Land-surface temperature (LST) is a key parameter in thephysics of land-surface processes on regional and global scales,combining the results of all surface–atmosphere interactionsand energy fluxes between the atmosphere and the ground. TheLST products retrieved from the data of Moderate ResolutionImaging Spectroradiometers (MODIS) (Salomonson et al.,1989) on the Terra (morning) and Aqua (afternoon) platformsare available in their current collection 4 (or Version 4) startingfrom early 2000 and mid 2002, respectively. The LST products,retrieved with the generalized split-window algorithm (Wan &Dozier, 1996) in most cases, have been validated within ±1 K inclear-sky conditions with in-situ measurement data collected infield campaigns over lakes, silt playa, grasslands and agricul-
tural fields (Coll et al., 2005; Wan et al., 2002b; Wan et al.,2004). However, it has been found since 2003 that the way ofusing the MODIS cloudmask product in the generation of V4LST products caused problems: the V4 products fill with cloud-contaminated LST values and miss valid LST values in areasunder apparently clear-sky conditions. Feedbacks from the LSTuser community request removal of the temporal averaging inthe V4 1 km level-3 LST product (MOD11A1). Therefore,revisions are needed. Since late 2005 the MODIS science teamhas worked on the refinements and tests of the V5 algorithmsand their product generation executives (PGE) for all standardproducts at levels 2–4. This paper discusses the lessons learnedfrom analysis of the MODIS Land-Surface Temperature/Emissivity (LST) products in the current (V4) and previousversions in Section 2, presents eight new refinements for V5PGE16 in Section 3, the test results with real Terra and AquaMODIS data in Section 4, validation results and uncertaintyanalysis in Section 5, and conclusions in Section 6.
60 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
2. Lessons learned from V4 and previous versions of theMODIS LST products
Two main lessons have been learned from V4 and previousversions of the MODIS LST products.
A lesson learned since the beginning of the generation of theMODIS LST product is the difficulty of discrimination of clear-sky pixels from cloud pixels. In the MODIS LST product, LSTis the radiometric (kinetic) temperature related to the thermalinfrared (TIR) radiation emitted from the land surface observedby an instantaneous MODIS observation. The land surface heremeans the top of canopy in vegetated areas or soil surface inbare areas. Because the instantaneous MODIS TIR signal at oneviewing angle does not contain information at other viewingangles due to weak scattering of the TIR signals in the clear-sky
Fig. 1. Daytime and nighttime LSTs in V3 and V4 MODIS LST products n
atmosphere and no surface modeling is applied in the LSTprocessing, the retrieved LST is only the radiometric temper-ature at the given viewing angle. Because TIR signals cannotpenetrate clouds, LST is retrieved from MODIS TIR data onlyin clear-sky conditions so that LST is not mixed with cloud-toptemperature. Cloudy pixels must be skipped in the LSTprocessing by using the MODIS cloud mask product(MOD35_L2 from Terra MODIS or MYD35_L2 from AquaMODIS).
Although a series of spectral tests with multiple MODIS bandswere used in the generation ofM⁎D35_L2, it is still not possible tomake correct cloud masking in all cases. In the V3 PGE16, LSTswere retrieved only for the clear-sky pixels at the highestconfidence (99%) over both land and lakes. This approachresulted in very few LSTs in Lake Tahoe, California, where there
ear the middle of Lake Tahoe (top) and Namco lake (bottom) in 2002.
61Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
are many clear-sky days each year. Lake Tahoe does not freeze inthe winter. There are also many clear-sky days in the Tibet Plateau,where Namco is a large lake (1940 km2) at elevation beyond4700 m above the sea level and it does not freeze during thesummer. Therefore, any LSTs significantly below the freezingtemperature (273.15 K) anytime in Lake Tahoe and Namco in thesummer would be unrealistic (called cloud-contaminated LSTs).As shown in Fig. 1, there are a small number of nighttime LSTs inLake Tahoe (top) in 2002 and one of them is below others by morethan 12 K. Many nighttime LSTs during the summer are wellbelow the freezing point in Namco (bottom). In the V4 PGE16,LSTs were retrieved for lake pixels in clear-sky condition at aconfidence of 66% or larger. Therefore, we have more LSTs inLake Tahoe in the V4 MODIS LST product, and more LSTs arecontaminated by clouds, especially in the high-altitude lake,Namco in Tibet. Because the thresholds used in the spectral testsfor the MODIS cloud mask product do not depend on surfaceelevation and the same thresholds are used for ocean and lakes,there are more problems over high-altitude regions in the MODIScloud mask product. This product also faces challenges in areaswith high reflectivity values in the visible and near-infrared such asdeserts and snow covers. Because we do not have a better choice,the MODIS cloud mask product will also be used in V5 PGE16.However, we have to decide how to appropriately use the cloudmask product in PGE16 and to use other methods for removal ofthe cloud-contaminated LSTs in the MODIS LST product.
It is easy to understand that the day/night LST algorithm ismore sensitive to the cloud contaminations because this algo-rithm uses pairs of daytime and nighttime MODIS observationdata. If the probability of clear-sky during the day and at night is70%, then the probability of clear-sky in a pair of day and nightobservations would be less than half. In the V4 PGE16, the bandemissivities retrieved in the previous days are used as initialvalues of the emissivities in new retrievals. If there is no cloudcontamination, this is a good approach. However, the effect ofcloud contaminations may propagate to the retrievals withMODIS data in clear-sky days through such initialization. Incases of contaminations, the values of retrieved band emissiv-ities in bands 31 and 32 may be as low as 0.92, approximately0.05–0.06 below the normal values.
Table 1Major refinements implemented in the V5 daily LST code (PGE16)
No. Specification/action In V4
1 Clear-sky pixels defined by MODIS cloudmask At 99% cAt 66% c
2 Temporal averaging in the 1 km LST product (M⁎D11A1) Yes3 Grid size of LST/emissivities in M⁎D11B1 retrieved from
day/night algorithm5 km×5
4 Number of sub-ranges of zenith view angles 5 for the5 Effect of topographic slope in the M⁎D11B1 grid Not cons6 Option of combined use of Terra and Aqua data in the
day/night algorithmNo
7 Incorporate the split-window method into theday/night algorithm
Partiallyand varia
8 Removing cloud-contaminated LSTs Not impl
Another important lesson learned from the field campaignsconducted in the last decade is the difficulty of LST validation.Because of the large spatial variation in LSTs, especially during thedaytime, we have to carefully select validation sites. In order tomeasure in-situ LSTs at the 1 km scale with an uncertainty wellbelow 1K,we have to select large homogeneous sites with sizes ofat least 5 km by 5 km and use high-accuracy TIR radiometers inmeasurements at multiple points. These requirements cannot bemet by the measurements at most conventional weather stations ortower stations designed for other purposes although any temper-ature data collected at weather stations are helpful to qualitativelycheck the trend of seasonal variations in LSTs. According toMasuoka et al. (1998), the instantaneous field of view (IFOV) ofMODIS pixels in TIR bands is about 1 kmby 1 kmat nadir, 1.3 km(along-track) by 1.6 km (across-track) at viewing zenith angle near40°, and 1.7 km by 3.3 km at viewing angle of 60°. Larger sites areneeded to validate the MODIS LST product at large viewingangles. Taking into account that the spatial response function ofMODIS detectors in the across-track direction is a triangle cov-ering twice of the IFOV (Barnes et al., 1998), the area contributingto the radiance of a MODIS pixel is approximately 1.3 km by6.6 km at viewing angle of 60°. It is even more difficult to findsuitable validation sites for the coarse-resolution MODIS LSTproduct that contains the LST and band emissivities at 5 km or6 km grids retrieved by the day/night LST algorithm. It will be apractical approach to indirectly validate the coarse-resolution LSTproduct by comparing with the aggregated values of the 1 kmLSTproduct once the 1 km LST product is validated.
