Estimating Hourly Land Surface Temperature From FY-4A AGRI Using an Explicitly Emissivity-Dependent Split-Window Algorithm

Land surface emissivity (LSE) has been roughly treated in the current split-window (SW) land surface temperature (LST) retrieval algorithms. This article extended the National Oceanic and Atmospheric Administration Joint Polar Satellite System enterprise algorithm to Feng Yun-4A/Advanced Geostationary Radiation Imager (AGRI) thermal infrared data by incorporating a daily LSE database for high-temporal-resolution LST retrieval. To improve the retrieval accuracy, day/night SW algorithm coefficients were calculated for different total water vapor content and view zenith angle conditions using a simulation database constructed by moderate spectral resolution atmospheric transmittance model version 5.2 and SeeBor V5.0 atmospheric profiles. The validation results show that the daily AGRI LSE has better accuracy than the LSE retrieved from the vegetation cover method (VCM), with average biases of −1.1×10−3 and −6×10−3 for channels 12 and 13. The accuracy of the AGRI LST retrieved using the daily AGRI LSE is slightly better than that retrieved using the VCM-retrieved LSE. The overall bias, MAE, and root mean square error of the AGRI LST retrieved using the daily AGRI LSE at 14 in situ sites are 0.11, 2.55, and 2.55 K, respectively, whereas these values are −0.11, 2.70, and 2.70 K, respectively, for the LST using the VCM-retrieved LSE. This study demonstrates that the daily LSE constructed from physically retrieved LSE can improve the accuracy of LST retrieved with the SW algorithm. The constructed daily LSE has high spatial coverage and dynamic emissivity information and can provide nearly complete spatial coverage if supplemented by the constructed eight-day or monthly AGRI LSE. It can also be applied to other LST retrieval algorithms that need LSE a priori.


