LST-Val station (Proc

Document Sample
LST-Val station (Proc Powered By Docstoc
					  Validation of Land Surface Temperature from MSG/SEVIRI with a
                        ground-truth station



                   Folke Olesen, Ewa Kabsch, Andreas Schmidt, Frank Göttsche

               Forschungszentrum Karlsruhe, Postfach 3640, 76021 Karlsruhe, Germany



                                                  Abstract
Land Surface Temperature (LS T) from Meteosat Second Generation (MS G) is routinely generated by
the Land Surface Analysis – Satellite Application Facility (LSA-SAF). This LST product is retrieved
with the generalized split-window method (Wan and Dozier, 1996) and is validated in this study by
means of comparisons with ground-truth radiometric measurements. The validation ground station
was set up in central Portugal, near E vora in 2005. It provides continuous in -situ LS T data derived
from meas urements with infrared radiomet ers. The validation reveals a negative bias for night MSG-
LST of about -1 K. In contrast, at daytime positive biases of MSG-LS T of up t o 4 K are observed. This
means t hat for night data the LS T retrieval algorithm yields LS T within the nominal accuracy of t he
algorithm (± 2 K ), but at daytime it does not reach the required accuracy.


INTRODUCTION

Land Surface Temperature (LS T) derived from satellites observations is a key parameter in many
environmental applications. Remote sensing measurements from s atellite platforms enable LS T
retrieval on a global scale. Before the radiance which leaves Earth’s surface reaches a satellite sens or
it must pass through the atmos phere; LS T retrieval algorithms account for t he attenuation by t he
atmosphere. Furthermore, Earth’s surface is not a perfect black body, so its emissivity must be tak en
into account. An emissivity correction is also included in the LS T retrieval algorithm. A major constraint
for LS T determination from satellites in the thermal infrared is clouds. Only for completely cloud-free
conditions LS T determination is feasible.
In order to guarantee its quality and reliability, a validation of t he MS G-LS T product is necessary. The
most meaningful and reliable method of LS T validation is a comparison with ground-truth data.
However, it has to be ensured that the ground meas urements are representative for the satellite
observations. Ground-based point measurements are only representative for a satellite pixel if the site
is flat and homogeneous. In case of LS T validation the homogeneity of the land cover in terms of its
surface temperature distribution must be ensured. Due to the fact that most natural surfaces are
heterogeneous, the selection of a suitable validation site is far from straightforward, especially if t he
corresponding satellite pixels are large.
Within the scope of LSA-SAF, a project initialized and financed by Eumetsat, satellite products from
Meteosat Second Generation Spinning Enhanc ed Visible and Infra-Red Imager (MS G/SEVIRI) are
operationally generat ed and disseminated to end users. LS T is ret rieved with the Generalised Split-
Window (GSW) algorit hm (Wan and Dozier, 1996); the GSW coefficients are determined from 2 m air
temperature and total column water vapour, both obtained from the European Centre for Medium-
Range Weather Forecasts (ECMWF), and from surface emissivity (also a LSA-SAF product) with
consideration of satellite viewing angles.
LST is a key product of LSA -SAF and, therefore, the consortium made an additional effort to
guarantee the high quality of this product. A permanent ground-t ruth station has been established in
central Portugal, near the town of E vora. The station provides continuous in-situ measurements with
radiometers in the spectral range of 8 – 14 µm, which are comparable to SEVIRI infrared channels.
Several field and satellite image studies (Ikonos, Landsat) have shown that this site is homogeneous
at the scale of a SEVIRI pixel (size at this location 6 x 6 km²).
Additional information about the LS T validation station in Portugal toget her with initial validation results
for MSG/SEVIRI can be found in Gajewska et al., 2006. In this contribution extended comparisons of
LST derived from parallel ground (validation station) and sat ellite measurements (MSG/SEV IRI) are
presented.


