HESSHydrology and Earth System SciencesHESSHydrol. Earth Syst. Sci.1607-7938Copernicus PublicationsGöttingen, Germany10.5194/hess-20-3263-2016Cloud tolerance of remote-sensing technologies to measure land surface
temperatureHolmesThomas R. H.thomas.r.holmes@nasa.govhttps://orcid.org/0000-0002-4651-0079HainChristopher R.AndersonMartha C.https://orcid.org/0000-0003-0748-5525CrowWade T.https://orcid.org/0000-0002-8217-261XHydrology and Remote Sensing Lab., USDA-ARS, Beltsville, MD, USAHydrological Sciences Laboratory, NASA Goddard Space Flight Center, Greenbelt, MD, USAEarth Science Interdisciplinary Center, University of Maryland,
College Park, MD, USAThomas R. H. Holmes (thomas.r.holmes@nasa.gov)11August20162083263327513April201621April20161July201621July2016This work is licensed under a Creative Commons Attribution 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by/3.0/This article is available from https://hess.copernicus.org/articles/20/3263/2016/hess-20-3263-2016.htmlThe full text article is available as a PDF file from https://hess.copernicus.org/articles/20/3263/2016/hess-20-3263-2016.pdf
Conventional methods to estimate land surface temperature
(LST) from space rely on the thermal infrared (TIR) spectral window and is
limited to cloud-free scenes. To also provide LST estimates during periods
with clouds, a new method was developed to estimate LST based on passive-microwave
(MW) observations. The MW-LST product is informed by six polar-orbiting
satellites to create a global record with up to eight observations per
day for each 0.25∘ resolution grid box. For days with sufficient
observations, a continuous diurnal temperature cycle (DTC) was fitted. The
main characteristics of the DTC were scaled to match those of a geostationary
TIR-LST product.
This paper tests the cloud tolerance of the MW-LST product. In particular,
we demonstrate its stable performance with respect to flux tower observation
sites (four in Europe and nine in the United States), over a range of cloudiness
conditions up to heavily overcast skies. The results show that TIR-based LST
has slightly better performance than MW-LST for clear-sky observations but
suffers an increasing negative bias as cloud cover increases. This negative
bias is caused by incomplete masking of cloud-covered areas within the TIR
scene that affects many applications of TIR-LST. In contrast, for MW-LST we
find no direct impact of clouds on its accuracy and bias. MW-LST can
therefore be used to improve TIR cloud screening. Moreover, the ability to
provide LST estimates for cloud-covered surfaces can help expand current
clear-sky-only satellite retrieval products to all-weather applications.
Introduction
Information about the land surface temperature (LST) is an important
element in the retrieval of many hydrological states and fluxes from
satellite-measured radiances. For example, the retrieval of soil moisture or
precipitation from passive-microwave observations requires a coincident
estimate of LST (e.g., Owe et al., 2008). In other
applications, the rate of change in temperature is contrasted with net
radiation to estimate evaporation as a residual of the surface energy
balance (e.g., Anderson et al., 2011).
The most direct way to estimate LST from spaceborne instruments is by
radiometers which measure within the thermal infrared (TIR) band of the
electromagnetic spectrum. Thermal emission within this frequency band can be related directly
to the physical temperature of the land surface and is more precisely
termed the ensemble radiometric temperature (Norman and
Becker, 1995). Spaceborne TIR radiometers allow for very high spatial
resolution imagery. Even at the height of geostationary platforms
radiometers can deliver 3 km spatial resolution, e.g., the Spinning Enhanced
Visible and Infrared Imager (SEVIRI). A drawback to the TIR technique is
that – at such wavelengths – clouds completely block the emission from the
land surface. This means that spaceborne TIR radiometers give no
information about the land surface below the clouds and instead reflect the
temperature and emissivity of the clouds. The result is that the quality of
the cloud screening directly affects the quality of a TIR-LST product.
An alternative, more cloud-tolerant technique is based instead on passive-microwave
(MW) observations. In particular the Ka-band (∼ 37 GHz) was shown to have a strong link with LST
(Owe and Van de Griend, 2001). Based on these
observations, linear-regression-based LST estimates were derived for the
Ka-band (Holmes et al., 2009), and variants
of these linear relations are currently used in soil moisture retrieval
(Jackson et al., 1999; Owe et al., 2008).
However, using a single linear regression across the globe ignores
potentially significant differences in microwave emissivity and can result
in large biases, especially in desert areas. It also cannot account for
large differences in the amplitude of the diurnal cycle between MW- and
TIR-based LST, which have been implicated in reduced soil moisture retrieval
skill during daylight hours
(Lei et al., 2015;
Parinussa et al., 2011). In contrast to these relatively simple linear
methods, a neural network method was developed by Aires et al. (2001) to estimate LST based on multiple microwave channels
besides Ka-band. By using atmospheric and surface information in addition to
TIR-LST in the training of the scheme, this method minimized systematic bias
in monthly mean temperatures. However, because the training is based on a
single polar-orbiting satellite, it cannot give diurnal temperature
information. Both of these methods were compared with station data by
Catherinot et al. (2011), giving strong
confirmation on the lack of sensitivity of microwave-based LST estimates to
cloud liquid water path. This method has recently been developed further to
allow for more continuous application to a more diverse suite of satellites
and overpass times (Prigent et al., 2016). One
drawback to all passive-microwave-based methods is relatively coarse spatial
resolution as compared to TIR sensors. At Ka-band, the smallest footprint
size currently achieved with polar-orbiting satellites is 10 × 15 km.
