Transcript Slide 1

SMOS Observations of the Gulf of Mexico and Caribbean Sea: Evaluating Surface Salinity Retrieval
and Roughness Correction Performance.
EGU2011-12371
XY 632
Abstract
Derek Burrage*, David Wang*, Joel Wesson*, Paul Hwang#, Stephan Howden&,
With the successful launch of ESA’s SMOS spacecraft, and the anticipated launches of NASA’s Aquarius/SAC-D and
SMAP satellites, a new global era in L-band microwave retrieval of Sea Surface Salinity (SSS) and Soil Moisture
commenced. The launch of SMOS on 2 Nov., 2009 and distribution of calibrated L2 products beginning 13 July, 2010,
opened up new avenues for mapping the global oceans and marginal seas, approaching within ~50 km of the coast.
MIRAS, the single scientific instrument payload of SMOS, is the first 2D microwave interferometric radiometer in space.
It represents an innovative design which allows imaging SSS and related parameters over a swath of ~1100 km, with a
pixel resolution ranging from 32 to 100 km, depending upon incidence angle. Considering the radiometric sensitivity of
the instrument, which is improved by spatio-temporal averaging, the goal is to map global SSS monthly with precision of
0.1 psu at spatial scales of 200 km. However, given the high signal to noise ratios of coastal areas, especially close to
the mouths of major rivers, one can trade off SSS precision to improve spatial resolution, and potentially map SSS
averaged over 3 to 7 day periods and 50 km spatial scales with a precision of ~2 psu.
The retrieval of SSS from L-band brightness temperature (Tb) observations requires several environmental
corrections. Among these, the sea surface roughness correction dominates the SSS retrieval error budget. Because
correction models for the influence of sea state on microwave emissivity are still evolving, the SMOS L2 processing
chain includes three operational roughness correction models.
In this work, SMOS observations over three western Atlantic ocean regions are presented to illustrate the potential
for monitoring SSS using L-band radiometry throughout the Atlantic domain. The temperature, wind and salinity data
available in the SMOS L2 products are compared with in situ measurements from ARGO drifters. Statistical estimates of
the bias and precision of SSS values retrieved from SMOS are computed within the regions, and the performance of the
three L2 roughness correction models is compared. The new roughness spectrum model of Hwang et al. is also tested
as an alternative to the one used in SSS2, and the performance of the resulting roughness correction model is evaluated
relative to the operational models (SSS1-SSS3).
Jordi Font@ and Carolina Gabarro@
Calibrated SMOS Products spanning the Amazon River Outfall
(a)
[Email: [email protected]]*
Remote Sensing Division, Naval Research Laboratory (NRL), Washington, DC, USA #
Department of Marine Science, University of Southern Mississippi, Stennis Space Center, MS, USA&
Marine Science Institute, CSIC, Dept. of Physical Oceanography & SMOS Barcelona Expert Center, Spain@
A SMOS Calibration Based on Monthly Statistics
Monthly regression coefficients, described in the section below, were used to calibrate selected SMOS overpasses (eg., Fig 4).
The Nov, 2010 regression plot (Fig. 4a) shows the expected tight relationship between ECMWF and ARGO Temperatures, while
the SMOS SSS retrievals reveal relatively high noise levels (not unexpected, given sensitivity constraints), and significant
departures from the mean ARGO S values in the region. The maps (Fig. 4b) show elevated S values near shore, which may be
due to land contamination effects. The calibrated maps, based on the regression coefficients, exhibit more uniform results from the
three roughness correction models, but SSS is unrealistically high both near and offshore (~38.5 psu). This could arise if surface
waters sensed by SMOS (in the top few cm) are fresher than the bulk mixed layer water measured by ARGO (at depths of 1-5m).
In the maps, relatively fresh water (35 psu) appearing east of the Lesser Antilles could be associated with an upwelling event, or
the outfall of the Orinoco River (Lat 10 N Lon 60 deg W). Aside from a remaining high SSS Bias, the calibration appears to
perform well, despite the relatively weak (~0.5) correlation and high SMOS L2 noise levels (residual standard deviation ~ 2 psu).
(a)
(b)
Fig 5 Calibrated SMOS overpasses spanning the Amazon Plume retroflection for Oct 1, 4 and 6 (a-c) .