THE LST VALIDATION STATION IN MITRA, PORTUGAL

A screening of one year of AVHRR data of Europe with respect to horizontal homogeneity and
cloudiness revealed several potential LS T validation sites for MSG (Dash et al., 2004). B ased on field
investigations, the final location of the station was chosen to be at Mitra, near the t own of E vora, in
central Portugal.
The land cover of the selected validation site consists of randomly distribut ed oak trees on a
background of grassland, which spreads over a large area. This uniform oak woodland covers more
than 2 x 2 SEVIRI pixels. Hence, moderate navigation errors are not expected to influence t he
validation analysis. A 28 m high t ower allows measurements from above the tree canopy and t he
satellite view can be well imit ated. Th e main instruments of the LS T validation station are several
infrared radiometers. On the satellite scale the region is homogenous because the surface leaving
radiance is integrated over a large number of tree crowns and grass within each satellite pixel.
However, for ground measurements these two landscape elements have to be treated separately.
Analyses of high resolution satellite images (Ikonos and Landsat) completed the characterisation of
the site. A subset corresponding to 2 x 2 MSG pixels of t he E vora site was selected and t he
percentage tree crown cover was determined. In summer the grass is completely dry, so the
Normalised Difference Vegetation Index ( NDV I) can readily be used to detect the remaining dense
vegetation, in this case the trees. Independent analyses of a Landsat image taken in 2003 (pixel size
30 m) and of an Ik onos image t aken in 2005 (pixel size 4 m) yielded a percentage tree crown cover of
47%. The remaining 53% is defined as grassland.
The main instrument of the LS T validation station is a rotating radiometer called “Rotrad”. Rotrad is
self-calibrating: during each of its measurement sequences it points its sensor at two black bodies
kept at different temperatures. Afterwards, Rotrad measures sky brightness temperature in t he
direction of MSG and then it samples a tree crown and two grass targets. Each measurement cycle
lasts 2 minutes and the diameters of the targets are 5 m for the tree crown, 6 m for grass-1, and 7 m
for grass-2.
As a back-up and to provide base-line measurements, a chopped K T15.82 radiometer from Heitronics
is mounted next to Rotrad. The K T15.82 provides simultaneous measurements of surface brightness
temperature and has a large field of view (FOV), which covers an area of about 18 m in diameter on
the ground. Thus, every 10 minutes the K T15.82 measures the mean brightness temperature of a
mixture of tree crown and grass. In order to minimize effects due to differences in view angle, ground
radiometric measurements are made in the viewing direction of MS G at the validation site.
Temperat ure differences between grass and tree crowns can be as large as 15 K. Only integrated tree
and grass radiometric measurements are representative for MSG satellite pixels. Integrated values
were obtained from Rotrad by weighing its measurements with the respective percentages of tree
crown cover and grass cover obtained previously from image analysis.
Under the assumption that LS T distribution is a stationary process, local surfac e temperature
fluctuations over short time intervals measured at a single location, e.g. the validation site, are
equivalent to random surface temperature fluctuations over space measured at a given time, e.g. a
single satellite image. In order to mak e the validation measurements more representative of SEV IRI
15 minutes slots, the ground LS T, which is sampled at high frequency, is smoothed wit h a 10 minutes
moving window.
For near ground measurements the attenuation of radiance by the few meters of atmosphere bet ween
surface and radiometer can be neglected. However, the use of radiometers requires an emissivity
correction of the measurements. In summer t he grassland at t he E vora site is completely desiccated,
resulting in completely different emissivities than in the rest of the year. Based on spectral libraries the
mean emissivity in the spectral range of the radiometer is estimated to be 0.98 for dens e vegetation
and 0. 95 for soil. Taking into account the two land cover area fractions at this site (47% trees and 53%
grass), the integrat ed emissivity during the summer months (May to September) equals 0.964 In
winter and autumn the grass is part of the green vegetation and is assumed to have an emissivity
equal to that of the trees, so that the integrated emissivity from October to April is 0.98.
DAT A PRE-PROCESSING

