, and evaluated the SST with the quality flag of 0, 1, and 2, and all quality levels using in situ SST and OSTIA SST, respectively. Then, detailed quality assessments are conducted primarily on the SST product with an excellent quality level. The FY-4B/AGRI level 1 GEO data and level 2 operational SST product can be downloaded from the website of the Fengyun satellite data center (https://satellite.nsmc.org.cn/portalsite/default.aspx, accessed on 31 January 2024).

2.1.2. In Situ SST

In situ temperature measurements with high quality are required for the systematic validation of satellite-estimated SST. The iQuam dataset [23], developed at the NOAA center for satellite applications and research, is used to assess the quality of the FY-4B/AGRI SST product. The current version of the iQuam dataset is 2.10, which includes the data from eight platform types and has undergone quality control. One month’s in situ data are stored as a single NetCDF file and can be available online (https://www.star.nesdis.noaa.gov/sod/sst/iquam/data.html, accessed on 17 February 2024). Considering the low accuracy of ship measurement [23], the in situ SST with a platform type of ship is excluded during the sample extraction in this study, and only the in situ SST with the highest accuracy, quality level 5, is used as the true value for assessing the satellite-derived SST. In the iQuam dataset, the platform types include commercial ships, drifting buoys, tropical moored buoys, coastal moored buoys, Argo floats, high-resolution drifters, integrated marine observing system (IMOS) ships, and coral reef watch (CRW) buoys. The depths are available for Argo floats, IMOS ships, and CRW buoys, and unavailable for other platform types. Due to the limited temporal and spatial coverage of high-resolution drifters and CRW buoys, there are no data of these two platform types within the coverage range of FY-4B/AGRI. After removing the commercial ships and IMOS ships data from the iQuam dataset, the remaining data mainly include four types of observation data: Argo floats, drifting buoys, coastal moored buoys, and tropical moored buoys. Argo floats provide SST measurement data from different water depths. Argo quality flags are used to select the best quality near-surface data from above 10 m depth, which are further subjected to the standard iQuam quality control. In this study, only the Argo data closest to the surface are retained. According to Xu et al. [27], the measurement of Argo closest to the surface is usually at a depth of 4–5 m. The measurement depth of the drifting buoys is generally 10–20 cm, while the measurement depth of the coastal and tropical moored buoys is around 1 m [27,28].

2.1.3. OSTIA SST

The OSTIA system was developed by the Met Office to produce an estimate of foundation SST and sea ice concentration in near-real time [24]. The foundation SST is the SST free of diurnal variability. OSTIA uses satellite data from both thermal infrared and microwave radiometers, as well as in situ observation provided by EUMETCast, or the world meteorological organization’s global telecommunications system communication, to determine SST [29]. OSTIA produces daily analysis SST for the global ocean and some lakes at a grid resolution of 1/20° (~6 km). The OSTIA foundation SST data are made available in a NetCDF format following the group for high-resolution SST data specification. The SST data are available from the Copernicus marine service information (https://doi.org/10.48670/moi-00165, accessed on 1 February 2024). According to the validation results of Donlon et al. [24], the mean bias of OSTIA SST is close to zero, and the RMSE is about 0.57 K. However, Woo et al. [30] found that the bias and RMSE of OSTIA SST relative to buoy observations were 0.32 and 1.29 K in the coastal region of the Korean Peninsula. In this paper, the OSTIA SST is also used as a reference value to assess the quality of the FY-4B/AGRI operational SST products.

2.1.4. ERA5 Data

ERA5 is the fifth-generation global climate monitoring reanalysis data from the European center for medium-range weather forecasts, generated by the Copernicus climate change service (C3S) [25]. ERA5 replaces the ERA-Interim reanalysis. This reanalysis data provides a detailed record of the global atmosphere, land surface, and ocean waves from 1940 onwards. This study uses the ERA5 hourly reanalysis of U (eastward) and V (northward) component data of 10 m wind speed and total column water vapor (TCWV) data on single levels from 17 July 2023 to 31 January 2024, with a horizontal resolution of 0.25° × 0.25°. ERA5 hourly data can be downloaded from the C3S climate data store (https://doi.org/10.24381/cds.adbb2d47, accessed on 13 February 2024).

