*3.3. Validation Using In-Situ Observation Data*

To evaluate the the GK2A LST algorithm, we conducted a quantitative validation of the GK2A LST using in situ observation data. As mentioned in Section 2.1.2, the upward longwave radiation data observed at the Tateno station of the BSRN were used as validation data. Approximately 718 sets from July 2019 to October 2019 satisfying the spatial-temporal matching conditions with the GK2A LST were used for the validation of the GK2A LST. In addition, the performance of the GK2A LST algorithm was evaluated according to daytime and nighttime. The validation results of the GK2A LST with the data from Tateno station are shown in Figure 11. The GK2A LST is similar to or slightly warmer than the ground observed LST at Tateno station, regardless of the temperature and time (day and night). The total correlation coefficient, bias, and RMSE between the GK2A LST and the Tateno LST were 0.95, 0.523 K, and 2.021 K, respectively. However, the GK2A LST was warmer than the ground observed LST; in particular, the LST was greater than 305 K during the daytime. As shown in Figure 11, the GK2A LST shows a greater warm bias (0.84 K) and RMSE (2.13 K) during the day than at night (bias: 0.32 K, RMSE: 1.948 K).

**Figure 11.** Scatter plot between the GK2A LST and the LST at Tateno station from upward longwave radiation (red square symbol: daytime, blue cross symbol: nighttime, gray dotted line represents the 1:1 line).

#### **4. Discussion**

In this study, we developed a GK2A SW LST retrieval algorithm using two adjacent infrared channels in the atmospheric window. GK2A/AMI has three infrared channels (channels 13, 14, and 15) corresponding to the atmospheric window [58]. Channel 15 is more sensitive to water vapor than the other two channels. By comparison, channels 13 and 14 are relatively less sensitive to water vapor, but the sensitivities of the two channels to aerosol and water vapor are slightly different. To select two of the three channels that can be used to retrieve the GK2A LST, RTM simulation results under the same RTM input conditions using channels 13 and 15, as well as channels 14 and 15, were analyzed. The regression coefficients (*c*0~*c*6) for the LST retrieval Equation (1) using channels 13 and 15 and channels 14 and 15, respectively, were derived from the simulated dataset. The scatter plot results of the estimated LSTs from each dataset are shown in Figure 12.

**Figure 12.** Scatter plots between reference LST and estimated LST from RTM simulation using two channels: (**a**) channels 13 and 15, and (**b**) channels 14 and 15.

The two sets of GK2A LST algorithms estimated LSTs in the range of 240 K to 330 K, but the correlation coefficient and RMSE using channels 13 and 15 showed better results than those using channels 14 and 15. In the scatter plots, the distribution used channels 14 and 15 showed a wider spread than that by channels 13 and 15 at over 300 K. This result is similar to that of a previous study using Himawari-8 [60]. Therefore, channels 13 and 15 were selected for this study.

In addition, several forms of linear and non-linear equations can be selected in the LST retrieval formula of the SW method [26,27]. When retrieving LST, the linear equation of SW LST algorithms showed a large error in wet and hot atmospheric conditions, therefore, non-linear equations of SW LST algorithms have been developed [26,47,77,82,83]. In this study, linear and non-linear equations of the SW LST algorithms were developed and applied to real GK2A data to compare the accuracy of the algorithm. The simulation conditions for the RTM are shown in Table 2, and the coefficient of *c*<sup>3</sup> in the LST retrieval equation (Equation (1)) is set to zero to represent a linear algorithm. In addition, the coefficients of LST retrieval algorithms were derived by dividing into day/night and dry/normal/wet conditions using the same thresholds of SZA and BTD, as described in Section 3. Table 6 shows the results of quantitatively comparing the GK2A LST and MODIS LST for one month from the linear algorithm and non-linear SW LST algorithm. As a result of verifying the GK2A LST calculated with linear and non-linear algorithms for the September 2019 case with the MODIS LST, the correlation coefficient between the GK2A LST and MODIS LST was very similar in linear and nonlinear algorithms, but the bias and RMSE showed better results in nonlinear algorithms.

**Table 6.** Comparison results of the GK2A LST data and the MODIS LST product (Collection 6) using a linear algorithm and non-linear algorithm according to the daytime and nighttime in September 2019.


