Introduction
Mean merged, clear-sky surface skin temperature values
from GOES-East, GOES-West, Meteosat-9, MTSAT-2, and INSAT-3D, October 2015.
Surface skin temperature (Ts) is a critical quantity for characterizing
the exchange of energy between the Earth's surface and the atmosphere.
Consistent land and ocean measurements of Ts are essential for regional
and global climate assessment and weather model data assimilation. Surface
energy balance and top-of-atmosphere (TOA) radiative budget calculations rely
on the accuracy of these surface parameters (Bodas-Salcedo et al., 2008). In
addition to surface flux analyses, Ts retrievals are used to minimize
model prediction uncertainty by updating model state values with observations
at regular time steps – an important consideration for climate and numerical
weather prediction (NWP) models (Garand, 2003; Tsuang et al., 2008; Reichle
et al., 2010; Ghent et al., 2010; Guillevic et al., 2012; Draper et al.,
2015). The modeling community could benefit significantly from the provision
of frequent, spatially contiguous, global land and ocean Ts data
(Rodell et al., 2004; Bosilovich et al., 2007). Many other uses of Ts as well as
the status and future of Ts retrievals are summarized by Li et al. (2013). It is clear that the need is growing for higher accuracy, global
coverage, and greater temporal and spatial resolution of Ts retrievals
from satellite imager data.
Satellite-based Ts retrieval, validation, and modeling studies originate
from a variety of sources, e.g., the National Environmental Satellite, Data,
and Information Service (NESDIS) or the National Oceanic and Atmospheric
Administration (NOAA) via the Advanced Very High Resolution Radiometer
(AVHRR) series and the Geostationary Operational Environmental Satellite
(GOES) sensors (Prata, 1993, 1994; Coll and Caselles, 1997; Sobrino and
Raissouni, 2000; Kerr et al., 2004; Sobrino et al., 2004; Yu et al., 2009,
2012a, b; Sun et al., 2012). Specifically, using two different
single-channel land surface temperature (LST) algorithms, Heidinger et al. (2013) and Minnis et al. (2016) found good agreement with the NOAA ESRL
Surface Radiation (SURFRAD) network in verification studies using LST
retrievals from GOES and AVHRR alone, respectively. Furthermore, near-realtime (NRT) LST is produced operationally from Meteosat Spinning
Enhanced Visible and Infrared Imager (SEVIRI) data, which offer continuous
coverage of Europe and Africa, and served as the focus of several LST
validation studies (Sobrino and Romaguera, 2004; DaCamara, 2006; Kabsch et
al., 2008; Trigo et al., 2008; Göttsche et al., 2013). Retrievals using
radiances from the Moderate Resolution Imaging Spectroradiometer (MODIS)
have been both the target and standard for a number of LST verification
studies (Wan et al., 2002, 2004, 2008; Coll et al., 2009; Jiménez et
al., 2012). Duan et al. (2014) used four daily observations from Terra and
Aqua MODIS to capture the diurnal cycle of LST, which is critical for full
characterization of the climate system. Wang et al. (2014) conducted a
three-way Ts comparison using MODIS, in situ ground observations, and
model simulations. They note the high importance of accurate cloud-clearing
and the inherit difficulties of resolution scaling when comparisons are
conducted between satellite data and point references – conclusions
supported in a similar MODIS daytime LST verification study conducted by
Williamson et al. (2013).
Average surface skin temperature from NOAA-19 AVHRR,
October 2013.
With more reliable calibrations, operational geostationary Earth orbiting
(GEO) and low Earth orbiting (LEO) satellite imagers are being used to
derive cloud and radiation properties in NRT, e.g., Minnis et al. (2008a).
The combination of GOES-East (GOES-13), GOES-West (GOES-15), Meteosat Second
Generation (MSG; Meteosat-9 or Meteosat-10), MTSAT-2 (recently replaced by
Himawari-8), and the Indian Space Research Organization INSAT-3D provides
high-temporal-resolution (1 h nominal) quasi-global Ts data produced
in NRT, with a shared single-channel retrieval algorithm (e.g., Fig. 1). The
methodology (Sect. 3) is flexible and easily transportable to other GEO
and LEO imagers, including the current AVHRR instruments on the NOAA and
EUMETSAT MetOp platforms. AVHRR Ts retrievals supplement the GEO data
and fill in missing measurements over polar regions (e.g., Fig. 2). This
same method is being applied to historic and current imager datasets,
particularly as part of the Satellite ClOud and Radiative Property retrieval
System (SatCORPS) analyses of AVHRR data for provision of a NOAA Climate
Data Record (Minnis et al., 2016), and for MODIS, GEO, and Suomi National
Polar-orbiting Partnership (S-NPP) Visible Infrared Imaging Radiometer Suite
data as part of the Clouds and Earth's Radiant Energy System (CERES) project
(e.g., Minnis et al., 2010).
This article highlights recent improvements made to the SatCORPS NRT
satellite Ts product (Scarino et al., 2013), via comparisons of GOES and
AVHRR Ts retrievals with established sea surface temperature (SST) and
LST reference datasets. The influence of NWP source on retrieved Ts
values is also examined. The main improvements over the earlier version are
enhanced pixel-level resolution output and hourly GEO retrieval time steps.
The SatCORPS Ts retrieved from GOES and AVHRR data is evaluated by
comparing with reference datasets based on in situ and satellite
measurements. Section 2 provides an overview of these product and reference
datasets, as well descriptions of ancillary validation sets and how the
reanalysis input is configured. Explanation of the single-channel Ts
retrieval algorithm is provided in Sect. 3.
The results and discussion are presented together in two sections: one for
SST and the other for LST. Section 4 focuses only on SatCORPS SST results
because the validation and angular corrections differ from those used for
LST. Validation of the LST and application of the theoretical model
developed by Vinnikov et al. (2012) comprise the main topics of Sect. 5.
Coefficients for the Vinnikov et al. (2012) model were determined
empirically from near-simultaneous measurements covering a limited viewing
range from a small collection of sites. Because those sites represent varied
surface and climate types, the resulting coefficients could potentially
serve as an effective initial step in the process of correcting LST for
angular dependency. The included subsections then highlight our independent
assessment of the Vinnikov et al. (2012) model over a large area, starting
with details on the origin and use of the correction model (Sect. 5.1), testing of
its broad influence (Sect. 5.2) with validation results/discussion relative to
satellite (Sect. 5.3) and ground references (Sect. 5.4), and an uncertainty discussion
related to the spatial homogeneity surrounding the ground reference (Sect. 5.5).
Finally, Sect. 6 summarizes the main conclusions. The combined GEO and
AVHRR retrievals allow for high-resolution temporal monitoring of the
Ts diurnal cycle, an essential state variable for numerical weather
model data assimilation and climate studies (e.g., Draper et al., 2015). The
Ts products and uncertainties described here should be valuable for
improving surface energy flux analyses and numerical weather prediction
due to their NRT global availability over land and ocean.
Data
Satellite data for surface skin temperature retrieval
Clear-sky surface skin temperature is retrieved from channel 4 (11 µm) radiances taken by the NOAA-18 or NOAA-19 AVHRR for the period
January–December 2008 or 2013, respectively, in the Global Area Coverage
(GAC) format. The nominal satellite equatorial crossing time is 13:30 LT. A GAC pixel radiance is formed by
averaging the radiances of four consecutive raw 1 km AVHRR pixels along the
scan direction. The process is repeated after skipping the fifth pixel and
so on to produce consecutive GAC pixels along the scan line. Two scan lines
are then skipped and the pixel averaging is applied again to the third scan
line. Thus, a GAC pixel nominally covers a 1 km × 4 km area (a
4 km2 pixel) but, because of sampling, represents a 3 km × 5 km area that yields an effective resolution of ∼ 4 km. The
AVHRR data were analyzed with the SatCORPS-A1 methodology (Minnis et al.,
2016) to retrieve cloud properties, TOA broadband fluxes, and clear-sky
surface skin temperature. Clear pixels are determined from the SatCORPS
cloud mask. Details of the skin temperature retrieval process are given in
Sect. 3.