3. New refinements for the V5 MODIS LST code
In order to address the issues involved with the lessonslearned from the V4 LST products, most new refinements weremade upon the current version of the daily LST code (V4PGE16) to create the next version (V5) of PGE16 code for thereprocessing and forward processing of collection-5 MODISLST products. The new refinements are summarized in Table 1.The objective of these refinements is to provide more high-quality LST/emissivity data in the LST products while keepingthe cloud-contaminated LSTs to a minimal.
In V5
onfidence over land At confidence of N=95% over land b=2000 monfidence over lakes At confidence of N=66% over land N2000 m
At confidence of N=66% over lakesNo
km (exactly 4.63 km) 6 km×6 km (exactly 5.56 km)
whole scan swath 2×8 for the whole scan swathidered Considered in the QA
Yes
with initial Ta and cwv,bles of em31 and em32
Fully with em31, em32, Ta and cwv as variablesin the iterations
emented Implemented for M⁎D11A1 and M⁎D11B1
62 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
The V4 PGE16 code consists of two parts, the granule-basedPGE16A and tile-based PGE16B. TheV5 PGE16 code consists ofthree parts: revised PGE16A and PGE16B, plus a new part ofPGE16C. Input data to the granule-based PGE16A includecalibrated radiance (MOD021KM) and geolocation (MOD03),cloud mask (MOD35) and atmospheric profile (MOD07), land-cover (MOD12), snow cover (MOD10), and BRDF parameters inband 7 (MOD43). The granule-based PGE16A retrieves LSTs forclear-sky land pixels in each granule containing five minutes ofMODIS data with the generalized split-window algorithm (Wan&Dozier, 1996), which takes a formula similar to the split-windowmethod used for AVHRR data (Becker & Li, 1990). The values ofemissivities in bands 31 and 32, and the root-mean-square (rms)values of their mean and difference used in the split-window LSTalgorithm are listed in Table 2. They are based on the results ofclassification-based modeling (Snyder et al., 1998) with somesmall modifications. The values used for broadleaf forest, woodysavanna, grass savannas, and shrubs are approximately the averagevalues in their seasonal cycles from green to senescent because it isnot easy to detect greenness change and to estimate its impact onthe surface emissivity at the pixel scale. Recent laboratorymeasurements of the reflectance/emissivity spectra of singleleaves indicate that as leaves become dry the emissivity variesonly weakly in the split-window (10–13 μm, covering bands 31and 32) but notably in the mid infrared range (3–5 μm). Thisimplies that the change in greenness may impact band emissivitiesin the split-window and NDVI at different paces and levels.PGE16A calculates the coverage of each land pixel on 1 km gridsand selectively cumulates the contribution of radiance in band 31 atthe pixel's LST retrieved by the split-window method to itsoverlapping 1 km grids in the sinusoidal projection (Masuokaet al., 1998). The selective accumulation of the contribution of
Table 2The values of emissivities in bands 31 and 32, and the root-mean-square (rms)values of their mean and difference in the last two columns used in the split-window LST algorithm are based on the classification-based emissivity model(Snyder et al., 1998) with some small modifications
And the values of constraints (δT) on the temporal variations in clear-sky LSTsare used in PGE16C.
radiance means that we only select the radiance at a higher LSTvalue to avoid temporal average and possible effects of cloudcontaminations. In the sinusoidal projection, each tile covers aportion (10°×10° at the equator, approximately 1100 km×1100 km) of the earth's surface, and contains 1200×1200 grids,each in a size of 0.93 km×0.93 km, called 1 km grid. Thesinusoidal projection is selected as the common projection for thelevel-3 MODIS land products because all the grids in this projec-tion have equal areas. Finally the selectively cumulated radiancevalue is converted to LST_Day_1 km or LST_Night_1 km in thelevel-3 1 km LST product (MOD11A1). The corresponding view-ing time and angle are also stored in MOD11A1 for each grid.Similarly, cumulated radiance values at 6 km grids are converted tovalues of LST_Day_6 km_(or LST_Night_6 km_)Aggre-gated_from_1 km in the level-3 6 km LST product (MOD11B1).Note that the LSTs in MOD11A1 and LSTs in MOD11B1 may bein different viewing times because the solar zenith angle is anadditional factor in the selection for MOD11B1. PGE16A alsoselectively cumulates the contributions of radiance values in sevenbands and other related information of pixels to the 6 km grids, andsaves or updates these values in interim UPD (short for updated)files. The data inUPD files are used in the daily tile-basedPGE16B.PGE16B retrieved LSTs and surface emissivities in bands 20, 22,23, 29, and 31–32 with the day/night LST algorithm (Wan & Li,1997), which is a further development of the temperature-indepen-dent spectral index (TISI) method (Li & Becker, 1993) for acapability to scale atmospheric parameters for a better solution ofthe non-linear problem of LST and emissivity retrievals. The re-gression method used in the development of the day/night algo-rithm to assign the initial values is no longer used in the operationalprocessing because the initial values may be provided by the split-windowmethod or estimated from the inputMODIS data products.It should be pointed out that the day/night algorithm is not simplya two-temperature method because the pair of day and nightobservations used in the day/night algorithm must meet somespecial requirements: the solar zenith angle in the day observationcould not be larger than 75°, the solar zenith angle in nightobservation must be larger than 90° (in other words, no directsolar beam is seen at the land surface), and the dates of day andnight observations must be within 32 days. Under theserequirements, the solar radiation in the day observation in the3.5–4.2 μmmid-infrared region, whereMODIS bands 20, 22 and23 are located, plays a role of active infrared signal for theretrieval of surface reflectance values in these three bands by acombined use of the land-leaving radiance values in the nightobservation. By applying some realistic assumptions about thesurface optical properties, surface emissivities and temperaturesin the day and night observations can be retrieved while treatingthe values of column water vapor and surface air temperature inthe day and night observations as variables in the non-linearproblem. Because of the above requirements, the day/nightalgorithm can be used only in limited periods of time during theyear in high-latitude regions. Since the theoretical bases andmathematical methods of the twoMODIS LSTalgorithms (Wan&Dozier, 1996;Wan & Li, 1997) have not been changed and a largenumber of references were already given in these two papers, andthese two algorithms were also highlighted in two validation
63Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
papers (Wan et al., 2002b; Wan et al., 2004), details will not berepeated here. The new part of PGE16C removes cloud-contaminated LSTs from the daily level-3 1 km and 6 km LSTproducts. It will be explained later when discussing the lastrefinement in Table 1.
