Hourly LST Monitoring with the Japanese Geostationary

Preparing to load PDF file. please wait...

0 of 0
Hourly LST Monitoring with the Japanese Geostationary

Transcript Of Hourly LST Monitoring with the Japanese Geostationary

Hourly LST Monitoring with the Japanese Geostationary Satellite MTSAT-1R over the Asia-Pacific Region
Kei OYOSHI1*, Shin AKATSUKA2, Wataru TAKEUCHI3 and Shinichi SOBUE1
1Earth Observation Research Center, Japan Aerospace Exploration Agency, 2-1-1 Sengen, Tsukuba, Ibaraki, Japan 2Yamanashi Institute of Environmental Sciences, 5597-1 Kenmarubi, Kamiyoshida, Fujiyoshida, Yamanashi 403-0005, Japan
3The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, 153-8505, Japan

Land Surface Temperature (LST) is a significant indicator of the energy balance at the Earth’s surface, and measurements of LST are required for a wide variety of climate, hydrological, ecological, and biogeochemical studies. As LST is temporally highly variable, an hourly LST map enables an improved understanding of the surface energy balances. In this study, hourly LST was estimated from thermal infrared data of the Japanese geostationary satellite, the Multi-functional Transport Satellite1R (MTSAT-1R). Numerical coefficients of the Generalized Split-Window equations were optimized based on the results of radiative transfer simulations. Model accuracies were less than, or equal to 1.09 K for a viewing zenith angle (VZA) lower than 40˚ and 1.73 K for a VZA of 60˚. Using a comparison of clear-sky images of six different climatic, land cover, and viewing angle regions (Australia, Indonesia, Japan, Mongolia, Papua New Guinea, and Thailand), the RMSE between MTSAT and MODIS LST over these regions was found to range from 2.32 K to 2.86 K, and the bias from -1.22 K to 1.46 K. However, a seasonal stability analysis on a daily basis in 2007 and 2008, demonstrated that consistency largely depends on the amount of cloud in each pair of MTSAT and MODIS LST images. RMSE were within 2 K or 3 K in clear-sky scenes comparison, but worse than 5 K in cloudy scenes. It is therefore considered that to utilize MTSAT LST more effectively and reliably, cloud contamination needs to be assessed, and a precise identification of cloud mask is required to eliminate MTSAT LST with high uncertainties.
Key words: LST, MTSAT, split-window algorithm, radiative transfer simulation, MODIS LST product.

1. Introduction
Land Surface Temperature (LST) is one of the most significant parameters used in understanding the energy balance, snow or ice melt, evapotranspiration, and vegetation growth at the Earth's surface. Hence, LST measurements are required for a wide variety of climate, hydrological, ecological, and biogeochemical studies (Parkinson et al., 2000). Until now, LSTs on a continental or global scale have mainly been retrieved using thermal infrared radiation measured by sensors onboard polar-orbiting satellites, such as NOAA Advanced Very High Resolution Radiometer (AVHRR) or Terra/Aqua Moderate Resolution Imaging Spectroradiometer (MODIS) (Price, 1984, Ulivieri et al.,
2014 AARS, All rights reserved. * Corresponding author: [email protected]
Phone: +81-50-3362-7982, Fax: +81-29-868-2961

1994, Wan et al., 1996). However, LST is highly variable temporally and has a large diurnal cycle, and therefore using a temporal resolution of 0.5-1 day is often inadequate for many applications such as numerical weather forecasting, urban heat island monitoring, ecosystem process modeling, or crop growth simulation modeling. Such applications require hourly LSTs or daily min/max/average LSTs for accurate execution. However, geostationary satellites can collect hourly data, and in the case of the Meteosat Second Generation (MSG) Spinning Enhanced Visible Infra-Red Imager (SEVIRI) satellite one image is acquired every 15 min. In addition, the Himawari-8/9 Advanced Himawari Imager (AHI), successor of Multi-functional Transport Satellite (MTSAT) is to be launched in 2014/ 2015, and has