Hourly channel-4 (10.8 µm) data from GOES-13 (GOES-East) and GOES-15
(GOES-West) taken during January, April, July, and October (hereafter, JAJO)
2013 are used to retrieve Ts for validation with surface and other
satellite surface skin temperature datasets. Furthermore, GOES-13 and
GOES-15 data are employed to test the angular anisotropy parameterization.
The nominal GOES imager resolution is 4 km. The pixels are sub-sampled,
however, to an effective resolution of 8 km during full disk and hourly
hemispheric scans. That is, every other pixel is skipped during realtime
full disk and hemispheric processing to improve computational speed and
produce more manageable output file size. The actual pixel measurements and
geolocation attributes, however, are still representative of a 4 × 4 km2 area. These data were analyzed with a version of SatCORPS-A1
adapted to the GOES channels as described by Minnis et al. (2008a).
Validation data
For validation comparisons, this study employs surface and satellite-based
references. The SatCORPS AVHRR SST values are compared to the daily
high-resolution blended SST analysis described by Reynolds et al. (2007). It
comprises the NOAA “Optimum Interpolation” SST (OI SST) version 2
high-resolution dataset, which consists of a global 0.25∘ × 0.25∘ grid of blended satellite (AVHRR two- and
three-channel algorithms and Advanced Microwave Sounding Radiometer (until
2011) data) and in situ measurements of daily SST. It covers the period from
January 1981 to the present.
The version 5 Aqua MODIS LST/Emissivity product (MYD11_L2;
hereafter, MYD11), which is derived from the generalized split-window
algorithm (Wan and Dozier, 1996; Wan and Li, 1997; Snyder and Wan, 1998), is
used to validate the SatCORPS AVHRR and GOES LST values. The dataset
includes values of LST retrieved from clear-sky 1 km MODIS pixels and
surface spectral emissivity values. Because MYD11 is derived from different
data using a different type of algorithm, and is accurate to ±1 K or
less (Wan, 2008; Wan et al., 2002, 2004), it serves well as an independent
reference for comparing with the GOES retrievals.
Surface radiometer measurements from the Atmospheric Radiation Measurement
(ARM) Southern Great Plains (SGP) Central Facility (36.3∘ N,
97.5∘ W) 11 µm upwelling/downwelling infrared thermometer
(IRT; Morris, 2006) serve as another LST validation source. The ARM IRT
ground-based radiation pyrometers provide measurements of the equivalent
blackbody brightness temperature for the 9.6–11.5 µm spectral band
every 60 s. From a 10 m height with 30.5∘ FOV, the
upwelling IRT, with a specified accuracy of ±0.5 K, measures the
effective ground radiating temperature, i.e., the temperature equivalent of
the ground infrared radiant energy assuming the surface emissivity
(εs) is equal to 1.0 (Morris, 2006). A true skin
temperature Ts can, therefore, be determined as
Ts=B-1BTo-1-εs×BTo↓εs,
where εs is from the CERES 11 µm database (e.g., Chen
et al., 2004) and the spectral downwelling narrowband brightness temperature
(To↓) is measured by a 2 m height up-looking IRT, which is
oriented so that the zenith view of the sky is reflected into the lens by a
gold mirror, and has a narrow 2.64∘ FOV (Morris, 2006). The Planck
function for the particular waveband is B(T), and To is temperature
equivalent to the surface-leaving blackbody radiance. Note that the ARM
downwelling IRT at the Lamont, OK, Central Facility was no longer operating
in 2013, and therefore To↓ was acquired from the nearby Lamont,
OK,
Extended Facility downwelling IRT, which operates in unison with the Central
Facility instrument. It is expected that there is negligible variation in
To↓ over the ∼ 200 m distance between the two
sites.
Supplementary ASTER data
Following the studies of Wang and Liang (2009) and Guillevic et al. (2014),
high-resolution Terra Advance Spaceborne Thermal Emission and Reflection
(ASTER) LST and emissivity (AST_08 and AST_05,
respectively) product data from 2001 through 2015 (available complete years)
are used to measure the spatial homogeneity of LST in both a 4 × 4 and 8 × 8 km2 area centered on the ARM SGP ground
station. The ASTER LST product has a 90 m spatial resolution at nadir,
derived from five infrared channels using the temperature–emissivity
separation (TES) method (Yamaguchi et al., 1998; Gillespie et al., 1998).
Each ASTER granule consists of 700 × 830 LST pixels, which can be
referenced to two 11 × 11 matrices of geocentric latitude and
geodetic longitude. Bilinear interpolation is used to estimate latitude and
longitude for each LST pixel. These LST values are subsetted into blocks of
45 × 45 pixels to simulate the spatial extent of a 4 km GOES pixel
centered on the Central Facility and to assess the spatial
representativeness of the site relative to the surrounding region. A subset
block of 89 × 89 pixels is meant to represent the worst possible
disparity between the satellite measurement and the ground station based on
the most extreme pixel-to-point matches possible, i.e., the ARM site being
situated in any corner of the GOES pixel.
Reanalysis input
Numerical weather model output parameters are used as input to compute TOA
brightness temperatures (TTOA). These include the model surface air
(Ta′) and skin (Ts′) temperatures, and vertical temperature and
humidity profiles. The realtime GEO retrievals employ National Centers for
Environmental Prediction (NCEP) Global Forecast System (GFS; EMC, 2003)
model forecasts accessed from the Man-computer Interactive Data Analysis
System (McIDAS; Lazzara et al., 1999). Non-realtime GEO studies utilize
either GFS or Modern-Era Retrospective Analysis for Research and
Applications (MERRA; Rienecker et al., 2011) reanalyses. The impacts of using one
reanalysis or the other are examined by analyzing the same satellite data
using each of the two reanalyses during the Ts retrieval.
MERRA data have a spatial resolution of 0.5∘ latitude × 0.66∘ longitude over the globe. The surface skin temperature is
available hourly, while the temperature and humidity profiles are provided
every 6 h. A total of 43 atmospheric layers are used. The version of GFS
used here has a 1.25∘ horizontal resolution and up to 11 levels in
the vertical, and it provides data every 6 h. No model values of
Ts′ are available in the GFS version over land, so Ts′ is estimated
from Ta′ as a function of local time and season.
Single-channel skin temperature retrieval
The method for calculating Ts from 11 µm TTOA observations is
an updated, higher-resolution version of that described by Scarino et al. (2013). Because some imagers (e.g., AVHRR-1, GOES-13) lack split-window
capabilities, the single-channel method best allows historical consistency
in application amongst many distinct sensors (Sun and Pinker, 2003;
Jiménez-Muñoz and Sobrino, 2010; Heidinger et al., 2013). The process to
determine Ts first employs the cloud mask algorithm developed for CERES
to classify pixels as cloudy or clear on a chosen grid (Minnis et al.,
2008b). The algorithm relies on comparisons of observations with estimates
of the clear-sky TTOA or reflectance at 0.65, 3.8, and 10.8 µm.
Those estimates are made using the CERES 10'-regional clear-sky albedo and land
surface emissivity databases (Chen et al., 2004, 2010), along with the
appropriate bidirectional and directional reflectance models, angularly
dependent sea surface emissivity models, predicted skin temperature, and
corrections for atmospheric absorption and emission (Minnis et al., 2011).
The emissivity for water surfaces is estimated using a wind-speed-dependent
model developed from theoretical calculations using the approach of Jin et al. (2006). A constant wind speed of 5 knots is assumed for all sea surface
pixels.