I. INTRODUCTION
L AND surface temperature (LST) is one of the key parameters in land-surface physical processes at regional and global scales, integrating the interactions and all energy exchanges between the atmosphere and the land [1], [2]. Due to its momentousness in research fields, such as surface energy balance [3], [4], drought monitoring [5], and urban heat island research [6], LST has been identified as a critical Earth System Data Record by the National Aeronautics and Space Administration and many other international organizations [7]. Considering the sparsely distributed in situ sites and inaccurate model simulations, remote sensing is a unique way to monitor LST with spatiotemporal continuity at regional and global scales [8], [9], [10].
Feng Yun-4 (FY-4) is the second generation of China's geostationary meteorological satellite series to replace the firstgeneration Feng Yun-2 (FY-2) series. FY-4A was launched on December 11, 2016, and began operation in 2018 [11]. The Advanced Geostationary Radiation Imager (AGRI) onboard FY-4A can measure the thermal infrared (TIR) radiation in four channels (channel 11: 8.0-9.0 μm, channel 12: 10.3-11.3 μm, channel 13: 11.5-12.5 μm, and channel 14: 13.2-13.8 μm) with 4000-m spatial resolution at the nadir, 15-min temporal resolution at full disk scan, and 5-min intervals at rapid scanning of China [12]. The accuracy requirement for prelaunch calibration and onboard calibration in AGRI TIR channels is less than 1 K at 300 K for temperatures between 180 and 330 K [12]. Field experiment data in Qinghai Lake indicated that the average temperature bias of AGRI channel 12 (13) was 0.12 K (0.61 K) on August 18 and −0.01 K (−0.48 K) on August 20 [12]. The cross-calibration result of AGRI TIR channels using the cross-track infrared sounder and infrared atmospheric sounding interferometer showed that the absolute brightness temperature bias for most of the AGRI TIR channels was less than 0.5 K [13]. The above-mentioned research studies indicated that the radiometric calibration of FY-4A/AGRI TIR channels works well and can provide high-quality input data for LST retrieval.
LSEs in the two TIR channels are essential inputs of the SW algorithm. Although the vegetation cover method (VCM) [33], the classification-based method [34], and the normalized difference vegetation index (NDVI) threshold method [35] have been widely used for estimating LSE, there are still some problems in estimating AGRI LSE using these methods. First, the classification-based method cannot characterize the seasonal variation of vegetation abundance, and inaccurate LSEs over bare soils can lead to large LST biases. Second, the NDVI threshold method relies on the accuracy of visible and nearinfrared (VNIR) channel reflectance and the empirical relationship fitted by using the emissivity spectra in the spectral library. On one hand, the limited emissivity spectra cannot represent a complex surface situation [36]. On the other hand, although the empirical relationship can be refitted using satellite data, given that the scattering of aerosol optical depth (AOD) is a critical component in atmospheric correction, it is difficult to obtain accurate land surface reflectance for the uncertainty of AOD estimation, especially over heterogeneous surfaces [37]. Third, although the VCM method can be improved by using the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Emissivity Dataset (GED), the data gap in the ASTER GED may result in the loss of valid AGRI pixels. In addition, the view zenith angle (VZA) of AGRI can reach up to 80°, and the above-mentioned LSE estimation algorithms cannot take into account the LSE angular effect. The angular variation of directional emissivity in the TIR channel centered around 11 and 12 μm could reach up to 0.01 [38], [39], which means an equivalent LST uncertainty of about 1.0 K with the LST equal to 300 K. The angular effect of emissivity on the retrieved LST with the SW algorithm may be up to 3.0 K over desert regions when VZA is greater than 65° [40]. Liu et al. [41] extended the improved TES (iTES) algorithm from Himawari-8/AHI to FY-4A/AGRI for obtaining the instantaneous AGRI LSE. Cross-validation results indicated the absolute biases and root mean square errors (RMSEs) were all less than 0.007 and 0.016 for channels centered at 8.5, 10.8, and 12.0 μm, and the eight-day mean value composited LSE database had achieved nearly full spatial coverage, which can be used for LST retrieval from SW algorithms.
The purpose of this study is twofold: 1) to make the SW algorithm more complete in principle, and 2) to improve the LST retrieval accuracy of FY-4A/AGRI. This article is organized as follows. In Section II, the SW algorithm, ground measurements, and satellite datasets used for LST&E retrieval and validation are described. In Section III, the evaluation results of the retrieved LST&E are presented. In Section IV, the discussion of this study is presented. Finally, Section V concludes this article.

A. National Oceanic and Atmospheric Administration Joint Polar Satellite System Enterprise Algorithm
A previous research study indicated that the National Oceanic and Atmospheric Administration (NOAA) Joint Polar Satellite System (JPSS) enterprise algorithm had comparable accuracy to other SW algorithms widely used [25]. Therefore, the NOAA JPSS enterprise algorithm was adapted to AGRI data for retrieving LST in this study. The enterprise LST algorithm is based on an SW technique that corrects for atmospheric absorption and applies surface emissivity explicitly in the retrieval, which can be described as follows [19], [42]: where T s is the LST; C i (i = 0-5) are the algorithm coefficients to be determined from simulation data; T 12 and T 13 are the brightness temperatures of AGRI channels 12 and 13, respectively; ε = (ε 12 + ε 13 ) * 0.5; Δε = ε 12 − ε 13 ; ε 12 and ε 13 are the LSEs of AGRI channels 12 and 13, respectively. Provided with LSE and two brightness temperatures, the LST calculation is straightforward.

