Comparison of Land Surface Temperatures from Near-surface Measurement and Satellite-based Product

Abstract

Land surface temperature ($T_s$) is a critical variable for understanding the surface energy exchange between land and atmosphere. Using the data measured from micrometeorological flux towers, three types of $T_s$, obtained using a thermal-infrared radiometer (IRT), a net radiometer, and an equation for sensible heat flux, were compared. The $T_s$ estimated using the net radiometer was highly correlated with the $T_s$ obtained from the IRT. Both values acceptably fit the $T_s$ from the Terra/MODIS (Moderate Resolution Imaging Spectroradiometer)satellite. These results will enhance the measurement of land surface temperatures at various scales. Further, they are useful for understanding land surface energy partitioning to evaluate and develop land surface models and algorithms for satellite remote sensing products associated with surface thermal conditions.

1. Introduction

Land surface temperature (T_s), the radiative skin temperature of vegetated ground, is a critical variable in climate, hydrologic, ecological, biophysical, and biochemical processes of the Earth (Jin, 2004). It is physically related to energy fluxes through landatmosphere interactions (Voogt and Oke, 2003). Understanding the spatial and temporal characteristics of Ts is heavily dependent on the development of remote sensing from space (Mildrexler et al., 2011).

Satellite remote sensing with thermal infrared (TIR) sensors has been widely implemented to estimate Ts with sufficient temporal and spatial resolution (Voogt and Oke, 2003). An optical sensor, such as that used in TIR measurements, is mostly recommended to detect Ts from the energy emitted by the land surface integrated over all viewing directions (Dash et al., 2002;Jones et al., 2009). Many algorithmsto estimate Ts from satellite TIR sensors have been developed in recent decades (Li et al., 2013; Park and Suh, 2013). One critical issue in these algorithms is that the determination of Ts from a space-based TIR sensor is largely affected by the absorption and emission of atmospheric water vapor and land surface emissivity (Prata et al., 1995; Park and Suh, 2014). However, few studies have attempted to validate the satellite-derived Ts, although it is a critically important process for assessing the uncertainty of data (Li et al., 2013).

Three methods have been proposed for validating satellite-derived Tss. First, the temperature-basedmethod compares the satellite-derived Ts with an in situ Ts measurement at the satellite overpass (Coll et al., 2005). Second, the radiance-based method obtains the accuracy from adjusting Ts, using the difference between the radiance measured at the top of the atmosphere (TOA) and the radiance simulated by an atmospheric radiative transfer model (RTM) (Wan and Li, 2008). The measured land surface emissivity and in situ atmospheric profiles at the satellite overpass must consider the RTM (Wan, 2008). Third, the crossvalidation method uses a well-validated Ts product as a reference (Trigo et al., 2008).

The radiance method and the cross-validation method are indirect methodsfor validating space-based Ts measurements. Only the temperature-based method involves the in situ comparison of Ts; however, it strongly depends on the accuracy of the groundmeasured Ts and the representative pixel size of the satellite image (Tang et al., 2015). Therefore, the ground-measurementmethod isimportantfor obtaining the true in situ value of Ts because thermometers can only contact a particular area of the surface, which includes various elements such as soil particles, plant organs, and water (Dash et al., 2002). In addition, they are sensitive to the view orsolar zenith angle (Jones et al., 2009). Further, given the large spatial and temporal variations in Ts, the data of a ground-measured Ts must be obtained fromcontinually operating sensorsinstalled at a homogenous surface within the satellite pixel size (Tang et al., 2015). It might be difficult to get an acceptable validation using a field-measurementsurvey over a short period.

Several micrometeorological flux towers, which measure the land-surface energy balance, such as sensible and latent heat fluxes, are located across a global network (Baldocchi et al., 2001). Some flux towers have a thermal-infrared radiometer (IRT) instrument as a type ofTIRsensor.Itshould be suitable for the temperature-based method of Ts validation. Furthermore, in thisstudy, two other ground-measured Ts data were applied to the validation.Anetradiometer, which measures upward longwave radiation, is installed in every micrometeorological flux tower, even if there is no IRT sensor. That upward longwave radiation can be converted to Ts using the StefanBoltzmann equation.Additional Ts data can be estimated by using the inverse method, based on the equation of sensible heat flux, wherein the heat flux is measured from the flux tower.

