A comprehensive evaluation of split-window and triple-window algorithms to estimate land surface temperature (LST) from Geostationary Operational Environmental Satellites (GOES) that were previously described by Sun and Pinker is presented. The evaluation of the split-window algorithm is done against ground observations and against independently developed algorithms. The triple-window algorithm is evaluated only for nighttime against ground observations and against the Sun and Pinker split-window (SP-SW) algorithm. The ground observations used are from the Atmospheric Radiation Measurement Program (ARM) Central Facility, Southern Great Plains site (April 1997–March 1998); from five Surface Radiation Budget Network (SURFRAD) stations (1996–2000); and from the Oklahoma Mesonet. The independent algorithms used for comparison include the National Oceanic and Atmospheric Administration/National Environmental Satellite, Data and Information Service operational method and the following split-window algorithms: that of Price, that of Prata and Platt, two versions of that of Ulivieri, that of Vidal, two versions of that of Sobrino, that of Coll and others, the generalized split-window algorithm as described by Becker and Li and by Wan and Dozier, and the Becker and Li algorithm with water vapor correction. The evaluation against the ARM and SURFRAD observations indicates that the LST retrievals from the SP-SW algorithm are in closer agreement with the ground observations than are the other algorithms tested. When evaluated against observations from the Oklahoma Mesonet, the triple-window algorithm is found to perform better than the split-window algorithm during nighttime.
Land surface temperature (LST) is an important parameter for formulating surface–atmosphere interactions (Dickinson 1996; Sobrino et al. 2003a,b, 2004). It controls the upward terrestrial radiation and surface–atmosphere sensible and latent heat fluxes, and it is an indicator of the energy balance at the earth’s surface (Sellers et al. 1988). LST retrievals are more challenging than sea surface temperature (SST) retrievals because of surface heterogeneity and the spatial variability of surface emissivity.
Attempts to derive LST from satellite observations have been ongoing for several decades. The focus was on polar-orbiting systems, such as the Advanced Very High Resolution Radiometer (AVHRR) and the Moderate Resolution Imaging Spectroradiometer (MODIS), because of global coverage and high spatial resolution. Much less attention has been given to geostationary satellites. Price (1984) applied the AVHRR SST split-window algorithm to agricultural land and found the retrieval accuracy to be about 3 K. Because the land surface emissivity is not equal to 1 and has spectral dependence (Becker 1987), Becker and Li (1990) developed a local split-window method (referred to here as BL90) with coefficients dependent on surface emissivity. Wan and Dozier (1996) optimized the split window algorithm developed by Becker and Li (1990) by computing coefficients that vary with satellite zenith angle (SZA), water vapor, and lower boundary air temperature [known as the “generalized split window (GSW) algorithm”, also referred to here as WD96]. Retrievals of LST from satellites are made in atmospheric windows that have relatively small water vapor absorption. To fully avoid such effects, several studies have added atmospheric correction to the GSW algorithms (Prata 1993; Sobrino et al. 1994; Becker and Li 1995; Francois and Ottle 1996; Coll and Caselles 1997). Such modifications improved the standard error in the temperature estimates to approximately ±1 and ±1.5 K, with no significant bias (Sun et al. 2002). Qin et al. (2002a,b) used the National Oceanic and Atmospheric Administration (NOAA) AVHRR observations to study the sand-dune region across the Israel–Egypt border. They found that the 3-K across-border difference during summer is due to differences in emissivities of biogenic crust and bare sand (being 0.972 and 0.954, respectively). Bhattacharya and Dadhwal (2003) also retrieved and evaluated land (soil–vegetation complex) surface temperature from NOAA/AVHRR thermal images over India. They found that in May the retrieved LST values are closer to the soil temperature. This was attributed to the lower fractional vegetation cover as characterized by the normalized difference vegetation index. To study the effects of different surface emissivities and the three-dimensional structure of topography on LST, Chen et al. (2004) indicate that using different emissivity values for soil and leaves will produce a 1-K difference in LST estimates. Ouaidrari et al. (2002) found that quadratic form is more accurate than the linear form, especially for high water vapor contents.
