Geophysical Research Letters

Disaggregation of GOES land surface temperatures using surface emissivity

Authors


Abstract

[1] Accurate temporal and spatial estimation of land surface temperatures (LST) is important for modeling the hydrological cycle at field to global scales because LSTs can improve estimates of soil moisture and evapotranspiration. Using remote sensing satellites, accurate LSTs could be routine, but unfortunately the only instruments available to provide diurnal cycle observations have coarse spatial resolution (4 km). One approach that may help overcome the spatial resolution constraint is to disaggregate geostationary LST data using visible to thermal infrared information provided by single time of day MODIS 1 km observations. These higher-resolution observations are correlative with observations at 4-km scales, and thus can be used to estimate 1-km LST values throughout a day. Inamdar et al 2008, for example, showed how GOES 10 imager and MODIS data could be combined to produce accurate half-hourly, 1-km LST values. However, the method disaggregated coarse LST values using Normalized Difference Vegetation Index (NDVI) data and was sometimes highly inaccurate when considering heterogeneous terrain. This problem can be greatly reduced with an alternative approach, whereby MODIS land cover emissivity data sets supply the needed 1-km information. In a study of LST estimation over the US Southwest, diurnal disaggregation models using emissivity data were significantly more accurate than a comparable NDVI-based model. This alternative approach which directly employs 8-day composites of MODIS 1 km emissivity is a simple and fast method.

1. Introduction

[2] Spatial and temporal sampling of the land surface temperature is vital to determining the land surface energy exchange with the atmosphere [Dickinson, 1992], in the assessment of the earth's hydrological cycle and optimum utilization of water resources in agriculture [Price, 1982]. Satellite-based land surface temperature (LST) retrievals have been accomplished using both polar orbiting sensors [Prata and Platt, 1991; Wan and Dozier, 1996] as well as geostationary orbiting sensors [Sun and Pinker, 2003]. Polar orbiting satellites provide a more uniform global view of the earth, a restricted range of view zenith angles, shorter atmospheric path lengths, better accuracy and higher spatial resolution (∼1 km) than feasible with geostationary satellites. However LST retrievals from the accurate MODIS sensors are available at most twice a day. Geostationary satellites, on the other hand, are capable of providing frequent measurements of LST, but due to their large instantaneous field of view have ∼4 km spatial resolution and are generally inadequate for distinguishing different land cover types.

[3] The potential advantages of both polar orbiting and geostationary sensors have been combined in a recent study [Inamdar et al., 2008, hereinafter referred to as I-8], by employing the MODIS (Moderate Resolution Imaging Spectroradiometer) instrument on board the EOS (Earth Observation System) TERRA and AQUA satellites as a calibration source for the GOES (Geostationary Environmental Satellite) satellite, to yield a diurnal cycle of LST at 1 km scales.

[4] The disaggregation procedure in I-8 utilized the negative relationship between LST and the NDVI at both 4 and 1 km scales. Although the inverse relationship between NDVI and LST is well documented in the literature [Badeck et al., 2004; Tucker et al., 2001], its utility in disaggregation of GOES-derived LST has limitations, such as its restricted validity over heterogeneous terrain and its poor correlation at times other than mid-day [Sun and Kafatos, 2007]. Agam et al. [2007] reported success with their algorithm for sharpening thermal imagery to higher resolutions over a corn/soybean production area in central Iowa, but extension of their algorithm would likely not be successful when considering much larger domains such as the entire US southwest [Karnieli et al., 2006; Sun and Kafatos, 2007].

[5] Although NDVI is correlated with the surface emissivity values in the thermal infrared [Momeni and Saradjian, 2006], emissivity is strongly coupled with LST through the radiative transfer equation, and has a direct influence on the satellite-observed brightness temperature. Thus estimation of LST through emissivity constitutes a more physical approach than through empirical NDVI-LST relationship. The surface emissivity values in the thermal infrared bands of MODIS have been derived from small-scale component measurements, land cover classification [Snyder et al., 1998] as determined from the quarterly land cover data, dynamic and seasonal factors, and have been shown to yield more accurate LST estimates when used in the generalized split window scheme [Wan and Dozier, 1996]. Thus they form a better predictor of LST than NDVI.