In this study, the well-validated Terra/MODIS (Moderate Resolution Imaging Spectroradiometer) satellite-derived Ts was compared with three types of ground-based Ts: one from the IRT sensor of a tower, one from a net radiometer instrument installed on an eddy-covariance tower, and one from the inversion of a sensible heat flux equation. Given the difficulties of Ts observation on the complex surface of various elements, these comparisons will improve the understanding, not only of Ts itself, but also of variousscientific issues associated with surface energy exchanges (e.g., evapotranspiration, surface dryness conditions, and urban heat islands).

2. Data and Methodology

1) Eddy covariance measurement site

Brooks Field Site 10 (Br1), one of the AmeriFlux network’s experimental sites, is located at 41° 58′ 29.64″ N and 93° 41′ 26.16″ W in the upper Midwest, USA (Corn Belt) (ameriflux.lbl.gov). This farming field is used for soybean/corn production. It has relatively close-to-idealsurface conditions(e.g.,flat and homogeneous land) for applying the eddy covariance technique. The experimental area has humid winter and hot summer seasons. The site is located at an elevation of 313 m above sea level. Measurements were taken by eddy covariance systems installed on a micrometeorological tower. We used only daytime flux measurements, collected between January 2006 and December 2007, to avoid condition of weaker turbulent mixing that frequently occurs at night

A net radiometer (CNR1; Kipp and Zonen, Delft, Holland) and a precision infrared temperature sensor (IRTS-P;Apogee, Logan, UT, USA) were installed on the tower. CNR1 has wavelength ranges of 0.305-2.8 μm for shortwave radiance (W m-2) and 5-50 μm for longwave radiance (W m^-2 ). The IRTS-Pmeasured the radiance at 8-14 μm wavelengths, at a height of 5 m above ground level for corn and 3 m for soybeans. In addition, the soil heat flux was measured using soil heat flux plates at 6-cm depth. Soil-temperature thermocouples were installed within the topsoil layer.

2) Estimation of land surface temperature

Sensible heat flux (H; W m^-2 ) can be represented as the temperature difference between a surface and the surrounding air. It is inversely proportional to the transfer resistance:

\(H = -\rho C_p {T_a-T_s \over r_H}, r_H ={u \over u^{*2}}\)       (1)

where T_a and T_s are the air and surface temperatures (°C), respectively. ρ is the density of air (kg m^-3 ) and varies with temperature and humidity; ρ = 1.2 kg m-³ was used. C_p isthe heat capacity of air (J kg-¹ °C-¹ ); it is approximately 1010 J kg^-1 °C^-1 . The transfer resistance r_H (s m-1) depends on the wind speed and surface characteristics.rH can be estimated by simply using the wind speed (u; m s-¹ ) and friction velocity (u^* ; m s-1 ).

T_s in Eq. 1 represents the aerodynamic surface temperature. It is often accepted as the radiometric surface temperature when estimating H in the energy balance. Using the inverse of Eq. 1 with H, T_s can be estimated simply as: 

\(T_{s_H} ={ 1\over{\rho C_p}} {u \over u^{*2}} H + T_a\)       (2)

On the other hand, according to the definition of T_s, it can be directly calculated from upward longwave radiance data (R_↑lw;Wm^-2 ), which can be obtained from net radiometer sensors. Net radiometers usually are installed on eddy covariance towers for net radiation (R_n; W m^-2 ) measurement.

\(T_{s_ {rad}} =( {R_{\uparrow lw}\over \varepsilon \sigma })^{1/4} - 273.15\)       (3)

where ε is the emissivity (0.98 in this study) and σ is the Stefan-Boltzmann constant (5.6703 × 10^-8 W m^-2 K^-4 ). The footprint area of the installed net radiometer is usually small. It is different from the footprint area of H measured by eddy covariance.

The wavelength range of an IRT sensor is well utilized for measuring T_s. T_{s_IRT} corrected the sensor temperature error using a modified Stefan-Boltzmann equation. On the other hand, in thisstudy, the effects of the background reflected temperature and atmospheric emissivity were not included in the estimation of T_{s_IRT} and T_{s_rad}, but those are considered in the LSTalgorithm for MODIS.The T_s measured from IRT(T_{s_IRT}) and T_{s_rad} have similar footprint area; however, the T_{s_H} based on an aerodynamic method is much larger than these. Indeed, although T_{s_H} might be conceptually similar to T_{s_IRT} and T_{s_rad}, T_{s_IRT} and T_{s_rad} are fundamentally different from T_{s_H} derived from the equation for H. T_{s_H} reflects the bulk surface temperature because it is derived fromaerodynamic surface fluxes, whereas T_{s_rad} and T_{s_IRT} are radiometric skin temperatures based on solar geometries.