Hourly LST Monitoring with the Japanese Geostationary Satellite MTSAT-1R over the Asia-Pacific Region

the capability of acquiring one image every 10 min using 16 bands. Obtaining such hourly LST data would enhance our understanding of thermal conditions at the Earth’s surface. Although land-surface monitoring by geostationary satellites has been limited because of poor sensor specifications, recent improvements in this respect have facilitated the use of these sensors, not only for atmospheric monitoring for use in weather forecasts, but also for land-surface monitoring (Fensholt et al., 2006). Accordingly, diurnal LST variations can now be retrieved from images obtained by geostationary satellites such as the Geostationary Meteorological Satellite (GMS) (Oku et al., 2004, Prata et al., 1999), the Geostationary Operational Environmental Satellite (GOES) (Pinker et al., 2008), and the MSG satellite (Sobrino et al., 2004).
In this study, we present a methodology for use in retrieving hourly LST from the Japanese geostationary satellite MTSAT, which covers the western Asia-Pacific area. Figure 1 shows the framework for use in retrieving LST from MTSAT. Our methodology progresses as follows: firstly, a split-window algorithm for polar-orbiting satellites data was optimized for the MTSAT sensor, and was based on at-sensor measurements of radiative transfer simulations. The optimized split-window algorithm was then used to estimate LST over six selected Asia-Pacific regions from thermal infrared data obtained from MTSAT, and from emissivity data of MODIS LST products. Finally, MTSAT LST products were compared with MODIS LST products, and the usefulness of hourly LST monitoring with MTSAT was demonstrated.

2. Methodology
2.1 Data and Study Area
MTSAT-1R is a Japanese geostationary satellite that was launched on February 26, 2005, and although it has been non-operative since 2011 it remains in standby mode. It covered the Asia-Pacific region from 35,800 km above the equator at 140 degrees east longitude, and obtained images every thirty minutes for the northern hemisphere (once an hour for a full-disk image) in five spectral bands including one visible band with spatial resolutions of 1 km, and four infrared bands with resolutions of 4 km. MTSAT-2 which carries a similar sensor to MTSAT-1R is now currently in operation at 145 degrees of east longitude. In this research, we use the full-disk imagery of MTSAT-1R. Table 1 lists the sensor specifications of MTSAT-1R.
MTSAT-1R uses five spectral bands, including two thermal infrared bands (IR1 and IR2) which are applicable for the estimation of LST. The spatial resolution of the MTSAT

TaTbalbele11. :CChhaarraacctteerriissttiiccssooffththeeMMTSTASTA-T1R-1sRensseonr.sor.

Band VIS IR1 IR2 IR3 IR4

Wavelength (m) 0.55 - 0.90 10.3 - 11.3 11.5 - 12.5 6.5 - 7.0 3.5 – 4.0

Spatial Res. (km) 1 4 4 4 4

Figure 1. Framework for LST retrieval from MTSAT IR1/IR2 data. The coefficients of generalized split-window (GSW) equation were optimized by using the result of a radiative transfer simulation. The LST was then retrieved from MTSAT IR1/
IR2 data and MODIS emissivity products with the GSW equation. Finally, MTSAT LST products were compared with MODIS LST products.

Asian Journal of Geoinformatics, Vol.14,No.3 (2014)

1R field of view and VZA, and the six study areas. Table 2 lists the location and approximate VZA of each study area.
2.2 LST Retrieval with Split-Window Algorithm

Figure 2. MTSAT-1R field of view and viewing angle, and the location of six study areas: (a) Australia, (b) Indonesia,
(c) Japan, (d) Mongolia, (e) Papua New Guinea, and (f) Thailand.
instrument varies with the spectral band, and ranges from 1 km to 4 km at nadir, and although it is coarse compared to polar orbiting satellites it has an excellent temporal resolution (tens of minutes to one hour). In this research, we obtained MTSAT IR1 and IR2 data from Institute of Industrial Science, University of Tokyo. All utilized data were radiometrically calibrated to brightness temperature (K), and geometrically converted to the Plate Carree projection (Takeuchi et al., 2010).
Six regions in Australia, Indonesia, Japan, Mongolia, Papua New Guinea, and Thailand, which are all located in different climate zones with differing land cover (including desert, tropical forest, croplands, urban, and grassland areas) and contain widespread viewing zenith angle (VZA) areas, were selected as study areas to confirm the spatial robustness of the proposed methodology. Figure 2 illustrates the MTSAT-