The logic behind the first refinement listed in Table 1 comesfrom visual inspection of the V4MODIS LST product and its inputproducts. Visualization of LST images reveals two problems withthe use of the 99% confidence defined in the MODIS cloud maskproduct as the criterion for clear-sky land pixels in the LSTprocessing. In the first problem, there are very sparely distributedblack pixels in the LST images of level-2 MODIS LST products insome daytime granules. These black pixels do not have a valid LSTvalue because they are skipped in the LST processing due to theflag value of Science Data Cloud_Mask in MOD35_L2 orMYD35_L2 indicating a 95% confidence of clear-sky for thesepixels. However, the radiance images in bands 31 and 32 indicatethat these pixels are in clear-sky conditions. In the second problem,a persistent contrast exists among lakes and their surrounding landsin the LST images over Tibet Plateau, especially at nights in thewinter. The lakes with valid LSTs are surrounded by land areaswithout valid LST value where the flag value of Cloud_Maskindicates 66% confidence of clear-sky. In order to avoid these twoproblems, the V5 LST processing is extended to more land pixels,which are in clear-sky conditions at a confidence of 95%and higherover land at elevations not beyond 2000m, or 66% and higher overland at elevations above 2000 m. Increased number of grids withcloud-contaminated LSTs in the level-3 LST products due to thisextension will be removed by execution of the new PGE16C.
The second refinement is related to the temporal averaging inthe 1 km level-3 LST product (MOD11A1 from Terra MODS dataand MYD11A1 from Aqua MODIS data). Temporal averagingwas made in the generation of the V4 1 km LST product but it willno longer be made in V5 per request of the user community. Afterthis change, LSTs in 1 km and 6 km products are all from instan-taneous MODIS observations. This change is also necessary tomake it easy to remove cloud-contaminated LSTs in the V5 1 kmLST products. It may be impossible to remove the cloud-con-taminated LSTs in the V4 MOD11A1 and MYD11A1 productsbecause the effect of cloud contaminations is reduced by thetemporal average.
Refinements 3–7 in Table 1 are related to the day/nightalgorithm.
Refinement 3 changes the grid size for LST/emissivity resultsretrieved from the day/night algorithm (Wan & Li, 1997) to6 km×6 km (exactly 5.56 km) from 5 km×5 km (exactly 4.63 kmin V4) so that it has the same 0.05° size of climate model grids inlatitude used for most CMG (Climate Modeling Grid) landproducts to avoid unnecessary uncertainties from re-sampling.
Refinement 4 increases the number of sub-ranges of viewingangles to 16 from 5 (in V4) in the interim data (UPD) file for theday/night algorithm without significantly increasing the file sizeby optimizing the data structure. In the geolocation file (MOD03),science data set SensorZenith gives MODIS viewing zenith anglein respect to the normal of Earth surface. The maximum viewingzenith angle is about 65° at the swath edges. In the UPD file, thedata structure has three dimensions, the first for rows (in latitude
direction), the second for columns (in longitude direction), and thethird for bins of viewing zenith angles. Each bin, corresponding toa sub-range of viewing zenith angles, contains a set of data for onedaytime observation (including the brightness temperatures ofradiances in seven bands) and a set of data for one nighttimeobservation. After increasing the number of sub-ranges, a max-imum of 16 sets of daytime and nighttime data can be stored foreach 6 kmgrid. This larger data set will reduce the chance of cloudcontaminations together with two special procedures in the tile-based PGE16B, one for selecting pairs of daytime and nighttimeobservations with higher brightness temperatures in band 31, andanother for setting the initial values of band emissivities at fixedvalues rather than with retrieved values which may be contam-inated by cloud effects.
Refinement 5 is to consider the effect of topographic slopesin the 6 km grids when assigning the QA of the LST/emissivityresults from the day/night algorithm.
Refinement 6 is to add an option for combined use of Terraand Aqua MODIS data in the day/night LST algorithm so that itis possible to use pairs of day and night observations at nearlyequal zenith angles in the same azimuth plane for the day/nightalgorithm because the Terra daytime orbit and Aqua nighttimeorbit are in the same descending node while the Terra nighttimeorbit and Aqua daytime orbit are in the same ascending node.
Refinement 7 in the V5 tile-based PGE16B code is to fullyincorporate the viewing-angle dependent generalized split-win-dow method into the day/night algorithm, using the variables ofemissivities in bands 31 and 32, the column water (cwv) and lowboundary air temperature (Ta) in the iterations of solution of theday/night algorithm. In this way, the split-window method is usednot only as constraints but also as a close component of the day/night algorithm, effectively increasing the weights of the highest-quality data of bands 31 and 32 for a better simultaneous retrievalof surface emissivity and temperatures. Note that the viewing-angle dependent generalized split-window method was alsoincorporated partially into the day/night algorithm in V4PGE16B code. A partial incorporationmeans that the incorporatedsplit-windowmethod always used the initial values of cwv and Tathe same as in the independent split-window algorithm in thegranule-based PGE16A.
With the last refinement implemented in the new part ofPGE16 code (PGE16C), cloud-contaminated LSTs in the V51 km product (MOD11A1 and MYD11A1) and the 6 km product(MOD11B1 andMYD11B1)will be removed by using constraintson the temporal variations (δT) in clear-sky LSTs in a period of32 days. The value of δT depends on land cover types as shown inthe first column of Table 2. There are three major steps in theremoving scheme. In step 1, remove the worst LSTs that aredifferent from the 32-day maximum by more than 4 times the δTvalue or different from the 16-day maximum bymore than 3 timesthe δT value. In step 2, remove the LSTs that are different from the8-day maximum by more than 2 times the δT value, then calculatethe 8-day average value of the remaining LSTs. In step 3, removethe LSTs that are different from the 8-day average value by morethan the δT value. In order to consider the larger natural temporalchanges in clear-sky LSTs in the growing and drying seasons, andin cold regions, the δTvalues in Table 2 are adjusted on the basis of
Fig. 2. Comparisons of the V4 (top row) and V5 (row 2) daytime (left) and nighttime (middle) LSTs and emissivity RGB images (right) in the MOD11B1 products intile h25v05, retrieved from Terra MODIS data acquired on 21 January 2003. In row 3 are NDVI (left), snow cover in red (middle) and landcover (right). The landcovertypes (referring to Table 2) are in color table: magenta for 0, yellow for 6, green for 7, red for 10, white for 15, and coral for 16. At the left side of bottom row is thesurface elevation image (the range of the color scale for surface elevation is from 900–6372 m). The geographic locations of the four corners (clockwise starting at thetop left) of tile h25v05 are (40° N, 91.372° E), (40° N, 104.427° E), (30° N, 92.380° E) and (30° N, 80.380° E). The geographic coverage of the tile is given by placingits corners on the geographic map at the bottom (Courtesy of Microsoft Mappoint online service).