[6] The present study outlines an approach, whereby 8-day composite images of MODIS 1 km gridded emissivity data are employed in the determination of 1 km LST in lieu of the NDVI-based disaggregation process along with the I-8 diurnal cycle model and an improved scheme for filling of data gaps. The resulting 1 km diurnal cycle of LST data are validated and employed in conjunction with MODIS Land Cover Class and NDVI data in an investigative study of feasibility of disaggregation schemes based on NDVI.

2. Data Sets

[7] The primary streams of data used in the present study comprise (1) the brightness temperatures in 3 infrared bands (3.9 μm, 11 μm and 12 μm) of the GOES-10 Imager (positioned at 135°W longitude covering the US southwest region and part of the Pacific Ocean), (2) the 1 km LST and surface emissivity in the thermal infrared window bands 31 (11 μm) and 32 (12 μm) from MODIS (MOD11L2), (3) the 8-day composites of the 1 km surface emissivity in bands 31 and 32 from MODIS (MOD11A2), (4) the 16-day composites of the 1 km MODIS NDVI data (MOD13A2), and (5) the land cover type data from MODIS (MOD12Q1).

[8] Both the GOES imager data comprising the brightness temperatures in the near infrared and thermal bands, and the 8-day composites of the emissivity data (MOD11A2) are reprojected using bilinear interpolation to US Albers 1 km equal area grid, in order to preserve a 1 km pixel size.

[9] The present study hinges critically on the quality of the MODIS L2 emissivity data in bands 31 and 32. The MODIS LST is produced by two independent algorithms: the day/night algorithm which yields both LST and emissivity at 5 km resolution (MOD11B1) and the generalized split-window algorithm (MOD11_L2) which relies on the land cover classification-based surface emissivity values [Snyder et al., 1998]. In a recent study [Wan, 2008], it was demonstrated that the two algorithms produce compatible LSTs at different spatial resolutions. Emissivities retrieved by the day/night method compared well with the surface emissivity spectra measured by a sun-shadow method in field campaigns. Further, comparison of MOD11_L2 LSTs (used in the present study) with in-situ measurements [Wan, 2008] showed that MODIS LST product is accurate to better than 1 K suggesting confidence in both the LST and the emissivity values in bands 31 and 32.

3. Methodology

[10] The focus of our study is the southwest region of the United States. The processing of primary data sources to obtain the desired half-hourly 1 km LST product consists of several modules: cloud-screening of GOES data, collocation of GOES and MODIS pixels, using a generalized split-window scheme, and application of a semi-empirical diurnal model filter to fill data gaps and derive a seasonally stabilized set of LST values. These modules, described in detail in I-8, will be reviewed below.

3.1. Cloud-Screening

[11] One of the important requirements of successful LST retrieval from satellites is the effective removal of cloudy radiances. But cloud-clearing over land surfaces is a challenging task, due to the wide variety of underlying land cover types and to the dynamic characteristics required for accurate spectral screening. We employ here a combination of schemes: initial evaluation of the GOES brightness temperatures in 11 μm for spatial coherence at small and large scales [Závody et al., 2000], followed by application of the four-step Bispectral Threshold and Height method (BTH) [Haines et al., 2003].

3.2. Matching and Merging of GOES and MODIS Data

[12] The initial matching and merging of GOES/MODIS data for the development of the split window relation [Wan and Dozier, 1996] follows I-8. The MODIS L2 1 km LST and emissivity data are first aggregated to the native GOES imager resolution. The cloud-cleared GOES pixels collocated with the aggregated MODIS pixels, yield for each matched domain a set of brightness temperatures in the 11 and 12 micrometers, together with MODIS LST and surface emissivity in the MODIS bands 31 and 32. The merged GOES/MODIS parameters for each day are input into a regression scheme similar to the generalized split window method [Wan and Dozier, 1996], allowing us to “calibrate” the GOES brightness temperatures in terms of MODIS LST.