Oku and Ishikawa (2004) used images from the Geostationary Meteorological Satellite Visible/Infrared Spin-Scan Radiometer to estimate diurnal variations of LST distributions over the Tibetan Plateau. LST retrieval from a Geostationary Operational Environmental Satellite (GOES-8) was addressed by Faysash and Smith (1999), Sun and Pinker (2003), and Sun et al. (2004, 2006a,b).
Evaluation of LST algorithms is complicated because LST is highly variable in space and time and ground measurements are not compatible with the spatial resolution of satellites. Observations from high-resolution satellite platforms (e.g., Landsat) are desirable; however, they are of low temporal resolution (about once every 16 days) and, as such, are of interest as a “proof of concept.” There is a strong demand for information on LST for evaluation of numerical weather prediction models at resolutions of about 40 km. As such, there is a need to evaluate the quality of the LST information derivable from satellites at scales of interest to the modeling community. The disparity between the ground and satellite observations is unavoidable, and the best possible matchup between ground and satellite observations is unable to resolve this intrinsic problem. The objective of this study is to address this complex issue, namely, How well do satellite grid averages of LST compare to available observations of LST?
We present a comprehensive evaluation of a split-window LST algorithm developed by Sun and Pinker (2003) (SP-SW) against ground observations as available from several sources and from independent algorithms. Results are compared with the NOAA/National Environmental Satellite, Data, and Information Service (NESDIS) operational daytime algorithm over the United States (Wu et al. 1999), the same algorithm implemented with an independent cloud screening approach developed at the University of Maryland for daytime and nighttime (Li et al. 2007), the previously published split-window algorithms with emissivity correction (Price 1984; Prata and Platt 1991; Ulivieri and Camizzaro 1985; Ulivieri et al. 1992; Vidal 1991; Sobrino et al. 1994, 1993; Coll et al. 1994), the widely used generalized split-window algorithms (Becker and Li 1990; Wan and Dozier 1996), and the Becker and Li (1995) algorithm with water vapor correction. The evaluation method is presented in section 2. In section 3 the data used for evaluation are described. The algorithms used are introduced in section 4. Section 5 presents the evaluation results for the different algorithms. Errors of LST estimation are addressed in section 6. Section 7 presents conclusions.
The split-window algorithm of Sun and Pinker (2003) as well as the respective triple-widow algorithm will be evaluated against ground truth and against several existing algorithms. To make the algorithm intercomparison consistent and relevant for the satellite observations used, it is necessary to rederive relevant coefficients of the algorithms in a systematic manner using the same inputs as used for the SP algorithms. The coefficients from the previously published algorithms are rederived using the same forward simulations as performed in this study with the Moderate Resolution Atmospheric Transmission (MODTRAN) code for GOES-8.
In what follows, the simulations that were performed and the algorithms that were used are described. A table is presented with all of the algorithms and the newly derived coefficients.
a. Simulations performed
To derive regression coefficients that relate satellite-observed brightness temperature to LST, we perform forward simulations for GOES-8 using the MODTRAN4.0 radiative transfer model. To make the LST simulation results applicable on a global scale, the National Centers for Environmental Prediction (NCEP) reanalyzed global surface temperature dataset with matching atmospheric profiles with 17 vertical layers at a resolution of 2.5° is used for forward simulations. The solar zenith angle was calculated according to date, time of day, latitude, and longitude for each point in this dataset. The column water vapor (CWV) was integrated from the profiles; the air temperature was also obtained from the NCEP reanalysis. The range for CWV (precipitable water) is from 0.0 to 7.0 cm. Surface–air temperature differences range from −36.1 to 30.0 K. The Department of Geography, University of Maryland (UMD), generated a 1-km-resolution global land cover product (Hansen et al. 2000; http://glcf.umiacs.umd.edu/data/landcover/index.shtml). These products include the following 14 International Geosphere–Biosphere Programme (IGBP) classes: 1) water, 2) evergreen needle-leaf forest, 3) evergreen broadleaf forest, 4) deciduous needle-leaf forest, 5) deciduous broadleaf forest, 6) mixed forest, 7) woodland, 8) wooded grassland, 9) closed shrubland, 10) open shrubland, 11) grassland, 12) cropland, 13) bare ground, and 14) urban and built-up. In addition to these 14 IGBP classes, we add snow/ice to replace the bare ground at high latitudes, thus totaling 15 classes. The land classification was aggregated first to the resolution of the global simulations at 2.5° and subsequently to the analyzed data resolutions of 0.5° and 0.125°. In the aggregation process, the land cover type in each grid box was assigned on the basis of the dominant surface type and the surface emissivity was calculated according to a linear mixing with weighted sum of the land cover percentage times the emissivity of this surface type; it varied from 0.58 to 0.98 for the 3.9-μm channel (Fig. 1 of Sun and Pinker 2003).