Even though the non-linear SW LST algorithm was used, the GK2A LST algorithm showed significant errors during the daytime compared to the MODIS LST. The MODIS Collection 6 LST product, used as validation data, is an improved version of the MODIS Collection 5 LST product by Wan (2014) [32]. One of the improvements in the MODIS Collection 6 LST over the Collection 5 LST is that the MODTRAN simulation is performed by dividing into day and night for the bare soil area and adjusting the emissivity difference in MODIS bands 31 and 32 over bare soil surfaces [32]. In addition, a term including the quadratic difference between the brightness temperatures of bands 31 and 32 was added to the MODIS' generalized SW algorithm. Even though the MODIS Collection 6 LST product had many improvements compared to Collection 5, a cold bias still appeared from −1.4 to −3.7 K during the daytime when compared with in situ measurements [81]. According to the validation study, MODIS Collection 6 LST showed the RMSE of daytime LSTs as 2.59 K, 2.86 K, and 3.05 K for the Gobi area, desert steppe region, and sand desert area, respectively [80]. Considering that the cold bias of the MODIS Collection 6 LST is strong during daytime over bare soil and desert regions, the warm bias of the GK2A LST algorithm can be regarded as normal rather than a serious problem. However, a detailed analysis of bare soil and desert areas using more validation data is needed. In the four-month verification results, the errors in September and October were systematically larger than those in July and August. So, we tried to find the cause but, unfortunately, we could not. In addition, a relatively strong warm bias appeared during the day at Tateno station. This seems to be related to the fact that the land cover at the Tateno point is grass, but most of the area around this point is urban.

To compare the accuracy of the MODIS LST, it was validated using the LST at Tateno station for the same period as the GK2A LST. The validation results of the MODIS (MOD11/MYD11\_L2) LST with those from Tateno station are shown in Figure 13. The MODIS LST was slightly colder than Tateno LST, regardless of the daytime and nighttime. The total correlation coefficient, bias, and RMSE between MODIS LST and Tateno LST were 0.925, −1.047 K, and 2.985 K, respectively. Although the number of samples was small, the MODIS LST showed a cold bias compared to the in situ LST in both daytime and nighttime. In particular, the daytime bias (−1.402 K) of the MODIS LST was nearly twice that of the nighttime (−0.767 K). In the comparative validation results of the GK2A LST and the MODIS LST, the reason why the warm bias of GK2A is large during the daytime is also considered to be related to the cold bias of MODIS LST during the daytime.

The number of on-site observation points in the GK2A observation area is not only limited but also the number of data accessible over the internet is small, so we used observation data from the Tateno station. The in situ measured radiation represents a narrow area, whereas the retrieved LST from the satellite is the average temperature corresponding to satellite resolution (2 km × 2 km), so there is a limitation in the spatial representativeness of in situ observation for the satellite. Since the period for retrieving and validating the GK2A LST is as short as four months, there is a limit to evaluating the integrated level of the LST retrieval algorithm.

When retrieving LST from a satellite, the split-window method assumes that the LSE of both channels is known. In this study, the GK2A LSE data derived in real-time using the modified VCM method were used [68]. The fractional vegetation cover of a given pixel was calculated using the GK2A vegetation index (VI) data generated by the maximum value composite with a consecutive 8-day VI [63]. Based on the calculated fractional vegetation cover, the LSEs were retrieved using the look-up table according to the land cover and daily snow cover of each pixel. The VI, land cover classification database, spectral emissivity look-up table, and daily snow cover were calculated from each algorithm, so they contain errors which also affect the accuracy of the retrieved LST. Therefore, to improve the accuracy of the GK2A LST, it is necessary to improve the accuracy of these algorithms.

**Figure 13.** Scatter plot between the MODIS (MOD11\_L2/MYD11\_L2) LST and the LST at Tateno station from upward longwave radiation (red square symbol: daytime; blue cross symbol: nighttime; gray dotted line represents the 1:1 line).