B. Determining the Coefficients of the Enterprise Algorithm
To obtain the algorithm coefficients, the moderate spectral resolution atmospheric transmittance model version 5.2 (MOD-TRAN 5.2) [43] was used to simulate atmospheric upward thermal radiance, downward thermal radiance, and transmittance of AGRI channels 12 and 13, with the SeeBor V5.0 atmospheric profile database [44] as input. After excluding atmospheric profiles with latitudes larger than 75°or less than −60°and removing atmospheric profiles that were considered cloudy based on the research of Galve et al. [45], 2515 daytime atmospheric profiles and 2088 nighttime atmospheric profiles were selected over the land according to the local sunrise and sunset times. As shown in Fig. 1, the bottom layer temperature (T0) in atmospheric profiles and the total water vapor content (TWV) vary from 200.16 to 318.53 K and from 0 to 7.0 g/cm 2 , respectively.
The atmospheric transmittance (τ (θ)) and the atmospheric upward radiance (L ↑ (θ)) at 15 VZAs between 0°and 70°a t an interval of 5°and the atmospheric downward radiance (L ↓ ) at 53°were simulated using MODTRAN 5.2 for each atmospheric profile. Then, the spectral results were convolved with the spectral response function of AGRI channels 12 and 13 to obtain the band-integrated atmospheric parameters. For realistic simulation under complex and variable atmospheric conditions using finite atmospheric profiles, we followed the work of Jiménez-Muñoz et al. [46] and set the LSTs as follows: T0 − 5, T0, T0 + 5, T0 + 10, and T0 + 20. Moreover, 98 emissivity spectra were selected from the ASTER spectral library [47] and MODIS UCSB spectral library [48], including 19 vegetation, 5 water/ice/snow, 35 rock, and 39 soil spectra. Using the 2515 daytime (2088 nighttime) atmosphere profiles, 5 LSTs, 98 emissivity spectra, and 15 VZAs, the brightness temperatures of AGRI channels 12 and 13 were simulated. After obtaining the simulation database, the algorithm coefficients in (2) are determined by statistical regression. To improve the fitting accuracy, the daytime/nighttime algorithm coefficients in different TWV and VZA subranges were estimated using the simulation data. Following the Algorithm Theoretical Basis Document for VIIRS LST production, we divided the TWV into four subranges: TWV < 1.5 g/cm 2 , 1.5 ≤ TWV < 3 g/cm 2 , 3.0 ≤ TWV < 4.5 g/cm 2 , and 4.5 g/cm 2 ≤ TWV, respectively, representing very dry atmosphere, dry atmosphere, moist atmosphere, and very moist atmosphere [19]. Based on the solar zenith angle (SZA) of each pixel, different coefficients were used to calculate the AGRI LST, and the daytime coefficients and nighttime coefficients were used when SZA < 80 • and 100 • < SZA, respectively. Moreover, for the dawn and twilight period (80 • ≤ SZA ≤ 100 • ), the AGRI LST of these pixels was done by calculating the linear average of the daytime and nighttime LST [30]. As shown in Table I, the RMSEs increase with the VZA and TWV values. The RMSEs are smaller than 1.0 K when the TWV < 3.0 g/cm 2 and the VZA < 65°, but RMSEs are larger than 1.0 K when the TWV > 3.0 g/cm 2 and the VZA > 50°. This suggests that the SW algorithm should be carefully implemented for LST retrieval of geostationary satellite data at higher VZA and TWV values.

