Land surface temperature (LST) represents one of the key climate variables retrievable from space-based remote sensing platforms, offering insight into a range of environmental processes and linking multiple disciplines across the natural and physical sciences. Apart from being a fundamental variable in quantifying elements of the surface energy budget [1
], LST has been used to study ocean–atmosphere interactions [2
], to track global warming and climate change impacts [3
], as well as being widely used in studies of vegetation monitoring [5
], drought persistence [6
] and urban climate assessments [7
]. LST also plays a critical role in linking the water and energy cycles through its relationship with surface heat fluxes [8
]. Indeed, thermal infrared (TIR) observations represent a fundamental element in efforts to map the spatial distribution of evaporation [9
] as well as in efforts to constrain land surface model simulations [10
]. Given the role that LST plays across broad aspects of earth and environmental sciences, determining its spatial and temporal variability is of considerable interest [11
]. However, accurately determining its absolute value, in addition to describing its spatial and temporal development, is challenging given that LST varies considerably throughout the diurnal cycle as a function of the surface radiative balance, as well as expressing a broad range of spatial and temporal variations due to changing land surface and atmospheric conditions [12
Thermal infrared observations from space offer the only plausible option for mapping LST over large areas, since in-situ measurements are unable to provide the density of coverage for adequate spatiotemporal information. A range of space-based thermal infrared platforms are available to retrieve the LST, including sensors onboard polar orbiting satellites such as the Moderate Resolution Imaging Spectroradiometer (MODIS) [14
], the Visible Infrared Imaging Radiometer Suite (VIIRS), Gaofen 4, the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Sentinel-3, Landsat [15
], as well as several geostationary platforms. The Landsat series of satellites offers one of the most complete and informative long-term records of surface temperature, providing a unique high spatial resolution (i.e., 60–120 m) thermal sequence that began with the launch of Landsat 4 in 1982 [16
]. Although most of this extensive record only provides radiance data from a single infrared channel, the most recent addition to the Landsat fleet provides two longwave bands: Band 10 (10.60–11.19 µm) and Band 11 (11.50–12.51 µm), at a 100 m ground resolution and a 16 days temporal resolution.
Direct estimation of LST from thermal infrared measurements is a challenging task since the radiances measured by the satellites are influenced by both surface and atmospheric effects [17
]. By using two TIR bands, atmospheric correction of the thermal data can be achieved using well developed split-window approaches [18
] or temperature–emissivity separation techniques [19
]. The split-window method, proposed by McMillin [20
], was based on differential absorption across adjacent spectral windows and allowed for an estimate of sea surface temperature to be obtained by a linear combination of spectrally adjacent brightness temperatures. An adaptation of the McMillin [20
] approach for application to land surfaces was presented by Becker and Li [18
]. More recently, both Jimenez-Munoz et al. [21
] and Rozenstein et al. [22
] proposed split-window algorithms for the Landsat 8 thermal infrared sensor (TIRS). Using simulated thermal infrared data, the studies estimated mean errors of less than 1.5 K and 0.93 K, respectively. These methods were validated using simulated data from the Thermodynamic Initial Guess Retrieval and Standard Atmospheres included in MODTRAN code [23
] and simulated radiance for a mid-latitude summer atmospheric profile with known LST and land surface emissivity (ε
]. The synthetic nature of these studies requires ongoing evaluation against in-situ measurements to further assess the performance of the LST algorithms over a range of land cover types and conditions. The temperature–emissivity separation (TES) approach of Gillespie et al. [19
], developed for ASTER data, uses an iterative technique to remove reflected sky radiance. The TES algorithm hybridizes three established algorithms and estimates normalized emissivity first, while calculating emissivity band ratios afterward. Using simulated ASTER data, Gillespie et al. [19
] recovered temperatures within ±1.5 K and emissivities within ±0.015. Wang et al. [15
] used the same temperature–emissivity separation algorithm for Landsat 8, combining the two thermal bands with atmospheric profile data from the National Centers for Environmental Prediction (NCEP), and determined the average absolute error between their estimated LST and four surface radiation sites [24
] to be 1.74 K, varying from 1.04 to 2.56 K.