The observed or modeled radiance at the TOA can be represented as
BTTOA=∏i=n1tiBTo+1-t1BT1+∑i=n21-tiBTi∏j=i1tj,
where To is the surface-leaving radiant energy equivalent brightness
temperature, which comes from Ts based on the following relationship
using the narrowband surface emissivity:
Ts=B-1BTo-1-εs×L↓εs.
L↓ is the downwelling radiant energy at the surface:
L↓=1-tnBTn+∑i=1n-11-tiBTi∏j=ni+1tj.
The subscripts i and j denote an atmospheric layer, where 1 and n refer to the
layers at the TOA and just above the surface, respectively (e.g.,
B(T1)≡B(TTOA)). The atmospheric layer temperature is Ti, and
B is evaluated at the central wavelength of the 11 µm band. B-1 is
the inverse Planck function. The layer transmissivity (ti) is determined
using the correlated k distribution technique, which accounts for gaseous
absorption within the spectral band of a given channel. This technique is
described in detail by Goody et al. (1989) and Kratz (1995), who depict
the discrete version of the spectral-mean transmission tΔω(u,p,Θ) as
tΔωu,p,Θ≅∑i=1nwiexp-kip,Θu,
where ki(p,Θ) is an absorption coefficient as a function of pressure
p and temperature Θ for a particular wavenumber ω, u is a
pathlength, and wi is a weighting factor for which the summation over
n calculations must equal 1. Although the technique does not explicitly
account for the details of the spectral response function, the
transmissivity from the surface to the TOA is the same with and without the
details of the spectral response function for the 11 µm band included
in the calculations (Kratz, 1995).
The surface temperatures and atmospheric profiles are linearly interpolated
temporally to the satellite image time and spatially to the center of each
0.5∘ × 0.5∘ AVHRR or 1.0∘ × 1.0∘ GEO grid box. These grid boxes of interpolated NWP sounding
data are called regions. For AVHRR retrievals, the regions have resolutions
up to 1.5∘ × 1.5∘ near the poles but are
nominally 0.5∘ × 0.5∘ everywhere else. The same
Ts retrieval methodology is used for all resolutions. The specific logic
of the cloud mask algorithm can be found in Minnis et al. (2008a, 2010,
2016) and Trepte et al. (2010), which describe cloud tests for different
scenarios (e.g., scenes over snow or desert, sun-glint-influenced ocean,
scenes with smoke or thin cirrus). It is important to note that although the
NWP skin temperature Ts′ is used as a seed value in the initial
application of the cloud mask, decisions based solely on the difference
between 11 µm observations and model values occur for only 2.3 %
(5.3 %) of the pixels over land during the day (night). Therefore, the
initial influence Ts′ is significantly diminished.
After the cloud mask is applied, the mean 0.65 µm reflectance and
mean 3.8 and 10.8 µm TTOA (i.e., < TTOA >) values are computed from the observed values for the
clear and cloudy pixels for each region. The data are then analyzed in pixel
groupings called tiles. For AVHRR, the tiles are 8 × 12 pixels in
area, and for GEO the tiles are the same resolution as the gridded region,
or 1.0∘ × 1.0∘. The different tile sizes are
employed to facilitate optimal processing speed. The GEO data are analyzed
in NRT, while the AVHRR data have been used for climate studies, which do
not have the same time constraints as NRT applications. If at least 20 %
of the pixels within the tile are considered clear, the mean observed
clear-sky temperature replaces the original NWP-based clear-sky temperature
for the region and the cloud mask is repeated using the observed clear-sky
mean brightness temperature. The 20 % criterion is used to minimize the
influence of cloudy pixels on the final temperature value while still
allowing sufficient sample size. If fewer than 20 % of the pixels are
clear, then the original clear-sky estimate Ts′ and cloud mask are
retained and no value Ts is retrieved.
For those tiles satisfying the 20 % criterion, a value of Ts for each
pixel is determined using a two-step process. First, the tile mean value
Ts (i.e., < Ts> ) is determined by solving Eq. (2)
from the inverse of Eq. (3) (i.e, To′ solved from Ts′). Then, the mean
observed 11 µm clear-sky < TTOA > is used to
adjust Ts′ based on the difference between the < TTOA > and the modeled TTOA′ for each tile. That is, a
correction is applied to the model Ts′ and temperature/humidity profiles
such that TTOA′ computed with Eqs. (2) and (5) equals < TTOA >,
thereby yielding < Ts >. For
the AVHRR retrievals, the tile average < TTOA > represents an area that is smaller than the area represented by
TTOA′, which is a regional value originating from the region-scale MERRA
Ts′. Thus, all AVHRR tiles with their center within a given MERRA
region use the same model profiles and Ts′. For the GEO Ts retrieval,
both the observed < TTOA > and the modeled
TTOA′ correspond to a 1.0∘ × 1.0∘ region,
because tile area matches region area for GEO imagers.
To save computational time, a value of Ts is estimated for each pixel in
the tile as
Ts=B-1RTBTTOA,
where RT is the ratio
RT=BTsBTTOA,
and TTOA is the observed clear-sky brightness temperature for the pixel.
This approach assumes that the atmospheric attenuation and contribution to
the exiting radiance is proportionally the same throughout the region. It
yields Ts pixel values that differ by -0.04 ± 0.20 K from
Ts computed using Eqs. (2) and (5) for each pixel.
July 2008 (a) AVHRR SST, (b) NOAA OI SST, (c) SST
difference, and (d) scatter density analysis of ∼ 3 million
daily matched grid cells.
July 2013 (a) GOES-13 SST derived, in part, from GFS-based
atmospheric corrections, (b) NOAA OI SST, (c) SST difference, and (d) scatter
density analysis of ∼ 1 million daily matched grid cells.
Land surface temperature angular anisotropy correction
Satellite-observed LST depends on the viewing and illumination conditions
because shading, vegetation conditions, soil type, and topography affect the
radiance exiting the scene (Lagouarde et al., 1995; Minnis and Khaiyer,
2000; Minnis et al., 2004). This thermal radiation anisotropy can result in
the retrieved LST varying by 6 K or more for some areas (Rasmussen et al.,
2010, 2011; Guillevic et al., 2013). From experimental measurements,
Sobrino and Cuenca (1999) and Cuenca and Sobrino (2004) found a viewing
zenith angle (VZA) dependence of LST that depends on soil type. Pinheiro et al. (2006) developed a physical model to estimate the variation of LST as a
function of canopy coverage, solar zenith angle (SZA), VZA, and relative
azimuth angle (RAA) for a savanna. Rasmussen et al. (2010, 2011) developed
and applied a similar model to predict the LST that would be retrieved by
Meteosat over Africa. Vinnikov et al. (2012) constructed a generalized model
to convert satellite-measured, VZA-, SZA-, and RAA-dependent LST into a
direction-independent equivalent physical temperature, which for general
application, requires many sets of matched measurements from different angle
sets to construct coefficients for the necessary kernels. Addressing the
anisotropic effects and thereby reducing the Ts uncertainties could
improve climate monitoring and be of significant benefit to data
assimilation and numerical weather prediction needs (Reichle et al., 2010;
Guillevic et al., 2013; Draper et al., 2015).
July 2013 (a) GOES-13 SST derived, in part, from
MERRA-based atmospheric corrections, (b) NOAA OI SST, (c) SST difference, and
(d) scatter density analysis of ∼ 1 million daily matched grid
cells.
Accounting for 3-D radiance anisotropy for a global retrieval methodology
will require the development of regional and seasonal kernels for a
universal model (e.g., Vinnikov et al., 2012) or developing canopy
configurations globally for physical models (e.g., Rasmussen et al., 2010).