From the simulations, the coefficients for the GSW algorithm (Wan and Dozier 1996) are derived as follows: 1) the atmospheric temperature profiles are aggregated according to the 11-μm brightness temperature, T11 ≤ 280 K or > 280 K; 2) the atmospheric CWV is aggregated into intervals of 0.5 cm; 3) day/night is separated using solar zenith angle θs (θs ≤ 87.5° is daytime and > 87.5° is nighttime; and 4) viewing angles are separated into intervals of 2°. The emissivity needed for the formulation of the split-window algorithm is assumed according to surface types following the approach proposed by Snyder et al. (1998).
b. Algorithms used in evaluation
1) Sun and Pinker (2003) split-window algorithm
To remove atmospheric effects for high viewing angles when such effects are amplified, McClain et al. (1985) added a satellite viewing angle correction term (secθ − 1) to the SST split-window algorithm equation. This correction is also used in the following formulation:
where i is the index of surface types (15 classes described in section 3) and θ is the satellite viewing angle, T11 and T12 are the brightness temperatures at 10.8- and 12.0-μm channels, a0–a4 are the coefficients, and Ts is the derived skin temperature. Table 1 lists the coefficients for some surface types derived from the MODTRAN simulations. Because of insufficient samples for certain surface types (e.g., urban), coefficients are not provided.
2) Sun and Pinker (2003) triple-window algorithm (nighttime only)
The following was proposed:
where T and ɛ are the brightness temperature and emissivity, subscripts 3.9, 11.0, and 12.0 refer to GOES-8 channels at 3.9, 10.8, and 12.0 μm, and other variables are the same as in Eq. (1). Because of solar contamination in the middle-infrared channel (3.9 μm), the Sun and Pinker triple-window (SP-TW) algorithm will only be applied and evaluated for nighttime.
3) The NOAA/NESDIS operational algorithm
Original version using channel-1 cloud detection
An algorithm developed by Wu et al. (1999) has been used at NOAA/NESDIS since December of 1998 to estimate LST over the United States [referred to hereinafter as the NOAA daytime algorithm (NOAA-O)]. Because cloud detection is based on channel 1, LST can be estimated only during daytime. LST is given as
where dT = T11 − T12 and A0–A3 are the coefficients.
Original version using multichannel cloud detection
To allow the implementation of the NOAA/NESDIS operational algorithm during both daytime and nighttime, we use a four-channel coupled cloud and snow detection algorithm (CCSDA) for both daytime and nighttime (Li et al. 2007) and couple it with the original NOAA/NESDIS algorithm to rederive LST (NOAA-R algorithm). The CCSDA cloud detection method also has the capability to detect clouds over snow (Pinker et al. 2007).
4) Split-window algorithms with emissivity correction
Wan and Dozier (1996) optimized the local split-window algorithm developed by Becker and Li (1990) by computing coefficients that vary with SZA, water vapor, and lower boundary air temperature (known as the generalized split window), presented in Table 2 as No. 9. In this formulation, atmospheric attenuation with optical path is accounted for. In accord with this, the coefficients vary with cold/warm conditions, day/night, viewing angles, and column water vapor.