C. Determination of LSE
LSE is an important parameter for generating high-quality LST products from the SW algorithms. The LSE used in this research was the daily emissivity composited from the valid clear-sky pixels of the AGRI LSE product, which was retrieved by the iTES algorithm. The modified water vapor scale method and a recalibrated empirical function over vegetated surfaces [49] were adopted in the iTES algorithm [41]. First, the "3σ-Hampel identifier" [50] was used to robustly remove outliers in the AGRI LSE product. Then, the mean, standard deviation, and the number of AGRI LSEs used in composition were calculated for each pixel. The daily composited LSE was incorporated into the enterprise algorithm to retrieve AGRI LST, abbreviated as EA-ites later.
VCM is a simple and effective method that has been integrated for producing LST products, for e.g., VCM has been used to produce the LSE from the SEVIRI [51], VIIRS [19], and Advanced Along Track Scanner Radiometer [52]. Therefore, as a reference, the ASTER GED product and Multisource data Synergized Quantitative remote sensing production system (MuSyQ) fractional vegetation cover (FVC) product [53] were also used to estimate the AGRI LSE using VCM with the following equations: where ε AGRI is the VCM-retrieved AGRI LSE for channels 12 and 13, ε Aster is the ASTER GED LSE for channels 13 and 14, ε bare,Aster is the ASTER bare soil component emissivity for channels 13 and 14, ε veg,Aster is the ASTER vegetation component emissivity, which was set as 0.981 and 0.983 for channels 13 and 14 [54], [55], respectively, ε bare,AGRI is the AGRI bare soil component emissivity for channels 12 and 13 that has been spectrally adjusted using (6), ε veg,AGRI is the vegetation component emissivity, which was set as 0.983 and 0.989 for channels 12 and 13, respectively, ε bare,Aster13 and ε bare,Aster14 are the ASTER bare soil component emissivities for channels 13 and 14, respectively, f v,Aster is the estimated ASTER FVC using the corresponding NDVI [56], NDVI max and NDVI min are NDVI values of highly dense vegetation and bare soil, which were set as 0.5 and 0.2, respectively [57], f v,AGRI is the estimated AGRI FVC using the resampled MuSyQ FVC product. The VCM-retrieved AGRI LSE was also incorporated into the enterprise algorithm to retrieve the LST from AGRI data for the purpose of comparison and named EA-vcm in the text. The spatial coverage of daily, eight-day, and monthly composited LSE is shown in Fig. 2. The spatial coverage of the daily AGRI LSE on January 15, 2019 is relatively complete, with most of the STDs less than 0.006 and most of the number of    there are still gray areas without valid clear-sky coverage, and thus the eight-day AGRI LSE and the monthly AGRI LSE were also composited based on mean value composition. Except for midlatitude regions in summer, the eight-day AGRI LSE can achieve full spatial coverage, whereas the monthly AGRI LSE can achieve nearly full spatial coverage in any region.

D. Ground Measurements
The surface upward and atmospheric downward longwave radiations collected from 14 sites at the Qilian Mountains integrated observatory network and OzFlux ecosystem research network were selected for evaluating the performance of the EA-ites. The Qilian Mountains integrated observatory network includes the Heihe Watershed Allied Telemetry Experimental Research network [58], [59], [60], [61], Qinghai Lake integrated observatory network [62], [63], and Cold and Arid Research Network of Lanzhou University. The ground measurements at those sites have been widely used for evaluating satellite-derived LST, surface longwave radiance, and evapotranspiration products [64], [65], [66], [67]. The geographical locations of the 14 sites are shown in Fig. 3, and details of the validation sites can be found in Table II.
The in situ LSTs were estimated from the surface longwave upward/downward radiation measured by the Kipp and Zonen CNR1/CNR4 net radiometer using the following equation: Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. where T s is the LST, F ↑ and F ↓ are the surface upward and atmospheric downward longwave radiations, respectively, ε b is the broadband emissivity (BBE), σ is the Stefan-Boltzmann constant (5.67 × 10 −8 Wm −2 K −4 ). The BBE was extracted from the global eight-day 5-km Global LAnd Surface Satellite (GLASS) BBE product. GLASS BBE was estimated from a nonlinear function of advanced very high-resolution radiometer VNIR reflectance data. Cross-validation result showed that the absolute biases and RMSEs were less than 0.003 and 0.014 [68], respectively.

E. Ground LSE Data
The ground-measured LSE comes from two sources [88], [90], and the detailed information can be found in Table III. The land cover type at all 14 sites is desert. Those ground measurements had been used to validate VIIRS LSE and MODIS LST products [19], [70]. We obtained the LSE data from their papers and converted them to AGRI channels 12 and 13 by spectral adjustment based on their similar spectral ranges, as shown in Fig. 4.