2.2. Methods

2.2.1. Matchup Samples Extraction

The matchup samples are extracted by combing the FY-4B/AGRI SST with the corresponding data quality flag, satellite zenith angle, solar zenith angle, in situ SST, OSTIA SST, wind speed, and TCWV on the grid resolution of 0.05°. Here, the wind speed is calculated using the U and V components of an hourly 10 m wind. Firstly, the full disk of 15 min satellite data is projected into equal longitude and latitude projection with 0.05°. The nearest neighbor interpolation algorithm is used in projection processing. Then, the in situ data with a quality level equal to 5 and within 15 min consistent with the satellite observation time are chosen from the iQuam dataset to be resampled to the 0.05° grid resolution with the nearest neighbor interpolation method. When there are multiple observation data in the same grid, the average value is calculated. The hourly wind speed and TCWV are resampled to the grid resolution of 0.05° with a linear interpolation method. Moreover, the wind speed and TCWV data are interpolated linearly in space and time to match the satellite observation pixels within a time window of 15 min. Finally, we generate the matchup dataset from the satellite data and the referenced data at the same time and location.

In some studies [31,32,33], outliers are removed from the matchup samples to ensure the accuracy of satellite-derived SST products. This study evaluates the accuracy of the FY-4B/AGRI operational SST product based on the data quality level; no additional outlier removal is performed. The total number of samples is 330,732, with 140,835 samples during the day and 189,897 samples at night.

2.2.2. Error Evaluation

The conventional error evaluation indicators, including mean bias (Bias), RMSE, unbiased RMSE (ubRMSE), and correlation coefficient (R), are used to inspect the accuracy of the FY-4B/AGRI operational SST products. The ubRMSE is different from RMSE in that it excludes bias and only considers the amplitude difference of variable variations. Moreover, the Nash–Sutcliffe efficiency (NSE) and Kling–Gupta efficiency (KGE) [34,35] are employed to evaluate the fitting degree of satellite-derived SST and observed SST. The formula of each indicator is as follows:

$$Bias=\frac{1}{N}{\sum }_{i=1}^N\left(X_i-Y_i\right)$$

(2)

$$\mathit{RMSE}=\sqrt{\frac{1}{N}{\sum }_{i=1}^N{\left(X_i-Y_i\right)}^2}$$

(3)

$$\mathit{ubRMSE}=\sqrt{\frac{1}{N}{\sum }_{i=1}^N{\left[\left(X_i-\overline{X}\right)-\left(Y_i-\overline{Y}\right)\right]}^2}$$

(4)

$$R=\frac{{\sum }_{i=1}^N\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{{\sum }_{i=1}^N{\left(X_i-\overline{X}\right)}^2}\sqrt{{\sum }_{i=1}^N{\left(Y_i-\overline{Y}\right)}^2}}$$

(5)

$$NSE=1-\frac{{\sum }_{i=1}^N{\left(X_i-Y_i\right)}^2}{{\sum }_{i=1}^N{\left(Y_i-\overline{Y}\right)}^2}$$

(6)

$$KGE=1-\sqrt{{\left(R-1\right)}^2+{\left(\frac{\overline{X}}{\overline{Y}}-1\right)}^2+{\left(\frac{{\sigma }_X/\overline{X}}{{\sigma }_Y/\overline{Y}}-1\right)}^2}$$

(7)

where i is the sample number, N is the total number of samples, X is the satellite-derived SST, $\overline{X}$ and ${\sigma }_X$ are the average and SD of satellite-derived SST, Y is the observed SST, and $\overline{Y}$ and ${\sigma }_Y$ are the average and SD of observed SST. In this study, the observed SST comes from the in situ SST and the OSTIA SST, respectively.

Moreover, the three-way error analysis method [36] is also adopted to simultaneously assess the errors between FY-4B/AGRI SST, in situ SST, and OSTIA SST. The error variance ${\sigma }_i$ for SST type i (where i = 1, 2, and 3) can be descripted as follows:

$$\left\{\begin{array}{c}{\sigma }_1=\sqrt{\left(V_{12}+V_{31}+V_{23}\right)/2}\\ {\sigma }_2=\sqrt{\left(V_{23}+V_{12}+V_{31}\right)/2}\\ {\sigma }_3=\sqrt{\left(V_{31}+V_{23}+V_{12}\right)/2}\end{array}\right.$$

(8)

$$V_{ij}=\frac{1}{N}{\sum }_{i=1}^N{\left({SST}_{ij}-\overline{{SST}_{ij}}\right)}^2$$

(9)

$${SST}_{ij}={SST}_i-{SST}_j$$

(10)

where SST_ij is the difference between SST types i and j, $\overline{{SST}_{ij}}$ is the average of SST_ij, V_ij is the variance of SST_ij, and subscripts 1, 2, and 3 refer to the FY-4B/AGRI SST, the in situ SST, and the OSTIA SST, respectively.