The above algorithms are summarized in Table 2 where the following applies:
3. Data used
a. Satellite observations
Satellite information needed for implementing the various algorithms includes cloud cover fraction and brightness temperatures for clear sky radiances in the 11.0- and 12.0-μm channels from GOES-8. These are produced as a by-product from the operational NOAA/NESDIS GOES surface shortwave radiative flux product (Pinker et al. 2003) developed in support of the Global Energy and Water Cycle Experiment (GEWEX) Continental Scale International Project (GCIP) and the GEWEX Americas Prediction Project as archived at the University of Maryland (http://www.atmos.umd.edu/~srb/gcip/). This information is at hourly time scales over the domain area of 125°–70°W and 25°–50.5°N, at 0.5° and at 0.125° spatial resolutions. To avoid cloud effects in LST retrievals, 10% cloud thresholds are used (if cloud amount is less than 10%, LST will be retrieved). Because cloud detection criteria are very strict, the compromise of a 10% threshold was necessary to ensure that a sufficiently sizeable sample of observations is available for the retrievals (experiments were performed with different percentages before the 10% value was selected).
b. Ground observations
Ground observations of surface skin temperature needed for the evaluation of satellite-based estimates are not readily available. The observations differ from each other, and a certain adjustment is needed to make them compatible with each other. Some of the observations are based on infrared radiometers while others measure upwelling longwave radiative fluxes. Some of the radiometric observations used have been converted to skin temperatures by data providers while others are converted in this study using information on surface emissivity.
1) ARM/GCIP NESOB-97 30-min skin temperature composite
The Atmospheric Radiation Measurement Program (ARM)/GCIP Near-Surface Observation Dataset-1997 (NESOB-97) came from the ARM–Cloud and Radiation Test Bed site (34°–39°N and 94.5°–100.5°W) during the period 1 April 1997–31 March 1998. The surface skin temperature used in this paper is observed at 30-min intervals at the Central Facility (36.6°N, 97.48°W) of the Southern Great Plains (SGP) site (Table 3). The instrument used to observe the skin temperature is the multifilter radiometer (MFR). It is a sensing head of the multifilter shadow band radiometer that is mounted to look downward to measure upwelling spectral irradiances. The filters of the MFRs can drift significantly and induce errors of about ±5%. The upwelling irradiance is converted to skin temperature using a NOAA/Atmospheric Turbulence and Diffusion Division (ATDD) algorithm (http://www.arm.gov/instruments/instrument.php?id=mfrsr).
2) The Surface Radiation Budget Network
NOAA established the Surface Radiation Budget Network (SURFRAD) in 1993 (Hicks et al. 1996). Its primary mission is to support climate research with accurate, continuous, long-term measurements of the surface radiation budget over the United States. The SURFRAD stations include Bondville, Illinois, Table Mountain, Colorado, Fort Peck, Montana, and Goodwin Creek, Mississippi. Desert Rock, Nevada, and The Pennsylvania State University (Penn State), Pennsylvania, were installed in 1998, and Sioux Falls, South Dakota, was installed in 2003 (Fig. 1). In this paper, data are used from the Illinois, Montana, Mississippi, Nevada, and Pennsylvania sites only (Table 3). The primary SURFRAD measurements are upwelling and downwelling solar and infrared radiation; direct and diffuse solar radiation; photosynthetically active radiation; ultraviolet radiation (UV-B, 290–320 nm); spectral solar radiation; and meteorological parameters. The data are available in daily files that are distributed in near–real time (http://www.srrb.noaa.gov). Observational data from the SURFRAD network are used for evaluating satellite-based estimates of surface radiation and for validating hydrologic, weather prediction, and climate models (http://www.srrb.noaa.gov/surfrad/surfpage0.html). The upwelling and downwelling longwave radiative fluxes are measured with the precision infrared radiometer, which is sensitive in the spectral range from 3000 to 50 000 nm. The SURFRAD observed upwelling (F↑) and downwelling (F↓) radiative fluxes are converted to temperature as follows:
where ɛIR is the surface broadband emissivity assigned by surface type, σ is the Stefan–Boltzmann constant and is equal to 5.669 × 10−8 J 2m−2 2s−1 2°−4, and
The evaluation of the various algorithms against ground observation is done at monthly time scale over the diurnal cycle.