The US Geological Survey (USGS) recently reported a calibration problem in the TIRS caused by stray light, resulting in a higher bias in one of its two bands (4% in Band 11, and 2% in Band 10) [25
]. As such, relying on split-window algorithms for the estimation of LST from Landsat 8 becomes increasingly problematic, as the introduction of errors may prove difficult to quantify [26
]. An alternative approach is to estimate LST using a single band or single-channel (SC) algorithm [27
]. Three different single-channel methods are commonly found in the literature, including: (1) the radiative transfer equation (RTE) [27
]; (2) the mono-window algorithm developed by Qin et al. [30
]; and (3) the generalized single-channel (GSC) method developed by Jiménez-Muñoz and Sobrino [31
]. These methods use the radiance measured by one of the Landsat TIRS bands and then correct the radiance using atmospheric profile data or total column water content. Atmospheric profiles can be obtained from ground-based atmospheric radio-soundings, satellite vertical sounders, or from meteorological forecasting models and reanalysis data. Unfortunately, ground-based radio-soundings are often difficult to obtain at the time of the satellite overpass and are sparsely recorded, making them suitable only for validation at specific sites [32
]. To circumvent the spatial limitations of ground-based sounding data, atmospheric profiles derived from satellite platforms provide a large spatial extent, but can be compromised by relatively poor temporal resolution [28
]. To overcome this spatiotemporal constraint, reanalysis data from weather prediction models provide a practical alternative to the spatial limitation of radio-soundings, while offering a flexible temporal resolution that can match retrieval requirements. Given the increasing availability of ground-based and reanalysis sources, a greater understanding of the limitations and uncertainties involved in using such data in LST retrieval schemes is required.
An example of a single-channel method that utilizes reanalysis data is a web-based correction tool proposed by Barsi et al. [34
] that uses the NCEP modeled atmospheric global profiles as input to the MODTRAN radiative transfer model [23
]. This procedure allows one to obtain the atmospheric transmission, upwelling radiance, and downwelling radiance values needed to convert sensor radiance into surface brightness temperature. McCarville et al. [35
] expanded on this process by using the NCEP North American Regional Reanalysis (NARR) database as inputs. In a more recent study involving Landsat 8, Tardy et al. [36
] evaluated MODTRAN using an automated Python tool that relies on the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-Interim product [37
] as the source of atmospheric profile data. The authors evaluated the tool in southwestern France and central Tunisia by comparing the retrieved Landsat LSTs against in-situ LST measurements, resulting in an RMSE of 2.55 K for France and an RMSE of 1.8 K for Tunisia. These and related investigations have tended to examine the estimation of LST from Landsat using a single atmospheric profile source [31
]. Importantly, relatively little attention has been paid to characterizing the sensitivity of LST to different atmospheric profile sources. Likewise, studies examining the atmospheric correction of Landsat 8 TIR imagery in arid lands are rather limited, despite the increased availability of cloud-free Landsat scenes in these environments that may help improve insights into spatiotemporal variability.
In addition to the source of atmospheric profile data, another potential source of uncertainty relates to the role that aerosols play in the prediction of LST. While the atmospheric correction in the thermal infrared is often focused on adequately accounting for the influences of water vapor and other trace gases [33
], a key consideration in many regions of the world is the impact of aerosols and the role that dust plays in attenuating the satellite observed signal. The Saharan desert and its surrounds produce 50% of the mineral dust in the world [38
], and dust is a recurrent feature throughout many arid and semi-arid environments [39
]. While the impact of dust on the optical region of the electromagnetic spectrum has been well studied, understanding the role it plays on the thermal portion of the spectrum [38
] and in dryland systems has not been examined as thoroughly [40
]. Previous investigations have illustrated that even small increases in aerosol concentrations can significantly affect radiative fluxes [41
]. A recent study focusing over the Red Sea and Arabian Peninsula estimated that the simulated domain average contribution of mineral dust to the total aerosol optical depth (AOD) was 87% [42
]. Clearly, understanding the impact of the aerosol effect on retrieved surface temperature in such environments is an area of needed research and is one aspect explored herein.
The overall objective of this work focuses on the atmospheric correction of Landsat 8 thermal data over dryland systems [43
], a climate zone that is relatively poorly represented in satellite-based temperature investigations, despite their importance in global energy budgets [44
] and food production [43
]. We explore this topic by characterizing the impact of using five different atmospheric profile records as input to MODTRAN. The atmospheric profile sources include satellite-based data from the Atmospheric Infrared Sounder (AIRS) onboard the Aqua satellite [45
], as well as the Moderate Resolution Imaging Spectroradiometer Atmospheric Profile Product from the Terra platform (MOD07) [46
], global reanalysis data from the ECMWF ERA-Interim [37
] and the NCEP National Center for Atmospheric Research (NCAR) Reanalysis 1 [47
], and in-situ balloon-based radiosonde releases from the Integrated Global Radiosonde Archive (IGRA) [48
]. Furthermore, to evaluate the impact of input parameter errors on the atmospheric correction of Landsat 8 TIR data, a sensitivity analysis on the satellite and reanalysis atmospheric profiles was performed by applying randomly distributed errors into profile parameters such as temperature and relative humidity, as well as to ozone, CO2
, and aerosol optical depth.