3. Result

3.1. Comparison with In Situ SST

The errors of the FY-4B/AGRI SST relative to in situ SST are calculated based on different data quality levels during the daytime and nighttime periods, respectively. Daytime and nighttime SST data are distinguished by the solar zenith angle. If the solar zenith angle is less than 85°, the SST data are identified as daytime data. If the solar zenith angle is greater than 85°, the SST data are identified as nighttime data. The accuracy of the FY-4B/AGRI SST is expressed as the bias, RMSE, and ubRMSE with respect to the in situ SST. The fitting effect between the FY-4B/AGRI SST and in situ SST is expressed by R, NSE, and KGE.

Table 2. Error information of FY-4B/AGRI SST relative to in situ SST. Where DQ presents data quality levels of 0, 1, 2, and all quality levels, N is the number of samples.

	DQ	N	Bias (°C)	RMSE (°C)	ubRMSE (°C)	R	NSE	KGE
Daytime	0	82,052	−0.19	0.66	0.63	0.99	0.99	0.97
	1	37,365	−0.47	0.88	0.74	0.99	0.98	0.98
	2	21,418	−2.11	2.65	1.60	0.94	0.66	0.87
	All	140,835	−0.55	1.23	1.10	0.98	0.95	0.97
Nighttime	0	106,550	−0.15	0.70	0.68	0.99	0.98	0.98
	1	28,617	−0.46	0.86	0.73	0.99	0.98	0.98
	2	54,730	−3.01	3.48	1.74	0.95	0.61	0.80
	All	189,897	−1.02	1.97	1.68	0.96	0.89	0.92

presents the overall error information of the FY-4B/AGRI SST relative to in situ SST from July 2023 to January 2024.

The accuracy of the FY-4B/AGRI SST varies greatly among the different data quality levels. The FY-4B/AGRI SST with an excellent quality level has the highest accuracy, with bias, RMSE, and ubRMSE of −0.19, 0.66, and 0.63 °C during the day, and −0.15, 0.70, and 0.68 °C at night. These results meet the accuracy requirement of 0.5–0.8 °C for the thermal infrared radiometer on board geostationary satellites [37]. And these accuracy statistics are better than the evaluation results of the FY-4A/AGRI SST product [19,33,38,39]. The improved accuracy of the FY-4B/AGRI SST may be related to the improved calibration accuracy of the FY-4B/AGRI thermal infrared instruments. In addition, the SST retrieval algorithm for FY-4B/AGRI uses the real-time daily SST from OSTIA as the first-guess SST data, while the SST retrieval algorithm for FY-4A/AGRI uses the 30-year daily average climate SST data from 1982 to 2011 as the first-guess SST data [19]. The accuracy of the FY-4B/AGRI SST with a good quality level has decreased, with the bias exceeding −0.46 °C, which is more than twice the bias of the SST with an excellent quality level. During the day, the bias, RMSE, and ubRMSE are −0.47, 0.88, and 0.74 °C, respectively, while, at night, they are −0.46, 0.86, and 0.73 °C. This result is close to that of the FY-4A/AGRI SST product [19,39]. The error statistics of both daytime and nighttime data show that the FY-4B/AGRI SST with a bad quality level has very poor accuracy, with a minimum bias exceeding −2.11 °C and a minimum RMSE exceeding 2.65 °C. This has also led to a decrease in the overall accuracy of the FY-4B/AGRI SST. A comparative analysis shows that there has always been a negative bias between the FY-4B/AGRI SST of various quality levels and in situ SST, both during the day and at night. As the data quality decreases, the negative bias of the FY-4B/AGRI SST becomes larger.

The R, NSE, and KGE quantitatively reveal the fitting degree between the satellite-derived SST and in situ SST through numerical magnitude. The FY-4B/AGRI SST with excellent and good quality levels exhibits strong correlations with in situ SST, with R greater than 0.99, NSE greater than 0.98, and KGE greater than 0.97. This result indicates that there is a good fit between the FY-4B/AGRI SST with excellent and good quality levels and in situ SST. The R, NSE, and KGE between the FY-4B/AGRI SST with a bad quality level and in situ SST are significantly reduced, indicating poor fitting results between the FY-4B/AGRI SST and in situ SST at this data quality level. Comparing the changes in R, NSE, and KGE at different data quality levels, it can be found that NSE has the strongest response to poor-fitting relationships.