In this several day sequence (Fig 5) for Oct 1,4, and 6 of 2010, SMOS spanned what appears to be meanders of the Amazon River
plume, following the North Brazil Current retroflection. These SSS retrievals, based on roughness model SSS2, were calibrated using
the ARGO v’s SMOS regression coefficients for Oct., 2010. The intervening (nearly daily) SMOS overpasses overlapping this region
(not shown) reveal the same meander pattern, but the meander was noticeably dispersed in the overpass of 7 Oct. According to MullerKarger et al. (1995) “The Amazon {River} plume flows around the North Brazil Current retroflection near 5-10 deg N, and is carried
eastward in the meandering North Equatorial Countercurrent.” The anomalously high SSS near shore is not related to winds, which
were relatively calm, and its noticeably linear boundary (Fig. 5c) has not been explained.
Comparison of Roughness Correction Models
(c)
The retrieval of SSS from L-band radiometer signals requires corrections to be performed for various geophysical factors that influence
the observed brightness temperature. The required corrections are usually derived using semi-empirical/analytical models. Factors
include surface temperature and roughness, atmospheric absorption, emission and polarization/Faraday rotation, and reflected sky and
extra-terrestrial (solar, cosmic and galactic) radiation. Among these, the deviation of the brightness temperature from the value
observed over a flat sea, due to the influence of surface roughness on emitted (outgoing) radiation is considered the largest error
source. The roughness concomitantly influences the reflectivity of the surface, and hence reflection of incoming radiation.
After extensive testing and calibration during the SMOS Commissioning Phase, using cold sky observations and analysis
techniques such as the Flat Target Transformation, operational L2 data were supplied to the SMOS Calibration and Validation
Team. During the commissioning period, problems due to Radio Frequency Interference (RFI) and different instrument
responses to ascending- or descending-pass coastal crossings, were identified. After some remedial action, remaining
spatially-dependent biases, evident even over the deep ocean, were addressed using an Ocean Target Transformation (OTT)
developed by J.Tenerelli et al. This was followed by production of an experimental L3 product at the SMOS-BEC, aimed at the
original GODAE mission goal of mapping SSS monthly and globally on ~200 km spatial scales.
In this paper, we investigate the reliability of the operational L2 Ocean Salinity (OSUDP2) product supplied during the first 8
months of the operational phase, with the intention of using it to map stronger SSS signals appearing near coasts and in
proximity to major rivers. Most of the L2 data analyzed here were processed by ESA using version 3.16 of the L2 processor
and were computed using the original SSS1, 2 and 3 roughness correction models established prior to launch. At NRL, we
have reprocessed some of these data using the v3.16 L2OS processor, in order to test the alternative roughness correction
model based on the SSA/SPM emissivity model and the new wind-wave spectrum of (Hwang (2008). The SSS2 and SSS3
models have been modified using empirical techniques and data obtained from the early operation phase. These new versions
were only implemented operationally in mid-March, 2011, so they are excluded from the analyses reported here.
Bias and Regression Statistics – Seasonal Trends
S2
aT
S3
S3
aT
M
M
7
8
9
10
11
12
13
SMOS Monthly Bias
vs
ARGO
for
NWA
Month
bS1
bS2
25
14
7
8
2
9
10
11
12
13
SMOS Monthly Bias
vs
ARGO
for
CWA
Month
bT
-2
8
9
10
11
12
13
bS1
bS2
25
14
7
8
S1
S2
S3
2
9
10
11
12
13
SMOS Monthly Bias
vs
ARGO
for
SWA
Month
bT
bS3
0
14
aS
35
bS1
bS2
14
bS3
0
-2
-2
7
sT
30
bS3
0
7
8
9
10
Month
11
12
13
7
14
8
9
10
11
12
13
14
Month
Month
Fig 2 Monthly SMOS mean and bias relative to ARGO T and S for the NW (a) and CW (b) regions.
Months 7 and 14 correspond to July 2010 and Feb 2011, respectively.
Fig 6 Difference between standard SPM/SSA roughness model driven by K (a) and H spectrum (c), and the original SSS3
empirical model for SMOS overpass in the NWA region on 14 Feb, 2011 at 0900 Z. Wind speed (b) is shown for reference.