### **5. Conclusions**

We have developed an operational LST retrieval algorithm for the GK2A viewing area using GK2A/AMI data from Korea's next-generation geostationary satellite. To develop the GK2A LST algorithm, the split-window method was used, and the nonlinearity of water vapor was considered in the algorithm. The RTM simulation data were constructed taking into account various factors affecting the LST calculation in MODTRAN 4 (atmospheric profiles, diurnal variation of LST and air temperature of the boundary layer, and the LSE variations of the two channels). From the RTM simulation data set, regression coefficients were derived according to the actual water vapor (GK2A BTD: dry/normal/wet) during the day and night. The GK2A LST was retrieved using the developed GK2A LST algorithm, and their accuracies were evaluated using MODIS LST and field observations as validation datasets.

As a result of evaluating the output level of the LST calculated by the RTM simulation using the prescribed LST, there was a correlation coefficient of 0.998, a bias of 0.01 K, and an RMSE of 0.767 K. When the BTD value is larger than 6 K and the satellite's VZA is large, the RMSE is large, but the error is relatively smaller than the result of using the linear algorithm of the previous study [61].

Using the GK2A LST algorithm developed in this study, LST was calculated from GK2A data for four months from July 2019 to October 2019. As a result of comparing the GK2A LST with the MODIS LST, the spatial correlation coefficient of the two LSTs was 0.969, the bias was 1.227 K, and the RMSE was 2.281 K. Compared to MODIS LST, GK2A LST shows a warm bias greater than 1.8 K during the day, but a relatively small bias of less than 0. 7 K at night. In particular, the warm bias of the GK2A LST was higher than that of MODIS LST in desert and barren areas during daytime. MODIS LST Collection 6, used as validation data, seems to be influenced by characteristics of cold bias in the desert and on bare soil [80,81]. The results of validation with data from the Tateno station of the BSRN, which were the field observation data, showed that the correlation coefficient is 0.95, the bias is 0.523 K, and the RMSE is 2.021 K. Compared to the Tateno LST, the day bias is +0.5 K greater than the night bias. The reason that the GK2A LST tends to be warmer than the Tateno LST during the day is that the temperature at Tateno station is the temperature of the grass, while the GK2A LST is the temperature of the surrounding urban area.

The GK2A LST algorithm developed in this study uses various outputs of GK2A (cloud detection, VI, snow cover, and LSE) and ancillary data to calculate LST. Therefore, it is possible to improve the accuracy of the GK2A LST by improving the algorithm that produces the basic input data for LST calculation. In addition, it is necessary to evaluate the output level of the GK2A LST algorithm using verification data for a longer time because the surface temperature and atmospheric characteristics differ depending on the season. As other satellite data (Visible Infrared Imaging Radiometer Suite LST, Sentinel-3 Sea and Land Surface Temperature Radiometer LST) and additional field observation data (HiWATER) are starting to be released, it is considered that a cross-comparison study with these data is necessary for globally continuous LST calculation. In addition, if the accuracy of the GK2A LST is improved, it will be able to contribute to the establishment of long-term climatological LST data retrieved from satellites, such as the CCI (Climate Change Initiative) LST project underway by the ESA (European Space Agency) [84].

**Author Contributions:** Conceptualization, M.-S.S.; Methodology, M.-S.S.; Software, Y.-Y.C.; Validation, Y.-Y.C.; Formal analysis, Y.-Y.C. and M.-S.S.; Investigation, Y.-Y.C.; Data curation, Y.-Y.C.; Writing—original draft preparation, Y.-Y.C.; Writing—review and editing, M.-S.S. and Y.-Y.C.; Visualization, Y.-Y.C.; Supervision, M.-S.S.; Project administration, M.-S.S.; Funding acquisition, M.-S.S.; All authors contributed extensively to the work presented in this paper. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported by "Development of Scene Analysis and Surface Algorithms" project, funded by ETRI, which is a subproject of the "Development of Geostationary Meteorological Satellite Ground Segment (NMSC-2019-01)" program funded by the NMSC (National Meteorological Satellite Center) of the KMA (Korea Meteorological Administration). This work was also funded by the Korea Meteorological Administration Research and Development Program under Grant KMI2018-06510.

**Acknowledgments:** We thank the Korean Meteorological Administration's National Meteorological Satellite Center for providing the GK2A/AMI dataset and cloud mask.

**Conflicts of Interest:** The authors declare no conflict of interest.