64 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
65Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
statistical values of mean and standard deviation of LSTs in fourperiods of 8 days for each landcover type in the tile underprocessing. The selection of δT values in Table 2 was justified byempirical testing for different land cover types with one year of V5data in several tiles: at right choice of δT values, most of theremoved LSTs should be the grids near or within cloudy areaswhere there is no valid LST, and the grids within large contiguousareas of valid LSTs should not be removed. MOD11A1 andMOD11B1 files will be archived as final output products afterexecution of PGE16C.
MOD11A1 data in a period of 8 days will be used as input tothe 8-day LST PGE31 to composite and average the daily LSTvalues for the generation of the LSTs in the 8-day LST product(MOD11A2). The information ofwhich days have clear-skyLSTswill be stored in bit flags at the grid level. The logic for theMOD11A2 product is that some applications need spatialdistributions of LSTs as fully as possible at a coarser temporalresolution and the MODIS repeats its orbit patterns every 16 days.
The daily MOD11B1 data will be used as input to the dailyCMGPGE32 to generate the LSTs at climatemodeling grids (0.05°latitude×0.05° longitude each) in the daily CMG LST product(MOD11C1). In PGE32, LST_Day_6 km in MOD11B1 that isretrieved from the day/night algorithm is used as the primary datasource for scientific data set LSTDay_CMG. If LST_Day_6 km isnot available, LST_Day_6 km_Aggregated_from_1 km, which isoriginally from the split-window algorithm, will be used assecondary source. Similarly for scientific data set LST_Night_CMG in MOD11C1. If the values of LST_Day_CMG and LST_Night_CMG come from the aggregated data, the band emissivitiesinMOD11C1 have valid values only in bands 31 and 32, originallyestimated from land cover types. The dailyMOD11C1will be usedas input to CMG PGE58 and PGE59 to generate the 8-day andmonthly CMG LST products MOD11C2 and MOD11C3, respec-tively. Because all the 8-day andmonthly LST products are derivedfrom the daily LST products, we will focus on the refinementresults of the V5 daily PGE16 in this paper.
Some other minor changes were also made in the V5 code. Forexample, the surface air temperature will be calculated by inter-polation (as in v3.0.0 PGE16) from the atmospheric temperatureprofile in MOD07 because Surface_Temperature in new MOD07has been changed to regression-based retrieval of (land surface)skin temperature from the GDAS surface (air) temperature in theold MOD07. The values of bit flags for LSTand emissivity errorsinQuality Indicators QC_Day and QC_Night of the V5 6 kmLSTproduct (MOD11B1 and MYD11B1) are set in a better way withthe difference between the LSTs retrieved by the day/night and theincorporated split-window methods.
In the V5 LST processing, the cloud-contaminated LSTs in thelevel-2 product MOD11_L2 or MYD11_L2 are not removedbecause of the difficulty in dealing with both daily variation anddiurnal variation in the clear-sky LSTs. This issue will be resolvedin the future processing.
4. Test results of V5 PGE16 code
The V5 PGE16 code was tested with Terra MODIS data inJanuary 2003 in tile h25v05 (covering Tibet Plateau). The
refinement of considering surface elevation in deciding pixels forLST processing made a significant improvement in the spatialcoverage of the LSTs in daily M⁎D11A1, M⁎D11B1 and 8-dayM⁎D11A2 products, especially in highlands, as shown in thecomparisons of daytime and nighttime LSTs in the V4 and V5MOD11B1 products retrieved from Terra MODIS data acquiredon 21 January 2003 in tile h25v05 in the first two rows in Fig. 2,where the grids in black are clouds. In the top row, images of V4daytime and nighttime LSTs, and the emissivity RGB image arefrom left to right. There is no valid value in 42% of the total area inthe V4 daytime LST image, nearly 46% in the nighttime LSTimage, and 23% in the emissivity image. The correspondingimages of the V5 product on the day are in the second row. Thepercentages of areas without valid values are 5.2%, 8.2% and 0%,respectively. It should be pointed out that the scheme forremoving cloud-contaminated LSTs was applied to the V5 databut not to the V4 data. This clearly indicates that the refinementsignificantly improved the spatial coverage of LSTs, especially inhighland regions. The spatial correlations among these imagesand the NDVI, snow cover, land cover images (in the third row)and surface elevation image (at the bottom) indicates that theresults in the V5 LST product are reasonably good. The surfaceelevation is above 2000 m in the entire tile except two smallportions at the upper left and right corners. The geographic map(Courtesy of Microsoft Mappoint online service) at the bottomright side is included to show the geographic coverage in theh25v05 tile.
The V5 PGE16 code was also tested with Terra and AquaMODIS data in July and August 2002 in two tiles, h08v05mostly in dry weather condition, and h09v05 partly in dry in thewest portion and in wet weather conditions in the east portion.The daytime LSTs retrieved from Terra MODIS data on 25August 2002 in these two tiles are shown in Fig. 3. Thegeographic coverage of these two tiles can be explained byplacing their four corners on the geographic map (courtesy ofMicrosoft Mappoint online service) at the bottom.
Fig. 4 shows the correlation between the retrieved emissiv-ities and land cover types and NDVI. The emissivities in bands22 and 29 are shown in Fig. 4(a) and (b). Fig. 4(c) shows thecolor composite with emissivities in bands 22, 29 and 31 asRGB components. Fig. 4(d) is the color composite with thesame components, each enhanced with the histogram equaliza-tion method before composition. The image of land cover typesis shown in Fig. 4(e) and image of NDVI in the period of days ofyear from 193–208 in Fig. 4(f). From Figs. 2 and 4, we can seethat the emissivities retrieved by V5 PGE16 are more reliableand well correlated with landcover and NDVI although NDVI isnot used as input of PGE16. The bright areas in the emissivityRGB images (representing high emissivities in all three bands)are forests, lakes, and dense vegetation areas. Similar resultswere also found in additional tests made with MODIS data inNorth Africa and Europe. Emissivities in bands 22 and 29 varyin wide regions from lower than 0.85 up to 0.985 in bare soiland rock areas and open shrublands.