F. Satellite Data
The AGRI L1 and L2 products were used to retrieve LST and were downloaded from the China Meteorological Administration National Satellite Meteorological Center (NSMC). The AGRI L1 products provide radiance, geolocation, VZA, and SZA data, whereas the L2 products provide TWV and cloud mask data. The AGRI LSE product retrieved from the iTES algorithm was used to calculate the daily average emissivity, and the AGRI LST was retrieved from the enterprise algorithm.
After spatial-temporal matching, the Collection 6.0 MODIS LST&E products (MYD21) were used for cross validation of AGRI LSE. As the LSE remained stable for a short period of time, the multiple MODIS LSE products acquired throughout the day were used to evaluate the AGRI LSE. To ensure that the VZA of the spatial-temporal collocated pixel was similar, we used the following equation to control the observation angle [72] cos (VZA AGRI ) where VZA AGRI and VZA MODIS are the AGRI VZA and MODIS VZA, respectively. The angular threshold of 0.01 corresponds to a maximum difference of 1% in atmospheric path length [73].

A. LSE Evaluation With Ground Measurements
Although there are no synchronous in situ LSE data, the referenced AGRI LSE calculated by the ground measurements presented in Table III was used to check the performance of the daily AGRI LSE, and the VCM-retrieved LSE was also evaluated. The evaluation results for the daily AGRI LSE in 2019 are shown in Table IV.
The absolute values of the average biases of the daily AGRI LSE for channel 12 are within 1.0×10 −2 at most sites. The daily AGRI channel 12 LSEs at Tenggeli2, Gebi, Huazhaizi, and Jichanghuangmo sites have excellent agreement with the referenced AGRI LSEs, with average biases of approximately zero. The absolute values of average biases at Badanjilin1, Badanjilin2, and Shenshawo2 sites are larger than 1.1×10 -2 . As for the results of AGRI channel 13 LSE, average biases at Kubuqi1, Shenshawo1, and Tenggeli1 are approximately zero, whereas the values are within 1.0 × 10 −2 for other sites except  The above-mentioned evaluation results indicate that the accuracy of the daily AGRI LSE is slightly better than the VCMretrieved LSE, with an absolute bias reduction of 8.7 × 10 −3 for channel 12 and 3.2 × 10 −3 for channel 13. The small improvement can be explained by the following reasons.
1) The daily AGRI LSE has higher temporal resolution than that of the VCM-retrieved LSE, which may capture the potentially changing characteristics of the surface, e.g., wind erosion and precipitation. 2) In arid/semiarid regions with low water vapor content, there is no significant difference in the accuracy of LSE retrieved with the iTES and TES algorithms, and therefore, the above-mentioned improvement is slight. 3) ASTER GED is a multiyear average emissivity that reduces the effect of outliers, whereas the daily average emissivity is affected by random errors with a high probability. Therefore, the performance of the daily AGRI channel 12 LSE at Badanjilin2, Mingshashan, is worse than that of the VCM-retrieved LSE. Table V shows the validation results of the AGRI LST retrieved by the EA-ites, EA-vcm, and iTES at all sites. Considering the effect of cloud and cloud shade on the retrieved LST, the "3σ-Hampel identifier" [50] was used to robustly remove outliers that existed in the retrieved AGRI LST. In addition, the daytime validation result for the retrieved AGRI LST, including the valid AGRI LST acquired during the dawn and twilight periods.