Satellite data (LST-MSG)
Radiomet ric measurements from E vora station as well as the LS T-MSG product were pre-processed
before comparing them in order to validate LS T-MS G. The main task was to resample ground and
satellite data to a common temporal and spatial resolution. Only simultaneous and spatially co-
registered data obtained for the same surface area can be directly compared. Furthermore, in the pre-
processing the “best” data were selected in order to avoid artefacts.
The main requirement for satellite measurements to be selected is a completely cloud-free sky wit hin
all satellite pixels relevant to the site. E ven thin clouds can affect surface temperatures derived from
satellite dat a. Although a cloud mask is provided by LSA-SAF, there are still some undetected clouds
in the LS T-MS G product.
LSA-SAF defines three accuracy levels of the LS T product: “below nominal”, “nominal” and “above
nominal”. Pixels t hat might be contaminated with clouds have a quality flag “below nominal”. In t he
validation only pixels with quality flags “nominal” and “above nominal” were considered, for which the
nominal accuracy of LS T-MS G is better than ±2 K.
A satellite image analysis of the site revealed a homogeneous area corresponding to more than 2 x 2
MSG pixels with the ground station near its south-eastern corner. These 4 MSG pixels are expected to
have nearly identical surface temperature. Therefore, the average surfac e temperature of the 4 MSG
pixels is used for further calculations. In order to avoid artefacts, only those LS T-MS G slots are
considered, for which standard deviation over the 4 selected MSG pixels is less than ±2 K;
furthermore, all pixels have to be cloud free.
Outliers in t he LS T satellite product over the E vora site were eliminated as follows: the maximum
allowed difference between two consecutive LS T-MS G slots (time interval 15 minutes) was set to be 3
K. Each pair of slots was compared and if this condition was not fulfilled the lower value was
eliminated. Furt hermore, isolated slots (missing slots before and after) were eliminated. In this way the
amount of cloud cont aminat ed pixels was reduced. This filtering process is intended to remove only
coarse disagreements in the data, but to retain natural scatter.
Finally, only thos e days were selected for validation, for which at least 80% of MS G slots per day (96)
were classified as cloud free. This limited the validation to so-c alled “golden” days (practically cloud-
free days), which are most suitable for validation purposes.
The nominal time of the MSG slots refers to the start of the scan. Sinc e each scan proceeds from t he
southern t o the northern hemisphere and lasts 15 minutes (including retrace and calibration), t he
actual time of measurement at Evora is MSG nominal time plus 10 minutes. Thus, satellite data
acquisition at E vora takes place every 10t h, 25th, 40th and 55th minute of each hour.

Ground-truth data
The ground LS T meas urements are carried out with higher temporal resolution than the satellite
measurements. Furthermore, the spatial res olution of the two kinds of measurement is very different.
Thus, the ground measurements register much of the local short -time surface temperature fluctuations
that are not noticeable on the scale of the satellite pixels. Based on stationary process theory, Rotrad
measurements (temporal resolution 2 minut es) were smoothed with a 10 minute moving window: the
resulting values are averages of 5 values each. In order t o exclude dat a affected by high local
temperature variations (mostly due to fluctuations in solar radiation), also standard deviation within t he
moving window was calculated and data with standard deviations in excess of 2 K were excluded from
further analysis. Then t he logging time of the ground measurements was synchronized with t he
satellite scan time by resampling the ground-truth data to MSG temporal resolution. This was achieved
by averaging the station measurements over 5 minutes before and after each 15 minut e MSG slot.
In this study it is assumed that MS G/SEVIRI scans the same fraction of shadow as the radiometers of
the station, because the radiometers measure in the view direction of MSG. Furthermore, the
characterisation of the site yielded the percentage tree crown cover over the selected MS G pixels; this
was used to weigh the respective Rotrad measurements (47% trees and 53% grass). Only the Rotrad
measurements were int egrated (weighed and summed) in this way, since the K T15.82 does not
measure tree and grass separately. However, long-term analysis shows that the integrat ed Rotrad
data match K T15.82 measurements well, which can be explained by the fact that the KT15.82 views
about a 50% tree crowns and 50% grass, which is quite close to the Rotrad weights. The remaining
differenc es of ±1 K are expected to be mainly caused by differences in shadow fractions. E ven though
the KT15.82 is not self-c alibrating, night data from K T15.82 and Rotrad show that it is stable over
longer periods of time and also provides reliable measurements for validation.
Corrections for reflected down-welling LW radiation were performed using the sky brightness
temperatures from Rotrad.


VALIDATION RESULTS

Operation of the ground station started in summer 2005. However, due to some technical problems
there are several gaps in the data. Similarly, the LS T-MSG satellite product is not continuously
available, because of interruptions of service. A nd even though the station is loc ated in a relatively
sunny region of Portugal, there is a surprisingly high frequency of cloudiness and many data had to be
excluded from the validation. In spite of the cloud masking applied to the LSA -SAF products, there are
still many MSG slots which are not sufficiently masked. This problem is depicted in Figure 1, which
shows a night -time LS T-MSG (y ellow circles) well below the brightness temperature obt ained from t he
ground station measurements. Furthermore, at night-time the sky brightness temperature is close to
surface brightness temperature, which strongly suggests clouds over the validation site. Undetected
clouds are a weakness of the LS T-MSG product and cause large errors.