T_s measured by satellite can be defined as the average temperature of the fractional cover of an inhomogeneous surface. To compare T_s derived from ground-measured data with satellite T_s, one day of scaled Ts with a 1-km resolution from Terra/MODIS Collection 6 (MOD11A1) was used. The TIR sensor on a satellite can produce Ts via various methodologies (Li et al., 2013).In a 1 × 1 km² area ofthe selected pixel that included the location of the micrometeorological flux tower(Br1), the land surface wasmostly composed of soybeans and corn in the footprint area of the eddy covariance system. These ground-measured data were analyzed to meet the scan time (at around 11 a.m.) of the Terra satellite (Table 1).

Table 1. List of Ts estimates used in this study

This study assumed that the characteristics of the footprint areas for IRT, net radiometer, and eddy covariance instruments were similar because crop plants are homogeneous in large-scale study areas. In addition, although the one-pixel area of MODIS T_s (T_{s_MODIS}) is dramatically different for the footprint area of an eddy covariance, a net radiometer, or an IRT at the Br1 site, the pixel including the Br1 site was considered to have properties close to the land of the Br1 site because contiguous land is considered to be similar to Br1 site.

3. Results and Discussion

1) Comparison among ground-based T_ss

Fig. 1 shows the relationship among the three T_ss (i.e., T_{s_H}, T_{s_rad}, and T_{s_IRT}) from the flux measurement tower at the crop field (Br1 site). Based on the results obtained for two years, data of all T_ss were generally well correlated with each other.Because only T_{s_IRT} was obtained forthe purpose of measuring T_s, we compared T_{s_H} and T_{s_rad} to T_{s_IRT}. The correlation coefficient (R) values of T_{s_H} and T_{s_rad} were 0.85 and 0.92,respectively, while the root mean square errors (RMSEs) were 4.32°C and 3.61°C, respectively (Fig. 1). T_{s_rad} best fitted to T_{s_IRT}, although the utilized wavelengths were different (5-50 μm for net radiometer and 8-14 μm for IRT).

Fig. 1. Comparison of estimated T_{s_H} and T_{s_rad} to measured T_{s_IRT}. In the relationships of T_{s_H} and T_{s_rad} to T_{s_IRT}, the slopes are 0.98 (1.08) and 1.05 (1.05), respectively, and the intercepts are 1.67 (0.84) and 0.58 (0.31), respectively. Values of R are 0.85 (0.91) and 0.92 (0.94), respectively and the RMSE values are 4.32°C (4.05°C) and 3.61°C (2.41°C), respectively. Values in brackets show the results of data × markers.

Further, a linear correlation between T_{s_rad} and T_{s_IRT} wasslightlymore established underthe lower downward shortwave radiation (R_↓sw) (≤ 300 W m^-2). Strong radiation may incite the effect owing to differences in the observable angle and sensor sensitivity. However, the linear correlations of T_{s_H} to T_{s_IRT} fitted better under lower r_H and higher u^*; the latter variables are often used as indicators of atmospheric turbulence conditions. The measurement errors of H from the eddy covariance technique may cause the under- and over-estimated T_{s_H} values against the radiance-based T_{s_IRT}. Indeed, according to Eq. 2, the values of T_{s_H} arithmetically decreased with lower rH and higher u^* values. Under lower r_H (≤ ~15) and higher u^* (≥ 0.5), T_{s_H} was mostly underestimated against T_{s_IRT}, but less scattered.

2) T_ss related to surface energy closure

Because the eddy covariance measurement does not usually indicate a closure ofthe surface energy balance, we evaluated the performance comparison of T_{s_H} to T_{s_IRT} and energy balance closure ratio (LE + H) / (R_n – G)separately. Fig. 2 showsthe distribution pattern of T_{s_H}/T_{s_IRT} against (LE + H) / (R_n – G) with R_↓sw. Most plots were located in the area with larger available energy (R_n – G) than in the area with turbulence fluxes (LE + H), particularly for R_↓sw > 700 W m^-2.

Fig. 2. Scatterplots of T_{s_H}/Ts_IRT against the energy balance closure (LE + H) / (R_n – G).