Figure 2 shows the density scatter plots between the in situ SST and the FY-4B/AGRI SST with different data quality levels during daytime and nighttime, respectively. When the data quality levels are 0 (excellent) and 1 (good), the correlations between the FY-4B/AGRI SST and in situ SST are very high, and their matchup samples are closely distributed along the diagonal. However, when the data quality level is 2 (bad), the correlation between the FY-4B/AGRI SST and in situ SST deteriorates, and the matchup samples deviate from the diagonal and significantly lean towards the in situ SST side, with a more severe deviation in nighttime data. The relationships between the satellite-derived SST and the in situ SST are consistent with the performances of R, NSE, and KGE in Table 2. Moreover, it can also be observed in the density scatter plots that the matchup samples are more concentrated at high temperatures.

Figure 3 displays the histogram distribution of the FY-4B/AGRI SST minus in situ SST for daytime and nighttime. The histograms display a bin size of 0.2 °C. The red dashed lines represent the Gaussian distributions defined using the mean value and SD of the FY-4B/AGRI SST minus in situ SST. The error distribution of the FY-4B/AGRI SST with an excellent quality level is most consistent with the shape of the Gaussian distribution. The histograms show that the FY-4B/AGRI SST is always colder than in situ SST. As the data quality deteriorates, the underestimation of the FY-4B/AGRI SST becomes more severe, and the asymmetric distribution of matchup samples differences between the FY-4B/AGRI SST and in situ SST also becomes stronger. This result is consistent with the results displayed in Table 2 and Figure 2.

3.2. Comparison with OSTIA SST

Table 3. Error information of FY-4B/AGRI SST relative to OSTIA SST.

	DQ	N	Bias (°C)	RMSE (°C)	ubRMSE (°C)	R	NSE	KGE
Daytime	0	82,052	−0.10	0.64	0.63	0.99	0.99	0.97
	1	37,365	−0.38	0.82	0.73	0.99	0.98	0.98
	2	21,418	−2.06	2.63	1.63	0.93	0.66	0.87
	All	140,835	−0.48	1.21	1.11	0.98	0.95	0.97
Nighttime	0	106,550	−0.13	0.68	0.67	0.99	0.99	0.97
	1	28,617	−0.42	0.82	0.71	0.99	0.99	0.98
	2	54,730	−3.01	3.48	1.74	0.95	0.61	0.80
	All	189,897	−1.00	1.96	1.69	0.96	0.89	0.93

presents the accuracy of the FY-4B/AGRI SST compared to OSTIA SST for daytime and nighttime. Since the evaluated data come from the previous matchup dataset, the number of samples used is the same as for the in situ SST. The general error characteristics of the FY-4B/AGRI SST with different quality levels are consistent with the results in Table 2. The FY-4B/AGRI SST with an excellent quality level has the highest accuracy, with bias, RMSE, and ubRMSE of −0.10, 0.64, and 0.63 °C during the day, and −0.13, 0.68, and 0.67 °C at night. Compared with the results in Table 2, the errors during the day and night are somewhat reduced. Meanwhile, R, NSE, and KGE also indicate a good fit between the FY-4B/AGRI SST and OSTIA SST at this quality level. The accuracy of the FY-4B/AGRI SST with a good quality level is inferior to that of the FY-4B/AGRI SST with an excellent quality level. The bias of the FY-4B/AGRI SST with a good quality level exceeds −0.38 °C, which is more than three times of the bias of the SST with an excellent quality level. In addition, the RMSE of the FY-4B/AGRI SST with a good quality level is also greater than 0.82 °C, which is at least 0.14 °C higher than that of the SST with an excellent quality level. But, from the performance of R, NSE, and KGE, it still fits well with the OSTIA SST. The accuracy of the FY-4B/AGRI SST with a bad quality level significantly deteriorates, with a minimum bias exceeding −2.06 °C and a minimum RMSE greater than 2.63 °C. Meanwhile, from the changes in R, NSE, and KGE, the fitting effect between the OSTIA SST and FY-4B/AGRI SST with a bad quality level significantly decreases. A comparative analysis shows that the negative bias between the FY-4B/AGRI SST and OSTIA SST still exists, and as the data quality level deteriorates, the negative bias also increases.

Figure 4 shows the density scatter plots between the FY-4B/AGRI SST and the OSTIA SST, and Figure 5 shows the histograms of the FY-4B/AGRI SST minus OSTIA SST. The density scatter plots and histograms intuitively reflect the correlation characteristics of the satellite-derived SST and the OSITA SST. The error distribution characteristics of the FY-4B/AGRI operational SST relative to the OSTIA SST are basically consistent with those of the in situ SST. The fitting degree between the OSTIA SST and the FY-4B/AGRI SST with excellent and good quality levels is good, but there is an obvious deviation from the fitting of the SST with a bad quality level. Just because the bias of the FY-4B/AGRI SST minus OSTIA SST decreases, the histogram distribution of the satellite-derived SST with an excellent quality level is more in line with the normal distribution, which is most significant on daytime data.