Such endeavors require many different matched datasets for a sufficiently
large configuration of viewing/illumination angle combinations across many
scene types and all seasons. Therefore, at present, we choose to employ the
Vinnikov et al. (2012) universal empirical model for angular anisotropy
correction. The model, built from varied surface and climate property
observations, can serve as a baseline for angular anisotropy correction
despite its development from a small number of ground sites across a limited
viewing range. Therefore, the goal of this section is to independently test
the efficacy of the model through use of large-area satellite and
independent ground site LST comparisons. If effective, a universal
anisotropic correction model such as this is certainly beneficial to NRT
global retrieval of satellite-based LST. As with Vinnikov et al. (2012), our
initial assessment will start regionally, i.e., within the GOES-East and
GOES-West satellite domains.
AVHRR (2008) and GOES-13 (2013) SST accuracy and
precision relative to NOAA OI SST. For the GEO retrievals, the atmospheric
correction is based on either GFS or MERRA reanalysis. Atmospheric
corrections for AVHRR retrievals are strictly based on MERRA.
Nadir-normalization model for LST retrieval anisotropy
Vinnikov et al. (2012) formulate the skin temperature at a given set of
viewing and illumination angles as
Tsθ,θo,φ=Tn1+a1-μ+bψθ,θo,φ,
where θ,θo, and φ correspond to VZA, SZA, and
RAA, respectively. The LST at nadir is Tn, and μ=cos(θ). In
sequence, the three terms within the brackets are referred to as the isotropic,
emissivity, and solar kernels. Vinnikov et al. (2012) chose the functional
form of the solar kernel as
ψθ,θo,φ=sinθcosθosinθocosθo-θcosφ,
while coefficients a and b are determined empirically. Physically, the solar
kernel attempts to account for the impact of solar intensity and shadowing,
as well as the hotspot effect. At night, the solar kernel is defined as 0, i.e.,
ψ(θ, θo≥ 90∘, φ) ≡ 0.
The emissivity kernel accounts for the VZA dependence of the effective
emissivity, which can be due to the VZA variation of the emissivity of a
pure surface, the changing combination of scene components (e.g., grass,
rocks, tree canopy, mountain slopes, valleys) and their respective
temperatures as VZA changes, or a combination of the two.
Vinnikov et al. (2012) first estimated the emissivity kernel by determining
the coefficient a in Eq. (8) at night by matching nearly simultaneous GOES-E
and GOES-W Ts measurements with five ground site measurements of LST.
The differences in the VZAs for the two satellites, covering a range of
43 to 66∘, provided the variation in μ needed to
perform the regression fit. The solar kernel coefficient for each site was
then determined in the same manner using the daytime GOES measurements with
the assumption that the emissivity kernel is the same for any hour of the
day. This follows if one considers that when the solar kernel is integrated
over the entire range of RAA (0–360∘), it reduces
to zero. Thus, the solar kernel, in effect, represents deviations from the
emissivity kernel. Thus, features such as the hotspot, which occurs in a
solar backscatter position when θo=θ, are
compensated by lower values at a different RAA, typically in shadow, or over
a range of RAAs at the same value of θ (e.g., Minnis et al., 2004).
It is possible, therefore, that the emissivity kernel, or VZA correction
model, could be determined during the daytime by taking measurements over a
sufficient range of SZAs and RAAs at a given VZA. Doing so, however, would
require a significantly large number of matched datasets in order to achieve
sufficient sample size for every interdependent VZA, SZA, and RAA
configuration across different surface types and seasons, and thus is much
simpler to accomplish at night.
GOES-East/West LST comparison
To test the efficacy of the Vinnikov et al. (2012) three-kernel anisotropic
correction, differences between the hourly GOES-East (GE) and GOES-West (GW)
LST retrievals from July 2013 were computed before and after applying the
angular adjustments. Prior to differencing, the 15 min discrepancy in the
image retrieval at the 3 h synoptic times (00:00, 03:00, …, 21:00 UTC) was mitigated by adjusting the GE LST, which is based on images
beginning 15 min before the UTC hour, to that UTC hour when the GW image
scan began. This approach accounts for the specific GE and GW scanline time
discrepancies. The GE data were linearly interpolated to the GW time using
the nearest surrounding synoptic hours. When those surrounding hours crossed
the sunrise terminator, no correction was applied because of the day-night
discontinuity in LST that occurs shortly after sunrise. Data taken near the
terminator (SZA between 80 and 100∘) were not used.
The image times at the non-synoptic hours are nearly identical, so no
temporal normalization was required. To minimize calibration differences,
the average nocturnal LST difference, 0.08 K, between GOES-13 and 15 within
0.5∘ longitude of 105∘ W, which is the bisector of the two
views, was computed and added to all GOES-15 values.
July 2013 day, night, and combined matched GOES-East minus
GOES-West mean clear-sky surface skin temperature difference (K) for regions
east and west of 105∘ W, without and with anisotropic correction.
The sample-weighted average bias is shown in the bottom row.
Combined
Day
Night
Longitude
Without
With
Without
With
Without
With
correction
anisotropic
correction
anisotropic
correction
anisotropic
correction
correction
correction
< 105∘ W
2.07
1.10
2.38
1.39
1.18
0.35
> 105∘ W
-0.53
0.01
-0.51
0.04
-0.55
-0.03
All
0.85
0.59
0.98
0.73
0.43
0.18
Viewing zenith angles for (a) GOES-West and (b) GOES-East, and (c) their differences over the matching domain.
Mean regional GOES-East–GOES-West LST differences for
July 2013. (a) Day without correction, (b) day with anisotropic correction,
(c) night without correction, and (d) night with anisotropic correction.
Figure 7 plots the VZAs for GW (Fig. 7a), GE (Fig. 7b), and the GE–GW VZA
differences (Fig. 7c). Although the differences are generally less than
±30∘, the largest VZAs are up to 70∘ or more. At
night, the emissivity kernel from Eq. (8) would suggest large LST differences
for pairs matched at the higher VZAs in this domain. All the retrieved
values of normalized LST for both satellites were adjusted to nadir to
account for the anisotropic dependence.
The mean regional differences, i.e., DTs= LST(GE)-LST(GW), are
shown in Fig. 8 for the matched July 2013 data. During daytime, DTs for
the unadjusted values (Fig. 8a) is mostly positive east of 105∘ W
and negative to the west. Notable exceptions include the positive values in
the west corresponding the highest mountain ranges in Colorado, Utah,
Mexico, Washington, Wyoming, Idaho, and New Mexico. After adjusting to nadir
(Fig. 8b), the same patterns remain, but they are mitigated considerably with
DTs values closer to zero. Also, the corrected differences for some of
the regions at extreme VZAs in the far northeast remain relatively large,
perhaps because the viewing dependence increases nonlinearly for large VZA.
At night, the unadjusted differences (Fig. 8c) are relatively small, i.e.,
|DTs| < 2, in most regions. The positive
differences are no longer evident over the high mountains. Applying the
anisotropic correction further reduces |DTs| to values
less than 1.0 K in nearly all cases (Fig. 8d).
Table 1 summarizes the GE–GW results. Over the eastern and western halves
of the domain, |DTs| drops by 0.99 and 0.54 K,
respectively, during the day with the application of the anisotropy
adjustment. The mean regional differences are much smaller than before
correction, especially for the western region where the difference is near
zero. Similarly at night, the corresponding regional differences decrease by
comparable amounts and are much closer to zero than without the corrections.
Furthermore, the mean absolute biases for both day and night, which are
determined by the east–west sample-weighted region differences (not
shown), are much closer after correction – reduced by a factor of 2 or
more. Overall, the mean bias for the entire domain after correction over all
non-terminator hours is 0.59 K.
Mean hourly, regional GOES-East–GOES-West LST
differences for January, April, July, and October 2013, (left) without and
(right) with anisotropic correction. The vertical dashed lines indicate the
terminator transitions to night (blue) and day (red) at
37.5∘ N, 105∘ W.
