An assessment of remotely sensed land surface temperature
Isabel F. Trigo,1,2 Isabel T. Monteiro,1 Folke Olesen,3 and Ewa Kabsch3
Received 27 February 2008; revised 30 May 2008; accepted 16 June 2008; published 4 September 2008.
[1] The satellite application facility on land surface analysis (Land SAF) generates,
archives, and disseminates land surface temperature (LST) in an operational basis. LST is
estimated from the spinning enhanced visible and infrared imager (SEVIRI) onboard
Meteosat, making use of a generalized split-windows algorithm. Here SEVIRI LST is
compared with retrievals from the moderate resolution imaging spectroradiometer
(MODIS), collocated in space and time, for three 10

10
areas (Iberian Peninsula,
Central Africa, and the Kalahari), and for six 7-day periods between July 2005 and May
2006. Overall, SEVIRI LSTs are warmer than MODIS values, with maximum
discrepancies generally observed for daytime. The mismatches between the two satellite
retrievals are then analyzed in terms of (1) satellite viewing angle differences, (2) surface
orography, and (3) surface type. Daytime discrepancies are strongly impacted by
differential heating rates of elements within a pixel (e.g., vegetation types, bareground),
leading to a relatively wide range of MODIS-SEVIRI LST differences, with strong
dependency on the MODIS view zenith angle. In contrast, average nighttime discrepancies
are generally below 2
C. The intercomparison between MODIS and SEVIRI LST is
complemented with in situ observations taken at Evora ground station (southwestern part
of the Iberian Peninsula). The differences between ground and satellite-derived values
show high variability for daytime for both sensors, with a systematic overestimation of in
situ values by SEVIRI LST. In the case of nighttime observations, both sensors tend to
underestimate local measurements, with estimated bias over all events under study of
1.7
C and 2.6
C for SEVIRI and MODIS LST, respectively.
Citation: Trigo, I. F., I. T. Monteiro, F. Olesen, and E. Kabsch (2008), An assessment of remotely sensed land surface temperature,
J. Geophys. Res., 113, D17108, doi:10.1029/2008JD010035.
1. Introduction
[2] Land surface temperature (LST) is an important
variable for the assessment of the surface energy budget.
LST controls the surface emitted long-wave radiation, being
also important for the estimation of the sensible [e.g., Hall et
al., 1992; Sun and Mahrt, 1995] and latent [e.g., Dickinson
et al., 1991; Sellers et al., 1997; Viterbo and Beljaars, 1995;
van den Hurk et al., 2000] heat fluxes between the surface
and the atmosphere. Remotely sensed thermal data consti-
tute the best source of information for LST estimation over
large areas, and within these, data obtained from geosta-
tionary satellites are the only capable of fully characterizing
the daily cycle of LST.
[3] Several algorithms have been developed to obtain
LST from space, most of them making use of data from
sensors onboard polar-orbiters [Price, 1984; Becker and Li,
1990; Prata, 1993; Sobrino et al., 1994; Wan and Dozier,
1996; Liang, 2001]. Similar methodologies [Madeira 2002;
Go¨ttsche and Olesen, 2002; Sun and Pinker, 2003; Sun et
al. 2004; Oku and Ishikawa, 2004, Pinker et al., 2007] have
also been applied to data from geostationary satellites, with
lower spatial resolution but higher temporal samplings.
Most LST algorithms [Dash et al., 2002] have in common
the use of one or more channels within the thermal infrared
(TIR) atmospheric window (8–13 mm). Among these, the
generalized split-window algorithms [Wan and Dozier,
1996; Madeira, 2002] assume that LST may be obtained
through a semi-empirical regression of top-of-atmosphere
(TOA) brightness temperatures of two pseudocontiguous
channels (the split-window channels) where (1) the atmo-
spheric correction is a function of the differential absorption
in the two channels and (2) channel surface emissivities are
known a priori.
[4] This article presents an assessment of the LST gen-
erated by the Satellite Application Facility on Land Surface
Analysis (Land SAF [DaCamara, 2006]). The Land SAF
LST is obtained through the application of a generalized
split-window (GSW) algorithm to data from the Spinning
Enhanced Visible and Infrared Imager (SEVIRI) on board
Meteosat Second Generation (MSG) satellites [Schmetz et
al., 2002]. The Land SAF LST, retrieved, archived and
disseminated on an operational basis since February 2005,
JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 113, D17108, doi:10.1029/2008JD010035, 2008
1Instituto de Meteorologia I.P. Land SAF, Lisbon, Portugal.
2Faculdade de Ciencias de Lisboa, Instituto Dom Luiz/CGUL, Lisbon,
Portugal.
3Institute for Meteorology and Climate Research, Forschungszentrum
Karlsruhe, University of Karlsruhe, Karlsruhe, Germany.
Copyright 2008 by the American Geophysical Union.
0148-0227/08/2008JD010035
D17108 1 of 12