Figure 1: An example of the LST-MSG (circles) contam inated w ith clouds at night. Surface brightness temperature
measured at the ground station [Rotrad integrated (black line) and KT15.82 (pink line)] are almost equal to sky
brightness temperature (purple line), which strongly suggests an overcast sky.

The main aim of t his study is to validate the LS T algorithm and not the cloud mask. In order to
eliminate the influence of undetected clouds, the validation was limited to “golden days” (see above).
Figure 2 shows averaged LS T-MSG slots against corresponding resampled surface temperatures
corrected for emissivity for selected days in 2005 and 2006. The red line represents equality bet ween
the two plotted temperat ures whereas the green line shows the actual linear regression.
Figure 2: LST-MSG plotted against integrated Rotrad temperatures (left graph) and KT15.82 temperatures (right graph).
The data only encompasses those days in 2005 and 2006 on which a m inimum of 80% of the MSG slots were classified
as cloud-free. All ground measurements were corrected with a seasonally variable em issivity. Red line represents
equality between measurements , the green line shows the linear regression.

The left scatter plot in Figure 2 shows the integrated Rotrad data, which were obtained by adding t he
three Rotrad targets with weights of 47% for “t ree”, 26.5% for “grass-1”, and 26.5% for “grass-2”.
Shadowed grass areas were included in the analysis, since the satellite also integrates over
shadowed areas. MSG is centred at 0° latitude above the Equator. Thus, MSGs view zenith angle at
the E vora site is a constant 47° (apart from fluctuations of less than 1°). Its view azimuth angle is also
constant and MS G always scans the E vora site in the south-east direction. The amount of shadow in a
given pixel observed by a geostationary satellite is only a function of the position of the sun. Hence,
MSG observes a maximum of shadowed areas in the morning and in the afternoon. Around noon t he
tree shadows at the E vora site are directly underneath t he tree-crowns and c ontribute little to t he
satellite signal. Due to the fact that the station’s radiometers measure in the same direction as MSG,
the fraction of shadowed area seen by the radiometers approximates that seen by the satellite.
It would be even more realistic, if the ground measurements at the E vora site would consider the three
target types tree, grass, and grass in shadow. Toget her with information about the vegetation structure
these measurements could be used to estimate the amount of shadow in a scene and the integrat ed
LST using a geometric-optical radiation transfer (GORT) model (Pinheiro et al., 2004 and Pinheiro et
al., 2006). However, because of practical limitations at E vora station there is no s eparate “grass-in-
shadow” target, i.e. no grass target for a fixed radiometer could be found which remains the whole day
in the shadow, and this 365 days per year. In order to meet this requirement, a radiometer has to be
able to follow the shadow. Since currently no such instrument is available, the estimation of shadow
fraction and its contribution to the integrated signal is performed using the Rotrad grass targets.
Figure 2 shows that the radiometric surface temperature meas urements corrected for emissivity from
Rotrad and K T15.82 have a high correlation with LS T-MSG of more than 0.985. The best agreement
is for lower surface temperatures. Low surface temperatures occur mostly at night, when atmospheric
correction of satellite data is less erroneous. At night, the difference bet ween surface and air
temperature is usually small, becaus e the land surface is not heated by direct solar irradiance.
Furthermore, at night tree crown and grass are nearly at the same t emperature, so it is also much
easier to ensure that the ground targets are represent ative. Additionally, there is no shadow fraction to
be accounted for.
It can also be seen in Figure 2 that high surfac e temperatures (e.g. at noon) are overestimated by the
split-window algorithm used for LS T-MSG retrieval. The average absolute difference between LS T-
MSG and LS T measured by Rotrad and K T15.82 equals 1.6 ±1.5 K and 1.4 ±1.3 K, respectively.
These results are slightly worse than the nominal accuracy of the LS T-MSG product of ±2 K. However,
this might still be partially due to some remaining cloud contamination in the validation data despite
the filtering.
In order to investigate the diurnal agreement bet ween satellite and ground LS T, differences (satellite
minus ground LS T) over selected validation data were c alculated. Furt hermore, mean differences and
standard deviations for each MSG slot (every 15 minutes) were determined in order to visualise t he
variation of these values over the diurnal temperature cycle. Figure 3 illustrates the results of these
analyses. At night the mean differences have a slightly negative bias of about -1 K and a standard
deviation of ±1 K: the LS T-MS G algorithm underestimates LS T, but its results are still within the error
margin of the LS T-MSG product (± 2 K).
In cont rast, the daytime positive biases are up to 4 K with standard deviations of up to ±3 K. This
means that daytime LS T estimated by the LSA-SAF s plit-window algorithm substantially exceed t he
nominal error margin of the LS T-MSG product (±2 K). Furt hermore, the large standard deviations
indicate strong data variations.