Atmospheric absorption and emission causes a large difference between the brightness temperatures measured insitu and those measured by satellite. Therefore, in order to retrieve accurate LST from thermal infrared data provided by satellite measurements, attenuation caused by the interaction between radiation from the Earth’s surface and atmospheric contents (mainly water vapor) requires elimination. Many researchers have previously developed algorithms for LST retrieval, and the split-window algorithm using two thermal infrared bands in the 10.5 to12.5 µm wavelength have been widely used (Price, 1984, Sobrino et al., 2000, Ulivieri et al., 1994), as this algorithm utilizes the differences in water vapor absorption between 11 µm and 12 µm to eliminate the effect of water vapor. Prata et al. (1999) used a GMS-5 Visible and Infrared Spin Scan Radiometer (VISSR) to retrieve LSTs, and found that the accuracy validated against ground measurement obtained from two Australian sites was approximately 2-3 K. Sobrino et al. (2004) optimized a two-channel algorithm for a European geostationary satellite, namely MSG SEVIRI (Spinning Enhanced Visible and Infrared Imager), and the theoretical accuracy in LST estimation was found to be lower than 1.5 K for VZAs lower than 50˚. However, the algorithm developed by Sobrino et al. (2004) requires column water vapor for LST retrieval, and there are currently no available water vapor products able to fulfill the temporal resolution (hourly) and spatial resolution (1-4 km) required for MTSAT data. Pinker et al. (2008) evaluated LST from GOES data with various algorithms for LST retrieval and found the accuracy to be about 1-3 K compared with six ground station data located in the United States.
The Generalized Split-Window (GSW) algorithm was originally developed by Wan et al. (1996) for retrieving LST from AVHRR and MODIS data, and was modified from the split-window algorithm developed by Becker et al. (1990) by optimizing coefficients with the VZA, atmospheric water vapor, and lower boundary temperature. Wan et al. (1996) reported that the GSW algorithm was able to estimate LST from MODIS thermal infrared data with an accuracy of less

Table 2: Selected study areas: name, location, and viewing zenith angle (VZA) of MTSAT-1R.

Table 2. Selected study areas: name, location, and viewing zenith angle (VZA) of MTSAT-1R.

Australia Indonesia Japan Mongolia Papua New Guinea Thailand

Upper Left
4S, 125E 4S, 104E 39N, 137E 50N, 104E 4S, 143E 17N, 98E

Lower Right
9S, 130E 9S, 109E 34N, 142E 45N, 109E 9S, 148E 12N, 103E

VZA (deg)
20-30 30-40 30-40 60-70 10-20 40-50


Hourly LST Monitoring with the Japanese Geostationary Satellite MTSAT-1R over the Asia-Pacific Region

than 1 K for the full range of VZAs across the swath (up to at the lowest boundary of each atmospheric model. Ouaidrari

55˚ from nadir). As MTSAT is a geostationary satellite, it has et al. (2002) reported a CWV of 7.0 g/cm2 in the tropics,

a large variety of VZAs; hence, the GSW algorithm is although on average it was found to usually vary between 0.3

suitable for MTSAT because this algorithm has robustness to 5.5 g/cm2. Hence the value of 5.0 g/cm2 was used for the

for VZA variation. In this study, we therefore applied the upper value of CWV in this study.