Although it significantly reduces the GE–GW differences, the anisotropic
correction does not eliminate all of the disagreement between the two
satellite retrievals. Also, although sign difference between the means over
the eastern and western domains essentially disappears for both day and
night with the correction, the remaining east–west difference suggests other
factors aside from angular anisotropy affect the observed temperatures. It
is possible that the solar azimuthal dependence seen in earlier studies
(e.g., Minnis et al., 2004; Vinnikov et al., 2012) is not balanced out for the
configurations seen here. The azimuthal dependence includes effects from
both the relative solar azimuth angle and the azimuthal orientation of the
terrain and vegetation. Moreover, the heating/cooling rates probably differ
between the eastern and western domains because of humidity and altitude
differences. Downwelling longwave radiation might play a greater role in the
diurnal cycle of Ts in the eastern domain, perhaps diminishing the
solar-induced anisotropy. It would be instructive to derive a
daytime-specific emissivity kernel over the entire range of RAA in order to
test these theories, but as alluded to in the previous subsection, such an
endeavor is beyond the purview of this paper, which merely is meant to
assess the current anisotropy model, as given by Vinnikov et al. (2012),
applied to a large satellite dataset.
We can, however, explore how DTs changes over the course of a day and
how much the anisotropic correction diminishes those differences. To that
end, the differences were averaged for each UTC for each of the four months
and are plotted in Fig. 9. The July results corresponding to Fig. 8 are
plotted in Fig. 9e and f as lines connecting the means at each hour. Over
the western domain (red line), the uncorrected DTs (Fig. 9e) gradually
approaches zero at 09:00 UTC from near -1 K after 03:00 UTC, when the sun has set
over the entire domain. At 12:00 UTC, it rises rapidly to a peak of 2.5 K near
16:00 UTC and drops precipitously after 17:00 UTC to -3 K at 22:00 UTC before
increasing until 03:00 UTC. In the east (blue line), DTs drops slowly toward
zero after 01:00 UTC but only reaches 0.4 K at 06:00 UTC before increasing again.
It only increases significantly after 12:00 UTC, maximizing at 3.5 K (17:00 UTC)
before decreasing to 1.3 K at 21:00 UTC, when it levels off. The behavior is
rather different for the corrected values (Fig. 9f), with the two curves
being much closer together between 03:00 and 16:00 UTC, while also being much
closer to zero overall than without the angular correction. The corrected
western domain DTs rises to near 1.0 K from 09:00 to 11:00 UTC and then
drops slightly to about 0.3 K at 12:00 UTC before gradually rising to 1.0 K
again by 17:00 UTC. At 17:00 UTC and after, the curves diverge significantly with
the eastern data varying more extremely (rapid and continuous increase from
13:00 through 21:00 UTC) than their western counterparts, suggesting different
heating/cooling rates. The bias for the entire domain (black line) shows
definitively that the afternoon points are mainly responsible for the
overall positive bias in Table 1. The results for the other months show that
the model generally reduces the mean absolute DTs at most hours. Some
exceptions are seen at night.
Seasonal and diurnal calibration gains for 2013 based on a
satellite ray-matching calibration technique described by Minnis et al. (2002). The gain coefficients derive from the ratio of mean GOES-13 and
Aqua MODIS time-, space-, and angle-matched radiances, later converted to
brightness temperature (BT). A MODIS-consistent GOES-13 BT is attained by
multiplying the appropriate gain with GOES-13 BT values. The gains include
adjustment for the spectral band difference between the GOES and MODIS
channels following the technique described by Scarino et al. (2016). Those
spectral band adjustment factor (SBAF) slope and offset values were applied
to MODIS BT values (converted from radiance) during the cross-calibration in
order to yield MODIS BTs that were spectrally consistent with GOES-13.
January
April
July
October
Day
Night
Day
Night
Day
Night
Day
Night
Calibration gain
0.9998
1.0018
1.0003
1.0027
1.0001
1.0030
1.0007
1.0015
SBAF slope
1.004
1.004
1.006
1.006
1.005
1.005
1.006
1.005
SBAF offset
-0.694
-0.708
-1.012
-0.994
-0.867
-0.811
-1.029
-0.952
Even with different heating/cooling rates, it is expected that DTs would
approach zero after correction for anisotropic effects as the surface air
and skin temperature equilibrate. Instead of going to zero after 03:00 UTC,
DTs drops to roughly -0.8 K for the entire domain by 04:00 UTC until about
06:00 UTC, before rising more rapidly to about 0.8 K at 09:00 UTC and remaining
relatively flat until 12:00 UTC. This odd behavior is likely an artifact of the
sun–satellite configuration, which causes a change in the infrared channel
calibrations at satellite midnight and for 3–4 h afterward. Yu et al. (2013) found that the GOES-11 and GOES-12 10.7 µm (channel 4)
brightness temperatures were biased by -0.5 K relative to their daytime
calibrations for 3–4 h after satellite midnight, even after an
operational correction for the midnight effect had been applied. A smaller
bias was evident for a couple of hours prior to midnight. This residual bias
could explain the unexpected variation in DTs seen between 03:00 and 12:00 UTC,
if GOES-13 and 15 suffer from a similar UTC bias. Assuming then that the
calibration biases are -0.40 and -0.80 K 2 h before and for 4 h after midnight, respectively, for GE, and the same for GW, then
DTs would almost follow the black curve in Fig. 9f exactly (assuming that
DTs= 0 in a perfectly calibrated system). By 06:00 UTC, DTs would
still be near -0.4 to -0.8 K because only GE is influenced by the midnight
effect. By 07:00 UTC, the smaller GW pre-midnight bias would partially offset
the GE bias causing DTs to rise until 09:00 UTC, when only GW is affected.
After 12:00 UTC, the daylight in the eastern half of the domain would overwhelm
any remaining bias. Of course the results discussed here only represent one
domain during one month, although DTs diurnal cycles are shown for other
seasonal months in Fig. 9. The midnight calibrations and the
viewing/illumination angles vary with time of year. Thus it is clear that a
much more comprehensive study would be needed to fully assess the angular
anisotropy dependence of the retrieved Ts values in this context.
Overall, however, application of the three-kernel model nets meaningful
reduction of |DTs| and can perhaps be improved further
by incorporating terrain considerations to account for differential
heating/cooling rates.
Bias and SDD values (K) for the GOES-13 and the MYD11
Aqua MODIS product comparison for day, night, and all times combined
separated by 2013 seasonal month, without and with the Vinnikov et al. (2012) three-kernel anisotropic correction applied. The numbers in
parentheses indicate the sample size for that month.
January (920)
April (1992)
July (2401)
October (1615)
Without
With
Without
With
Without
With
Without
With
correction
correction
correction
correction
correction
correction
correction
correction
Combined
0.88
0.44
0.85
0.24
1.25
0.25
0.84
0.38
Bias
Day
1.34
0.72
1.30
0.49
1.89
0.52
1.69
0.98
Night
0.25
0.08
0.29
-0.07
0.24
-0.13
-0.03
-0.27
Combined
1.79
1.49
1.76
1.28
1.42
1.11
1.90
1.46
SDD
Day
2.05
1.71
2.05
1.37
1.07
1.06
2.19
1.45
Night
1.11
1.06
1.05
1.08
1.01
1.03
1.08
1.13
Probability distributions of LST differences from
GOES-13 and the MYD11 Aqua MODIS product for day, night, and all times
(combined) (a) without and (b) with the Vinnikov et al. (2012) three-kernel
anisotropic correction applied, for 2013.