3) The Oklahoma Mesonet
The Oklahoma Mesonet is an automated network of over 110 remote, meteorological stations across Oklahoma (Brock et al. 1995; McPherson et al. 2007) (http://www.mesonet.org). During the period of this study, 10 sites measured surface energy fluxes using the eddy correlation method. These sites are referred to as the Oklahoma Atmospheric Surface-Layer Instrumentation System (OASIS) Super Sites. The sensors at OASIS Super Sites allow estimating net radiation (shortwave and longwave components) and sensible, latent, and ground heat fluxes (Basara and Crawford 2002). Surface skin temperature is measured using an infrared temperature sensor (IRT) manufactured by Apogee Instruments, Inc., mounted at 2 m (Fiebrich et al. 2003) and LST is converted from the upwelling longwave radiation with unit emissivity while surface reflection of downwelling longwave radiation is ignored (Fiebrich et al. 2003). Sun et al. (2004) discussed the effects of these two factors and found that the total effect may be a slight underestimation of the skin temperature. Surface skin temperature is measured at an additional 79 sites designated as OASIS Standard Sites.
Observations such as skin temperature and longwave radiation are referred to as research measurements. The research measurements do not have the routine quality assurance applied to other parameters (McPherson et al. 2007). Even so, for this study, a combination of automated and manual tests was applied to the data provided before use to ensure the skin temperature observations are of research quality.
a. Model evaluation resulting from MODTRAN simulations
To select algorithms for comparison with ground truth, simulation results for several independent algorithms were first analyzed (Figs. 2 –5). From Figs. 3 and 4 it is evident that the Ulivieri2 algorithm with emissivity difference term shows improvement when compared with the Ulivieri1 algorithm that uses only an average emissivity correction term, especially for small viewing angles (<5°). The Vidal, Coll, and Ulivieri2 algorithms show similar results; the Sobrino2 algorithm with a nonlinear term of brightness temperature difference between the two split-window channels shows improved bias error for small satellite viewing angles (<6°) while the Price algorithm performance is below the others. As evident from Fig. 4, the Becker and Li (1995) (referred to as BL95) algorithm with water vapor correction shows further improvement as compared with the Sobrino2 algorithm for satellite viewing angles of less than 5°, and the GSW algorithm that considers viewing angle effects shows improvement at large viewing angles (>5°) as compared with the Ulivieri1, Ulivieri2, Vidal, Coll, Price, and Sobrino2 algorithms. The SP-SW algorithm with viewing angle correction and nonlinear terms shows smaller errors at both small and large viewing angles as compared with the other algorithms. Because the Ulivieri2, Vidal, and Coll algorithms are similar in structure and give similar results, the later Coll algorithm was selected as a representative of this group. Included in the evaluation against ground truth are the Price, Ulivieri1, Coll, Sobrino2, GSW, BL95, and SP-SW algorithms.
b. Evaluation of algorithms
The retrieved and ground-measured LST data are stratified as daytime and nighttime according to the solar zenith angle (≤87.5° is daytime, and >87.5° is nighttime). The retrieved LST in a grid closest to the observing site is used to match with ground observations. Then the data are aggregated for each season and the root-mean-square (RMS) error, bias, and standard deviation (STD) are calculated.
1) Evaluations at the ARM SGP site
Detailed error statistics for different algorithms at the ARM SGP site are presented in Table 4. When compared with other algorithms, the LST retrieval errors from the SP-SW algorithm are the smallest, followed by GSW, Coll, Ulivieri1, Price, NOAA-R, BL95, and Sobrino2 algorithms. The algorithms selected for further evaluation are GSW, NOAA-R, BL95, and SP-SW.
The cloud-screening algorithm applied to the satellite data allows the detection of snow. Under snow conditions, the SP-SW algorithm as appropriate for snow conditions used in the simulations is utilized; for the other algorithms, snow emissivity (ɛ11 = 0.996; ɛ12 = 0.962) is assigned. Figure 6 presents evaluation results for the daytime LST retrievals from the SP-SW algorithm against ARM SGP observations for different seasons of 1997. This algorithm has low bias and RMS errors in winter, summer, and autumn, with the largest errors in the spring. Figure 7 compares the RMS error differences between retrievals aggregated at 1/2° and 1/8° resolution during nighttime as evaluated against ARM SGP observations in 1997. For all four seasons, the results from the SP-SW algorithm are in best agreement with observations. Figure 8 compares the bias and residual standard errors evaluated at Bondville during daytime in winter for 1996–2000. The NOAA, GSW, and SP-SW algorithms do not show large differences in LST errors at different resolutions, and the BL95 results show larger errors at higher resolution.