B. LST Validation With Ground Measurements
The overall bias, RMSE, and unbiased RMSE (uRMSE) of the EA-ites-, EA-vcm-, and iTES-retrieved AGRI LST exhibit similar characteristics. The overall bias, RMSE, and uRMSE between the EA-ites-retrieved AGRI LST and the in situ LST are about 0.11, 2.55, and 2.55 K, respectively, whereas these values are −0.11, 2.70, and 2.70 K for the EA-vcm-retrieved AGRI LST, respectively, and 0.34, 2.58, and 2.56 K for the iTES-retrieved AGRI LST, respectively. The bias, RMSE, and uRMSE between the EA-ites-retrieved nighttime AGRI LST and the in situ LST are −0. 35 To further illustrate the accuracy of the AGRI LST at each site, Fig. 5 shows the validation results of the daytime/nighttime AGRI LST at 14 sites. The absolute values of daytime (nighttime) biases are less than 1.85 K (2.09 K) at most Qilian networks, except for the EA-vcm-retrieved AGRI LST at the HM site, with a daytime bias of 3.33 K and a nighttime bias of 2.83 K. The EA-ites-, EA-vcm-, and iTES-retrieved AGRI LST have considerable RMSEs and uRMSEs at most sites, with RMSE and uRMSE differences less than 0.34 (0.15) and 0.29 K (0.17 K) during daytime (nighttime), respectively. As for the validation results at the OzFlux network, daytime (nighttime) biases are between −1.05 K (−1.54 K) and 2.23 K (1.12 K) for the EA-ites-retrieved AGRI LST, −1.27 K (−1.92 K) and 1.54 K (0.40 K) for the EA-vcm-retrieved AGRI LST, and 0.33 K (−0.51 K) and 2.89 K (1.23 K) for the iTES-retrieved AGRI LST, respectively. The validation metrics between the EA-itesand EA-vcm-retrieved AGRI LST and in situ LST exhibit similar characteristics at most sites, except for the HM, LZ, DU, and SP sites covered by bare soil, cropland, savanna, and grassland, respectively. According to the statistical results, the lower and upper quartiles of the VCM-retrieved LSE are between 0.876 (0.869) and 0.989 (0.993) for channel 12 (13)     on January 1, 2019, and July 24, 2019, and underestimation on April 4, 2019, and October 17, 2019. The LST series can be used to build a diurnal temperature cycle model for obtaining theoretical clear-sky LST, which is a prerequisite for obtaining all-weather LST.
The above-mentioned validation results show that the accuracy of the EA-ites-retrieved AGRI LST is slightly better than that retrieved by the EA-vcm and iTES.

C. Comparison With the MODIS LSE Product
Comprehensive validation of the estimated LSE is challenging due to limited ground measurements, especially on vegetated surfaces. As shown in Fig. 4, the spectral ranges of MODIS channel 31 (32) are similar to those of the AGRI TIR channel 12 (13); therefore, the MYD21 LSE product retrieved from the TES algorithm [74] was used to assess the quality of the daily AGRI LSE, and the evaluation results are shown in Figs. 7 and 8. Since the single MODIS image can only cover part of the FY-4A image, the error bars in Figs. 7 and 8 represent the daily standard deviation between the daily AGRI LSE and the MYD21 LSE.
The evaluation results show a good agreement between the daily AGRI LSE and MYD21 LSE, with average biases (RMSEs) equal to 2.1 × 10 −3 (1.1 × 10 −2 ) and 0 (1.1 × 10 −2 ) Fig. 9. Biases and RMSEs of LST differences between the EA-ites-retrieved and iTES-retrieved AGRI LST in 2019. The error bars denote the monthly standard deviation between the EA-ites-retrieved AGRI LST and the iTES-retrieved AGRI LST.  Fig. 9 shows the monthly average biases and RMSEs of LST differences between the EA-ites-retrieved and iTES-retrieved AGRI LST in 2019. The biases and RMSEs of LST differences change from −0.90 to −0.18 K and from 1.20 to 2.14 K in 2019, respectively, whereas the corresponding standard deviations range from 0.20 to 0.32 K and from 0.14 to 0.27 K, respectively. The evaluation results indicate that EA-ites-retrieved AGRI LST is in good agreement with iTES-retrieved AGRI LST, with an absolute monthly average bias (RMSE) of less than 0.50 K (1.67 K), except for June, July, and August.

D. Comparison With iTES-Retrieved AGRI LST
Taking the evaluation result at 03:00 UTC on July 22, 2019, as an example, we try to explain the possible reason for the larger biases. Fig. 10 shows the LSE, LST, and TWV results extracted from AGRI data. The overall bias and RMSE are −1.56 K and 2.51 K, respectively. The apparent difference between the EA-ites-retrieved AGRI LST and the iTES-retrieved AGRI LST is mainly due to the following reasons:  Table I, with the increase in TWV and VZA, the SW algorithm may introduce uncertainty. Thus, the SW algorithm should be carefully implemented for LST retrieval of geostationary satellite data at higher VZA and TWV values.