Fig. 5 shows the comparisons between daily LSTs retrieved bythe day/night method and the incorporated split-window methodwith MODIS data collected on 4th July 2002 (day of year 185) in
66 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
tiles h08v05 and h09v05. The mean and standard deviation of thedifferences are −0.022 K and 0.14K for daytime LSTs,−0.051 Kand 0.22 K for nighttime LSTs in tile h08v05. In h09v05, they are−0.014 K and 0.20 K, −0.07 K and 0.33 K. It is interesting topoint out that the distributions shown in Fig. 5 are not close toGaussian, rather leptokurtic (their values of kurtosis Kurt(x)=E[(x−μ)4] /σ4, where μ and σ are mean and standard deviationof variable x, are larger than 9). These distributions of the dif-ferences have extremely high peaks near zero difference, indi-
Fig. 3. Daytime LSTs in MOD11B1 retrieved fromMODIS data on 25 August 2002 i(in h08v05) or clouds (in both). The geographic locations of the four corners (cloclongitude pairs of (40° N, 130.545°W), (40° N, 117.492°W), (30° N, 103.937°W) anof tile h08v05. The longitude values of the corners of tile h09v05 on the right edgedefined by placing these corners on the geographic map at the bottom (Courtesy of
cating the incorporated split-window method performs very wellwithin the day/night algorithm when it uses not only the variablesof em31 and em32 but also the variables of Ta and cwv in theiterations. These good test results also indicate that the newprocedures for selecting pairs of daytime and nighttime observa-tions and setting initial values of band emissivities at fixed valueswork well in the V5 PGE16B code. The maximum standarddeviation of the differences in LSTs retrieved by the day/nightmethod and the incorporated split-window method is less than
n tiles h08v05 (upper left) and h09v05 (upper right). The areas in black are oceankwise starting from the top left corner) of tile h08v05 are defined by latitude/d (30° N, 115.485°W). The left edge of tile h09v05 is connected to the right edgeare 104.438° and 92.388° W. The geographic coverage of these two tiles can beMicrosoft Mappoint online service).
Fig. 4. Surface emissivities retrieved from MODIS data in August 2002 (a–d, see details in text), image of land cover types (e), and image of NDVI in the period ofdays of year from 193–208 (f). The range of scale for emissivities in (a) and (b) is from 0.492 to 1.0. The land cover types in (e) are expressed in color table: magentafor 0, yellow for 1, green for 5, red for 6, white for 7, and coral for 8, forest for 9, plum for 10, cyan for 12, turquoise for 13 and blue for 16. The range of color scale forNDVI in (f) is from −0.2 to 0.85.
67Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
0.4 K for all days in July and August 2002. This indicates that therefined day/night algorithm consistently gives reliable results. Ifthe portion of incorporated split-window method is deleted fromthe V5 PGE16B code, the day/night algorithm becomes lessreliable and the errors in the retrieved emissivities and tem-peratures get larger, the retrieved daytime and nighttime LSTswould be overestimated by 1–2 K and the retrieved emissivityvalues in all bands would be underestimated by 0.02–0.03 inmostcases.
5. Validation and uncertainty analysis
There are two kinds of methods in the validation of the LSTproduct retrieved from satellite thermal infrared data: conven-tional temperature-based method and advanced radiance-basedmethod.
In the temperature-based method, multiple TIR radiometersare used to measure the surface radiometric temperature. Effectsof surface emissivity and reflected atmospheric radiation arecorrected to obtain the in-situ measured LST using the emissivityvalue based on landcover and/or sample measurements andatmospheric radiative transfer simulations. Comparisons betweenthe in-situ LSTs and MODIS LSTs give the accuracy estimate ofMODIS LSTs. This method has been used in validation field
campaigns over large lakes, grasslands and agricultural fields(Coll et al., 2005; Wan et al., 2002b;Wan et al., 2004). Due to thelarge spatial variation in LSTs, especially during daytime, it isvery difficult to make ground-based measurements of LSTs at the1 km scale reaching a required accuracy (say, better than 1K)withsingle or small number of TIR radiometers in most sites includingweather stations.
In the radiance-based method, radiosonde balloons arelaunched to measure the atmospheric temperature and watervapor profiles at a validation site around the MODIS overpasstime. Based on the measured atmospheric profile, MODIS LST(the LST value at the validation site in the MODIS LST product)and the spectral emissivity value measured in the field orestimated from landcover and/or sample measurements, anatmospheric radiative transfer simulation is made to calculatethe top of atmospheric (TOA) radiance at the MODIS viewingangle in MODIS band 31. According to the difference betweenthis calculated TOA radiance and the MODIS radiance value,land-surface (kinetic) temperature is changed for anotheratmospheric radiative transfer simulation that is used to calculateone more TOA radiance value. The theoretically correct LSTvaluemay be estimated by a linear interpolation of these two TOAradiance values versus corresponding LST values at the MODISradiance value. The difference between the MODIS LST value
Fig. 5. Comparisons of LSTs retrieved from Terra MODIS data collected on 4 July 2002 (day of year 185) by the day/night and incorporated split-window methods intiles h08v05 (top) and h09v05 (bottom).
68 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
and the theoretical LST value is the accuracy estimate of theMODIS LST. This theoretically calculated LST is also calledLSTin-situ in the radiance-based approach because it is based onthe in-situ measurement of the atmospheric profile. The mainadvantage of this method is that it works for both daytime andnighttime because in-situmeasurements of LSTs are not needed inthis method and it is relatively easy to find validation sites withsmall spatial variations in land-surface emissivity. The radiance inMODIS band 31 is used in the calculation of LSTin-situ because theeffect of variations in atmospheric water vapor and temperatureprofiles, and solar radiation on the TOA radiance is smallest in thespectral range of band 31, and the uncertainty in surfaceemissivity is also smaller in this band. This method requiresaccurate atmospheric temperature and water vapor profiles and anaccurate atmospheric radiative transfer code such as MOD-TRAN4.0 (Berk et al., 1999).
Surface emissivity spectra were measured with a sun-shadowmethod in two field campaigns, one in Railroad Valley playa,
Nevada in late June 2003, another in a grassland larger than 10 kmby 10 km inDallamCounty, Texas inApril 2005. The coverage ofgrass is larger than 80% in the grassland. A thermal infraredinterferometric spectroradiometer was deployed in the centralportion of Railroad Valley playa (38.4817°N, 115.6905°W) or atthe middle of grassland (36.2992° N, 102.5706° W) to measurethe surface-leaving radiance spectra in the spectral region from3.5–14 μm under sunshine and sun-shadow conditions. Twoblackbodies, one at ambient temperature and the other at highertemperature were used to calibrate the radiance to accuracy betterthan 0.1 K for the brightness temperature of the radiance. Thesolar and environmental downward irradiance under sunshine andsun-shadow conditions were measured with a sand-blastedaluminum plate placed above the surface target. This aluminumplate reflects the solar radiation like a diffuse reflector as long asthe solar incident angle is not far from the nadir. The spectralreflectivity of the plate was measured in laboratory. Averagedradiance values in sevenMODIS bands (bands 20, 22, 23, 29, and
69Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
31–33) were calculated from themeasured radiance spectra of thesurface target and aluminum plate under sunshine and sun-shadow conditions with the relative spectral response functions ofthe MODIS bands as weights. Based on these averaged radiancevalues, the temperatures of the surface target under sunshine andsun-shadow conditions, and the band emissivities of the surfacewere retrieved by a sun-shadow method, which is similar to butsimpler than the day/night method used in the retrieval of theMODIS LST product with pairs of daytime and nighttimeMODIS observations. After the surface temperature under sun-shadow condition was determined, spectral emissivity of thesurface can be retrieved from themeasured radiance spectra of thesurface and aluminum plate under sun-shadow conditions. Thequality of the retrieved emissivity spectra depends on the stabilityof the atmospheric condition during the whole measurementprocess of the radiance spectra of the surface target and thealuminumplate under sunshine and sun-shadow conditionswhichtakes about 10 minutes. The error in the measured emissivityspectra in the 10.5–12.5 μm range is less than 0.005 in the normalclear-sky weather conditions. More details of the sun-shadow
Fig. 6. Emissivity spectra measured by sun-shadow method in Railroad Valley (top) aproduct are shown for comparisons.