Monthly time series of the mean of ARGO and SMOS data (Fig 2, upper panels) are presented for the NWA (a), CWA (b) and SWA
(c) regions over the full 8-month analysis period (July, 2010 - Feb, 2011). Corresponding bias estimates for SMOS relative to ARGO
are also shown (lower panels). The statistics were computed from the matching data falling in each month. Results for ARGO
temperature (aT) and the SMOS ECMWF auxiliary data (sT) are plotted on the same scale as the ARGO (aS) and SMOS (S1,S2,S3)
salinities, even though their units differ. For clarity, however, in the mean plots, T is offset upward by 10 C (i.e., it is actually10 deg.
cooler). The means reveal weak monthly variability for T and modest variability for S. The results for the three roughness correction
models track each other closely. The SMOS biases computed relative to ARGO reveal a significant seasonal trend; weaker for T, but
quite strong for S, and varying between 1 and 2 psu throughout the period. The SMOS S bias tends to be weak, or slightly negative,
in the NH summer and more strongly negative during the winter months.
(b)
S1
S2
4
S3
2
0
S1
S2
4
S3
2
0
7
8
9
10
11
12
13
14
T
S1
S2
S3
2
0
7
100
Fig 1 NWA (a), CWA (b) and SWA (c) Regions showing representative SMOS SSS passes and locations of matching ARGO drifters (+)
T
SMOS vs ARGO Regression Stats for SWA
Slp
T
(c)
SMOS vs ARGO Regression Stats for CWA
Slp
Slp
4
8
9
10
11
12
13
14
7
8
9
10
11
12
13
14
7
8
9
10
11
12
13
14
20
0
7
8
9
10
11
12
13
0
-20
14
8
9
10
11
12
13
14
7
8
9
10
11
Month
12
13
14
R2
0.5
R2
0
2
0
7
0.5
R2
0.5
Int
Int
Int
4
-100
Statistics were computed for each month of operation, so significant seasonal variation would be apparent. The main statistics were
the mean of all matching ARGO and SMOS data for the month, and their differences, which measure the SMOS bias, assuming the
coincident ARGO data are ‘True’, and representative of the area spanned by a SMOS pixel. An ordinary linear least squares
regression was performed, with SMOS and ARGO the dependent and independent variable, respectively. This approach is
considered appropriate, since SMOS data exhibit significantly higher random noise levels than ARGO, which may be regarded as
essentially deterministic. Regression coefficients (slope and intercept) and corresponding measure of fit, R 2 were computed. The
slope, intercept and R2 were plotted as a monthly time series, along with the mean bias. After transformation to make ARGO S the
dependent variable, the regression coefficients were then used to calibrate selected SMOS overpasses. We present the calibration
results first, then discuss the seasonal trends evident in the regression statistics.
S2
30
SMOS vs ARGO Regression Stats for NWA
The standard SMOS L2OS SST and SSS products were compared over the 8-month period (July, 2010 – Feb, 2011) with nearsurface data from ARGO drifter trajectories lying within the three western Atlantic regions. The ARGO data were selected from the
depth range 0 to 5 m, but only the shallowest datum from each matching profile was used for comparison with SMOS. Matches were
determined by applying a maximum temporal separation of 2 days, and a maximum spatial separation of 0.5 deg (56 km). In Fig.1,
pale blue dots are ARGO drifter trajectories during the 8-month period, while black crosses represent ARGO drifter locations
matching the SMOS overpass shown. Corresponding matching T, S values were analyzed statistically.
S1
40
(a)
SSS derived from the MIRAS L-band brightness temperatures (Tb) represent the key ocean surface observation of SMOS. Sea
Surface Temperature (SST), also included in the SMOS data files, is an auxiliary product of ECMWF analyzed meteorological fields
derived from a coupled atmospheric/ocean model, which assimilates in situ data. As may be expected, it correlates highly with the
ARGO drifter temperatures. However, it is a key input to L2 OS retrieval. We use it here as a check on our statistical analysis and to
reveal certain oceanic features.
aS
35
B
B
sT
B
25
(c)
45
40
bT
SWA
S1
30
Salinity and temperature data from the SMOS L2OS products were compared with near-surface data from ARGO drifters in
three regions of the western Atlantic ocean. The north-west Atlantic, NWA, region (Fig 1a), which is temperate, spans the Gulf
Stream and eastern US and Canadian continental shelves. The central-western, CWA, region (Fig 1b) spans the tropical and
subtropical regions of the Caribbean Seas and Gulf of Mexico, and includes the Mississippi River and Amazon/Orinoco River
outfalls, while the temperate southern-western, SWA, region (Fig 1c) lies off South America and includes the Plata River outflow.