3.3. Comprehensive Evaluation

Figure 6 shows the target diagrams for the errors of the FY-4B/AGRI SST relative to in situ SST and OSTIA SST, respectively. The target diagrams can provide summary information on how the pattern statistics and biases, respectively, contribute to the RMSE [40]. Here, the bias and ubRMSE are used together to describe the distribution patterns of errors along the vertical and horizontal axes. The target diagram fully demonstrates the underestimation characteristics of the FY-4B/AGRI SST products. The SST products with an excellent quality level have the closest distance to the center point and the highest accuracy. The error of SST products with a good quality level is within the range of 1.0 °C, but the bias is significantly greater than that of SST products with an excellent quality level. However, the error of SST products with a bad quality level is far away from the center point, resulting in a decrease in the whole accuracy of SST products.

Table 4. Error variances of FY-4B/AGRI SST, in situ SST, and OSTIA with three-way error analysis.

DQ	Daytime				Nighttime
	N	AGRI (°C)	In Situ SST (°C)	OSTIA (°C)	N	AGRI (°C)	In Situ SST (°C)	OSTIA (°C)
0	82,052	0.59	0.22	0.23	106,550	0.64	0.22	0.19
1	37,365	0.69	0.27	0.22	28,617	0.67	0.29	0.23
2	21,418	1.60	0.06	0.30	54,730	1.73	0.21	0.15
All	140,835	1.09	0.20	0.25	189,897	1.67	0.20	0.21

presents the estimated error variances for the FY-4B/AGRI SST, in situ SST, and OSTIA SST through the three-way error analysis method [36], which allows the simultaneous inspecting of the accuracy of three different types of observation data [33]. For the FY-4B/AGRI SST, the accuracy of the SST with an excellent quality level is still the best, with an error variance of 0.59 °C. The accuracy of the SST with a good quality level declines slightly, with an error variance of 0.69 °C. The accuracy of the SST with a bad quality level is the worst, with an error variance of 1.60 °C. The accuracy of the in situ SST and OSTIA SST is better than that of the FY-4B/AGRI SST. Since the ship measurements have been eliminated from the iQuam dataset, the accuracy of the in situ SST is also superior to previous inspection results [33]. As a product of reanalysis and processing, the OSTIA SST consistently exhibits a high accuracy and stable data quality.

4. Discussion

The verification results of the previous sections indicate that the data quality of the FY-4B/AGRI operational SST products with an excellent quality level reaches the accuracy level of other similar satellite-derived SST products [9,41,42], which can meet the needs of operational applications. However, we still need to further discuss the error characteristics of this SST product on the spatiotemporal scale, as well as the dependence relationship with factors such as the satellite zenith angle, water vapor content, wind speed, and in situ SST. The subsequent quality assessments are based on the FY-4B/AGRI operational SST products with an excellent quality level.

4.1. Spatiotemporal Distribution of FY-4B/AGRI SST Error

Figure 7 shows the spatial distributions of bias, RMSE, R, and sample number N for the full disk of the FY-4B/AGRI SST minus in situ SST in a 2° × 2° grid for daytime and nighttime. In the tropical regions around the equator, the bias, RMSE, and R are not calculated due to the lack of matchup data. A negative bias can be observed in the mid–low-latitude regions, mainly concentrated in the northern hemisphere. However, a positive bias can be observed in the areas near the edge of the full disk, especially in the waters around and south of Australia, where overestimation is widespread [16]. The spatial distribution of the bias may be related to the use of the same set of regression coefficients for the full disk in the FY-4B/AGRI operational SST retrieval algorithm. The observation path in the edge region is longer, and the thermal infrared data are more severely affected by water vapor. However, the FY-4B/AGRI NLSST algorithm, using global SST regression coefficients, cannot completely eliminate the bias caused by regional atmospheric attenuation [19]. The spatial distributions of RMSE and R also have a certain correlation with the spatial distribution of the in situ data. The larger RMSE and smaller R are mainly distributed near the equator and the edge of the disk, where the in situ data are scarce.