GSW algorithm to MTSAT data. The GSW equation for MTSAT data is expressed as:

In order to determine the emissivity ranges for the simulation, the band-averaged emissivities of MTSAT IR1 and IR2

where LST is land surface temperature (K), IR1 and IR2 are bands calculated from representative land-surface materials

such as vegetation (dry grass, and the leaves of maple, oak,

and pine), water (seawater, distilled water, ice, and snow),


= (a + a

1−ε + a

∆ε ) IR1+ IR2 + (b




soil∆(εsoilIR901P−4I7R6S2 a)

and +c






1 2ε

3 ε2


1 2 ε so3ilε, s2oil samp2le 1 of Death Valley soil, and soil sample 1 of

Page Arizona Sandy Soil), and man-made materials (black

+ a ∆ε ) IR1+ IR2 + (b + b 1− ε + a ∆ε ) IR1− IR2 + c ( 1 )

3 ε2


1 2ε

3 ε2


asphalt, cobblestone pavement, and life concrete tile) are shown in Figure 3. The emissivity data were obtained from MODIS UCSB Emissivity Library (http://www.icess.ucsb.

brightness temperatures (K) of MTSAT IR1 and IR2; and ai, bi (i = 1, 2, 3), and c are numerical coefficients. ɛ1, ɛ2 are surface emissivity of IR1 and IR2; ɛ and ∆ɛ are defined by ɛ
= (ɛ1 +ɛ2)/2, ∆ɛ = ɛ1 − ɛ2, respectively.

Because the coefficients depend greatly on the sensor specifications, the coefficients of the GSW algorithm require optimization for MTSAT. This study therefore also determines coefficients, based on the results of simulated atsensor measurements, by using the MTSAT Spectral Response Function (SRF). The simulation was implemented by MODTRAN4.3 (Berk et al., 2003), which is a widespread radiative transfer code for wavelengths ranging from the visible to the thermal infrared region.At-sensor measurements of MTSAT brightness temperature were simulated under various observational, atmospheric, and land-surface conditions to derive coefficients that were applicable to the entire Asia-Pacific region. Table 3 lists the parameters used in MODTRAN simulations: these ranges were determined based on the previous research of Ouaidrari et al. (2002) and Wan (1999). In this study, we performed forward simulations for MTSAT using the conditions from five atmospheric models (tropical, mid-latitude summer/winter, sub-arctic summer/winter) bundled in MODTRAN, with LST ranging from Tair−5 K to Tair +15 K, cumulative water vapor (CWV) from 0.5 to 5.0 g/cm2, emissivity from 0.94 to 1.00, and VZAs from 0 to 60˚. Tair represents the surface air temperature

Figure 3. The band-averaged emissivities of MTSAT IR1 and IR2 bands calculated from the emissivities of
representative land-surface materials. The middle line represents the 1:1 line. Upper and lower lines indicate that the differences in IR1 and IR2 emissivities are −0.02 and
+0.02, respectively.

Table 3: Parameters used in MODTRAN simulations with spectral response function (SRF) of the MTSAT1R sensor.
Table 3. Parameters used in MODTRAN simulations with spectral response function (SRF) of the MTSAT- 1R sensor.




Atmospheric Model Column Water Vapor (CWV)

Tropical, Mid-Latitude Summer/Winter


Sub-Arctic Summer/Winter


0.5- 5.0 (g/cm2)


Emissivity 12

0.94 - 1.00


Viewing Zenith Angle (VZA)

0 - 60 (degree)


Land Surface Temperature (LST)

Tair–5 - Tair+15 (K)



Asian Journal of Geoinformatics, Vol.14,No.3 (2014)

edu/modis/EMIS/html/em.html). The emissivities are distributed from approximately 0.94 to 1.00. In addition, the differences between the IR1 and IR2 emissivities are distributed around the 1:1 line and the absolute values of the differences are almost all less than 0.02. Therefore, we executed the simulations under the conditions that emissivities range from 0.94 to 1.00 and absolute values of the emissivity differences of IR1 and IR2 are less than 0.02.
These conditions were used in the simulation and the results were then used to optimize the numerical coefficients of the GSW algorithm by using the Levenberg-Marquardt algorithm for non-linear minimization. Since the coefficients highly depend on the VZA (Wan et al., 1996), the optimized coefficients were calculated for each VZA. We then composed a coefficient look-up table (LUT) with respect to VZAs.
2.3 Emissivity Map
The GSW equation requires the emissivity maps of IR1 and IR2 bands to estimate LST, and the quality of emissivity map strongly affects the accuracies of the LST retrieval. However, land-surface emissivity has high spatial variation and is therefore difficult to estimate, and in addition seasonal changes cause temporal emissivity variations because of alterations in land-surface conditions (in particular, deciduous forest or cropland areas) (Liang, 2004). In this study, the MODIS emissivity map (MOD11A1, collection-5) derived from land cover maps (Snyder et al., 1998), MODIS SRF, and emissivity data of various materials were used for the LST retrieval from MTSAT data. However, in order to apply the MODIS emissivity maps to MTSAT LST retrieval, it was necessary to confirm the consistency of the emissivities from the MTSAT and MODIS thermal bands. Therefore, by calculating the average emissivities of MTSAT (IR1, IR2) and MODIS (band31, band32) from the SRFs of two thermal infrared bands, and from the MODIS UCSB Emissivity Library data used in section 2.2, we confirmed that the differences caused by SRF configurations are less than 0.005. According to Wan et al. (1996), to achieve an accuracy of 1 K, the uncertainty of the emissivity should be around 0.005, and Snyder et al. (1998) used this criterion to evaluate emissivity products. Accordingly, it was found that MODIS emissivity products are applicable to LST retrieval from MTSAT data.