Because of the complementary nature of TIR- and MW-based LST, there is a
clear interest in merging these two independent technologies. For example,
TIR-LST-based evaporation retrievals would benefit from observational data
during cloudy periods (e.g., Anderson et al.,
2011). On the other hand, microwave soil moisture retrievals from Soil Moisture Active Passive (SMAP) have
the goal of 9 km spatial resolution, and this poses a resolution challenge to
MW-LST inputs if TIR-LST cannot be leveraged. Reconciling the systematic
differences in diurnal temperature cycle (DTC) between TIR and Ka-band is
the first step towards an ultimate merger into a diurnally continuous
LST product. To do this, Holmes et al. (2015) developed
a method to scale the diurnal characteristics of a multi-satellite dataset of
Ka-band observations to a TIR-LST product with 15 min temporal
sampling from geostationary satellites. This scaling was able to account for
biases in characteristics of the DTC related to Ka-band emissivity, sensing
depth and atmospheric effects (Holmes et al., 2015).
By explicitly taking account of systematic differences in DTC between TIR
and Ka-band, this method is able to estimate LST at any time of day from
sparse Ka-band observations. Note that a similar pixel-by-pixel approach was
applied by André et al. (2015) over the Arctic region
where a single satellite can provide diurnal sampling.
The aim of this paper is to evaluate the new global MW-LST dataset in
comparison with existing TIR-LST data over clear-sky days and particularly
to test the assumption that MW-LST is tolerant to high levels of cloud
coverage. Ground observations provide a common benchmark to test the
relative accuracy of the two satellite products. Because the diurnal MW-LST
product (Holmes et al., 2015) was scaled to TIR-based
LST as produced by the Land Surface Analysis Satellite Application Facility
(LSA-SAF; see http://landsaf.meteo.pt), the evaluation is mostly concerned
with temporal precision, not with absolute bias. In previous work it was
shown that relative aspects of a coarse-scale product can be evaluated using
sparse in situ observations
(Holmes et al., 2012).
For a thorough discussion of absolute accuracy, readers are referred to
papers detailing validation exercises for LSA-SAF-LST
(e.g., Ermida et al.,
2014; Göttsche et al., 2013, 2016).
After establishing the accuracy of MW-LST relative to TIR-LST for a
particular site, the stability of the precision of MW-LST (relative to
ground data) for increasing cloudiness will be tested. Previous work showed
indications of cloud tolerance of MW-LST in comparison to TIR-LST
(Holmes et al., 2015), but the analysis used proxies
for both cloud cover and LST quality. In this paper we use a more direct
estimate of cloudiness and provide a more detailed look at the validation
statistics for different levels of daytime cloudiness with the ground
station as the reference. The hypothesis we test is that clouds affect a
satellite-measured LST by introducing an error (E). If E is consistent in
sign throughout the measurement period (e.g., if clouds always lower the
satellite LST estimate), this will introduce a systematic bias that will
increase with cloud cover. If on the other hand the sign of E varies, it will
increase the random error in LST, but not necessarily a systematic bias.
Only if we do not see a systematic bias with increasing cloud cover, nor an
increase in random error, can we reject the hypothesis that clouds affect the
satellite LST.
TIR-LST is available from many sources, including both low Earth-orbiting
satellites and geostationary satellites. Because of our interest in the
diurnal features of LST, used in surface energy balance evaluations, this
study focuses on TIR-LST products developed from geostationary satellites.
The first product is based on the SEVIRI aboard the Meteosat Second
Generation (MSG-9) satellite. MSG-9 is positioned over the Equator at
0∘ longitude. It has geographic coverage of Africa, Europe and the
east coast of South America (with incidence angles below 70∘).
The Land Surface Analysis, Satellite Application Facility (LSA-SAF) produces
operational LST products based on split-window observations (channels
centered at 10.8 and 12.0 µm) of MSG-9. LSA-SAF-LST is
originally produced at 3 km spatial resolution. For this study, the data are
aggregated to match the 0.25∘ resolution of MW-LST. If two-thirds
of the 3 km observations are masked, then the sample average is rejected for
that location and time.
For North America, NOAA operates the Geostationary Operational Environmental
Satellites (GOES). GOES Surface and Insulation Products (GSIP) are produced
by the Office of Satellite Data Processing and Distribution (OSPO), National
Environmental Satellite Data and Information Service (NESDIS), NOAA,
and include LST (V3 was used for this study). Unlike MSG-based LST, GOES
GSIP LST is based on a dual-window technique (3.9 and 11.0 µm) rather
than the preferred split-window technique due to the lack of a 12 μ
thermal channel on the current generation of the GOES imager. An operational
hourly LST at 0.125∘ spatial resolution is available from 2
April 2009 onwards. For this study, we averaged the nine 0.125∘
nodes that cover the edge and center of the 0.25∘ MW-LST
product grid cell. In order to further reduce possible cloud contamination,
a particular data point is only used if all nine 0.125∘
pixels covering the 0.25∘ grid box contain (non-cloud-masked)
data.
Satellite LST estimates: microwave
The MW-LST product is based on vertical polarized Ka-band (36–37 GHz)
brightness temperature (TKa) as measured by microwave radiometers
on six satellites in low Earth orbit. These satellites include the Advanced
Microwave Scanning Radiometer on EOS (AMSR-E) to October 2011 and its
successor on AMSR2 from July 2012. Also included are the Special Sensor Microwave
and Imager (SSM/I) on platforms F13, F14 and F15 of the Defense Meteorological Satellite
Program; the Tropical Rainfall Measurement Mission (TRMM) Microwave Imager (TMI); and Coriolis-WindSat.
These observations are combined to create a global record with up to eight observations per day for
each 0.25∘ resolution grid box. The data are binned in 15 min
windows of local solar time (00:00–00:15 is the first window of the day). The
brightness temperatures are intercalibrated using observations from the
TRMM satellite (with an equatorial overpass) as a transfer reference.
Individual 0.25∘ averages are masked if the spatial standard
deviation of the oversampled Ka-band observations exceeds a prior determined
threshold for a given grid box. Both the intercalibration and quality
control procedures are described in detail in Holmes et al. (2013a).