method will be presented in another paper. The emissivity spectraof the playa and the grassland measured by the sun-shadowmethod are shown in Fig. 6. The band emissivities retrieved by theday/night algorithm from Terra MODIS data are also shown forcomparisons. The difference between emissivities retrieved fromMODIS data and measured by the sun-shadow method in thesetwo field campaigns is less than 0.0075 in the 10–12.5 μm range.A part of the difference attributes to the different spatial scales.
Six radiosonde balloons were launched to measure the tem-perature and water vapor profiles in the Railroad Valley fieldcampaign in 2003. Three radiosonde balloons were launched inthe grassland field campaign in April 2005. A radiosonde system(that consists of receiver/processor PP15, antenna RM20, andradiosonde sensor RS80) manufactured by Vaisala was used tomeasure the atmospheric temperature and water vapor profiles.The technical specifications of the radiosonde system are sum-marized as follows: pressure measurement resolution 0.1 hPa andprecision 0.5 hPa in the range of 3–1060 hPa, temperaturemeasurement resolution 0.1 °C, and precision 0.2 °C in the rangefrom −90 to 60 °C, relative humidity (RH) measurement
nd the grassland in northern Texas (bottom). Band emissivities in the MOD11B1
Table 3Radiance-based validation of LSTs in the V5 level-2 Terra (T) and Aqua (A) MODIS LST products
The δT value in the parentheses after MODIS LST is the standard deviation of LSTs in the four pixels surrounding the validation site.
Fig. 7. Comparisons of LST values in the V5 MOD11_L2 and MYD11_L2products with the in-situ values in Coll et al. (2005); Wan et al. (2002a,b); Wanet al. (2004), and radiance-based validation in Table 3.
70 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
resolution 1% RH and precision 3% RH in the range of 0–100%RH. A temperature-dependent correction was made to theradiosonde RH measurements from Vaisala RS80-A thin-filmcapacitive sensors due to the problem at cold temperature(Miloshivich et al., 2001). Based on the measured atmosphericprofiles and the surface emissivity spectrum, and the LST valuesat the location of the validation sites in the V5 MODIS level-2LST products, radiative transfer simulations were made with theMODTRAN4 code to calculate the TOA radiances for theradiance-based validation of theMODISLSTproduct. The detailsof validations for eight cases of Terra MODIS and nine cases ofAqua MODIS are listed in Table 3. The major sources ofuncertainties in the radiance-based validation include effects oferrors in the surface emissivity spectrum measured by the sun-shadow method, in the measured atmospheric temperature andwater vapor profiles, in the radiative transfer simulations, and inthe MODIS calibrated radiance in band 31. The effect ofuncertainties (about 0.005) in the surface emissivity spectra isabout 0.3 K. In a typical atmospheric profile (column water vaporat 2.2 cm and lower boundary temperature at 310K), the effects ofincreasing column water vapor by 10% and air temperatures by1 K on the brightness temperature of TOA radiance in band 31 are−0.2 K and 0.2 K, respectively. Following the same approach oferror analysis in the estimate of calibration accuracy of MODISTIR bands (Wan et al., 2002a), the root sum squares (RSS) oferrors is 0.5 K if 0.3 K is taken as the error related to surfaceemissivity and 0.2 K is considered four times (two times for theuncertainties in atmospheric profiles and two more times for theuncertainties in radiative transfer simulations). It is reasonable toassume that the major source in the differences is due to thetemporal variations in the atmospheric profiles, especially in thewater vapor profiles. Typically it takes about 2 h for a radiosondeballoon to reach the elevation level about 20 km above thesurface. Therefore, the measured profiles may not be the realatmospheric conditions at the times of instantaneous MODISobservations. If we believe that the temporal variation in
atmospheric profiles within 2 h is smaller at night than duringthe daytime, then the smaller differences (around −0.3 K)betweenMODIS LSTand LSTin-situ values in the nine night cases(four cases for Terra MODIS and five cases for Aqua MODIS)would be the representative of the accuracy of the MODIS LSTproduct in homogeneous areas.
The V5 PGE16A code was used to generate the level-2 andlevel-3 1 km LST products from V5 input data of MODIScalibrated radiance, geolocation, cloudmask and atmosphericprofile products on the dates when in-situ LSTs are available.The difference between V4 and V5 calibrated radiance values inthe seven bands used by the LST algorithms is less than 0.1 K inmost cases. The effect of the difference between V4 and V5
71Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
atmospheric profile products on the LST product is less than0.2 K in general. The MODIS data granules used in Table 2 ofvalidation paper (Wan et al., 2002b), Table 2 in another paper(Wan et al., 2004), and Table 7 in the third validation paper(Coll et al., 2005) were processed with the V5 PGE16 code.Note that the first case in Table 2 of Wan et al., 2002b is notconsidered here because two of the four TIR radiometers used inthe in-situ measurements are too close to Paoha Island in themiddle of Mono Lake. The comparisons between the MODISLST values and the in-situ LST values are shown in Fig. 7. Asshown in Table 3 and Fig. 7, the LST values in the MODIS LSTproduct are underestimated in most cases because the effect ofaerosols above the average loading corresponding to the defaultsurface visibility of 23 km was not considered in the devel-
Fig. 8. Histograms of differences in LSTs in the daily 6 km LST product (MOD11B1)of January 2003 in tile h25v05 (bottom).
opment of generalized split-window algorithm and day/nightalgorithm. Comparisons between V5 LSTs and in-situ values in47 clear-sky cases (in the LST range from −10 °C to 58 °C andatmospheric column water vapor range from 0.4 to 3.5 cm)indicate that the accuracy of the MODIS LST product is betterthan 1 K in most cases (39 out of 47) and the root of meansquares of differences is less than 0.7 K for all 47 cases or 0.5 Kfor all but the 8 cases apparently with heavy aerosol loadings.
