**1. Introduction**

Land surface temperature (LST) can be interpreted as the skin temperature of Earth's land and is derived using the upward longwave radiation measured by a satellite sensor [1]. The LST retrieved from satellites represents the surface temperature of ground pixels that are not contaminated by clouds and is affected by many factors, such as land use/cover, vegetation, soil moisture, snow, etc. [2,3]. Land information is used as input and verification data for numerical/climate models. These are important factors to understand the Earth's system and land–atmosphere interactions [4–6]. LST has various research applications, such as studying land use/cover changes [7–9], drought monitoring [10–12], energy balance estimation [13–15], analyzing urban heat islands [16–18], studying evapotranspiration [19] and so on [20–22]. As various applied studies using LST have been conducted, the demand for long-term LST data with high spatial and temporal resolutions as well as good accuracy is increasing [23,24].

LST is one of the climate elements with very different spatial and temporal variability depending on the ground condition (land cover, vegetation condition, soil moisture, etc.) of the observation point. Due to these characteristics, more detailed spatial and temporal field observations are necessary, but this is not practical in economic and technical terms. So, at present, special observation is only performed in some areas where the ground condition is relatively homogeneous. The LST observed

in a field is a value that can represent only a few tens of meters of the radius of the observation point, therefore, there is a limitation of spatial representation. To overcome this limitation, studies on retrieving LST from satellites with spatial resolutions from hundreds of meters to several kilometers have been conducted for more than half a century [25–27]. LSTs retrieved from polar orbit satellites have a relatively high spatial resolution but long revisit cycles; they help to understand a momentary phenomenon in detail for a relatively narrow space, such as urban heat islands [28–32]. Moreover, LSTs retrieved from geostationary satellites have a relatively coarse spatial resolution compared to polar orbit satellites, but they can continuously retrieve the same observation area in a short observation period. Owing to these advantages, LSTs retrieved from geostationary satellites can be used to analyze phenomena occurring in large areas over a long period of time and can also fill in the blank areas of field observation [33–37].

Many studies have been conducted and different methodologies have been developed to retrieve high-quality LST data from satellites. For LST retrieval from the radiation observed by thermal infrared sensors, cloud detection is essential, and it is necessary to know the surface condition and atmospheric effects [38–40]. LST retrieval methods, assuming that the land surface emissivity (LSE) is known a priori, can be roughly divided into three groups: single-channel methods [31,41], multi-channel methods [26,27,33], and multi-angle methods [42,43]. There are also several methods for retrieving LST and LSE at the same time when the LSE is unknown: classification-based emissivity methods [44,45], normalized difference vegetation index (NDVI)-based emissivity methods [46,47], day/night temperature-independent spectral indices-based methods [38,48], two-temperature methods [49,50], and temperature emissivity separation methods [28,51,52]. Among these various LST retrieval methods, a commonly used one for retrieving LST from a geostationary satellite is the split-window (SW) method using two adjacent thermal infrared channels with different absorption capabilities for water vapor and other substances [53–55]. The split-window method is relatively simple, efficient, and convenient to apply to most sensors, but it is assumed that the LSE is accurately known in such cases. In addition, the split-window method has various types of algorithms and thus has different performance characteristics according to the type of algorithm. Another characteristic of the split-window method is that the accuracy is degraded in a specific region where the total column water vapor is high or where the satellite viewing zenith angle (VZA) is large [23].

With improvements in sensor performance on satellites, next-generation geostationary meteorological satellites (Himawari-8, Geostationary Operational Environmental Satellite (GOES)-16/17, GEO-KOMPSAT-2A (GK2A)) have started to conduct their observations with high performances in time and space resolution [56–58]. In addition, the Meteosat Third Generation (MTG) series will be launched in 2021 as the next system after the Meteosat Second Generation (MSG) series of geostationary satellites [59]. With improved sensor performance, the space-time resolution of level 2 products has also been improved. The LST product was designated as the official level 2 product in GOES-16 and GK2A as in previous satellites (GOES-13/15, Communication, Ocean and Meteorological Satellite (COMS)). In addition, the European Space Agency (ESA) has adopted LST as an official product on MTG satellites following MSG satellites, and LST retrieval research is being conducted to prepare for it. In the case of Himawari-8, a study was conducted to retrieve (LST) for research purposes [55,60]. In the Asia-Oceania region, the COMS was retired on March 31, 2020. The COMS' follow-up satellite, GK2A, has been officially operating since 25 July 2019, and replaced COMS' services [61]. Therefore, it is necessary to develop an LST retrieval algorithm using GK2A/AMI (Advanced Meteorological Imager) data.

In this study, we developed an operational LST retrieval algorithm for the GK2A observation area using GK2A/AMI data from next-generation geostationary satellites in Korea. The contents of this paper are as follows. The properties of the data are described in Section 2.1, and the process of developing the simulation dataset using the radiative transfer model (RTM) and LST retrieval process and methodology are described in Section 2.2. The RTM simulation results and the GK2A LST retrieval results are presented in Section 3. In addition, the accuracy, problems, and future improvements of the GK2A LST retrieval algorithm are presented in Section 4.