Figure 3: Mean differences and standard deviation between LST-MSG and station LST calculated for each MSG slot
(every 15 m inutes). Rotrad measurements have been integrated over all three targets according to the percentage tree
crown cover and corrected for em issivity. Only selected data (m inimum 80% of MSG slots cloud-free per day) from
2005 and 2006 were considered in the analysis. The largest deviations occur at daytime, whereas at night-time LST-
MSG and station LST are in good agreement.

Tests showed that the high maxima of LS T-MSG are not due to uncertainty in emissivity. Assuming
higher emissivity values reduced the difference between satellite and station LS T only for night-time
data (in some cases to z ero); no comparable reduction was observed for daytime data for any
reasonable choice of emissivity.
The percentage tree crown cover is used as a weight in the integrat ed Rotrad LS T. The uncertainty of
the determined perc entage tree crown cover is a few percent; additionally, the slight location
uncertainty of MSG pix els increases this uncertainty. The uncertainty in tree crown cover has a higher
influence on ret rieved LS T when the difference in tree and grass temperature is large. At daytime,
local temperature variations are generally larger and any error in the determination of t he
representative targets may contribute to the deviations. At night-time trees and grass have almost the
same temperature and, therefore, differences in cover fractions do not alter the integrated Rotrad
temperature. Thus, at night -time this error is negligible.


COMPARISONS WITH LST FROM METEOSAT 7

For t he period when Meteosat 7 and MSG provided overlapping service, LS T from both s atellites and
from E vora ground-truth data were c ompared wit h eac h other. Since Meteosat 7 has only one infrared
channel (10.5 – 12.5 µm ), only this band was used for LS T retrieval. LS T from Meteosat 7 was
retrieved with a neural network trained with data from a single-channel atmospheric correction method
based on MODTRAN. At the investigated times the mean of the four MS G and Meteosat 7 pixels
representative for the E vora site are compared. The selected data for June to October 2005 is shown
in the left scatter plot in Figure 4. The lower surface temperatures (night-time) from both satellites
match well, but for higher surface temperatures a bias is observed: LST derived from MSG is
systematically higher than LS T from Meteosat 7.




Figure 4: Left graph: LST-MSG plotted against LST from Meteosat 7 from June to October 2005 for mean LST over 4 pixels at the
Evora site. Right graph: ground-truth LST and the LST derived from Meteosat 7 and MSG satellites on 28th of September 2005.

                                                              th
The right hand side of Figure 4 compares LS T on the 28 of Sep 20005 derived from both satellites as
well as from ground measurements: LS T from Meteosat 7 clearly matches ground-truth station data
better than LS T from MS G. This is also observed in the large majority of the investigated days and
gives confidence in the LS T derived from ground meas urements. Therefore, it is concluded that the
too high LS T-MSG around noon are caused by the ret rieval algorithm.