The methodology to estimate LST from this record of Ka-band observations is
described in Holmes et al. (2015) and summarized
below. For days with suitable observations (a minimum of four, including at
least one within a third day length from solar noon) and no TKa<250 K (an indication of frozen
soil), a continuous DTC is fitted. The DTC model used is based on Göttsche and Olesen (2001) with slight
adaptations to limit the number of parameters. This implementation (DTC3) is
fully described in Holmes et al. (2015). DTC3 summarizes
the DTC with two daily parameters (daily minimum T0 at the start and
end of the day, and diurnal amplitude A) together with diurnal timing
(φ), which is assumed a temporal constant (Holmes
et al., 2013b). The daily mean is defined by the daily minimum and the
amplitude (T‾=T0+A/2). The Ka-band
DTC parameters for individual days (T‾dKa, AdKa)
are scaled to match the long-term mean of TIR observations:
AdMW=AdKa/δ,T‾dMW=β0+β1T‾dKa.
The scaled parameters are indicated with the superscript “MW”. The parameter
δ represents the slope of the zero-order least-squares regression
line for estimating the amplitude of AdKa from TIR-LST (AdTIR). The intercept (β0) and slope (β1) to correct
the mean daily temperature (T‾dKa) for systematic differences
with TIR-LST (T‾dTIR) are determined with a constrained
numerical solver, as in Holmes et al. (2015). The
constraint is based on radiative transfer considerations and assures that
the scaling of the mean is in agreement with the prior scaling of the
amplitude (Eq. 1). These scaling parameters were determined for each
0.25∘ grid box based on data for the period 2009–2012. The scaling
(Eqs. 1 and 2) is applied to every day for which estimates of T‾Ka and AKa are available. Together with the timing of the
diurnal cycle of TIR-LST, φTIR, as determined based on
(Holmes et al., 2013b), we then calculate the diurnal
MW-LST based on the same DTC3 model:
MW-LST=DTC3φTIRT0MWAMW.
Comparing actual Ka-band observations to estimates provided by the fitted
DTC model provides a valuable means of quality control. The root mean square
error (RMSE) of the misfit between the DTC3 model and the sparse TKa
observations is used to flag days where the assumptions imposed by
the shape of clear-sky DTC are not valid or individual Ka-band observations
have a large bias.
Besides the continuous MW-LST product, we can also evaluate the product at
the actual Ka-band observation times (thus weakening our reliance on the
DTC3 model). To do this, we project the difference between the original MW
data and the DTC model fit onto MW-LST. This product is referred to as
MW-LST-Sparse:
MW-LST-Sparse=DTC3φTIR,T0MW,AMW+1δTKa-DTC3φKa,T0Ka,AKa.
In MW-LST-Sparse the impact of the DTC3 model is limited to providing the
minimum and amplitude of the diurnal. The differences between the
observations and the diurnal model at the actual observation time are
preserved. The difference between the continuous MW-LST and the
MW-LST-Sparse is illustrated in Fig. 1 (lower panel).
An example of the available data at station B, showing 8-day time
series of station-measured shortwave incoming radiation
(S↓, or SW, top) and LST (bottom graph).
In the top graph S↓ (black lines) is
compared to the clear-sky expected value,
Sclear↓ (Eq. 7, red dash), to
illustrate the computation of the cloud cover proxy
(Acloud, Eq. 8, values in blue text). In the lower
graph the station LST is compared to the TIR and MW satellite LST products.
The dashed line is an occurrence where the MW-LST is masked due to a high
misfit between sparse observations and the diurnal model.
Ground observations
FLUXNET is a worldwide network of meteorological measurement towers (flux
tower) with common measurement protocols
(Baldocchi et al., 2001). Each
flux tower includes an instrument positioned above the vegetation canopy to
measure net radiation. This instrument is made up of two pyranometers to
measure up- and downwelling shortwave radiation and two pyrgeometers to
measure up- and downwelling longwave radiation. The radiometric surface
temperature (T) can be derived from the longwave radiation measurements
(upwelling: L↑, W m-2; downwelling: L↓,
W m-2) using the following equation:
L↑=εLσT4+1-εLL↓,
where εL is the broadband emissivity over the spectral range
of the pyrgeometer (4.5–42 µm) and σ is the Stefan–Boltzmann
constant (σ= 5.670 × 10-8 W m-2K-1).
List of ground validation targets detailing geographic location and
International Geosphere-Biosphere Programme (IGBP) vegetation type. Local parameters determined for each station are
weight (W) in spatial average, longwave emissivity and clear-sky
absorption.
* Sites are located in neighboring grid cells.
Reference: C: Marek et al. (2011).
Longwave emissivity
An estimate of εL is obtained for each site based on
measurements of L↑ together with additional measurements of
screen level air temperature (Ta), sensible heat flux (H) and wind speed
(u). This estimate is based on three assumptions: (1) H is directly
proportional to the near-surface temperature gradient, (2) the difference
T-Ta represents that temperature gradient and (3) H=0 when there is
no vertical temperature gradient (i.e., T=Ta). With these three
assumptions in mind we then iterate over εL to find the
solution where the regression of H against (T-Ta) goes through zero
and the squared errors are minimized. When measurements of wind speed are
available, they are used to select atmospheric conditions where the
relationship between H and the near-surface temperature gradient is strongest
(u > 2 m s-1). For forest sites, the direct relationship between H and T gradients
breaks down. In those cases, the simpler assumption is used that the long-term average of T and Ta are equal:
〈T〉=〈Ta〉. For more discussion and examples
of this method see Holmes et al. (2009). In this study we apply this method
to determine monthly εL for each site individually and then
use the median value of εL (listed in Table 1) to calculate
T based on Eq. (5). The standard deviation of the monthly measurements of
εL is also listed in Table 1 and provides an indication of
both uncertainty and seasonal variation in εL.
