• Oklahoma MESONET

The Oklahoma MESONET is an automated network of over 110 remote, meteorological stations across Oklahoma (http://www.mesonet.org). The surface types are predominantly grassland/wooded, grassland/cropland. In 1999, infrared temperature (IRT) sensors (Apogee Instruments, Inc.) were installed at 89 of the MESONET sites. A combination of automated and manual tests was applied using simultaneous soil and atmospheric measurements to inter-compare observations and ensure that the skin temperature observations are of research quality [44]. The measurements collected by the MESONET provided a unique opportunity to inter-compare observations. Fiebrich et al. [45] provide an evaluation of 5-min-resolution field measurements collected using the sensors. This sensor was chosen for use because it is water resistant and was designed for continuous outdoor use. Sensor accuracy is approximately ± 0.2 ◦C from 15 ◦C to 35 ◦C and 0.3 ◦C from 35 ◦C to 45 ◦C. The sensor is installed at a height of 1.5 m and has a field of view of a diameter circle of 0.5 m. The energy detected by the sensor is converted to a temperature using the Stefan–Boltzmann law and an assumed surface emissivity of 1.0. Slight underestimation is caused because the true emissivity of the land surface is less than 1.0. In addition, slight overestimation is caused by reflected longwave radiation from the target [46]. While surface reflection of downwelling longwave radiation is ignored, Sun et al. [32] discussed the effects of these two factors and found that the total effect may be a slight underestimation of the skin temperature. Generally, estimated impact of uncertainty in relevant parameters on in situ LST are as follows: radiometric calibration uncertainty of ±0.2 to 0.5 K can impact LST as much as 0.2 K; emissivity uncertainty of ±1% can impact LST by as much as 0.3 K; downwelling atmospheric radiance uncertainty of ±10% can impact LST by as much as 0.1 K [47].

#### *2.7. Cloud Detection*

For cloud masking, various ancillary data are needed such as surface type, snow-free channel-1 radiance, and a pixel position information, all in the same dimension and location as the satellite images. The land cover data used in this study for cloud screening implementation, are generated at 1-km resolution [48]. This product includes 14 International Geosphere–Biosphere Programme (IGBP) classes and the underlying surface types are aggregated according to the IGBP classification.

A Coupled Cloud and Snow Detection Algorithm (CCSDA) that was developed initially for use with GOES-8 satellite is adjusted as appropriate for each GOES satellite is used. The algorithm is described in [49,50]. Variants of the approach were tested and evaluated in several publications [51]. In the case of the GOES-8 imager four channels were used to detect clouds, snow, and to perform background analysis for each hour of the diurnal cycle. Beginning with GOES-12, Channel 5 is no longer available (Table 1) [52].

The CCSDA algorithm is capable of producing its own snow analysis using an algorithm that applies three tests using three GOES channels. Alternatively, there is a switch to allow the use of a snow analysis from a different source. The advantage of using the snow analysis generated by the CCSDA algorithm is that it is updated hourly, which provides a more accurate analysis of the expected background when applying the cloud tests. If a daily snow analysis is used, the snow conditions cannot change for each hour of the cloud analysis, and this may introduce error.

In Table 2 we present a description of the cloud screening tests used for GOES-8, along with an explanation of how the tests are assembled to determine a clear probability.


**Table 2.** Cloud Screening Tests for GOES-8.

Note: Reflectance Gross Contrast Test (RGCT); Channel-2 Albedo Test (C2AT); Thermal Gross Cloud Test (TGCT); Three Minus Five Test (TMFT); Four minus Five Test (FMFT); Uniform Low Stratus Test (ULST); Cirrus IR Test (CIRT).

The ultimate clear probability (*P*) can be assembled in various ways on the basis of individual test results. In this method:

$$P = \sqrt[n]{\prod\_{1}^{n} P\_{i}} \tag{7}$$

*Pi* is the clear probability from each individual cloud screening test and n is the total number of cloud screening tests. This assembly method guarantees that the target pixel is cloudy (*P* = 0) if any individual test identifies it as cloudy (*P<sup>i</sup>* = 0). Otherwise, it compiles the confidence levels from all of the individual tests to obtain an overall clear probability.

Referring to Table 2, there was only one cloud screening test that required Channel 5, namely, the FMFT test. In the cloud screening algorithm for GOES-12 and beyond, the FMFT test will not be used, and the test assembly method described for GOES-8 is implemented with one less cloud test.

For each cloud test, threshold levels are used to differentiate between clear and cloudy pixels. For each new satellite, it is necessary to test the thresholds and modify as needed. Here, the cloud mask method applies two spatial tests and one threshold test on an 11–3.7 µm difference image. This fourth test compares the temperature from the 11 µm channel to a 20-day clear-sky composite of 11 µm temperatures, and labels the pixel as cloudy if the difference is greater than the threshold. The pixel level data were gridded to 0.05◦ and compared to the Pathfinder Atmospheres—Extended (PA TMOS-X) product [34]. Agreement above 95% for various times of the day was found. The only region which showed slight disagreement between the two independent cloud masks is over areas of complex terrain in the Western US, but even over these regions the two cloud masks agreed over 85% of the time.