where n is the number of samples; LST is the true LST;
and LSˆT is the MTSAT LST from the optimized GSW algorithm. RSEs range from 0.86 K to 1.73 K, and tend to increase with the VZA. The slant path length through the atmosphere increases with the VZA, and therefore the simulation results exhibit large variance. However, the derived GSW equations were able to estimate LST with an accuracy of approximately 1 K when the VZA was less than or equal to 40˚, and are thus considered to be acceptable for use with MTSAT LST data for many land applications (Prata et al., 1999, Wan et al., 1996).
3.2 Comparison with MODIS LST Product
This section examines the results from actual MTSAT data. MTSAT LST was estimated by applying the optimized GSW algorithm with the observed brightness temperatures of MTSAT IR1 and IR2, and MODIS emissivity maps. The GSW coefficients for each VZA were calculated by linear interpolation from the LUT of the GSW coefficients with respect to the VZA (0˚, 20˚, 40˚, and 60˚).
Although validation of MTSAT LST with in-situ measurements is preferable, it is difficult to find spatially homogeneous areas suitable for validating coarse resolution satellite images such as MTSAT data. Even if it was possible, the number of in-situ measurement sites would be limited spatially and temporally. Therefore it would be a practical way to compare with the MODIS LST product, which has been well-validated by in-situ data and an accuracy of more than 1 K has been confirmed in many areas including lakes, grassland, rice fields, and snow covered areas (Wan et al., 2002). Furthermore, MODIS LST covers the entire globe on a daily basis. In this study therefore, we confirmed the consistency of MTSAT with MODIS LST MOD11A1 collection-5 product (Wan, 1999). Matchup LST data were observed at almost the same time (at about 10:30 local time) and observation time differences were within 30 minutes. Before the comparison, MODIS LST products were aggregated from 1 km to 4 km of spatial resolution by

3. Results and Discussions

Table 4: Theoretical accTuarbalcey4(.RTehsiedouraeltiSctaalnadcacrudraEcryro(rRs)esoifdMuaTl SSAtaTndLaSrdT Eesrtriomrsa)teodf by the optimiz







MTSAT LST each viewing

estimated by zenith angle

the optimized (VZA).



Equations Optimized for MTSAT

for each viewing zenith angle (VZA).

We firstly evaluated the theoretical accuracies of the derived GSW equations. The numerical coefficients given by Eq. (1) were determined with respect to each VZA. All the coefficients showed statistical significance (p < 0.05), and the residual standard errors (RSE) of each VZA are listed in Table 4. The RSE is defined as:

VZA (degree) 0 20 40 60

RSE (K) 0.86 0.90 1.09 1.73


Hourly LST Monitoring with the Japanese Geostationary Satellite MTSAT-1R over the Asia-Pacific Region

averaging 4 × 4 pixels.