CWA
aS
35
Matching ARGO and SMOS SST and SSS
NWA
sT
(b)
SMOS vs ARGO Monthly Means for SWA
45
40
(a)
(c)
SMOS vs ARGO Monthly Means for CWA
SMOS vs ARGO Monthly Means for NWA
45
2
(c)
(b)
(a)
M
Subsequent sections describe the methodology for the investigation. The results of a statistical regression analysis of matching
ARGO drifter data and SMOS SSS retrievals is presented, along with some calibrations of representative ESA L2 OS retrieval
products. The alternative roughness correction model is then briefly evaluated.
Due to uncertainties in the roughness corrections, three separate correction models are used for SMOS SSS retrieval: SSS1, the Twoscale model (TSM) of Yueh (1997) which uses a Gaussian distribution of long-wave slopes; SSS2, the SSA/SPM model of Voronovich
(1994), driven by the Kudryavtsev et al (1999), roughness spectrum (K) and the empirical model of Gabarro et al. (2004), based on
optimal multi-linear regression of airborne and in situ data. In this work, we have tested an alternative roughness correction model
implemented by replacing the original K spectrum by the H spectrum of Hwang (2008) as the input to SSA/SPM. The Hwang (2008)
spectrum was recently compared with in situ data by Hwang et al. (2011). The corresponding Look-up Table (LUT) logically replaces the
SSS1 TSM LUT in the L2 OS processor. We then reprocessed selected SMOS L1c products using the new model.
Fig 4 Calibration of SMOS pass through CWA on 15 Nov at 2300 Z, based on Nov, 2010 regression coefficients for region NWA.
Regression of SMOS v’s ARGO for Nov. ,2010 (a) original (b) and calibrated (c) product. Note reduced S scale range in (c).
aT
(b)
(c)
Ocean Sciences Division, Naval Research Laboratory (NRL), Stennis Space Center, MS 39529, USA
Introduction
(a)
(b)
0
7
8
9
10
11
Month
12
13
14
0
7
8
9
10
11
12
13
14
Month
Fig 3 Monthly ARGO v’s SMOS T and S regression slope, intercept and R2 for the NW (a) CW (b) and SW (c) regions (month 7  July, 2010)
Time series of the regression statistics, a (slope), b (intercept), where SMOS S= a * ARGO S + b and corresponding R 2 (Fig 3 a, b,
c) were computed for region NWA, CWA and SWA over the 8-month period for surface temperature and salinity. The plots indicate
significant temporal variability in the accuracy and precision of the SSS retrievals from SMOS for these regions. SWA results are
restricted to a 5 month period, since we do not yet have access to SMOS L2 products in this region. At least three years of mission
data would be needed to reliably estimate the seasonal dependence, and give some indication of inter-annual variability, but the
results suggest some seasonal trends. For clarity in Fig. 3, the R2 value for T is offset downward by 0.5 (it is actually just under
1.0). Hence, the SMOS ECMWF auxiliary temperature, T, is highly correlated with ARGO T (R>0.9). It shows a tendency for weaker
slope, intercept and R2 values, particularly in NWA, but is otherwise relatively stable.
SMOS SSS show significant monthly variations in all three regions and the models track each other quite closely. Noticeable
exceptions occur for SSS2 in the NWA, which could be subject to more intense wind forcing than the other regions. The correlations
are statistically significant everywhere except in SWA in Oct and Nov. However, the regression fit (R 2) for SSS is quite low. The
correlations, R, based on R2 as shown, tend to be higher in the NH and SH summers, when winds are presumably weaker. The
low correlation cannot simply be ascribed to the roughness correction models. They could also be due to other factors such as
residual along-track instrumental biases in L1c Tbs manifested as along-track striping in the SSS maps, RFI in proximity to the
coasts, evident in diagonal ripple patterns (eg. In SWA, Fig 1 c), and other geophysical correction model errors. However, bias
changes can be related to differences in the roughness correction models used, while other factors remain equal.