Validation with independent MODIS LST, MYD11
The JAJO 2013 GOES-13 LST values are compared with the independent MYD11
product between 60∘ N and 60∘ S to determine if the
Vinnikov et al. (2012) angular parameterization improves the consistency of
the two products. The GOES-13 10.8 µm channel was first
cross-calibrated as in Minnis et al. (2002) against its Aqua MODIS
counterpart, band 31, for day and night and the JAJO seasonal months in
order to minimize any calibration differences. As part of the
cross-calibration, spectral differences were accounted for via convolution
of Infrared Atmospheric Sounding Interferometer (IASI) hyperspectral
brightness temperature measurements over the GOES-13 and Aqua MODIS
11 µm channel spectral response functions, as thoroughly detailed by
Scarino et al. (2016). The diurnal/seasonal calibration coefficients and
spectral band adjustment factors (SBAF) are provided in Table 2. For each
Aqua overpass, the MYD11 pixel LST values are converted to pixel radiance
and are averaged on the 1∘ × 1∘ GOES-East
domain. The mean radiance values are then converted to mean LST and matched
to within 15 min of the GOES-13 hourly scans, provided there are at
least 150 valid MODIS and GOES-13 pixels per grid cell. To eliminate any
differences due to surface emissivity discrepancies, the GOES-13 LST was
retrieved using the MYD11 band 31 emissivity values. To effect the
comparisons, both the GOES-13 and MYD11 LST values were normalized to the
nadir view using Eq. (8).
Figure 10 shows histograms of the differences between the GOES-13 and MYD11
LSTs without (Fig. 10a) and with (Fig. 10b) the Vinnikov et al. (2012)
anisotropic correction. Without correction, the GOES LSTs tend to be greater
than their MYD11 counterparts, especially during the day. The SDD is
greatest during the day at 1.82 K, and the day and night GOES biases are
1.62 and 0.19 K, respectively, resulting in a combined (both day and
night) 0.98 K bias. After applying the anisotropic correction, the daytime
SDD drops to 1.33 K, with the bias decreasing by 1.0 down to 0.62 K. The
nocturnal bias drops to almost -0.12 K, while its SDD increases slightly
from 1.07 to 1.09 K. Although the nocturnal bias changes sign, the
magnitude is less than that prior to the correction.
The anisotropic correction serves the comparison well, for the most part,
across all seasons. Table 3 summarizes the bias and SDD adjustments for day,
night, and combined times across the JAJO seasonal months. For all months,
the daytime and combined bias and SDD values reduce substantially following
nadir normalization. At night, however, the SDD changes rather subtly and
sometimes does not constitute an improvement. In October, for example, the
SDD increases from 1.08 to 1.13 K at night, and this is also the only
month when the absolute bias increases (-0.03 to -0.27 K). Overall, the
angular anisotropy adjustments reduce the bias by ∼ 0.7 K, a
value resulting from a 1.0 K reduction during the day and ∼ 0.1 K (0.3 K) absolute (total) reduction at night. The overall SDD dropped
by 23 % comprised of a 27 % daytime reduction and a 2 % increase at
night. Thus although the emissivity kernel is based on nighttime data, these
results indicate that its use to derive the daytime adjustment coefficients
is built on a sound assumption of diurnal applicability. The inconsistency
at night, especially in October, may be a result of the limited range of
viewing angles used to construct the emissivity kernel, although it is perhaps
more likely due to nighttime calibration artifacts (e.g., discussion of Fig. 9) that have not yet been fully resolved (Yu et al., 2013).
Similar results (not shown) were found for the GOES-13 LST values retrieved
using GFS instead of MERRA. Unlike the SST comparisons (Figs. 4 and 5), the
GFS-derived GOES LST bias and SDD values are comparable to those based on
the MERRA profiles. Without applying the anisotropic correction, the daytime
and nocturnal biases for GOES / GFS retrievals relative to MYD11 are 1.69 ± 1.99 and 0.22 ± 1.12 K, respectively, which are higher, but
not significantly worse than the corresponding MERRA values. After applying
the VZA adjustment, the day and night biases are 0.76 ± 1.43 and
-0.11 ± 1.13 K, respectively. Although the GFS results over land,
compared to those over ocean, are much closer to those from MERRA, the
MERRA-based results are slightly more accurate, relative to MYD11, than
their GFS counterparts.
Mean bias and SDD values (K) based on results relative to
the ARM IRT before and after anisotropic correction using only GOES data,
only 2013 AVHRR data, and using combined 2013 GOES and AVHRR results (All).
GOES SatCORPS retrievals are based on MERRA (top) and GFS (bottom) input.
GOES
GOES corrected
AVHRR
AVHRR corrected
All
All corrected
MERRA
Bias
SDD
Bias
SDD
Bias
SDD
Bias
SDD
Bias
SDD
Bias
SDD
Combined
-1.13
1.84
0.16
1.79
-0.80
2.20
-0.02
2.23
-1.07
1.94
0.11
1.91
Day
-0.93
2.30
0.16
2.24
-0.88
2.75
-0.13
2.79
-0.94
2.41
0.09
2.38
Night
-1.32
1.19
0.15
1.18
-0.74
1.75
0.06
1.75
-1.18
1.37
0.13
1.34
GFS
Bias
SDD
Bias
SDD
Combined
-0.98
1.76
0.32
1.75
Day
-0.85
2.19
0.24
2.17
Night
-1.10
1.29
0.39
1.29
Scatterplots of clear-sky surface skin temperatures from
2013 GOES-13, GOES-15, and AVHRR imagery, matched with ARM SGP IRT
temperatures (a) without and (b) with anisotropic correction.
Validation with ARM SGP infrared thermometer
To obtain estimates of the LST bias and SDD relative to ARM IRT
measurements, only confidently clear pixels that include the ARM SGP site
were selected from the 2013 GOES-13, GOES-15, and NOAA-19 AVHRR retrievals.
In order to minimize the chance of cloud mask errors and edge effects, all
adjacent pixels were required to be clear. Figure 11 shows the scatterplots
of LST retrieved from the ARM SGP IRT and from matched GOES and AVHRR data.
The IRT is a down-looking instrument that measures the upwelling radiating
temperature of the ground surface, so it is considered to have a nadir view
for this comparison. The points (Fig. 11a) tend to parallel the line of
agreement but are mostly above it. The IRT values are 1.07 K greater than
their satellite counterparts, on average, with an SDD of 1.94 K. The points
corrected for anisotropy (Fig. 11b) are scattered about the line of
agreement and the average difference is 0.11 K with SDD = 1.91 K. Given
that agreement improves comparably for day as well as night further supports
the assumption that the night-based emissivity kernel is valid during all
hours of the day.
Probability distributions of LST differences between
satellite (GOES-13, GOES-15, and AVHRR) and ARM IRT (a) without and (b) with
anisotropic correction.
This improvement for both halves of the diurnal cycle is easier to see in
Fig. 12, which plots histograms of the differences, SatCORPS–IRT, before
(Fig. 12a) and after (Fig. 12b) anisotropic correction. The daytime bias
approaches zero, moving from -0.94 ± 2.41 to 0.09 ± 2.38 K,
while the nocturnal bias changes from -1.18 ± 1.37 to 0.13 ± 1.34 K. When only the GOES data are considered, the
corrected data yield 0.16 ± 2.24 and 0.15 ± 1.18 K for day and night, respectively, and
when only the AVHRR results are considered, the respective day and night
corrected data yield -0.13 ± 2.79 and 0.06 ± 1.75 K. The GOES
data were also analyzed using the GFS atmospheric profiles and yielded
smaller absolute biases compared to their MERRA counterparts (-0.85 and
-1.10 K for day and night, respectively) for no anisotropic correction.
With the correction, the day and night measurements exhibit biases of 0.24
and 0.39 K, respectively, and when combined yield an overestimate of 0.32 ± 1.75 K.