Spatial representativeness and site selection
The tower-based estimate of T, from (Eq. 5), directly represents only the
immediate tower surroundings within a radius of approximately 50 m. Clearly
this is a very small spatial sampling of the 0.25∘ grid box
(∼ 25 × 25 km) represented by the satellite LST estimates used
in this study (see Sects. 2.1 and 2.2). As a consequence, we typically find
large systemic differences between the station data and the areal average.
Given that overall weather conditions are relatively homogenous over
distances of 25 km, these differences can be attributed to the land cover
type at the station location in comparison to that over the entire grid box
(for examples of this, see Holmes et al., 2009). The representation of the spatial
average by ground observations can be improved significantly if more than
one station is available in the same grid box and the towers are situated in
thermodynamically contrasting land cover types (forest and
cropland/herbaceous). In that case the land cover associated with the tower
site (subscript, s) determines the weight (W) according to the spatial
fraction of that land cover type within the 0.25∘ grid box
(MCD12C1; Friedl et al., 2010). We use this information to
estimate the grid average LST as the weighted average of site-measured T
according to Eq. (6):
LST=∑s=1nWsTs∑s=1nWs.
For example, site DE-Hai of location A is located in a forest and represents
16 % of the pixel. Site DE-Geb is located in croplands and represents
80 % of the pixel that has low vegetation cover or bare soil. Urban, open
water or wetland accounts for the remaining 4 %, which does not affect
the weighting. We only considered locations where this rest fraction is
below 5 % of the grid coverage. Another criterion for site selection was
that the land cover at the site must represent more than 75 % of the
pixel. Sites in mountainous areas are also excluded. For the period of
2009–2012 this means there are 13 locations with at least 2 years of flux
tower sites available for this study, and four of these locations contain
multiple stations. For pixels where only one station is available, LST is
set equal to the site measurement: LST =T. All the validation
targets are listed in Table 1, together with the geographic location of the
individual stations, the land cover type as reported by the flux tower
operators and the parameters W and εL as described above.
Cloudiness at tower location
The downwelling shortwave radiation (S↓) as measured at the
flux tower is strongly affected by the amount of condensed water in the
atmosphere. We can therefore use the reduction in site-measured daytime
S↓ relative to an expected value during clear skies as a
proxy for cloudiness. The clear-sky irradiance Sclear↓ is
estimated based on top-of-atmosphere solar irradiance (STOA), which can
be calculated based on geographic location and day of year (Van
Wijk and Ubing, 1963). Even on a clear day, atmospheric absorption reduces
the irradiance at the surface by 20 to 30 % from the top-of-atmosphere
value. We estimate this clear-sky absorption (Aclr) by calculating the
slope (β) of the zero-order linear regression between S↓ and
STOA for days that are in the highest quintile of S↓/STOA : Aclr=1-β.
These estimates of Aclr (listed in Table
1 for each individual site) range from 0.22 to 0.31 and show a good
agreement between stations of the same cluster. We use the minimum recorded
value for each cluster to calculate Sclear↓:
Sclear↓=STOA1-Aclr.
By using Sclear↓ to normalize measured S↓,
we account for solar zenith effects and can formulate a measure for
shortwave cloud absorption (Acloud), expressed in percentage:
Acloud=100Sclear↓-S↓Sclear↓.
This definition of Acloud is used as a measure of
cloudiness and calculated based on 3 h totals of insolation for the
daytime between 06:00 and 18:00. Obviously this definition of cloud absorption
does not apply when the Sun is below the horizon. For nighttime hours we
use the neighboring daytime window:
Acloud0-6=Acloud6-12,Acloud18-24=Acloud12-18.
Figure 1 gives an example of the site-measured S↓ and the
calculated Acloud for an 8-day summer period at
station B (top panel). The bottom panel shows the site-measured LST and
illustrates how the temporal sampling of the satellite products is affected
by clouds.
Statistical metrics
In the description of the results we make use of standard statistical
metrics. In terms of absolute error metrics we report bias, the long-term
mean difference between satellite product and in situ data, and the
RMSE. By removing the long-term mean difference, we can
calculate RMSE of the unbiased data (ubRMSE). These three metrics are
related as follows:
RMSE=ubRMSE2+bias2.
We further report standard error of estimate (SEE) as a measure of temporal
precision:
SEE=σ1-ρ2,
where σ is the standard deviation of in situ data and ρ is
Pearson's correlation coefficient.
Results
We acquired data for 17 field sites in 13 unique grid locations with data
records within the 4-year time period of 2009 to 2012. Table 2 lists the
amount of days with at least 12 hours of observations for either MW or
TIR-LST to give an overall sense of available validation data for this
study. In total we have 13 316 data days of in situ data (36 out of 44 data
years). Of these data days, 50 % also have MW-LST estimates, and 36 %
have TIR-LST observations. For MW, this percentage is negatively affected by
the gap between AMSR-E (radiometer turned off in October 2011) and AMSR2
(first observations in July 2012). The MW-LST product is heavily reliant on
these satellites with a midday overpass for constraining the diurnal
amplitude. For TIR, the percentage is particularly low for GOES due to a
more stringent cloud filter than employed for MSG.
Percentage coverage for two LST products.
MSG domain 2009–2011 Frost-free days SatAllAllAllClear-skyCloudyproductskyskyTIR3642479414MW5055647556
To better represent the relative data coverage that is possible with the two
LST retrieval techniques, we focus on the four station pairs in the MSG
domain and limit the time period to 2009–2011. We further focus on the days
where the minimum temperature (as measured at the station) stays above
freezing (the MW method is not applicable for subfreezing temperatures).
Within this smaller subset, we have 2506 data days of in situ data, and the
coverage of MW is 55 % in comparison with the 42 % coverage for TIR.
However, breaking this down by cloud cover reveals the big difference in
coverage resulting from the wavelength-dependent tolerance to clouds. During
clear skies, the coverage of TIR is 93 %, and MW comes in at 76 %
(mainly attributed to the 2-out-of-3-days revisit for ascending AMSR-E).