The work presented here contributes to the atmospheric correction of Landsat 8 in arid lands, a climate zone that is relatively poorly represented in satellite-based temperature investigations. The increased availability of cloud-free Landsat scenes in these environments provides a rich dataset that helps to improve insights into spatiotemporal variability. Results indicate that the major physical parameters controlling the accuracy of the satellite retrieved LST include the relative humidity and emissivity. An introduced uncertainty of 20% in relative humidity can result in LST errors as high as 1.5 K for bare soil and 1 K over alfalfa in irrigated arid environments, while an uncertainty in emissivity values of 0.01 can result in errors between 0.7 and 1 K for both land cover types. Typically, determining the emissivity with any degree of precision is a challenging task, so an error of 0.01 in emissivity is certainly not unusual and in line with the reported standard deviation in the emissivity values taken from the ASTER GED. Ozone and CO2 were found to have much smaller influences on the estimated LST for all the atmospheric profiles and were not seen to significantly affect the results relative to the other sources of error.
The AIRS profiles generally better performance over bare soil might be attributed to its capacity to represent the local atmospheric conditions better than reanalysis sources. Figure 5
suggests that the AIRS and ECMWF are consistent with the available radiosonde record. In contrast, MOD07 seems to be unable to accurately represent the conditions sensed by the radiosonde. This is particularly evident in DOY 288 (Figure 5
c), where MOD07 has serious discrepancies compared to the rest of the profiles. Landsat-based LST from the MOD07 profile was expected to provide better results, as the overpass time is within 30 min of that from Landsat. However, it was discovered that the MOD07 profile is located at the edge of the swath, affecting the quality of the data. Moreover, contrary to AIRS, MOD07 is not a sounding instrument and its precision might be lower than that of AIRS for the same conditions in the region.
Aerosol optical depth is of interest in arid lands since dust sources, regardless of size or strength, can usually be associated with topographical lows located in these regions [77
]. Furthermore, aerosols covering large areas in the region affect the satellite signal [72
], and AOD variability is seldom considered in the estimation of LST [28
]. Recent studies have found that the discrepancies between the Deep Blue algorithm and ground truth are somewhat larger over bright desert surfaces than over other land cover types [75
]. In particular, over arid sites and across 16 AERONET validation sites, 37% of the studied points fall outside of the reported error for the Deep Blue algorithm [62
]. The arid sites region includes the “Solar Village”, located approximately 200 km northwest of the Tawdeehiya farm. The Deep Blue algorithm for this site has an RMSE of 0.16, larger than the average of 0.145 for all the arid sites [62
]. In addition, 40% of the studied points fall outside of the reported error, suggesting that the AOD errors in this region could be larger.
The effect of AOD in the satellite signal is apparent in the Landsat-based LST validation of the atmospheric correction results, where LST from days with a high AOD were mostly underestimating the radiometer measurements (Figure 3
). While not necessarily a problem in many regions of the world, for arid land agricultural systems further attention should be focused on characterizing such influences, particularly in the light of potential increases in dust loadings [73
]. For the irrigated arid study area, errors of up to 0.25 K in the estimation of LST appear when AOD has an error of 20%, in line with recently reported errors [75
], suggesting that the parameter needs to be considered in accounting for LST uncertainty in arid environments. Despite substantial uncertainties associated with satellite-based AOD retrievals over this region [40
], further challenges arise, as MODTRAN seems unable to accurately simulate the aerosol conditions for the TIR bands. While the AOD influence on reflectance over arid lands is relatively well understood [40
], a better characterization of aerosol influence in the TIR bands for the single-channel method might be needed in this region.