Triple-window algorithm versus split-window algorithm
The Sun and Pinker triple-window (SP-TW) algorithm is applicable only during nighttime. It is of interest to compare it with the SP-SW algorithm for the same time period. As evident from Table 5, during nighttime, LST retrieval errors from the SP-TW are lower than those from the SP-SW algorithm.
2) Evaluations at the SURFRAD sites
Figure 9 shows the time series of bias and STD errors for daytime LST retrievals resulting from the different algorithms for the period from 1996 to 2000. The bias error shows overestimation of LST from the BL95 algorithm. The NOAA algorithm shows larger underestimation of LST when compared with the SP-SW algorithm (Fig. 9a). Among the tested algorithms, the SP-SW algorithm is in best agreement with ground observations, giving the lowest STD errors of about 1.0 K from 1996 to 2000 (Fig. 9b). During nighttime, the NOAA algorithm underestimates LST most (Fig. 9a). The STD errors are smallest from the SP-SW algorithm and largest from the BL95 algorithm (Fig. 9b). All the algorithms show larger underestimates of bias in spring and autumn. This may be because, as shown in Fig. 6, LST is in the range of 280–300 K during spring and autumn, when the water vapor is the highest (Sun and Pinker 2003).
The seasonal daytime and nighttime bias of LST retrievals evaluated at Fort Peck from 1996 to 2000 is shown in Fig. 10. The LST retrieval errors from the SP-SW algorithm are the lowest, and the average bias during the 5-yr period is about 1 K; the underestimate of LST from the NOAA algorithm can be as large as −8 K.
3) Evaluations at the Oklahoma Mesonet sites
Results from the evaluation of the various LST retrievals against the Oklahoma Mesonet measurements are given in Table 6. The NOAA/NESDIS algorithm using the multichannel cloud screening gives the lowest errors when compared with other algorithms, followed by the SP-SW algorithm, generalized split-window algorithm, and BL95 algorithm. The bias error (accuracy) of the SP-SW algorithm is positive, and the bias of the NOAA algorithm implemented with the CCSDA cloud screening is negative. The apparent “better” performance of the NOAA algorithm at the mesonet sites seems to be related to the fact that the measured skin temperature is converted from the energy detected by IRT using the Stefan–Boltzmann law and assumed surface emissivity of 1.0 (Fiebrich et al. 2003). The actual emissivity at the Mesonet sites should be lower than 1.0 and therefore, the IRT-measured skin temperature is lower than the actual temperature of the ground.
4) Evaluations at MODIS test sites
Limited observations are available from homogeneous sites where data were collected in support of the validation of MODIS LST products (Wan et al. 2002, 2004). To be specific, data from the following sites are available (Wan et al. 2002): two locations at Mono Lake, California; three locations at Bridgeport, California; and two rice fields in California.
The data are matched to the GOES overpass for seven cases, and errors are presented as a difference between the ground observation and the GOES estimated value. A similar pattern was followed for the SP-SW algorithm as shown in Table 7. The LST from the SP-SW algorithm show very close agreement with the ground observations.
5. Discussion of error sources
a. Observational error
The filters of the MFRs at the ARM SGP sites drift during windy weather conditions, resulting in errors in the observed upwelling irradiance that can reach ±5%. The instruments are installed at 10 and 25 m, and the skin temperature is obtained from the upwelling irradiance by applying the NOAA/ATDD algorithm. The skin temperature is converted from measured upwelling irradiance with an assumed emissivity value, which might result in some errors. Errors in the LST retrievals include scattering patterns induced by heterogeneous surfaces (Weng et al. 2004), uniform surfaces having smaller scattering effects than mountain areas. The accuracy of the IRT used by the Oklahoma Mesonet for skin temperature measurement is approximately 0.2 K from 288 to 305 K and 0.3 K from 278 to 318 K (Bugbee et al. 1998).