During cloudy days the coverage drops to 13 % for TIR, whereas for MW it
maintains 55 % coverage.
In the following section we want to answer two questions. How does the MW-LST
compare to TIR-LST in relation to ground data during days with clear
skies? And is the performance of MW-LST affected by clouds? We focus on
hourly average temperatures for days where the station data remain above
1 ∘C to avoid snow or frozen surface conditions.
Comparison between TIR-LST (x axis) and MW-LST (y axis) in terms
of their validation metrics with station LST for frost-free and cloud-free
days. From left to right the three panels show (a) ubRMSE, (b) mean bias
and (c) bias in Δ (LST). Each marker represents the statistics as
calculated for individual locations as identified by the letter (see Table 1
for definition). For the GOES domain the filled markers highlight the
stations used in the cloud analysis. Black lines provide visual support and
indicate targets (e.g., 1:1 line, cross at zero bias).
Clear-sky comparison of satellite LST products
The ground observations provide a common benchmark to test the relative
accuracy and bias of the two satellite products with the same in situ data.
Days with clear skies are selected based on the measure of cloudiness as
defined in Sect. 2.3.3, with a maximum accepted value of
Acloud=0.2. This is in addition to the cloud
screening performed in the generation of the TIR products (Sect. 2.1), the
quality control of the MW-LST (Sect. 2.2) and the selection of frost-free
days. Even though the spatial representativeness and uncertainty in
εL may insert systematic errors in the estimation of the
spatial average from the ground stations, they can be used as a reference to
compare different satellite LST products.
For each of the 13 validation targets we tabulate (see Table 3) ubRMSE and
bias (see Sect. 2.4 for definition). By excluding long-term bias, the
ubRMSE gives an indication of the overall data quality, which includes the
random error and errors resulting from a mismatch in variance (either
seasonal or diurnal). Errors in spatial representation of the sites affect
MW and TIR in the same way. In order to highlight the relative performance
of the two satellite products with respect to the common benchmark, we
compare their performance directly in Fig. 3.
Encouragingly, the multi-site average ubRMSE for the four European FLUXNET
sites shows the MW-LST (2.3 K) to be only moderately higher than TIR-LST
(2.1 K) for these frost-free and cloud-free observations. This is a positive
result for MW-LST because of the extra processing needed to correct MW data
for sensing depth differences with TIR. Both satellite products have a
higher ubRMSE with the AmeriFlux stations, but again the multi-site average
ubRMSE for MW-LST (3.0 K) is only slightly higher than that for TIR-LST
(2.8 K). Figure 2a compares the ubRMSE with in situ data directly for the two
satellite technologies. The high correlation between the two methods is an
indication that the spurious effect of spatial representation of the site
affects both methods to similar degrees. Of all the stations, MW-LST has a
lower ubRMSE at 5 of the 13 stations, and the only stations where we record
more than 0.5 K difference in ubRMSE between TIR and MW-LST are FLUXNET
station D (2.5 K for MW vs. 1.7 K for TIR) and AmeriFlux stations I (3.5 K
for MW vs. 2.9 K for TIR) and J (3.4 K for MW vs. 2.4 K for TIR). These
stations, together with K and L, all have dry conditions with low
vegetation. When there is less vegetation, the influence of soil emissivity
on the observed Ka-band brightness temperature becomes larger. Small changes
in soil moisture can affect the soil emissivity and will result in biases
for MW-LST when a constant emissivity is assumed (as in the current
implementation). This points to possible improvements when the scaling to
TIR is performed at shorter window lengths, perhaps in 3-month moving
windows.
Because MW-LST is scaled directly to TIR-LST, its bias is almost completely
determined by the bias between TIR-LST and the site (Table 3 and Fig. 2b).
The European FLUXNET sites fall within the MSG domain, and these data
years were part of the data on which the scaling of MW-LST is trained
(Holmes et al., 2015). Although the mean bias (Fig. 2b) is almost identical,
the bias in morning heating (ΔT, Fig. 2c) has more variation between
the two satellite products. It is interesting that generally the satellite
products overestimate ΔT compared to ground data: on average they
both overestimate the recorded heating at the stations by about 10 %.
RMSE and bias of satellite LST with regression-corrected in situ
data for five levels of cloud cover ((Acloud, Eqs. 8–10). From left to right
are locations A–G (see Table 1 for site information). Results for TIR-LST
(red) are contrasted with those for MW (blue). Markers indicate that more
than 15 days with data were available for a particular cloud cover bin.
Green dashed lines indicate the results for MW-LST-Sparse. For each site and
cloud interval the percentage coverage of the temporal record is depicted in
the top row with half-rounds in proportion to the number of data pairs. The
potential number of data pairs (grey) refers to the number of in situ data
points for each cloud bin. The actual number of data pairs is superimposed
on this for MW (blue) and TIR (red).
Cloud tolerance of satellite LST
To test the stability of the MW-LST for increasing levels of cloudiness, we
took a closer look at the four sites in Europe and three in the US (sites A–G). To
isolate the effect of clouds on the agreement between satellite and ground
observations, we first remove structural differences by fitting a linear
regression for each location, based on data with cloudiness below 20 %
(Γ – see Sect. 2c). We then divide the data into five equal bins of
increasing cloud coverage from 0 to 100 %. The RMSE and mean difference
(bias) between the satellite data and the regression-corrected in situ data
is then calculated for each 20 % cloud bin. The purpose is to test the
assumption that MW-LST is tolerant to higher levels of cloud coverage.
Figure 3 shows the result of this analysis for locations A–G (from left to
right). For each location the data coverage (top row), RMSE (middle row)
and bias (bottom row) are displayed for the five 20 % bins of cloudiness.