MTSAT LST were highly consistent with those of MODIS

LST. In addition, the LSTs of urban central business district Clear-sky scenes of six Asia-Pacific regions (Australia, areas such as Tokyo, Jakarta, and Bangkok were found to be Indonesia, Japan, Mongolia, Papua New Guinea, and higher than those of surrounding areas. As this indicates that Thailand) were selected to assess consistency. The MTSAT MTSTAT LST is able to identify urban heat islands, it can and MODIS LSTFmiguarpes 4o:f tChleeasri-xskryegiimonagsearceosmhpoawrinsoinn FofigMurTeSAT atnhderMefOoDreISbLeSTusoevder tAousmtraolniaito(Nr otvhe1r5m, a2l00c7o),nditions over 4. A visual interpretation sIhndoownedsiath(aJut lth1e4,s2p0a0ti7a)l, pJaapttaenrn(Ms oarf 1, 2007), Mongolia (Aug 11, 2008), Papua New Guinea

(May 25, 2007), and Thailand (Dec 17, 2008). Black pixels represent cloud mask or an area of no

data, and were excluded from the comparison.

(a) Australia (MTSAT)

(b) Indonesia (MTSAT)

(c) Japan (MTSAT)

(e) Australia (MODIS)

(f) Indonesia (MODIS)

(g) Japan (MODIS)

(h) Mongolia (MTSAT)

(i) Papua New Guinea (MTSAT)

(j) Thailand (MTSAT)

(k) Mongolia (MODIS)

(l) Papua New Guinea (MODIS)

(m) Thailand (MODIS)

Figure 5: Scatter plots of MODIS and MTSAT LST, and RMSE, bias, and number of samples over Australia
Figure 4. Clear-sky image comparison of MTSAT and MODIS LST over Australia (Nov 15, 2007), Indonesia (Jul 14, 2007),

Japan (Mar 1, 2007), Mongolia (Aug 11, 2008), Papua New Guinea (May 25, 2007), and Thailand (Dec 17, 2008). Black

pixels represent cloud mask or an area of no data, and were excluded from the comparison. 31


(Nov 15, 2007), Indonesia (Jul 14, 2007), Japan (Mar 1, 2007), Mongolia (Aug 11, 2008), Papua New GuineaA(sMiaanyJ2o5u,r2n0a0l 7o)f aGnedoTinhfaoirlmanadti(cDs,eVco1l.71,42,N00o8.3).(2014)

(a) Australia

(b) Indonesia

(c) Japan

(d) Mongolia

(e) Papua New Guinea

(f) Thailand

Figure 5. Scatter plots of MODIS and MTSAT LST, and RMSE, bias, and number of samples over Australia (Nov 15, 2007), Indonesia (Jul 14, 2007), Japan (Mar 1, 2007), Mongolia (Aug 11, 2008), Papua New Guinea (May 25, 2007) and Thailand
(Dec 17, 2008).

metropolises on an hourly basis.
Figure 5 illustrates scatter plots and statistics of MODIS and MTSAT LST, including RMSE, bias, and number of samples. The RMSEs of Australia, Indonesia, Japan, Mongolia, Papua New Guinea, and Thailand were 2.32 K, 2.85 K, 2.77 K, 2.69 K, 2.86 K, and 2.68 K, respectively, and the bias ranged from -1.22 K to 1.46 K. In a comparison of clear-sky images, all regions showed a similar consistency with those of MODIS LST, even though these regions differ widely in terms of climate zone, land cover, and VZA.
3.3 Seasonal Stability Assessment of MTSAT LST

The time series stability was assessed using daily LST data observed at approximately 10:30 (local time) throughout 2007 and 2008. Although cloud pixels in the MODIS LST were masked by the MODIS cloud mask product, thin or low altitude clouds remained, and these were difficult for the 3M2 ODIS spectral data to identify (Ackerman et al., 2006). LST cannot be retrieved precisely if the target pixel is obscured by cloud, and we considered that this would be a potential cause of large inconsistencies between MTSAT and MODIS LST. Therefore, to reduce the effect of any remaining cloud, MODIS or MTSAT pixels with a LST lower than 270 K were defined as cloud-obscured pixels and were excluded from comparison. In addition, to mitigate the co-registration error between MOIDS and MTSAT LST images, MTSAT


Hourly LST Monitoring with the Japanese Geostationary Satellite MTSAT-1R over the Asia-Pacific Region

LST images was shifted to certain offsets on the x and y axes 150-240), but the RMSE of less than 2-3 K in the dry winter

when MTSAT LST was overlaid with MODIS LST to season (DOY 0-60, 330-360). A similar tendency was found

calculate the RMSE or bias. The MTSAT LST image was in Australia, but in the opposite seasons from those in Japan.

then moved within ±5 pixels on both the x and y axes, and the offset values were determined when the RMSE between MTSAT and MODIS LST was minimized.