is freely available either in near real time or off-line (http://
landsaf.meteo.pt).
[5] The main goal of LST validation is the quantification
of its accuracy, putting into evidence optimal/suboptimal
conditions for the retrievals and helping their interpretation.
In this sense, validation is the only means of ensuring
the correct use of estimated parameters [e.g., Justice and
Townshend, 1994; Morisette et al., 2002]. The validation of
the Land SAF LST involves its comparison with similar
parameters retrieved from sensors on-board other platforms,
and in situ measurements. Both types of intercomparison
exercises have to deal with the generally high thermal
heterogeneity of land surfaces. Here the Land SAF
SEVIRI-based LST is compared with the equivalent MOD-
erate resolution Imaging Spectroradiometer (MODIS)
parameter over three 10
longitude-by-10
latitude areas
within the Meteosat disk. The impact of the observation
angles on the measurements, which is a direct consequence
of surface nonhomogeneity [e.g., Barroso et al., 2005;
Pinheiro et al., 2006], is analyzed for different types of
land cover and orography.
[6] The study is complemented with the comparison
between satellite-retrieved values and in situ measurements
collected at the Land SAF ground-truth site in Evora
(Southern Portugal). Again, the thermal heterogeneity of
land surfaces poses strong constraints to the suitability of
sites for LST validation, and thus to the relatively few
stations found in the literature [e.g., Coll et al., 2006]. The
choice of the Evora site for the ground-truth collection of
LST observations followed homogeneity criteria relevant
for both satellite (5 km for SEVIRI sampling distances
over South-Western Europe) and ground-measurement spa-
tial scales [Dash et al., 2004].
[7] The next section presents the description of SEVIRI/
Meteosat and MODIS LST products, including the Land
SAF algorithm, as well as of the in situ observations. The
results of the intercomparison between satellite-retrieved
LST values and their validation against ground-truth data
are analyzed in sections 3 and 4, respectively. Finally,
section 5 summarizes the main conclusions of this work.
2. Data
2.1. SEVIRI/Meteosat Land Surface Temperature
[8] The Land SAF LST is estimated using a Generalized
Split Window (GSW) algorithm [Madeira, 2002], with the
formulation first proposed by Wan and Dozier [1996] for
Advanced Very High Resolution Radiometer (AVHRR) and
MODIS data. Thus LST is estimated as a linear function of
clear-sky top-of-the-atmosphere (TOA) brightness temper-
atures measured by SEVIRI split-window channels centered
at 10.8 mm and 12.0 mm, respectively (Tb10.8 and Tb12.0):
LST ¼ A1 þ A2
1 e
e
þ A3
De
e2


T10:8 þ T12:0
2
þ B1 þ B2
1 e
e
þ B3
De
e2


T10:8  T12:0
2
þ C ð1Þ
where the regression coefficients depend explicitly on the
mean surface emissivity for the two channels (e) and on the
difference of emissivities for channel 10.8 mm ‘‘minus’’
12.0 mm (De). The par s A1, A2, A3, B1, B2, B3, and
C, have been empirically estimated for classes of total
column water vapor, and satellite zenith view angles (SZA).
The estimation of the GSW parameters relied on linear
regressions of synthetic Tb10.8 and Tb12.0, obtained from
radiative transfer simulations performed over a set of
40 (clear sky) atmospheric profiles extracted from the
TIGR-like atmospheric profile database [Chevallier et al.,
2000] and 50 (clear sky) radiossondes covering a wide
variety of atmospheric conditions: 2 m-temperature (T2 m)
varies between 250 and 310 K, while total column water
vapor ranges from 0 cm (very dry conditions) to 6 cm (very
wet conditions). Several simulations are then performed for
each of these profiles with different values assigned to
surface emissivities (e between 0.94 and 1; and De between
0.0135 and 0.022), satellite zenith angles (from 0
up to
60
), and surface temperatures. The latter range from T2 m
minus 15 K to T2 m plus 15 K. The assessment of the GWS
algorithm against an independent data set of radiative
transfer simulations showed that the algorithm is bias free,
although random errors tend to increase for high view
angles and with water vapor content.
[9] Channel surface emissivity is estimated as an average
of bareground and vegetation emissivities within each
scene, weighted with the pixel Fraction of Vegetation Cover
(FVC). The bareground/vegetation emissivities have been
previously assigned to each class of a land cover map [Peres
and DaCamara, 2005; Trigo et al., 2008], while FVC is
routinely retrieved by the Land SAF [Garcı´a-Haro et al.,
2005].
[10] The Land SAF LST product (Figure 1) is available
with a 15-minute frequency for all land pixels within
Meteosat disk, which have SZA lower than 60
, since
retrieval errors increase significantly for long optical paths.
The cloud mask makes use of the software developed by the
SAF on support to Nowcasting and Very Short-Range Fore-
casting (NWC SAF; http://nwcsaf.inm.es), and total column
water vapor forecasts are obtained from the European Center
for Medium-Range Weather Forecasts (ECMWF). Further
details on the Land SAF LST algorithm and product may be
found in the respective Product User Manual (available at
the Land SAF web site http://landsaf.meteo.pt/).
2.2. MODIS Land Surface Temperature
[11] The MODIS level 3 daily LST includes a pair of
observations (daytime and nighttime) per day, at 1km spatial
resolution. Here we use data from MODIS version 3.0. LST
is retrieved for all land pixels under clear sky conditions,
using a GSW algorithm with a formulation very similar to
equation (1), but applied to MODIS split-window bands 31
and 32 [Wan and Dozier, 1996; Wan, 1999]. As in the Land
SAF algorithm, the parameters in the MODIS GSW depend
on the SZA, column water vapor and also on the low
atmosphere boundary temperature. The band emissivities
rely on the classification-based method [Snyder et al., 1998]
according to land cover types in the pixel.
[12] In order to compare SEVIRI and MODIS LST’s,
these are both reprojected onto a common grid, close to the
coarser resolution of the former. Thus the original MODIS
level 3 LST is reprojected to a regular 0.05

0.05
grid,
by averaging all 1km pixels within each grid box. To avoid
cloud contamination, only MODIS pixels with the highest
rank quality flag are used. As a consequence all SEVIRI
D17108 TRIGO ET AL.: LAND SURFACE TEMPERATURE ASSESSMENT
2 of 12
D17108