[13] The reprojected GOES Albers 1 km grid pixels are also screened separately for clouds following the same procedure described in section 3.1. Brightness temperatures from these cloud-cleared GOES pixels, and MODIS emissivity—also reprojected to the same grid—are used as input to the split-window relation, yielding 1 km LST at half-hourly intervals. Data gaps due to lack of GOES observations and cloud-contaminated LST values are addressed through the help of a median composite filter and a diurnal model as described below.

3.3. Diurnal Modeling of Median Composites

[14] The half-hourly 1 km LST values are used in forming a monthly median composite and employed as input into a diurnal temperature cycle (DTC) model [Göttsche and Olessen, 2001; I-8], consisting of a day-time harmonic wave and a night time decay phase. This helps to generate a stabilized set of reference LST values that are cloud-free and can be used to fill GOES data gaps. In order to fill data gaps on a specific day for a given pixel, for example, we first determine a median offset between the monthly median DTC and the observed LST values for the pixel in question for the day, and then add this offset to the median DTC wherever GOES observations are lacking. If the number of observed good LST values is less than 5 for a given pixel for any day, then the data gaps are substituted by the median DTC for the pixel, similar to the procedure adopted in I-8.

4. Results and Discussion

[15] Prior to using the 1 km LST data set, in the assessment of the NDVI-LST relationship, it is essential to establish validity of the LST data product in order to gain confidence in our results and analyses.

4.1. Performance of the Model

[16] Performance of the present model was tested against ground truth by choosing two different surface types (a crop land and an arid site) and seasons (summer and winter): (1) the Southern Great Plains (SGP) Central Facility (C01) operated by the Atmospheric Radiation Measurement (ARM, http://www.arm.gov) program and (2) the Desert Rock (DRA) in Nevada, operated by the Network of Surface Radiation Measurement sites (SURFRAD, http://www.srrb.noaa.gov/surfrad).

[17] We use the upwelling and down-welling longwave flux components as measured by radiometers mounted at the sites and broadband surface emissivity to estimate the surface radiometric temperature or LST. The broadband surface emissivity has been determined from either the MODIS-derived NDVI employing the log-linear relationship or from the five thermal infrared bands of the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) as described in I-8.

[18] A set of clear-sky days has been selected during the summer (June 2002) and winter (January 2003) months for each site for comparison. The radiometric measurements at the site represent 15 minute averages, and thus presence of clouds can be detected as discontinuities or spikes in the time series. Data gaps in GOES were filled using the procedure outlined in Section 3.3. Comparison of model results and observations is shown here in the form of a scatter plot (Figure 1) of predicted diurnal LST values and ground measurements using separate symbols for each site. Results reveal a better performance for the Desert Rock site (diamond symbols) than the SGP site (plus symbols) with the overall mean bias and root mean square error of −0.9 K and 1.98 K respectively for the two sites combined.

Figure 1.

Scatter plot of predicted 1 km LST values versus ground-measurements of diurnal cycle LST values for selected clear sky days at the Central ARM SGP facility (shown by plus symbols) and Desert Rock site, Nevada (shown by diamond symbols).

4.2. NDVI–LST Correlation

[19] For an evaluation of the relationship between the vegetation index and LST, we employ the half-hourly 1 km LST data set, derived from the present emissivity method, in combination with the 16-day composites of the NDVI (MOD13A2). Separate analyses have been performed for the major land cover classes using the quarterly composites of the MODIS-derived land cover types available in the data set MOD12Q1.

[20] Over 75% of the land surface in the US Southwest consists of open shrub land (31%), grassland (28%) and cropland (12%). Figure 2 illustrates the LST-NDVI relationship for the three major land class types, in the Central time zone (95 W to 100 W), at a mid-day time (1:45 PM local time, or 19:45 hours UTC), which nearly corresponds to the peak temperature of the day. The linear regression line, the R2-value, the slope, d(LST)/d(NDVI), and the root mean square error (RMSE) are also shown in Figure 2. Important observations of this correlational study are as follows: (1) the quality of correlation between NDVI and LST is poor for all three land classes, as evidenced by the low R2 value (which represents the percent of explained variance) ranging from 0.11 to 0.38; (2) the correlation deteriorates away from the peak of the diurnal cycle (not shown), with the LST distribution displaying a nearly isothermal pattern with respect to NDVI, around the local sunrise hours; and (3) there are substantial changes in correlational slopes among the different land class types, with the open shrub land displaying the maximum slope and the crop land showing the minimum slope.