A. Comparison With LST Retrieval Using the Official LSE Product
Since the official LST product is only available after August 2019, we downloaded the official AGRI LSE product (2019/01/18-2019/12/31) from NSMC and tested its efficacy in retrieving LST. The official AGRI LSE was estimated using the physical-based optimization method [41], and cross validation with the MODIS LSE product showed that the absolute deviations of AGRI LSE for channels 12 and 13 are greater than 0.01 [75]. This comparison can avoid errors caused by inconsistent algorithms. The validation result of the AGRI LST retrieved from the enterprise algorithm is shown in Fig. 11.
As shown in Fig. 11, when compared to the in situ LST, the AGRI LSTs retrieved using the official LSE product are overestimated at all sites, with daytime biases ranging from 0.22 to 6.08 K and nighttime biases ranging from −0.75 to 5.27 K. The AGRI LSTs retrieved using the official LSE product are also overestimated when compared to the EA-ites-and EA-vcm-retrieved LSTs. The overestimation can be explained by the poor performance of the official AGRI LSE. On one hand, when using the official LSE product with a spatial resolution of 12 km for LST retrieval, the spatial mismatch between the official AGRI LSE and other data may lead to larger errors. On the other hand, coarse resolution causes official LSE to lose spatial detail [41], which may result in the poor performance of retrieved AGRI LST, especially on heterogeneous surfaces.

B. Comparison With Existing Studies
Wang et al. [76] adopted two SW algorithms developed by Becker and Li [77] and Kerr et al. [78] for AGRI LST retrieval. The in situ LST measured by the SI-411 infrared temperature sensor fixed at 1.5 m on three observation towers was used for validation. According to the research of Wang et al. [76], the RM-SEs of AGRI LST retrieved from two SW algorithms are 6.28 K and 5.55 K. To obtain land surface characteristic parameters and turbulent heat fluxes, Ge et al. [79] used the SW algorithm developed by Becker and Li [80] to retrieve the AGRI LST. The retrieved LST was validated using in situ LSTs from two sites on the Tibet Plateau, and the result shows that the MAE and RMSE are 3.24 K and 3.85 K, respectively. Fan et al. analyzed the accuracy of the official AGRI LST using in situ data observed by platinum resistance sensors and found an underestimation of −0.63 K and an uncertainty of 8.59 K in the official AGRI LST. The validation results of Liu et al. [41] showed that the official AGRI LSTs are systematically underestimated at AR and HM sites, with biases of −0.82 K (−0.92K) and −1.73 K (−2.06 K) during daytime (nighttime), respectively. The RMSEs of the official AGRI LST are 3.53 K and 2.95 K, respectively, for AR and HM sites during daytime, whereas the values are 3.15 K and 3.09 K, respectively, during nighttime, which are greater than the RMSEs presented in Sections III and IV.
As shown in Table VI, the EA-ites-retrieved LST has lower RMSE than the above-mentioned validation results. Therefore, the daily AGRI LSE can be used to improve the accuracy of the LST retrieved from the SW algorithm. The above-mentioned evaluation results may be explained by the following reasons: 1) the unstable radiometric performance of FY4A/AGRI VNIR channels. The NDVI-based method was used for estimating LSE in the study of Wang et al. [76] and Ge et al. [79], and thus, the radiometric performance of VNIR channels has an effect on the estimated LSE over nonvegetated surfaces. The study of Zhong et al. [81] indicated that the radiometric performance of FY4A/AGRI is less stable with significant fluctuations, which may lead to large RMSEs of AGRI LST. 2) According to the discussion presented in Section IV-A, there are differences in the performance of the AGRI LST retrieved using the daily AGRI LSE, VCM-retrieved LSE, and official AGRI LSE when maintaining the consistency of the LST retrieval algorithm but with different LSEs as input data. Thus, the different performance of the retrieved LST can be attributed to the discrepancy in LSE. The daily AGRI LSE inherits the advantage of the iTES algorithm and thus outperforms the official LSE product and the VCM.