The major difference between the day time and night time algorithms (Table 2) is that there are no Reflectance Gross Contrast Test (RGCT, when visible channel is missing) and Channel-2 Albedo Test (C2AT) during the night time. This will mostly affect the detection of reflective clouds. For GOES-12, the FMFT test could not be used. We have tested the cloud screening algorithm for use at nighttime. We applied the nighttime algorithm on daytime data and compared to results when the full daytime algorithm is used. Evaluation of LST estimates in each case is presented in Figure 1. As this figure shows, the daytime algorithm provides better agreement with ground observations than the nighttime one yet, the differences are small as illustrated in Figure 1.

ିଵ

− − − −

> = ඥ∏ ଵ

**Figure 1.** Evaluation of LST for 2004 against a SURFRAD/BSRN station at Desert Rock, NV (DRA) for: Left: daytime; Right: nighttime. (<sup>௦</sup> ) <sup>௦</sup> <sup>↑</sup> ↓

1

1

#### **3. LST Retrieval Algorithm Development for GOES Satellites**

Satellite observed radiance *R* ↑ *o* , can be expressed as

$$R\_o^\uparrow = \varepsilon B(T\_s) \mathcal{X} + R\_a^\uparrow + (1 - \varepsilon) R\_a^\downarrow X \tag{8}$$

where ǫ is surface emissivity, *B*(*Ts*) is blackbody emission at surface temperature *T<sup>s</sup>* , *X* denotes the atmospheric transmittance, *R* ↑ *<sup>a</sup>* and *R* ↓ *<sup>a</sup>* are atmospheric emission to space and surface, respectively. With known surface emissivity and simulated atmospheric emission and transmittance, the surface temperature can be retrieved 

$$T\_s = B^{-1} \left[ \frac{1}{\epsilon} \left( \frac{1}{X} \left( R\_o^\uparrow - R\_a^\uparrow \right) - (1 - \epsilon) R\_a^\downarrow \right) \right] \tag{9}$$

where *B* <sup>−</sup><sup>1</sup> denote the inverse of Planck function for GOES-12 channel 4.

Here, the approach is based on the Radiative Transfer for TOVS (RTTOV) model v11.2 [53–55] adjusted for the GEO characteristics and driven with MERRA-2 reanalysis fields. The CAMEL data are also implemented in the method. The advantage of this approach is that it is consistent with the retrieval approach used at JPL to generate the MOD21 product [56]. The processing sequence is described in Figure 2.

**Figure 2.** Flow-chart describing the derivation of LST from GOES observations.

Data processing sequence starts from raw digital counts from GOES satellites. Calibration is applied to all channels. Channel 4 (10.2 to 11.2 um) was used for LST retrieval. All channels except channel 3 (6.7 um) were used in the cloud detection algorithm. After cloud screening, GOES observations were resampled to a uniform grid of 0.05◦ resolution. The atmospheric radiation and transmittance were simulated with the RTTOV model using MERRA-2 fields as input. The MERRA-2 fields were first temporally interpolated to satellite observation time and then collocated to satellite locations. RTTOV calculated upwelling, downwelling radiances and atmospheric transmittance combined with CAMEL surface emissivity to retrieve the LST according to Equation (11).

#### **4. Evaluation of GOES-E Based LST Estimates**

We will present results of evaluation for UMD LST retrievals against MOD11 products, the BSRN/SURFRAD network over the USA, the ARM/SGP C1 site over the Southern Great Plains, and the MESONET network over Oklahoma. The issue of evaluation of satellite products of LST against ground measurements is complex, primarily, due to scale issues and known large spatial variability of LST. A comprehensive discussion on all aspects of validation issues are described by Guillevic et al. [43] and Göttsche et al. [57].

#### *4.1. Scale Issues Related to Satellite and Ground Observations*

The ground observations are point observations while the satellite LST product is at pixel level gridded to 0.05◦ . To assess the homogeneity of each site, we use the ASTER Global Emissivity Dataset at 1-km V003 (DOI: 10.5067/Community/ASTER\_GED/AG1km.003) available for the period of 2000–2008. It is based on observations from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Emissivity Dataset (GED) land surface temperature and emissivity data products using the ASTER Temperature Emissivity Separation (TES) algorithm with a Water Vapor Scaling (WVS) atmospheric correction method with MODIS MOD07 atmospheric profiles and the MODTRAN 5.2 radiative transfer model. The spatial distribution of the emissivity values is illustrated in Figure 3a and their frequency distribution is shown in Figure 3b. As shown, except for the DRA site, the 0.05◦ boxes show a high degree of homogeneity at the 1-km scale. As seen from Figure 3b, the emissivity values range between 0.965–0.980 with two distinct peaks of 0.965 and 0.975 with some lower values (0.948) at the DRA site. As also shown in the study of Hulley and Hook [58], who compared ASTER emissivity band 11 (8.6 µm) at 90 m spatial resolution to the same at 1 km, the agreement was very good. The spatial matching of ground and satellite observations is done by taking the weighted average of the pixels that fall in the cell box (0.05◦ × 0.05◦ ) around the target location of the station. The time matching is done by taking the averages of ±15 min around the start scanning time of GOES12 (this interval is selected based on the duration of the satellite scan).

**Figure 3.** (**a**) Spatial characterization of the sites used in evaluation of the LST products in terms of emissivity as obtained at 1 km spatial resolution using the ASTER Global Emissivity Dataset 1 km V003. (**b**) Frequency distribution of the emissivity values over sites used in evaluation as illustrated in Figure 3a.