Fig. 8 shows the histograms of the differences between dailydaytime LSTs in solid-line (and nighttime LSTs in dot-line)retrieved by the day/night algorithm and the independentgeneralized split-window algorithm with Terra MODIS data inall days of August 2002 in tiles h08v05 and h09v05 based on allV5 MOD11B1 products in the month in the top two plots. The
in all days of August 2002 in tiles h08v05 (top), h09v05 (middle), and in all days
72 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
mean and standard deviation of differences are less than 0.1 Kand 0.5 K. The similar histograms for tile h25v05 based onTerra MODIS data in all days of January 2003 are shown at thebottom, the mean and standard deviation of differences are lessthan 0.2 K and 0.5 K. The sources of the temperature differencesinclude the errors in landcover-dependent classification-basedemissivities in bands 31 and 32, errors in both LST algorithms.The main sources of errors in the day/night algorithms include:(1) besides the two bands (bands 31 and 32) used in the split-window algorithm, five more bands (bands 20, 22, 23, 29 and33) used in the day/night algorithm have larger calibrationerrors (Wan et al., 2002a) and noise equivalent temperaturedifference (NEDT) values (Wan et al., 2004); (2) the effect ofdifferent sub-pixel temperature components are larger in themid-infrared bands (bands 20, 22 and 23); (3) although MODISachieved a sub-pixel geolocation accuracy, approximately 50 m
Fig. 9. Daytime and nighttime LSTs in V5 MOD11A1 products near t
(1σ) at nadir (Wolfe et al., 2002), the error in TIR radiancevalues at equal-area grids in the sinusoidal projection dependson the spatial variation in LSTs because one location-fixed gridmay be covered by variable portions of multiple pixels withdynamical locations and variable sizes dependent on observa-tion time and viewing angles, and therefore, the radiance valueof a grid unavoidably contains the contributions from theportions outside the grid; (4) the surface emissivity at a givengrid may be different at the times of a pair of daytime andnighttime MODIS observations due to high frequency eventssuch as raining, snowing, and dew occurrence at night. It is hardto make a quantitative analysis of the last two error sources(3 and 4) without good knowledge of the LST and surfaceemissivity distributions at fine sub-pixel resolutions. In fact,these two error sources also affect the level-3 1 km LST product(M⁎D11A1) which originates from the split-window algorithm.
he middle of Lake Tahoe (top) and Namco lake (bottom) in 2003.
73Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
The actual spatial resolution of LSTs in M⁎D11A1 depends onthe viewing angles of the MODIS observations, it retains the1 km resolution in the areas observed by MODIS at and nearnadir, but it may be much coarser than 1 km at grids observed byMODIS at large viewing zenith angles near the swath edges ofMODIS granules. Because coarser grids at the size of 6 km by6 km are used for the temperature/emissivity product retrievedfrom the day/night algorithm, the relative percentage ofcontributions from variable portions outside the grid is smaller.The effect of dew occurrence or change in the surface moisturecondition on surface emissivity is smaller in the spectral rangeof 10–13.5μm covered by bands 31–33 because all land covertypes already have high emissivities in this spectral range evenat dry conditions. In order to reduce the effect of dewoccurrence on the performance of the day/night algorithm,smaller weights are assigned to bands 20, 22, 23, and 29 of thenighttime observation in the solution of the day/night algorithm.It is found from analysis of time series of M⁎D11B1 productthat the maximum random error in the retrieved emissivities inbands 20, 22 and 23 is approximately 0.04 and it corresponds toa maximum temperature error of 1 K in these bands. The error inthe retrieved nighttime LSTs may be reduced by using smallerweights in these mid-infrared bands. Small temperaturedifferences shown in Fig. 8 mean that the LSTs retrieved bythe split-window method using the classification-based emis-sivities in bands 31 and 32 are not far from the LSTs retrievedby the day/night algorithm using pairs of day and night MODISdata in seven TIR bands plus the information provided inMOD07, MOD10, MOD35 and MOD43 products. Althoughthese two algorithms give compatible LSTs at different spatialresolutions, only the day/night algorithm can retrieve emissivityin the mid-infrared range (covered by bands 20, 22 and 23) andband 29. Emissivities in these bands cannot be accuratelyestimated from land cover types because of their highvariations. Emissivity in the mid-infrared range is useful ingeology studies and in monitoring of water requirements bycrops and soil moisture conditions. Emissivity in band 29 isnecessary in geology studies and in the estimation of broadbandemissivity for calculation of long wave radiation (Wang et al.,2005).
Fig. 8 indicates that it is possible to use the validated LSTs inlevel-2 and level-3 1 km LST products (Wan et al., 2002b,2004) for indirect validation of the LSTs in 6 km LST productsthrough the scientific data set Aggregated_from_1 km in the6 km LST products in relatively homogeneous areas. However,it is not always possible to use this approach in V4 due to thetemporal average in the generation of 1 km LST products.
Fig. 9 shows the daytime and nighttime LSTs in the V5MOD11A1 product near the middle of Lake Tahoe and Namco(lake) in 2003. It is not used for a point by point comparisonwith Fig. 1 because different years of data are used in these twofigures. There are two reasons to use the 2003 data in Fig. 9.First, one year of V5 MOD11A1 product is available only in2003 at this moment. Second, Fig. 9 is used to show that it is notpossible to remove all cloud-contaminated LSTs. Although thethird part of the PGE16 was executed to remove the cloud-contaminated LSTs in the V5 MOD11A1 product, there is still a
nighttime LST below the freezing temperature in the middle ofnon-frozen Lake Tahoe. Obviously this low value is not truelake surface temperature. As mentioned in Section 3, thescheme for removing cloud-contaminated LSTs uses constraintson the temporal variation in clear-sky LSTs in a period of32 days. As shown in Fig. 9, there are only two nighttimetemperatures in the second 32-day period (from days of year 33to 64) for Lake Tahoe, one at 278.5 K on day 41 and another at267.9 K on day 59. The low value (267.9 K) is a single point notonly in the last 8-day period but also in the second 16-dayperiod of this 32-day period. Although this point may beremoved by considering spatial features of grids with validLSTs, the scheme for removing cloud-contaminated LSTs needsreal clear-sky LSTs as references so it would fail in cloud-prevailing seasons, especially in rainy seasons in Amazon andother tropical regions. Therefore, the quantity and the quality ofthe MODIS LST product depend on weather conditions becauseof the inherent limitation of the thermal infrared remote sensing.Note that the removing scheme cannot remove slightly andmodestly cloud-contaminated LSTs if the effect of cloudcontaminations is smaller than the normal temporal variationsin the clear-sky LSTs. The effect of cloud contaminations on theretrieved band emissivities may still exists even the removingscheme has been applied to the MOD11B1 product because theeffects of cloud contaminations at a sub-grid level may be smallon the LST retrieval but modest on the emissivity retrieval. It isbetter to check the spatial distribution before using the retrievedemissivities. If the emissivities exist over large contiguousareas, their quality usually is good. However, the uncertainty inretrieved emissivities may be large if the emissivities only existin small disconnected patches or near cloudy areas that are blackin the emissivity images. Because no action is taken for thecloud-contaminated LSTs in the V5 level-2 LST product(MOD11_L2 or MYD11_L2), users of this product should beaware of the cloud-contamination problem. It is wise to checkthe LST spatial distribution and values in the daily level-3 LSTproduct, 1 km MOD11A1 or 6 km MOD11B1, before using theLSTs of MOD11_L2 in serious applications.