First of all, the large increase in negative bias with increasing cloudiness
for the TIR-LST product stands out. At all stations we see a clear negative
bias in response to increasing cloudiness for TIR-LST, and the overall
agreement between stations is striking. At 40–60 % cloud cover, all
stations but one show a significant negative bias for TIR-LST. Above 60 %
cloud cover all stations (where TIR-LST is still available, presumably due
to failure of the cloud mask) show a negative bias of 2 K or more. This
clearly shows that for TIR-LST we have to accept the hypothesis that clouds
affect the satellite LST estimate, even after a cloud mask is applied. It is
well known that TIR observations are sensitive to clouds and that a failure
to mask for cloud conditions will result in an underestimated LST (for land
surface above freezing). Because of this systematic response to clouds, the
bias metric by itself is a good indicator of the effect of cloud
contamination in clear-sky TIR-LST products. The symbols in the top row show
the diminishing temporal sampling with increasing cloud cover. When we
contrast this with the size of the bias, it is clear that the cloud mask as
implemented in the LSA-SAF product (for sites A–D) is not sufficient at
removing cloud artifacts. The GOES product (for sites E–G) appears to remove
times with high cloud values more completely. Although investigating the
efficacy of cloud masks for TIR techniques is not the purpose of this paper,
it does help illustrate how cloud effects can be identified with these
ground stations.
Validation results by MW satellite (all data,
a.m. or p.m. overpass), aggregated for locations A–M (see Table 1).
2009 linear regression 2015 diurnal scaling SatelliteOverpassRMSESEEBiasRMSESEEBiasNAMSR-E4.530.73.22.60.26227a.m.4.41.933.12.1-1.12817p.m.4.62.7-1.53.32.61.43410AMSR24.42.41.631.90.8843a.m.51.73.72.62-0.7347p.m.3.81.6-0.33.21.61.9478WindSat3.52.30.83.12.5-0.63080a.m.3.32.11.43.12-0.21617p.m.3.71.90.232.1-0.91463SSM/I3.82.40.93.12.6-0.35401a.m.3.52.31.13.12.20.22740p.m.42.10.832.3-0.82661Average (all sites) 4.12.41.23.12.40.6Forest (sites: C, E, H) 5.12.14.3320.5Low vegetation 3.92.50.93.22.50.7
In clear contrast to the TIR-LST products, the response of MW-LST to
increasing cloudiness is much more muted and not as consistent across
stations. Stations A, B, C and E show no response in terms of bias, and
below 80 % cloud cover there is no station with a MW-LST bias of more
than 1 K. One station shows a negative trend (D), and two stations show a
positive trend (F and G). But only above 80 % cloudiness do these trends
result in bias error of greater than 1 K. Because we see both positive and
negative biases in the MW-LST analysis, we cannot rely solely on the
bias metric to assess the impact of clouds. If there are cancelling biases
affecting an individual station, this could suppress the bias. The increased
retrieval error would still be reflected in an increased RMSE. However, the
RMSE of MW-LST changes minimally relative to its baseline value at 0–20 %
cloudiness and mirrors the size of the bias. This indicates that there is
little potential for “hidden” biases behind these numbers. For MW-LST we can
therefore reject the hypothesis that clouds affect the satellite LST
estimate.
The MW-LST-Sparse product (Eq. 4) adopts the same scaling with TIR as the
diurnal MW-LST but has much less sensitivity to the imposed shape of the
diurnal model (DTC). For clear skies this distinction is negligible, as
apparent from the almost identical values of ubRMSE shown in Table 3. The
effect of the clear-sky model is likely to be higher on days with cloudy or
partially cloud-covered sky. And although the sparse set only has four–eight
observations per day, it allows more samples on days with complex
temperature changes. Such days are removed from the MW-DTC product if no
good match is found between the diurnal model and observations. We can
therefore use the MW-LST-Sparse product to test for undue influence of the
DTC model (and its related quality flags) on the relationships between LST
errors and cloudiness. The response in bias of MW-LST-Sparse to increasing
cloudiness is almost identical to the response of MW-LST for each station
(see Fig. 3). In terms of RMSE the sparse set shows values equal to or higher
than the diurnally-continuous MW-LST product, which is not surprising as it
does not have the smoothing and quality control associated with the DTC
model.
All-sky validation by satellite overpass time
The MW-LST record is a combination of different satellites. In the following
analysis the validation results of the MW-LST product are broken down by
time of day and satellite input record. All data pairs where the minimum
temperature at the station stays above freezing are included in this
analysis, regardless of cloud cover. It is interesting to compare these
results to the much simpler approach that uses a single linear regression
model globally (Holmes et al., 2009).
Table 4 lists RMSE, SEE and bias for the old and new approach. The
statistics are aggregated for all locations as listed in Table 1. The mean
scaling with TIR-LST results in a drop in bias for the MW-LST, reducing the
average RMSE by 1 K. Part of this reduced RMSE results from the improved
characterization of the amplitude of the diurnal cycle, which improves the
slope at all times of day and accounts for 0.2 K of the improvement in RMSE.
The impact on the precision (quantified here by SEE) is mixed – on average
there is no change. Biggest improvements in all metrics are recorded for the
forest locations (sites C, E and H).
Discussion
Considering all eight locations used in the cloud analysis, we see little to no
response to clouds in terms of bias and RMSE for MW-LST, and this allows us
to reject the hypothesis that clouds negatively affect its accuracy.
However, for three sites we do find weak and opposing biases at higher cloud
coverage which require an explanation. The wavelength of Ka-band (8 mm) is
2 orders of magnitude larger than a typical cloud droplet (10 µm).
Therefore, any effect of clouds on MW-LST would stem from changes in
associated meteorological conditions like atmospheric vapor content and
temperature profiles and their potential impact on Ka-band emission
processes. According to the zero-order radiative transfer model, an
increased atmospheric opacity (through increasing atmospheric water content)
increases the weight of the atmospheric contribution to the satellite-measured
brightness temperature, relative to the top of vegetation emission.