Figure 2.

Two-dimensional histogram plot of the monthly median composite 1 km LST and 16-day composite of NDVI for the (top) open shrub land, (middle) grass land, and (bottom) crop land for June 2002 for 1945 hrs. The R2 value of the regression fit, slope(dLST/dNDVI) and the root mean square error (RMSE) for each are shown.

[21] These observations are consistent with other recent studies [Sun and Kafatos, 2007; Karnieli et al., 2006]. Sun and Kafatos [2007] stipulate from their comprehensive evaluation of the NDVI-LST relationship that the slope of the NDVI-LST correlation depends upon the season and time of the day, and strong negative correlations are found only during the warm season and close to the maximum temperature of the day. The minimal slope for the crop land may be attributed to the wet bare soil surfaces, which are common in cultivated regions [Agam et al., 2007]. Also, rapid changes in slopes for different land classes poses significant difficulty in the disaggregation process over heterogeneous land cover classes such as a mix of shrubland, grass land and crop land. Hence, disaggregation techniques that rely on a strong negative relationship between NDVI & LST to derive higher spatial resolution LST data from coarser resolution data sets such as those from GOES imager, are feasible only over homogeneous land cover classes and close to the peak temperature of the day. Even when the best correlation of the day prevails, notice that the progressively increasing spread of LST values (see Figure 2) at lower NDVI, contribute to increased root mean square errors.

[22] In order to further demonstrate the relative performance of the two approaches, we examine the monthly median composite of the aggregated LST and the 16-day composite of the NDVI for the month of June 2002 over one of the SGP sites located in Oklahoma. Disaggregation of the median LST to 1 km scale is performed using the d(LST)/d(NDVI). Figure 3 compares the results of such disaggregation with those of the present model. The site is a crop land surrounded by a mix of shrub land and grass land. The NDVI value for the site is about 0.16 higher than the corresponding aggregated NDVI bounding the site. The monthly median DTC derived from 15-minute measurements on each day is also shown as a dotted curve. The 1 km DTC derived from the NDVI-based disaggregation (shown by plus symbols in Figure 3) is seen to be systematically and significantly lower than the measurements (mean bias of −4 K and rmse of 2.4 K). In spite of a slight overshoot near the peak of the diurnal cycle, the present method shows a much better agreement (mean bias of 0.6 K and rmse of 0.8 K) than the NDVI-based approach. The reasons for the bias in the NDVI method can be attributed to the fact that the NDVI-LST slope used for the disaggregation is weighted by the open shrub land and grass land and is thus higher, and causes a significant drop in LST over the site location from the corresponding aggregated LST value.

Figure 3.

Diurnal cycle of the monthly median LST at the ARM SGP extended facility site in Lamont derived from the NDVI-based disaggregation (plus symbols) and present method using emissivity (diamond symbols). Monthly median of in situ measurements is shown by the dotted curve.

5. Conclusions

[23] A re-evaluation of an earlier technique (I-8) to disaggregate geostationary LST data to higher spatial resolution based on inverse relationship between vegetation index and LST has been performed. Results of the study reveal that NDVI is an inconsistent and less correlative estimator of LST and such disaggregation schemes work only in a limited way. Further, the present study suggests an alternative approach where 8-day composites of MODIS emissivities in the thermal infrared channels are employed in the disaggregation process and yield more accurate LST values. The study also incorporates improvements in filling up of observational gaps.

Acknowledgments

[24] This research was supported by NASA grant EOS/03-0057-0459. ARM data were made available through the U.S. Department of Energy as part of the Atmospheric Radiation Measurement Program. SURFRAD data is made available through NOAA's Air Resources Laboratory/Surface Radiation Research Branch. The authors wish to thank the two anonymous reviewers for their valuable comments and suggestions which helped to improve the paper.

Ancillary