This bias is larger than the MERRA-based GOES Only retrievals, but
the SDD is reduced by 2.2 %. Thus, the overall accuracy is similar for the
two vertical profile sources for this location. Because the MERRA and GFS
biases differ significantly for SSTs (Fig. 6), use of the MERRA profiles for
retrieving Ts with a single IR channel is preferable. See Table 4 for a
summary of all results discussed in this section. For any post-adjustment
value in Table 4, with the exception of daytime ALL and AVHRR Only SDD, the individual
day and night MERRA- or GFS-sourced bias and precision values are within the
GOES-R specifications of 2.5 and 2.3 K, respectively (Yu et al., 2012b).
It is worth noting that Heidinger et al. (2013) reported very small changes
in LST relative to SURFRAD measurements as a function of VZA and concluded
that they are not a major concern. The VZA corrections developed by Vinnikov
et al. (2012) and employed here, however, improve the overall absolute bias
in all cases, and the SDD in most, relative to ARM. Exceptions for SDD
include the combined and daytime AVHRR Only analyses. The anisotropic correction, on
average, has a total bias influence close to 1.0 K, which amounts to a
significant net reduction of about 0.8 K, yielding an average absolute bias
of ∼ 0.1 K relative to the ARM surface site values. The
average SDD reduction after correction is only 1 %, resulting in an
average precision of about 1.9 K relative to ARM. This is a significantly
smaller change compared to the 23 % reduction in SDD seen in the MYD11
validation, which may be due to the smaller sample size for the SGP
analysis. Nevertheless, these small improvements, together with the better
satellite-to-satellite normalization in Figs. 10 and 8, demonstrate that
large-scale application of the three-kernel LST adjustment for anisotropic
dependencies will result in a more accurate and uniform product.
Spatial homogeneity of LST
The biases in the results can be due to many factors, including errors in
the assumed surface emissivities, the atmospheric profiles, and the surface
observations themselves. The representativeness of the site for the much
larger area is also potentially a large source of bias. This issue,
sometimes called the scaling, or up-scaling, problem (Wang and Liang, 2009;
Guillevic et al., 2012, 2014; Li et al., 2013), is a
concern for any ground-based satellite LST validation effort, but no attempt
is made here to up-scale the ground station point observations to fully
characterize the relatively large pixel area of the satellite product. The
potential impact of the large scale is important to mention, however. For
example, Wang and Liang (2009) conclude that it is not possible to compare
satellite-derived LST relative to a single ground LST measurement without
introducing bias. Guillevic et al. (2014) discuss that although ground-based
LST measurements are in most cases suitable for “well-defined and dedicated
sites”, investigative procedures into measurement performance should be
employed when spanning the full range of surface types and conditions
surrounding the site. In both studies, high-resolution ASTER data were used
to assess the degree of heterogeneity of LST around field stations and
evaluate the spatial representativeness for ground-based measurements.
Heidinger et al. (2013), who forewent up-scaling, bring attention to
potentially underestimated errors caused by the scaling uncertainty.
Furthermore, Fang et al. (2014), in their non-scaled validation study with
SURFRAD and ARM, acknowledge the need to better characterize the
uncertainties of comparing point measurements with pixel observations. This
need was also recognized by Pinker et al. (2009) and led Fang et al. (2014)
to suggest using ASTER or MODIS. Therefore, as similarly cautioned by
Heidinger et al. (2013), we advise users to be mindful of the scaling-based
uncertainties of these non-scaled LSTs. To that end, the remainder of this
section aims to quantify the spatial variability of LST surrounding the ARM
site.
Spatial homogeneity assessment of LST (K) using aggregates
of 45 × 45 (4 km) and 89 × 89 (8 km) ASTER pixels at 90 m
resolution centered on the ARM SGP Central Facility for day, night, and
combined times, using data from 2001 through 2015. The bias and SDD indicate
the mean difference and standard deviation of the difference, respectively,
of the ASTER central pixel relative to the 4.05 × 4.05 or
8.01 × 8.01 km2 average. Bias and SDD are based on granules
N, which indicates the number of aggregates where at least 95 % of the 90 m
pixels signified clear-sky conditions with no known defects and also
represents the number of ASTER granules used to determine the region's
minimum, median, and maximum values of LST standard deviation (SD).
Scale
N
Bias
SDD
Min
Median
Max
(km)
SD
SD
SD
Combined
4
92
-0.09
1.12
0.20
0.99
5.81
8
83
-0.17
1.12
0.22
1.05
5.45
Day
4
32
-0.12
1.69
0.20
2.24
5.81
8
27
-0.24
1.71
0.22
2.27
5.45
Night
4
60
-0.08
0.67
0.28
0.68
4.14
8
56
-0.14
0.69
0.30
0.79
5.16
Inspired by the techniques of Wang and Liang (2009) and Guillevic et al. (2014), we use high-resolution ASTER data to measure the spatial variance of
LST in both a 4 × 4 and 8 × 8 km2 area
centered on the ARM SGP Central Facility. Spatial homogeneity is evaluated
in two ways. First, the 4 or 8 km (actually 4.05 and 8.01 km) LST mean
(computed from mean radiance and AST_05 surface emissivity)
is compared to the central pixel. Second, the heterogeneity is assessed
using the minimum, median, and maximum standard deviations (SD) of the 45 × 45 or 89 × 89 pixel area, an approach similar to the
method Guillevic et al. (2014) employed for evaluating a 1 km region
centered on various SURFRAD sites. Note that it is not our aim to assess the
accuracy of the ASTER product but instead only the consistency of the large-area and
central pixel measurements centered on the ARM site. For such an evaluation,
therefore, measurements from ARM are irrelevant. The similarities of the
values from the various scales determine the scaling consistency relative to
a given ARM SGP measurement.
Results of the ARM spatial homogeneity analysis are presented in Table 5,
where N indicates the number of ASTER granules for which at least 95 % of
the pixels signified clear-sky conditions with no known defects. Whether
day, night, or combined, the magnitude of the mean difference between the
4 km mean LST and central-pixel LST is near 0.1 K (0.12, 0.08 and 0.09 K, respectively), with the central pixel being slightly cooler in all cases.
Although it is true that daytime LST exhibits higher uncertainty than that
of nighttime (as also observed by Wang and Liang, 2009, and Guillevic et al., 2014), the SDDs of 1.69, 0.67, and 1.12 K are within the respective
GOES Only day, night, and combined post-adjustment precision values of 2.24, 1.18, and 1.79 K. The same can be said for the median SDs, although the
daytime value matches the 2.24 K precision exactly. It is therefore
concluded that our GOES LST retrievals are too warm relative to the ARM
measurement by about 0.1 K on average, which is a minor bias adjustment,
with an associated uncertainty not in excess of the stated precision values.
These values are consistent to the results when considering the 100 %
clear granules, in which case N= 9, and the absolute mean bias and
uncertainty are 0.12 ± 0.75 K, and median SD is 0.57 K in a range
from 0.20 to 2.40 K (not shown). Interestingly, although the error is not
significantly large, reducing the anisotropy-corrected GOES LST by 0.1 K
would bring GOES more in line with the measurements from ARM with both day
and night mean biases closer to zero.
The 8 km spatial homogeneity analysis is meant to demonstrate the most
extreme cases of pixel-to-ARM matching disparity, e.g., theoretical
situations where the ground site is situated near the corners of the
containing GOES pixel. The result is a near doubling of the bias to -0.17 K
for the combined case, which includes a -0.24 K daytime contribution and a
-0.14 K nighttime contribution. The SDDs are comparable to those of the 4 km
assessment, which suggests that increasing the aggregation area surrounding
the ARM site from 4 × 4 to 8 × 8 km2 does
not significantly influence LST uncertainty. Although the biases are larger
in magnitude than those from the 4 km analysis, if they are representative
of the expected scaling error between pixel area and the point measurement,
then, as before, they serve well to adjust GOES LST to match the cooler, on
average, ARM measurements. Compared to the 4 km analysis, the median SD
values increased only slightly to 2.27, 0.79, and 1.05 K for the day,
night, and combined cases, respectively. Therefore, as expected, only the
daytime SD exceeds the uncertainty determined from the ARM validation
results. These results are not meant to suggest that a 4 × 4 or 8 × 8 km2 area LST is generally representative
of any given point measurement within that region. In fact, such cases are
certainly unlikely as concluded by Wang and Liang (2009) and Guillevic et al. (2014), even for well-known validation sources.