C. Merits, Limitations, and Possible Improvements
Although previous studies have shown that the MxD21 LSE product can provide accurate LSE over barren surfaces [82] and can improve the accuracy of the GSW algorithm [70], the MxD21 LSE product was not used in this research for the following two reasons: 1) As shown in Section III-C, the daily AGRI LSE has comparable accuracy with the MYD21 LSE but with more spatial integrity. With a temporal resolution of 15 min, the AGRI can acquire 96 images per day in theory, whereas the MODIS can only obtain four images per day. More images represent a higher probability of spatial integrity. 2) When the daily AGRI LSE was used for retrieving the AGRI LST, there was no spatial sampling/matching problem with AGRI. However, for the MxD21 LSE, spatial matching, temporal matching, and spectral adjustment are the first steps, which may introduce uncertainty into the estimation of the LSE and LST.
The original space coverage of the multiyear composited ASTER GED is 95.38% in the AGRI full disk area. When the ASTER GED dataset was used in the LST retrieval algorithm, the blank banding in the LST image was still very noticeable, even without the effect of clouds. The median spatial coverage of the constructed daily, eight-day, and monthly AGRI LSE is 85.02%, 99.24%, and 99.91%, respectively, whereas the corresponding upper and lower quartiles are 80.01% and 88.04%, 98.47% and 99.53%, and 99.83% and 99.93%, respectively. Although the constructed daily AGRI LSEs cannot provide full coverage, they can be supplemented with the constructed eight-day or monthly AGRI LSEs to achieve full spatial coverage, thus providing LSE input for LST retrieval of other satellites. In addition to FY-4A/AGRI, this approach can also be extended to other satellites, e.g., the FY-2 and FY-3 series. When applied to retrieve LST from other satellite TIR data, only spectral adjustments are needed, so even the Indian National Satellite-3D (INSAT-3D), which only has one visible band and two longwave TIR bands, can benefit from this approach for LST retrieval.
The daily AGRI LSE was validated using ground measurements, and the referenced LSEs were calculated by the spectral adjustment of the ground-measured VIIRS or MODIS LSEs, rather than the field-measured emissivity spectrums, which may add adjunctive uncertainty to LSE validation. When ground measurements were used for LST validation, it was a pointto-pixel comparison, not a pixel-to-pixel comparison that was implemented by upscaling the ground measurements to satellite pixel scale. Moreover, the angular effect was not considered in the cross validation of AGRI LSE and LST.

V. CONCLUSION
This study developed an explicitly emissivity-dependent SW algorithm for FY-4A/AGRI LST retrieval by combining the NOAA JPSS enterprise algorithm with a daily LSE database. The algorithm was named the EA-ites algorithm. The simulation database constructed by MODTRAN 5.2 and SeeBor V5.0 atmospheric profiles was used to derive the day/night SW algorithm coefficients under different TWVs and VZAs. The major findings can be summarized as follows.
1) The evaluation results using ground measurements show that the accuracy of the daily AGRI LSE is slightly better than the VCM-retrieved LSE, with an absolute bias reduction of 0.007 for channel 12 and 0.004 for channel 13. The evaluation results using the MYD21 LSE product demonstrate that the daily AGRI LSE can achieve comparable accuracy to the MYD21 LSE, with average biases of 1 × 10 −3 and 0 for channel 12 and −2×10 −3 and −1×10 −3 for channel 13 in January and October 2019. The constructed daily AGRI LSE has high spatial coverage, with a median of 85.02%. 2) Ground measurements collected at 14 in situ sites were used to evaluate the performance of the EA-ites-retrieved AGRI LST. Validation results indicate that the performance of the EA-ites-retrieved AGRI LST is slightly superior to the LST using the VCM-retrieved LSE and the iTES-retrieved LST, with a 0.15 K (0.15 K) and 0.03 K (0.01K) reduction in the RMSE (uRMSE). This study demonstrates the utility of physically retrieved LSE in the estimation of LST by the SW algorithm. The composited daily LSE can also be applied to the LST retrieval for other satellites using the SW algorithm and the multichannel LST retrieval algorithm that needs LSE a priori.