The temporal variation of errors is also a method for measuring the quality of satellite SST products. This study aims to determine if there is any seasonal variation in SST accuracy by calculating the monthly mean bias and RMSE. Figure 8 shows the monthly variation of the FY-4B/AGRI SST minus in situ SST through a box-whisker plot. The error trend of daytime and nighttime data remains consistent. From the box-whisker plot, the dispersion of nighttime samples is stronger than that of daytime samples. This may be related to the lower accuracy of nighttime cloud detection compared to daytime. Daytime cloud detection is recognized through a combination of visible and infrared bands [43], while nighttime only has infrared bands [44]. In addition, it can be observed that the median and mean values of the FY-4B/AGRI SST minus in situ SST during the day are greater than those at night. But the RMSE of daytime is less than that of nighttime. For daytime data, the RMSE fluctuates from 0.63 to 0.69 °C, and the bias (mean value of the FY-4B/AGRI SST minus in situ SST) fluctuates from −0.30 to −0.14 °C, while, for nighttime data, the RMSE fluctuates from 0.67 to 0.72 °C, and the bias fluctuates from −0.29 to −0.08 °C. The changing amplitudes in RMSE during the day and at night are 0.06 and 0.05 °C, respectively. Although the changes are not significant, it is possible to observe a weak fluctuation pattern of 3–4 months in the monthly variations. In the future, longer time-series data will be needed to continue monitoring the pattern of monthly variations.

4.2. Hourly Variation of FY-4B/AGRI SST Error

FY-4B/AGRI generates a full-disk SST product every 15 min. This high-frequency observation feature is helpful in studying the DVs in the SST. The hourly error variations can obviously reflect the accuracy of the satellite-derived SST [45]. Therefore, it is necessary to analyze the hourly error of the FY-4B/AGRI SST product.

Figure 9 shows the hourly variations of the difference between the FY-4B/AGRI SST, in situ SST, and OSTIA SST at UTC time. The FY-4B/AGRI SST and in situ SST suffer from daytime warming and nighttime cooling, whereas the OSTIA SST represents the foundation SST, which is free of DVs. From Figure 9, it can be observed that the FY-4B/AGRI SST and in situ SST have DV effects relative to OSTIA SST. In Figure 9b, the FY-4B/AGRI SST shows a peak around 4:00 UTC, which is just noon local time for the FY-4B satellite. For the FY-4B satellite, all oceans enter daytime around 4:00 UTC, some oceans enter nighttime around 7:00 UTC, and all oceans enter nighttime around 16:00 UTC. However, the peak of the in situ SST appears around 5:00 UTC in Figure 9c, with a lag of approximately one hour compared to the FY-4B/AGRI SST. This time difference is attributed to the inconsistency of the water depth measured by satellite and in situ instruments. According to Donlon et al. [37], the temperature measured by an infrared radiometer is the skin temperature at a depth of about 10 to 20 µm. However, the temperature measured by different in situ instruments is the depth temperature at depths ranging from 10⁻² to 10³ m. In addition, Figure 9a also shows that there is a very weak peak in the difference between the FY-4B/AGRI SST and in situ SST around 4:00 UTC. From 0:00 UTC until the peak appears, the heating amplitudes in Figure 9a–c are approximately 0.06, 0.21, and 0.18 °C, respectively. This peak seems to be related to the heating of seawater by sunrise [46], and the satellite-estimated SST increases to some extent, narrowing the gap between the observed SST and the foundation SST. Beginning a few hours after noon, the mixing of wind begins to mix downwards with the heat stored near the sea surface during the noon period, leading to the rapid cooling of the sea surface water [47]. The thermal infrared instrument on board a satellite explores the skin SST, but the FY-4B/AGRI SST is the temperature calculated by regression fitting with in situ measurements. Therefore, the hourly bias of the FY-4B/AGRI SST minus in situ SST does not significantly reflect the DVs.

There is another significant peak in the comparison of the FY-4B/AGRI SST with both in situ SST and OSTIA SST after approximately 16:00 UTC in Figure 9a,b. The maximum value of the hourly bias is approximately 0.02 and −0.01 °C. Because of the midnight effect [45], there will be a strong negative bias at midnight for all satellites. However, the FY-4B/AGRI SST exhibits the opposite feature. This unexpected feature will be the topic of our further research. Additionally, during the actual use of FY-4B/AGRI operational SST products, we found that cloud detection errors may occur around midnight, which could lead to missing SST products and changes in accuracy.