The effect of wave spectrum choice on SPM/SSA model performance is seen by comparing SSS retrievals from the standard
Kudryavtsev et al. (K) spectrum with that of Hwang (2008) (H) (Fig. 6). Values from the SSS3 empirical model of Gabarro et al.,
assumed to represent the ‘true’ SSS field, were subtracted from both retrievals to reveal deviations. The K model shows
significant wind effects near Nova Scotia and Newfoundland, but has low residuals offshore. The modified H model shows less
residual wind effect, particularly in the high wind area east of Newfoundland, but exhibits higher bias seaward. Work is in progress
to statistically evaluate the SSA/SPM model response, when driven by alternative wind-wave spectra over the full 8-month
period.
Conclusions
Statistical analyses of SMOS SSS retrievals obtained using the standard roughness correction models employed during the first 8
months of the operational period reveals significant seasonal variability in three different western Atlantic ocean regions. The
analyses, done monthly, were based on comparisons with near-surface data from the ARGO drifting buoy network. Calibration of
the SMOS retrievals based on monthly regression statistics suggest that discrepancies between retrievals using different
roughness correction models can be reduced, making the resulting SSS distributions appear more realistic. Evaluations of
roughness correction models driven by different wind-wave spectra reveal significant performance differences particularly in high
wind areas.
Acknowledgements
Many European, British and American government agencies, and academic and commercial organizations working in collaboration with the European Space
Agency, ESA, have contributed to the success of the SMOS mission, and their efforts are gratefully acknowledged. ESA provided the SMOS L1c and L2 OS
products for this project, through our membership of the SMOS Validation and Retrieval Team (SVRT) under ESA AO 2005, 2007 projects 3229 and 4667. ESA
contractors Paul Spurgeon (Argans Limited, UK) and Thomas Block (Brockmann-Consult, Germany) are thanked for helpful advice on L2 processing and
products. The ARGO drifter data were obtained from the Coriolis Data Center, IFREMER (France). Institutional support from the Naval Research Laboratory,
USA, The CSIC Marine Science Institute (Spain), University of Southern Mississippi (USA) is also acknowledged. Nicolas Reul IFREMER, provided code for the
SSA/SPM model and much helpful advice concerning roughness correction modeling. This work was supported by the Office of Naval Research as part of the
“Sea Surface Roughness Impacts on Microwave Sea Surface Salinity Measurements” research program under Program Element Number 0601153N.This is
NRL contribution NRL/AB/7330-11-583.
References
Gabarro, C., J. Font, A. Camps, M. Vall-llosera, and A. Julia, “A new empirical model for microwave sea surface microwave emissivity for salinity remote
sensing,” Geophys. Res. Lett., vol. 31, L01309, 2004.
Hwang, P. A., D. M. Burrage, D.W.-C. Wang, and J. C. Wesson (2011) An advanced roughness spectrum for computing microwave L-band emissivity in sea
surface salinity retrieval (Geoscience and Remote Sensing Letters, 2011 pp. 5 (In Press).
Hwang, P. A. (2008) “Observations of swell influence on ocean surface roughness” J. Geophys. Res., vol. 113, no. C12, C12024, doi:10.1029/2008JC005075.
Kudryavtsev, K., V. Makin, and B. Chapron, “Coupled sea surface atmosphere model, 2, Spectrum of short wind waves,” J. Geophys. Res., vol. 104, no. C4, pp.
7625– 7639, 1999.
Muller-Karger, F. E., P. L. Richardson and D. McGillicuddy (1995) On the offshore dispersal of the Amazon’s Plume in the North Atlantic: Comments on the paper
by A. Longhurst, “Seasonal cooling and blooming in tropical oceans” Deep-Sea Research I, Vol 42, No. 11/12, pp. 2127-2137.
Voronovich, A. (1994) “Small-slope approximation for electromagnetic wave scattering at a rough surface of two dielectric half spaces” Waves in Random Media,
v 4, pp 337-367.
Yueh, S. H. (1994) Modelling of Wind Direction Signals in Polarimetric Sea Surface Brightness Temperatures. Trans. Geoscience and Rem Sensing, v 35, no. 6,
pp 1400-1418.