The sign and size of the effect of a change in atmospheric opacity thus
depend on the contrast between the atmospheric temperature and the land
surface temperature times the effective emissivity. It is therefore possible
that this could explain the site-to-site differences in bias as shown in Fig
3. Analyzing the overall effect of the atmosphere on biases in MW-LST will
require more detailed atmospheric profile information coupled with a
radiative transfer model.
Another possible explanation is that the positive biases recorded at
locations F and G are related to scale differences between the site and the
0.25∘ grid cell. Spatial heterogeneity in LST is likely more
pronounced during clear-sky periods when spatially varying soil and
vegetation yield a strong influence on the daytime temperature gradients.
During cloudy periods the temperature gradients are not as pronounced and
more directly linked to the more uniform air temperature. If the mean
temperature at the station is generally higher than the areal mean LST, and
this bias diminishes with increasing cloudiness, this would be transformed
through our clear-sky training into a positive bias for the satellite
product at high cloudiness. However, this effect would affect both MW and
TIR to the same extent. We have tested this at locations with two stations
in contrasting land cover types (A–C). What we found is that indeed it is
possible to “rotate” the bias response by changing the weights of the
individual stations and that this rotation affects both MW and TIR-LST.
This effect of site representation can therefore explain the greater
variation in response from station to station for locations where only one
station was available (D–G).
Conclusions
In this paper, a recently developed satellite MW-LST product is compared to
ground station data and satellite TIR-based LST products. The MW-LST was
developed to complement TIR-LST with a coarser spatial resolution but at a
higher temporal resolution. The higher temporal resolution of MW-LST is
based on the assumption that MW has a relatively high tolerance to clouds,
which allows for observations at times when no TIR observations are
possible. This paper tests this assumption by looking at the precision with
respect to ground stations for increasing levels of estimated cloudiness.
Our analysis is performed at the 0.25∘ spatial resolution as
predicated by the MW-LST product. At this coarse spatial resolution, the
overall unbiased RMSE between TIR-LST and ground stations during clear-sky
days is 2.1 K for the four locations in the MSG domain, and 2.8 K for the
nine locations in the GOES domain. For the same locations we find that the MW-LST
is only slightly higher (+0.2 K for both domains).
With increasing cloudiness the RMSE increases significantly for TIR-LST,
caused by a matching negative trend in bias that is seen at all seven
locations. This demonstrated the known effect that clouds have on TIR
estimates of LST. The fact that these trends are so apparent highlights the
limitations of current cloud screening techniques as employed in TIR-LST
products that are in general use. In clear contrast to this we find a much
more limited response in both RMSE and bias for MW-LST. Because of this we
conclude that there is no significant direct impact of clouds on the
accuracy of the MW-LST product. However, at three stations the size and sign
of the response is such that further research is needed to identify the
exact causes introducing error in MW-LST. By taking into account the
atmospheric humidity and temperature profile, further analysis may
investigate the extent to which this mixed response can be explained by
atmospheric conditions associated with cloudiness. Alternatively, if a
greater database were available of locations with flux tower sites in
contrasting land covers, this could be used to isolate the role of scale
mismatch between station and the satellite product.
As an immediate outcome the result of this work highlights the utility of MW
technology for cloud screening of TIR-LST. This is something that will be
explored in future work. Ultimately, the goal is to find the best way of
combining MW and TIR technology for the estimation of LST from space.
Data availability
Time series of MW-LST and TIR-LST covering the locations and time period of
this paper are available upon request from the corresponding author. The
global source data for MW-LST are publicly available. They are aggregated
from several data centers, and we would like to thank Goddard Earth Sciences
(GES) Data and Information Services Center (DISC) for archiving and
distributing TRMM satellite as acquired by NASA's Earth-Sun System Division,
the National Snow and Ice Data Center for archiving and distributing Aqua-AMSR-E
data, and NOAA's Comprehensive Large Array-data Stewardship System (CLASS)
for dissemination of Defense Meteorological Satellite Program data. LSA-SAF
disseminates EUMETSAT products. This work further used data acquired by the
FLUXNET community (fluxnet.ornl.gov) and in particular by the following
networks: AmeriFlux and CarboEuropeIP.
Acknowledgements
This work was funded by NASA through the research grant “The Science of
Terra and Aqua” (13-TERAQ13-0181). We would further like to thank Li Fang
(NOAA) for preparation and interpretation of GOES LST.
Edited by: A. Loew
Reviewed by: C. Prigent and one anonymous referee
ReferencesAires, F., Prigent, C., Rossow, W. B., and Rothstein, M.: A new neural
network approach including first guess for retrieval of atmospheric water
vapor, cloud liquid water path, surface temperature, and emissivities over
land from satellite microwave observations, J. Geophys. Res.-Atmos.,
106, 14887–14907, 10.1029/2001JD900085, 2001.Anderson, M. C., Kustas, W. P., Norman, J. M., Hain, C. R., Mecikalski, J. R., Schultz, L., González-Dugo, M. P.,
Cammalleri, C., d'Urso, G., Pimstein, A., and Gao, F.: Mapping daily evapotranspiration at field to continental scales
using geostationary and polar orbiting satellite imagery, Hydrol. Earth Syst. Sci., 15, 223–239, 10.5194/hess-15-223-2011, 2011.André, C., Ottlé, C., Royer, A., and Maignan, F.: Land surface
temperature retrieval over circumpolar Arctic using SSM/I–SSMIS and MODIS
data, Remote Sens. Environ., 162, 1–10, 10.1016/j.rse.2015.01.028,
2015.
Baldocchi, D., Falge, E., Gu, L. H., Olson, R., Hollinger, D., Running, S.,
Anthoni, P., Bernhofer, C., Davis, K., Evans, R., Fuentes, J., Goldstein,
A., Katul, G., Law, B., Lee, X. H., Malhi, Y., Meyers, T., Munger, W.,
Oechel, W., U, K., Pilegaard, K., Schmid, H. P., Valentini, R., Verma, S.,
Vesala, T., Wilson, K., and Wofsy, S.: FLUXNET: A new tool to study the
temporal and spatial variability of ecosystem-scale carbon dioxide, water
vapor, and energy flux densities, B. Am Meteorol. Soc., 82, 2415–2434,
2001.Catherinot, J., Prigent, C., Maurer, R., Papa, F., Jiménez, C., Aires,
F., and Rossow, W. B.: Evaluation of “all weather” microwave-derived land
surface temperatures with in situ CEOP measurements, J. Geophys. Res.-Atmos., 116, D23105, 10.1029/2011JD016439, 2011.