6. Conclusions
It has been a tough job to generate more high-quality LSTsfrom TIR data and keep the cloud-contaminated LSTs to aminimal in an operational product. New refinements for V5LST product generation executive code (PGE16) were made toaddress the issues involved in the lessons learned in the past andtested with real Terra and Aqua MODIS data. These refinementssignificantly improved the spatial coverage of LSTs, especiallyin highland regions, and the accuracy and stability of theMODIS LST products. The mean and standard deviation of thedifferences between the LSTs retrieved from the day/nightmethod and the independent generalized split-window methodwith MODIS data in January 2003 in tile h25v05 (coveringTibet Plateau) are less than 0.2 K and 0.5 K. They are less than0.1 K and 0.5 K for the LSTs retrieved from Terra MODIS datain August 2002 in tiles h08v05 and h09v05 (covering southweststates of the US). Comparisons between V5 LSTs and in-situ
74 Z. Wan / Remote Sensing of Environment 112 (2008) 59–74
values in 47 clear-sky cases (in the LST range from −10 °C to58 °C and atmospheric column water vapor range from 0.4 to3.5 cm) indicate that the accuracy of the MODIS LST product isbetter than 1 K in most cases (39 out of 47) and the root of meansquares of differences is less than 0.7 K for all 47 cases or 0.5 Kfor all but the 8 cases apparently with heavy aerosol loadings.Emissivities retrieved by the day/night algorithm are wellcompared to the surface emissivity spectra measured by a sun-shadow method in two field campaigns. Evaluations of the V5MODIS LST product over two sites (Lake Tahoe in Californiaand Namco lake in Tibet) show that the quantity and quality ofMODIS LST products depend on clear-sky conditions becauseof the inherent limitation of the thermal infrared remote sensing.Other techniques such as passive microwave remote sensingand modeling methods are needed for applications requiringLSTs every day and everywhere.
Acknowledgment
This work was supported by EOS Program contract NNG04-HZ15C of the National Aeronautics and Space Administration.The author would like to thank three anonymous reviewers fortheir valuable comments and suggestions, which helped to im-prove the paper.
References
Barnes, W. L., Pagano, T. S., & Salomonson, V. V. (1998). Prelaunchcharacteristics of the Moderate Resolution Imaging Spectroradiometer(MODIS) on EOS-AM1. IEEE Transactions on Geoscience and RemoteSensing, 36, 1088−1100.
Becker, F., & Li, Z. -L. (1990). Toward a local split-window method over landsurface. International Journal of Remote Sensing, 3, 369−393.
Berk, A., Anderson, G. P., Bernstein, L. S., Acharya, P. K., Dothe, H., Matthew,M. W., Adler-Golden, S. M., Chetwynd, J. R., Richtsmeier, S. C., Pukall, B.,Allred, C. L., Jeong, L. S., & Hoke, M. L. (1999). MODTRAN4 radiativetransfer model for atmospheric correction. SPIE Proceeding, OpticalSpectroscopic Techniques and Instrumentation for Atmospheric and SpaceResearch III, 3756, 348−353.
Coll, C., Caselles, V., Galve, J. M.., Valor, E., Niclòs, R., Sánchez, J. M., et al.(2005). Ground measurements for the validation of land surface tempera-tures derived from AATSR and MODIS data. Remote Sensing ofEnvironment, 97, 288−300.
Li, Z. -L., & Becker, F. (1993). Feasibility of land surface temperature andemissivity determination from AVHRR data. Remote Sensing of Environment,43, 67−85.
Masuoka, E., Fleig, A., Wolfe, R. E., & Patt, F. (1998). Key characteristics ofMODIS data products. IEEE Transactions on Geoscience and RemoteSensing, 36, 1313−1323.
Miloshivich, L. M., Vömel, H., Paukknen, A., Heymsfield, A. J., & Oltmans, S. J.(2001). Characterization and correction of relative humidity measurementsfrom Vaisala RS80—A radiosondes at cold temperatures. Journal ofAtmospheric and Oceanic Technology, 18, 135−156.
Salomonson, V. V., Barnes, W. L., Maymon,W. P., Montgomery, H., & Ostrow, H.(1989). MODIS: Advanced facility instrument for studies of the Earth as asystem. IEEE Transactions on Geoscience and Remote Sensing, 27, 145−153.
Snyder, W. C., Wan, Z., Zhang, Y., & Feng, Y. -Z. (1998). Classification-basedemissivity for land surface temperature measurement from space. Interna-tional Journal of Remote Sensing, 19, 1−16.
Wan, Z., & Dozier, J. (1996). A generalized split-window algorithm forretrieving land-surface temperature from space. IEEE Transactions onGeoscience and Remote Sensing, 34, 892−905.
Wan, Z., & Li, Z. -L. (1997). A physics-based algorithm for retrieving land-surface emissivity and temperature from EOS/MODIS data. IEEE Transac-tions on Geoscience and Remote Sensing, 35, 980−996.
Wan, Z., Zhang, Y., Li, Z. -L., Wang, R., Salomonson, V. V., Yves, A., et al.(2002a). Preliminary estimate of calibration of the Moderate ResolutionImaging Spectroradiometer thermal infrared data using Lake Titicaca. Re-mote Sensing of Environment, 80, 497−515.
Wan, Z., Zhang, Y., Zhang, Y. Q., & Li, Z. -L. (2002b). Validation of the land-surface temperature products retrieved from Moderate Resolution ImagingSpectroradiometer data. Remote Sensing of Environment, 83, 163−180.
Wan, Z., Zhang, Y., Zhang, Y. Q., & Li, Z. -L. (2004). Quality assessment andvalidation of the global land surface temperature. International Journal ofRemote Sensing, 25, 261−274.
Wang, K., Wan, Z., Wang, P., Sparrow, M., Liu, J., Zhou, X., et al. (2005).Estimation of surface long wave radiation and broadband emissivity usingModerate Resolution Imaging Spectroradiometer (MODIS) land surfacetemperature/emissivity products. Remote Journal of Geophysical Research,110, D11109. doi:10.1029/2004JD005566
Wolfe, R. E., Nishihama, M., Fleig, A. J., Kuyper, J. A., Roy, D. P., Storey, J. C.,et al. (2002). Achieving sub-pixel geolocation accuracy in support ofMODIS land science. Remote Sensing of Environment, 83, 31−49.