In addition, T_{s_H}/T_{s_IRT} under strong solar radiation largely varied from 0.5 to 2.0. On the other hand, when R_↓sw was less than 300 W m^-2, T_{s_H} and T_{s_rad} had similar values, regardless of the energy balance closure condition.These characteristicsmight be highly governed by variations in aerodynamic heat and moisture transfer.

3) Comparison of ground-based T_ss to MODIS T_s

Fig. 3 shows the relationship between T_{s_MODIS} and the estimated T_ss. The plots of all ground-measured T_ss mostly fit T_{s_MODIS}. This might mean that the difference between the satellite pixel size and the footprint areas of the IRT, net radiometer, and eddy covariance instruments did not introduce critical problems in this case.

 

Fig. 3. Scatterplots of Ts_IRT, Ts_rad, and Ts_H versus Ts_MODIS

T_{s_rad} best correlated with T_{s_MODIS} (R = 0.90 and RMSE = 2.89°C) among the estimated T_ss (Table 2). Because Ts_IRT, T_{s_rad}, and T_{s_MODIS} were estimated from longwave measurements, they were expected to have similar performances. However, the correlation of T_{s_IRT} to T_{s_MODIS} (R = 0.67 and RMSE = 5.91°C) was unexpectedly lower than that to T_{s_rad} (R = 0.90 and RMSE = 2.89°C). This correlation behavior was also found in some other Ameriflux sites, particularly in crop fields (data not shown).

Table 2. Correlation of estimated T_{s_IRT}, Ts_rad, and Ts_H to measured T_{s_MODIS}

MODIS bands 20 (3.660-3.840 μm), 21 (3.929- 3.989 μm), 22 (3.929-3.989 μm), 23 (4.020-4.080 μm), 31 (10.780-11.280 μm), and 32 (11.770-12.270 μm) for brightness temperatures could be used in T_{s_MODIS} algorithms (Wan, 2008). Bands 20, 21, 22, and 23 overlapped with the longwave range of the net radiometer (5-50 μm), but not with IRTS-P (8-14 μm). Matching the band wavelength ranges in the TIR sensors can be critically important to the difference between estimated T_ss.

Ground measurement is generally considered to be a true value. However, as seen in the difference among T_{s_IRT}, T_{s_rad}, and T_{s_H}, ground measurements also include some error. Given that T_{s_MODIS} is a wellvalidated product among satellite-derived T_ss, T_{s_MODIS} is presumed to be a true value in this analysis. When T_{s_MODIS}, as the true T_s, was close to T_a, T_{s_rad} was well fitted to T_{s_MODIS}.

4. Conclusions

In this study, we evaluated T_ss estimated using measurement data from a micrometeorological flux tower. It was found that T_{s_IRT}, the surface temperature of the IRT, had the most similar pattern to T_{s_rad}, the temperature of the net radiometer. This result is encouraging because most towers have a net radiometer, but not an IRT. The T_{s_rad} could be used, not only to validate the satellite-derived T_s, but also for plant physiological applications, particularly when Ts is close to the vegetation canopy temperature under a closed canopy condition. Thus, T_{s_rad} can increase our understanding of CO₂ and H₂O fluxes based on plant physiological conditions.

In addition, meaningful relationships of T_{s_IRT} and T_{s_rad} with the satellite-observed T_s were observed. The data from the IRT installed in the tower and the net radiometer should be useful for validating satellitederived skin temperatures. However, given that both T_{s_IRT} and Ts_rad were not well harmonized with the tower-measured land surface energy balance, and that the satellite remote sensing algorithms such as evapotranspiration were based on the energy-balance equation, T_{s_IRT} and T_{s_rad} might not be suitable for evaluating and developing satellite remote sensing algorithms for other land surface properties.

The meaning of T_{s_H} is different from that of T_{s_rad} or T_{s_IRT} in terms of bulk surface temperature and skin surface temperature. Therefore, care should be taken when applying T_s. For example, the value of Ts calculated by the land surface model (LSM), which includes a module to check the equilibrium of all energy components, might not be appropriate to compare with the T_s of the IRT, net radiometer, or physical equation for H. In this case, Ts calculated using the energy balance equation, T_{s_H}, would be more appropriate to validate with T_s of an LSM, although T_{s_H} is not very close to T_{s_IRT}.

The findings of this study will enhance the measurement of land surface temperatures at various scales. However, these findings are based on data from a single micrometeorological flux measurement site. Further analyses with more measurement data are necessary to identify significant characteristics of each estimated T_s.