CONCLUSIONS AND OUTLOOK

Comparisons with LS T from the E vora ground-truth station have shown that the LSA-SAF Generalized
Split-Window algorithm for MSG/SEVIRI provides LST with mean absolute differences of about 1.5
±1.5 K. The largest discrepancies occur during daytime: then t he mean value of LS T-MSG minus
station LS T increases to 4 ±1.5 K. In cont rast, at night-time the LS T-MSG has a negative bias
compared to t he ground station and a mean difference of -1 ±1 K is observed. The determined night-
time bias is still within LSA-SAF’s nominal accuracy for the LS T-MS G product (±2 K) and choosing
higher emissivity values can reduc e this bias furt her.
In contrast, the large bias at daytime cannot originate from inaccurat e emissivity. The too high LS T-
MSG seems to be caused by the LSA -SAF ret rieval algorithm or by erroneous input dat a. In this study
the LS T station data were corrected using mean emissivity values from spectral libraries. LS T derived
from ground measurements and LS T retrieved from satellite data should use the same (integrated)
emissivity. However, the emissivity product provided by LSA-SAF is not operational yet and its quality
can not be guaranteed. There is a need to improve the LSA-SAF emissivity product as well as t he
Fractional Vegetation Cover (FV C) product, which is also derived from MSG/SEVIRI. FV C is used in
emissivity retrieval: without accurate FVC there will be no reliable emissivity values. LSA-SAF is
currently improving its FVC and Land Emissivity products and improved LS T products should be soon
available.
In order to ensure the reliability of satellite derived LS T, there is a need to set up further validation
stations for different land covers. Currently preparations are underway to establish validation stations
in Africa. These stations have the advantage to be locat ed near the centre of MS G’s coverage area
and are close to nadir. It is also much easier to find large homogeneous areas in Africa than in
Europe; however, the involved logistics are more complicated. New validation stations will be set up in
western (Senegal) and southern A frica (Namibia) by the end of 2007; by this time first in-situ
radiometric data from these stations are expected. They will be primarily used for LS T-MSG validation,
but also for validation of ot her MS G products. Furthermore, a new series of the successful Meteosat
satellites, Meteosat Third Generation, is under development: these are scheduled for launch in 2015.
The launc h of the first Eumetsat Polar System (EPS) satellites (Metop-1) in 2007 also broadened and
diversified the amount of satellite data available for environmental and meteorological applications.
This leads to an increased need for t he validation of sat ellite products. The existing station in Mit ra
(E vora) has a diverse set of instruments and can be used to validate other satellite products, e.g.
longwave and short wave downwelling fluxes. The new stations in Africa will be equipped with a similar
set of instruments and will also provide standard meteorological data. Once the technical infrastructure
is in place, it is feasible to add instruments for other missions to the stations. The validation activities
at the ground-trut h station in Mitra will be continued and it is planned to carry out a cross-validation
with LS T from the AATS R sensor on board ENV ISA T.


REFERENCES

Dash, P., Olesen, F. S. and Prata A. J. (2004) Optimal land surface temperature validation site in
Europe for MSG. Proceedings of E UME TSA T Meteorological S atellite Conference, P rague, 31 May –
4 June 2004, pp 248 – 264

Gajewska, E., Prata, F., Olesen, F. (2006) Ground-truth for MSG Land S urface Temperature.
Proceedings of Eumetsat Meteorologic al Satellite Conference, Helsinki, 12-16 June 2006.

Gajewska, E., Prata, F., Olesen, F. (2006) Validation of Land Surface Temperatures (LS Ts) derived
from MSG/SEVIRI with the E vora, Portugal ground-truth station measurements. Second Recent
Advances in Quantitative Remote Sensing. Ed. Jos é A. Sobrino. Servicio de Publicaciones.
Universidad de Valencia. Valencia, 2006.

Pinheiro, A. C. T., Privette, J. L., Mahoney, R. and Tucker, C. J. (2004) Directional effects in a daily
AVHRR land surface temperature dataset over Africa. IEEE Trans actions on Geoscience and Remote
Sensing, vol. 42, no. 9, pp 1941 – 1954

Pinheiro, A. C. T., P rivette, J. L. and Guillevic, P. (2006) Modeling the observed angular anisotropy of
land surface temperature in a S avanna. IEEE Transactions on Geoscience and Rem ote Sensing, vol.
44, no. 3, pp 1036 – 1047

Wan, Z., Dozier, J. (1996) A Generalized Split-Window Algorithm for Retrieving Land Surface
Temperat ure from Space. IEEE Trans action on Geoscience and Remote Sensing, vol. 34, no. 4, pp
276 – 281

				
DOCUMENT INFO
Shared By:
Categories:
Stats:
views:12
posted:12/25/2010
language:English
pages:8
pptfiles pptfiles
About