The RMSEs of Australia and Mongolia showed the highest variances among the six regions. Land cover in the study areas of Australia and Mongolia is mostly desert, and the

Figure 6 shows daily RMSEs and biases. Temporal profiles LST therefore changes dramatically over a short time period

of RMSEs and biases in Australia, Japan, and Mongolia because the thermal inertia of the desert is high. Therefore,

indicate large variances and seasonality in one year. In these the difference in observation times between MTSAT and

areas, MTSAT LST was consistent with MODIS LST within MODIS is also considered likely to contribute to

5 K of the RMSE in the dry season, but more than 5 K in the inconsistency in desert areas. Furthermore, the quality of

rainy season. For example, Japan showed the RMSE of more MODIS LST can cause inconsistencies: Wan (2013) reported

than 5 K inFtihgeurreai6n:y DtoahiliyghRlyM-hSuEmainddsubmiams beertsweaeesonnM(DTOSAYT and tMhaOt DMIOSDLISSTLoSvTeretrhroersixarsetuwdiythainrea±s1iKn 2in00m7oasntdcases, but


(a) Australia

(b) Indonesia

(c) Japan

(d) Mongolia

(e) Papua New Guinea

(f) Thailand

Figure 6. Daily RMSE and bias between MTSAT and MODIS LST over the six study areas in 2007 and 2008.


Asian Journal of Geoinformatics, Vol.14,No.3 (2014)

more than 2.5 to 4.5 K in desert areas.

of a clear scene comparison in Section 3.2 show a RMSE

In Mongolia, the LST was found to be so low in the winter season that most of the pixels were masked as cloud. There were therefore few comparison results available from this region in winter because a LST lower than 270 K was defined as cloud in this study. The RMSE of Mongolia was the largest among the six study areas, and this is considered to be attributed to both the land cover and the large VZA of 6070°. Indonesia, Papua New Guinea, and Thailand are located in the tropics, but the RMSE and bias of Papua New Guinea showed a greater stability than the other three study areas. The possible reasons for this are that the VZA is almost nadir (10-20°), and that the land cover of the study area is almost completely forested and therefore has a small temperature variation.

better than 2-3 K in all regions, at times the cloudy scene comparison indicates an RMSE greater than 5 K. This suggests that the consistency of MTSAT and MODIS LST tends to be dramatically decreased when cloudy image pairs are compared. As cloud pixels remain even with the MODIS cloud mask, a LST thresholding method was applied to eliminate the cloud- obscured pixels. However, the remaining clouds cause a large inconsistency because estimated LST from cloud-obscured pixels is very uncertain. In addition, atmospheric water vapor in a cloudy scene would be higher than that in a clear-sky scene, and large amounts of atmospheric water vapor increase the uncertainty in LST retrieval (Wan, 1999). Moreover, the difference in observation time between MTSAT and MODIS causes a difference in the cloud or water vapor distribution spatially

and temporally, and this observation time difference also

Figure 7 shows the number of clear pixels and RMSE in causes a bias error because LST rises or drops even within 30

2007 and 2008 calculated from daily pair images of MTSAT minutes. Such cloud and atmospheric-water related factors

and MODIS LST, where the RMSE is seen to greatly depend are the main sources of inconsistency, and the spatio-

on the number of clear pixels in a scene (the RMSE decreases Figure 7: The number of clear pixels and RMSE


t2e0m0p7oraanldpa2t0te0r8n.oEf athcehsepfoaicnttorsshroewsuslttshien vthaeluienconsistency

as the number of clear pixels increases). Although the results

calculated from daily pair images of MTSAT and MODIS LST.