4.3. Variation of FY-4B/AGRI SST Error with Satellite Zenith Angle

The satellite zenith angle is a key factor affecting the accuracy of the remote sensing retrieval SST [48,49]. The variation in the satellite zenith angle causes the change in the atmospheric path radiation. With the increase in the satellite zenith angle, the atmospheric path radiation path is extended, which results in an increase in surface radiation attenuation [50]. As shown in the fourth term on the right side of Equation (1), the influence of the satellite zenith angle is simply considered in the NLSST algorithm. However, the operational SST algorithm for FY-4B/AGRI does not further consider the impact of changes in the satellite zenith angle. FY-4B/AGRI performs SST retrieval only on pixels with a satellite zenith angle of less than 70°. Figure 10 shows the relationship between the satellite zenith angle and the difference of the FY-4B/AGRI SST minus in situ SST. The colorful density scatter plot displays the number of samples in each 1° bin. The black error bar represents the bias and SD in each 5° bin. The in situ SST data are scarce in the range of satellite zenith angles below 10°. The bias of FY-4B/AGRI SST notably increases with the decrease in satellite zenith angle. It can be observed that the difference of the FY-4B/AGRI SST minus in situ SST decreases first and then increases with the variation in the satellite zenith angle. This reverse change occurs around a satellite zenith angle of 45° [51]. As Figure 7 shows, the distribution of the matchup data is uneven, and there are few samples around the subpoint of the satellite and the equator. At the same time, the FY-4B/AGRI SST tends to be underestimated with a negative bias in mid–low-latitude regions, while it is overestimated with a positive bias in high-latitude regions and near the edge of the full disk. When the satellite zenith angle is small, its coverage is concentrated in the areas of SST underestimation. As the zenith angle increases, its coverage gradually expands into the SST overestimation areas. The mutual cancellation of both positive and negative biases leads to a decrease in the mean bias. Within the range of satellite zenith angle, the mean bias during the day ranges from −0.42 to 0.02 °C, and the mean bias at night ranges from −0.38 to 0.05 °C. The deviation trend of the FY-4B/AGRI SST with respect to the satellite zenith angle is consistent with that of the FY-4A/AGRI SST [39].

4.4. Influence of Water Vapor and Wind on FY-4B/AGRI SST Error

The thermal radiation received by satellite sensors is influenced by the atmosphere between Earth and the satellite, especially the absorption of water vapor in the atmosphere [52]. The correction of water vapor absorption in SST retrieval algorithms is usually achieved by using the difference between two split window channels [32]. The NLSST algorithm used by FY-4B/AGRI operational SST products also corrects the water vapor absorption through the difference between two thermal infrared channels. Figure 11 shows the relationship between the TCWV and the difference of the FY-4B/AGRI SST minus in situ SST. The colorful density scatter plots show the sample numbers in every 1 kg/m² bin. The black error bar represents the bias and SD for each 5 kg/m² bin. There is an obvious dependence relationship between the difference of the FY-4B/AGRI SST minus in situ SST and TCWV. As water vapor increases, the error between the FY-4B/AGRI SST and in situ SST gradually increases. When the water vapor content exceeds 45 kg/m², the negative bias rapidly increases in its magnitude with the increase in water vapor content. When the water vapor content is 60 kg/m², the biases during the day and at night are −0.70 and −0.78 °C, respectively. After the water vapor content exceeds 65 kg/m², the reliability of the error result is not high due to the limited number of samples. However, the trend of increasing error remains unchanged.

The amplitude of the cool skin effect strongly depends on the wind speed [53]. Figure 12 shows the relationship between the wind speed and the difference of the FY-4B/AGRI SST minus in situ SST. The matchup data are predominantly distributed within the wind speed range of 20 m/s, while the matchup data with a sample number exceeding 100 are concentrated within the wind speed range of 15 m/s. Although the FY-4B/AGRI SST product shows a general negative bias, its relationship with the wind speed is consistent with previous research findings [41,54]. During the day, under clear skies and calm conditions, thermal stratification may occur several meters above the top of the ocean [41,47]. It shows an increasing negative bias in the wind speed range from 0 to 10 m/s. The maximum of negative biases exceeds −0.2 °C. This is mainly due to the cool skin effect [46,47]. At high wind speeds, the ocean is completely mixed, and the SST obtained by the satellite should be close to the in situ measurement values. During the night, there is no sunlight, so the diurnal warming disappears [41]. It displays a small negative bias of about −0.1 °C within the wind speed range of 0 to 5 m/s. However, larger negative biases exceeding −0.3 °C appear when the wind speed is higher than 15 m/s.

4.5. Relationship between FY-4B/AGRI SST Error and In Situ SST

Figure 13 shows the dependency of the difference between the FY-4B/AGRO SST and in situ SST on SST distribution. As the SST increases, the error in the FY-4B/AGRI SST shifts from a positive bias to a negative bias. When the in situ SST is less than 10 °C, there is a significant overestimation with a positive bias for the FY-4B/AGRI SST, while, when the in situ SST is greater than 25 °C, there is a significant underestimation with a negative bias for theFY-4B/AGRI SST. This is consistent with the spatial distribution of bias in Figure 7a,e, with a significant positive bias observed in high-latitude sea waters and a significant negative bias observed in mid- to low-latitude sea waters. The negative biases increase in their magnitude with the increase in measured SST. The substantial number of matchup samples present in the high-temperature region is a significant factor contributing to the negative bias characteristics of the FY-4B/AGRI SST.