Ermida, S. L., Trigo, I. F., DaCamara, C. C., Göttsche, F. M., Olesen,
F. S., and Hulley, G.: Validation of remotely sensed surface temperature over
an oak woodland landscape—The problem of viewing and illumination
geometries, Remote Sens. Environ., 148, 16–27, 2014.
Friedl, M. A., Sulla-Menashe, D., Tan, B., Schneider, A., Ramankutty, N.,
Sibley, A., and Huang, X.: MODIS Collection 5 global land cover: Algorithm
refinements and characterization of new datasets, Remote Sens. Environ.,
114, 168–182, 2010.Göttsche, F.-M., and Olesen, F. S.: Modelling of diurnal cycles of
brightness temperature extracted from METEOSAT data, Remote Sens. Environ.,
76, 337–348, 10.1016/S0034-4257(00)00214-5, 2001.Göttsche, F.-M., Olesen, F.-S., and Bork-Unkelbach, A.: Validation of
land surface temperature derived from MSG/SEVIRI with in situ measurements
at Gobabeb, Namibia, Int. J. Remote Sens., 34, 3069–3083,
10.1080/01431161.2012.716539, 2013.Göttsche, F.-M., Olesen, F.-S., Trigo, I. F., Bork-Unkelbach, A. and Martin, M. A.: Long Term Validation
of Land Surface Temperature Retrieved from MSG/SEVIRI with Continuous in-Situ Measurements in Africa, Remote Sens., 8, 410, 10.3390/rs8050410, 2016.Holmes, T. R. H., De Jeu, R. A. M., Owe, M., and Dolman, A. J.: Land surface
temperature from Ka band (37 GHz) passive microwave observations, J.
Geophys. Res.-Atmos., 114, D04113, 10.1029/2008JD010257, 2009.Holmes, T. R. H., Jackson, T. J., Reichle, R. H., and Basara, J. B.: An
assessment of surface soil temperature products from numerical weather
prediction models using ground-based measurements, Water Resour. Res.,
48, W02531, 10.1029/2011WR010538, 2012.Holmes, T. R. H., Crow, W. T., Yilmaz, M. T., Jackson, T. J., and Basara, J.
B.: Enhancing model-based land surface temperature estimates using
multiplatform microwave observations, J. Geophys. Res.-Atmos., 118,
577–591, 10.1002/jgrd.50113, 2013a.Holmes, T. R. H., Crow, W. T., and Hain, C.: Spatial patterns in timing of the diurnal temperature cycle,
Hydrol. Earth Syst. Sci., 17, 3695–3706, 10.5194/hess-17-3695-2013, 2013b.Holmes, T. R. H., Crow, W. T., Hain, C. R., Anderson, M., and Kustas, W. P.:
Diurnal temperature cycle as observed by thermal infrared and microwave
radiometers, Remote Sens. Environ., 158C, 110–125,
10.1016/j.rse.2014.10.031, 2015.
Jackson, T. J., Le Vine, D. M., Hsu, A. Y., Oldack, A., Starks, P. J.,
Swift, C. T., Isham, J. D., and Haken, M.: Soil moisture Mapping at regional
scales using microwave radiometry: The Southern Great Plains hydrology
experiment, IEEE Trans. Geosci. Remote Sens., 37, 2136–2149, 1999.Lei, F., Crow, W. T., Shen, H., Parinussa, R. M., and Holmes, T. R. H.: The
Impact of Local Acquisition Time on the Accuracy of Microwave Surface Soil
Moisture Retrievals over the Contiguous United States, Remote Sens., 7,
13448–13465, 10.3390/rs71013448, 2015.Marek, M. V., Janouš, D., Taufarová, K., Havránková, K.,
Pavelka, M., Kaplan, V., and Marková, I.: Carbon exchange between
ecosystems and atmosphere in the Czech Republic is affected by climate
factors, Environ. Pollut., 159, 1035–1039,
10.1016/j.envpol.2010.11.025, 2011.Norman, J. M. and Becker, F.: Terminology in thermal infrared remote sensing
of natural surfaces, Agr. Forest Meteorol., 77, 153–166,
10.1016/0168-1923(95)02259-Z, 1995.Owe, M. and Van de Griend, A. A.: On the relationship between thermodynamic
surface temperature and high-frequency (37 GHz) vertically polarized
brightness temperature under semiarid conditions, Int. J. Remote Sens., 22,
3521–3532, 10.1080/01431160110063788, 2001.Owe, M., De Jeu, R. A. M., and Holmes, T. R. H.: Multisensor historical
climatology of satellite-derived global land surface moisture, J. Geophys.
Res.-Earth, 113, F01002, 10.1029/2007JF000769, 2008.Parinussa, R. M., Holmes, T. R. H., Yilmaz, M. T., and Crow, W. T.: The impact of land surface temperature on soil moisture
anomaly detection from passive microwave observations, Hydrol. Earth Syst. Sci., 15, 3135–3151, 10.5194/hess-15-3135-2011, 2011.Prigent, C., Jimenez, C., and Aires, F.: Towards “all weather”, long
record, and real-time land surface temperature retrievals from microwave
satellite observations, J. Geophys. Res.-Atmos., 121, 5699–5717, 10.1002/2015JD024402, 2016.
Van Wijk, W. R. and Ubing, D. S.: Radiation, in Physics of plant
environment., North-Holland Publ. Co., Amsterdam, 62–101, 1963.