b. Land cover classification
DeFries et al. (1995) indicate that land cover classification accuracy of 70%–90% can be achieved, and the correct typing probability for 1 km can be expected to be 80% (Loveland and Belward 1997). Errors from land cover classification will produce errors in the coefficients relating land cover types to emissivity variability (Sun and Pinker 2003).
c. Surface emissivity
Prata et al. (1995) showed that 1% emissivity uncertainty can affect LST values for each pixel by about 0.7°C. Yu et al. (2008) tested the sensitivity of 10 different split-window algorithms to uncertainties in surface emissivity and found that the split-window algorithms are very sensitive to emissivity values. Sun and Pinker (2003) performed model (MODTRAN) simulations with input errors introduced directly to emissivity or to land cover. The results indicated that the “error” in the simulated brightness temperature increases with the increase in emissivity error, whereas the simulated brightness temperature error in the split-window channels due to land cover error was very small (less than 0.2 K), even with 50% of pixels being misclassified. The SP-SW algorithm proposes to use land cover information with land cover–dependent regression coefficients and shows improvements to the LST algorithm if emissivities are used. The surface emissivity effect is considered implicitly through land cover types.
d. Atmospheric correction error
In the atmospheric window channels, water vapor is the major absorber of upwelling radiation, especially near the surface. Most of the water vapor in the atmosphere is for conditions within the temperature range of 280–305 K; it can vary from 0.25 to 7 cm (Sun and Pinker 2003). This may be the reason why negative bias errors (underestimates due to water vapor absorption) are usually higher at the LST range of 280–305 K as compared with those below 280 K.
e. Reliability of the evaluation process
Ground observations of LST are point measurements; satellites measure at pixel level. Land surface temperature is not homogeneous within a pixel area. When we match the ground observations to the satellite retrievals, we use satellite LST at the closest grid point at the same time as the ground measurements. The spatial difference in the satellite–ground match process may introduce errors.
f. Cloud contamination
To ensure that a sufficiently sizeable sample of observations is available for the retrievals, a 10% cloud threshold is used. This may include some cloud contamination and cause underestimation of LST or a negative bias. On the other hand, ground measurements are taken under all cloud conditions and may include cloud effects. All of these error sources are listed in sections 5a to 5f and may cause LST retrieval errors.
6. Summary and discussion
Because of the difficulty in obtaining ground observations of LST, the available observations that are directly and indirectly linked to LST are used in the evaluation process of satellite retrievals. Evaluation against the ARM and SURFRAD observations shows that the LST retrievals from the SP-SW algorithms are in better agreement with the ground observations than are the other split-window-type algorithms tested here. Evaluation against the mesonet observations shows that the NOAA algorithm is in closest agreement with the observations, followed by the SP-SW algorithm, GSW algorithm, and the BL95 algorithm. The apparent “better” agreement of the NOAA algorithms with the Mesonet site observations is due to the fact that the “observed” LST used was converted from the measured energy units using a surface emissivity of 1.0 (Fiebrich et al. 2003) as assumed in the NOAA/NESDIS algorithm. Overall, the SP-SW algorithm is in better agreement with ground observations than are the other algorithms that were tested, and the triple-window (SP-TW) algorithm shows improvement over the split-window (SP-SW) algorithm for nighttime LST retrieval.
Support from Grants NA86GP0202 and NAG59916 from the NOAA Office of Global Programs is appreciated. Data were obtained from the Atmospheric Radiation Measurement Program sponsored by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Environmental Sciences Division. The installation of the Oklahoma Mesonet’s skin temperature sensors was made possible, in part, by an NSF MRI Grant (ATM-9724594). Continued funding for maintenance of the network is provided by the taxpayers of Oklahoma through state taxes. In addition, support from the NOAA Office of Global Programs (NOAA Grant NA17RJ1227) was instrumental in the development of research-quality datasets. We appreciate the reviewers for their helpful comments to improve this manuscript.
Corresponding author address: Rachel T. Pinker, Department of Atmospheric and Oceanic Sciences, University of Maryland, College Park, College Park, MD 20742-0000. Email: firstname.lastname@example.org