(a) Australia

(b) Indonesia

(c) Japan

(d) Mongolia

(e) Papua New Guinea

(f) Thailand

Figure 7. The number of clear pixels and RMSE in 2007 and 2008. Each point shows the value calculated from daily pair images of MTSAT and MODIS LST.


Hourly LST Monitoring with the Japanese Geostationary Satellite MTSAT-1R over the Asia-Pacific Region

of seasonality and locality.
3.4 Hourly LST Observation by MTSAT
Figure 8 shows hourly MTSAT LST and daily MODIS LST (MOD11A1, MYD11A1) observed on March 1, 2007, in central Tokyo, Japan (N35.69˚, E139.76˚), although MODIS LST (MYD11A1) data for the afternoon observation (13:30 JST) was not available because of its poor quality flag. For a comparison, we selected a pixel of MTSAT LST data and aggregated MODIS LST data by averaging 4 × 4 pixels to standardize the pixel size. MTSAT LST was consistent with MODIS LST, but MTSAT was able to capture diurnal changes in greater detail, including minimum, maximum, and inflection points. Such hourly profiles are considered to be useful in gaining a better understanding of energy exchanges between the land and atmosphere.
Finally, monthly composites of hourly MTSAT LST from a full-disk image are shown in Figure 9. Full-disk LST were generated by using the GSW algorithm optimized for MTSAT with monthly global emissivity maps from the MODIS LST product (MOD11C3) to cover the MTSAT fulldisk area on a daily basis. Monthly composite images were generated by selecting the pixel with the highest LST from pixels observed at the same hour during August 2007. MTSAT was found to be able to capture the diurnal LST changes over the Asia-Pacific region.
3.5 Discussion
According to literature (Prata et al., 1999, Wan et al., 1996), RMSEs of 1–3 K are acceptable, accuracies of 1 K are very useful for gaining a better understanding of the entire Earth system on a global scale, and accuracies of 3 K are also useful for limited use. MTSAT LST has an excellent sampling frequency and is considered to be marginally useful for landsurface monitoring, if cloud contamination, atmospheric

water vapor, land cover, and VZA are taken into consideration. The accuracy of the LST product from a polar orbiting satellite (POS) like MODIS is better than 1 K in most cases (Wan, 2008). However, LST products from geostationary satellites (including GOES, MSG, or GMS) have been found to have generally an accuracy of between 1 and 3 K (Pinker et al., 2008, Prata et al., 1999, Sobrino et al., 2004), and the results of this study are therefore in agreement with those of previous literature.
The main source of inconsistency is confirmed as being related to remaining clouds, as discussed in Section 3.3. It is therefore considered necessary to evaluate cloud contamination over the target area when MTSAT LST is utilized, and a reliable cloud mask could be quite useful for eliminating MTSAT LST data with high uncertainties. However, in addition to cloud contamination, the optimization of the GSW is considered likely to cause inconsistencies. In this study, the LUT of the GSW coefficients was composed with respect only to VZA, whereas the LUT of the MODIS LST algorithm was composed with respect not only to VZA, but also to atmospheric profiles such as air temperature or atmospheric water vapor. The optimizations of the GSW equations by separating these atmospheric conditions is therefore considered likely to improve the accuracies of the GSW algorithm, although atmospheric profiles are needed to estimate LST. If atmospheric water vapor data is available, a LST retrieval algorithm which directly inputs water vapor to the LST retrieval formula (Sobrino et al., 2000) would also be applicable. However, it is difficult to generate a water vapor product that fulfills MTSAT coverage on an hourly basis with a similar spatial resolution.
Other discrepancies are considered to be related to coregistration errors and observation time differences between MTSAT and MODIS, and these differences are inevitable when MODIS LST is used for comparison. Although registration errors were offset by shifting the image on the

Figure 8. Hourly profile of MTSAT LST and three points of MODIS (MOD11A1, MYD11A1) LST at Tokyo, Japan (Mar 1, 2007, Japan Standard Time). The location of the sampled area in central Tokyo is N35.69˚, E139.76˚.
MtsatVzaLstModis LstMtsat Lst