Summary and conclusions
Accurate assessment of global climate and improvement of climate models, as
well as numerical weather forecasts, rely on consistent land and ocean
Ts measurements, among others. Atmospheric flux calculations depend on
the robustness of such surface variables, and NWP analyses are driven by
reliable and frequent state variable updates over large spatial domains.
Despite key downsides, satellite data are ideal sources of Ts given
their model-ready retrieval schedule and broad continuous areal coverage.
Thermal-infrared-derived Ts relies on accurate cloud clearing,
atmospheric adjustment, and angular anisotropy consideration. Therefore,
validation of satellite Ts relative to known standards is of critical
importance.
The SatCORPS provides a Ts product retrieved from GEO and AVHRR sources
using the same single-channel algorithm. The benefit of the single-channel
approach is that this method is more universally applicable to historic and
future satellite instruments compared to the split-window technique. Having
GEO and AVHRR Ts values derived from the same algorithm reduces relative
uncertainty and, hence, are better able to supplement one another.
Validation of SST retrieved from both satellites demonstrates consistent
accuracy and precision results of less than 0.1 and 0.6 K relative to NOAA
OI SST, respectively, for atmospheric corrections based on MERRA profiles.
If GFS temperature and humidity profiles are used to account for atmospheric
attenuation, however, the accuracy and precision values for the GEO SST
exceed 0.6 and 1.0 K, respectively. The larger negative bias and precision
relative to the MERRA-based results suggests that the GFS atmosphere is
drier than MERRA over the oceans, on average. This result is surprising in
that satellite (Tian et al., 2013) and radiosonde (Kennedy et al., 2011)
comparisons indicate that MERRA is too dry at altitudes below 500 hPa.
Daytime LST retrievals can be significantly influenced by satellite and
solar viewing geometry. One must therefore account for this 3-D radiance
anisotropy dependence on a global scale in order to create an accurate and
uniform product. Creating a universal model such as this, however, will
require the development of regional and seasonal kernels, which requires
many different matched datasets for a sufficiently large configuration of
viewing/illumination angle combinations across many scene types and all
seasons. Such an endeavor is left for future work. Here, we have employed
the Vinnikov et al. (2012) universal empirical model for angular anisotropy
correction. It was developed and tested using a very limited set of
measurements taken at only five sites over the United States but had not
been exercised over a larger scale prior to this study. This article has
highlighted independent tests of model effectiveness via large-area
satellite LST comparisons and ground site validation, which effectively
demonstrate the benefit of applying this anisotropic correction to LST
retrievals over much of North American in all seasons.
Land surface temperatures retrieved from July 2013 matched GOES-East and
GOES-West data over North America showed distinct VZA-dependent differences.
Normalization of the daytime LSTs to the nadir view using the Vinnikov et al. (2012) anisotropic correction model reduced the absolute bias by a
factor of 2. The remaining daytime differences are likely due to
differential heating/cooling rates and topographical orientations, which can
be potentially mitigated in the future by implementing terrain and
vegetation considerations into the correction model. The GE–GW average
nocturnal absolute LST difference is ∼ 0.4 K. Applying the
anisotropic correction reduces the mean absolute bias to ∼ 0.2 K. Overall, application of the three-kernel model nets meaningful
improvement despite a need for better terrain handling and more
comprehensive study of near-midnight calibration effects for GOES
satellites.
The SatCORPS retrievals from GOES-13 were compared to the Collection-5
Aqua MODIS LST product, a well-validated dataset. Normalization of both the
GOES and MODIS LSTs to the nadir view reduced the daytime bias and SDD by
1.0 K and 23 %, respectively. The daytime angle-corrected GOES data were,
on average, 0.62 ± 1.33 K greater than their MODIS counterpart. For
nighttime, the anisotropic correction had a smaller absolute effect on the
bias and resulted in a slightly increased SDD of 1.09 K. When combined, the
mean GOES and MODIS difference is 0.29 ± 1.28 K, reduced from 0.98 ± 1.67 K prior to anisotropic correction. Use of the GFS profiles in
place of their MERRA counterparts slightly degraded the combined bias and
precision to 0.34 and 1.36 K, respectively. Comparisons with LSTs from the
ARM IRT ground station provide further evidence of the validity of the
SatCORPS retrieval approach and the application of the anisotropic
correction, both for day and night. On average, MERRA-based atmospheric
corrections seem to perform slightly better than GFS-based attenuation for
LST retrievals compared to surface and other satellite LSTs. This finding,
however, should not restrict use of GFS for LST retrievals, as the
differences are rather small and not strictly better/worse in all scenarios.
For SST validation, the MERRA atmosphere is clearly preferred. The small
improvements in bias and SDD relative to both the ARM and MYD11 validation
efforts (1) demonstrate that large-scale application of the three-kernel LST
adjustment for anisotropic dependencies will yield a more accurate and
consistent product and (2) support the assumption of diurnal efficacy of the
night-based emissivity kernel.
Further investigation is necessary for the ARM ground-site validation
approach, particularly in terms of the up-scaling problem. However, a
spatial homogeneity analysis using ASTER data at 4 and 8 km scales
demonstrated that the average scaling error is small. Also, SD of LST surrounding the ARM measurement site only exceeded
determined precision values for daytime granules, which highlights the
importance of robust solar angle considerations in the satellite retrievals.
Regardless, disparity between pixel- and ground-station-observed surface
conditions and model sounding deficiencies are the likely contributors to
the surface–satellite differences. Beyond the outlier cases, however, the
corrected SatCORPS GEO and AVHRR Ts exhibit minimal mean bias along with
high precision. The anisotropic correction, with an adjustment magnitude of
∼ 1.0 K, affords reductions of 0.8 K and 1 % in absolute LST
bias and SDD, respectively. These small reductions yield mean bias and
precision values of 0.1 and 1.9 K, respectively, compared to the ground
site reference.
This study has examined data from only one small part of the Earth using a
single anisotropic model developed using a limited range of viewing angles.
It appears to work quite well for the larger domain (central North America),
which included the sites used in its development, but there remain several
areas for future testing and improvement. The impact of such corrections
should be tested over other areas of the globe having different vegetation
and terrain. The simple linear emissivity kernel under-corrects at higher
VZAs, indicating that a higher-order formulation may be needed. Biases in
mountainous areas stand out even after correction, suggesting that terrain
orientation and morphology may introduce additional complexity in the
anisotropy. Regional determination of the Vinnikov et al. (2012) model
coefficients may be ideal, but deriving those coefficient values would
require many matched datasets to achieve sufficient sampling at a large
variety of VZA, SZA, and RAA combinations across all seasons, a task that is
left for future work. Because land areas are viewed at fixed VZAs by GEO
imagers, the LST retrievals will suffer from VZA biases and, at a given
local hour, solar illumination biases. Removal of those biases will improve
the quality of LST monitoring and enhance the utility of these datasets for
assimilation into numerical weather models. Therefore, incorporating these
anisotropic corrections for LST into the near-global NRT retrievals, for
overlapping GEO and LEO imagers with robust cloud screening algorithms, will
benefit the data assimilation and climate research communities and hopefully
lead to improved forecasts and better understanding of the global climate
system.