The data derived from the Landsat 8 TIRS sensor have suffered a variety of calibration adjustments that impact the fidelity of retrievals [26
]. Starting in August 2013, discrepancies between Bands 10 and 11 were noted, resulting in water surface temperatures derived from TIRS data being warmer than measured temperatures ≥ 2 K. The source of these errors has been attributed to thermal energy from outside the normal field of view, also known as stray light. These excess energy leakages of 0.29 and 0.51 W/m2
·srad·μm result in a temperature error of 2.1 K and 4.4 K at 300 K in Band 10 and 11, respectively. Unfortunately, these discrepancies are not consistent across the focal plane, making a correction a challenging task [26
]. Since these errors were reported, several calibration and reprocessing efforts have been carried out [79
The Landsat Calibration-Validation Team (CVT) temporally adjusted the TIRS band’s radiometric bias to improve the absolute radiometric error for typical Earth scenes during the growing season. Irrigated arid lands are perhaps outside of the typical Earth scenes domain and this adjustment might not be adequate for the region. The estimated stray light error by the CVT was found to be 0.29 ± 0.12 W/m2
·srad·μm for Band 10. It is likely that errors caused by stray light could be larger in arid lands since the area surrounding the TIRS field of view typically has higher than average temperature values. In other words, the higher temperatures that the farm is subject to throughout the year, compared with many other agricultural regions in the world, could influence the amount of thermal energy that the sensor is observing due to the stray light discrepancy. This is apparent by the positive bias during the summer period in this study (Figure 3
). After this study was carried out, the CVT released a new stray light correction method for the TIRS that reduces the TIRS uncertainty to under 0.5% [81
], reducing the errors from 2 K @ 300 K with no correction to 0.3 K with the stray light correction for band 10. The study also found out that light was impinging on the detectors from a ring about 13° outside of the field of view, suggesting that the overall higher bias during the hot days might be attributable to hot desert soil around the study area reaching the TIRS. Currently, work is underway to assess whether the correction is adequate for split-window correction methodologies and the CVT still does not recommend the use of band 11 for split-window techniques.
The errors seen in the different seasons could be a combination of the stray light errors and atmospheric profile errors. Further studies with stray-light-corrected Landsat 8 TIRS imagery might be necessary to reassess the feasibility and choice of the source of the atmospheric profile for single-channel TIR atmospheric correction. However, the sounding capabilities of AIRS (despite the 3 h difference in overpass time relative to Landsat) might provide a better representation of the local moisture conditions (i.e., irrigation) of the farm than the ECMWF or NCEP profiles. Furthermore, the location of the farm in the Terra MOD07 profiles is far from ideal, having a large viewing angle due to the orbital co-geometry of the Landsat 8 and Terra satellites. In reality, the atmospheric conditions in the troposphere are the main driver for the atmospheric correction of TIR imagery, as most of the moisture in the atmosphere is located in this region and water vapor is the main absorber in the TIR [31
]. Following this, the higher vertical resolution from NCEP and ECMWF does not seem to be playing a significant role in the atmospheric correction results (Figure 3
and Figure 4
). An analysis showed that for a given day, LST calculated using ECMWF produced a 0.2 K difference when using the full profile (1 mbar) versus that only using up to 15 km altitude (125 mbar), suggesting that the precision of atmospheric conditions at the surface (AIRS) is preferred to enhanced vertical resolution.
The availability of Landsat-scale spatially consistent alfalfa in-situ LST was limited by the harvesting schedule of the field. After each harvest, LST measurements from the infrared radiometer are not consistent with the Landsat 8 TIRS imagery since the area covered by the infrared radiometer is not harvested. Therefore, the data corresponding to these days for the alfalfa LST validation were discarded following the implementation of an NDVI threshold approach. Despite using this data separation technique, the difference between in-situ and satellite LST values remain considerably large, presenting LST differences of up to 4 K. The large bias (Table 1
) might be explained by the different conditions that the unharvested patch in the alfalfa field has in relation to the whole field. Furthermore, the Apogee radiometers could be a source of inaccuracies if their calibration has deviated from their optimal operating parameters.
An obvious source of error could come in the form of discrepancies when comparing satellite pixels to point observations. Variations of subpixel surface geometry and shadows from the canopy are always present but rarely considered, resulting in emissivity and temperature variations that are combined into a single pixel [69
]. These variations can interact in nonlinear ways, as the combined radiance depends on the surface materials and on the temperature distributions of each of those materials. The effect of mixed-pixel response is apparent in the results, particularly over alfalfa (Figure 2
) where the Landsat-based LST consistently overestimates the in-situ LST (Figure 4
) as the Apogee radiometer is located near the center of the pivot. However, in-situ monitoring remains the best way to evaluate these high-resolution LST retrievals, despite their limitations.