5. Conclusions

This paper aims to assess the data quality of the first batch of FY-4B/AGRI operational SST products with in situ SST and OSTIA SST. The FY-4B/AGRI operational SST product provides data quality level flags on a pixel-by-pixel basis. Accuracy evaluations are conducted on the FY-4B/AGRI SST with different data quality levels.

The FY-4B/AGRI SST with an excellent quality level can be used in cases where the absolute accuracy of the SST is crucial, especially in climate applications, but the FY-4B/AGRI SST with good and bad quality levels may be used in applications that primarily focus on the SST coverage rather than the absolute accuracy. The FY-4B/AGRI SST with excellent and good quality level demonstrates a strong correlation with the in situ SST and OSTIA SST, with correlation coefficient R exceeding 0.99, NSE exceeding 0.98 and KGE exceeding 0.97. However, from the performance of the bias, RMSE, and ubRMSE, the accuracy of the FY-4B/AGRI SST with a good quality level is obviously inferior to that of the FY-4B/AGRI SST with an excellent quality level. Compared with the in situ SST, the FY-4B/AGRI SST with an excellent quality level has the mean bias, RMSE, and ubRMSE of −0.19, 0.66, and 0.63 °C for daytime data, and −0.15, 0.70, and 0.68 °C for nighttime data, respectively. This result is superior to the quality validation results of FY-4A/AGRI SST products [19]. Compared with the OSTIA SST, the FY-4B/AGRI SST with an excellent data quality level has the bias, RMSE, and ubRMSE of −0.10, 0.64, and 0.63 °C during the day, and −0.13, 0.68, and 0.67 °C at night. This result is slightly better than the comparison with the in situ SST. According to the RMSEs, ubRMSEs, and error variances of the three-way error analysis method, the daytime SST products of FY-4B/AGRI are slightly better than nighttime SST products. However, from the mean biases between the satellite-derived SST and in situ SST, the underestimation of SST products during the day is greater than that at night.

The error of the FY-4B/AGRI SST exhibits regional differences in the spatial distribution and fluctuation in the temporal variation. In terms of spatial distribution, the FY-4B/AGRI SST tends to underestimate the sea water temperature in the mid–low-latitude regions, while it tends to overestimate the sea water temperature in the high-latitude regions and near the edges of the full disk. In terms of the time variation, the RMSE of the FY-4B/AGRI SST shows a weak fluctuation over a period of 3–4 months, with fluctuation intensities of 0.06 °C during the day and 0.05 °C at night, respectively. However, due to the insufficient time series of the data, more significant patterns still need to be observed in the future. Hourly accuracy verification shows that the difference of the FY-4B/AGRI SST minus OSTIA SST can reflect the diurnal effect, with heating from sunrise to noon and cooling from the afternoon. It should be noted that the FY-4B/AGRI SST product may experience changes in SST accuracy around midnight due to cloud detection failure. The key attribute named “NOMQC” which characterizes the SST quality of the full disk should be used for identification.

There are multiple reasons why the FY-4B/AGRI operational SST product underestimates seawater temperature. The FY-4B/AGRI estimates the skin SST, and, due to the flux almost always flowing from the ocean to the atmosphere, it is expected that the FY-4B/AGRI SST is lower than the bulk in situ measurements [55,56]. The in situ SST has a bidirectional effect on the FY-4B/AGRI SST, overestimating the low SST and underestimating the high SST. As shown in Figure 13, many effective samples concentrate in the high-temperature range; the negative biases increase in their magnitude with an increase in measured SST. The water vapor content, wind speed, and satellite zenith angle also play an undeniable role in the negative bias of the FY-4B/AGRI SST. In addition to the factors mentioned in this paper, many other factors affect the accuracy of the satellite-derived SST, such as clouds, sea fog, aerosol, dust [57,58], etc. Even the type of observation platform affects the accuracy of the satellite-retrieved SST [23,41]. It is precisely for this reason that the ship measurements in the iQuam dataset are excluded from this study. Moreover, changes in the status of satellite sensors, such as equipment aging and calibration updates [33], lead to changes in the SST accuracy as well.

Future research might explore the errors caused by different in situ platform types and the effect of other environmental conditions (e.g., atmospheric aerosol and cloud contamination) on the quality of the retrieved SST in different seasons. Conducting a bias correction to improve the accuracy of the SST products is also a future research work.