Walton McBride U.S. Naval Research Lab, Stennis Space Center MS Robert Arnone University of Southern Mississippi, Stennis Space Center MS Jean-François

1

Walton McBride U.S. Naval Research Lab, Stennis Space Center MS

Robert Arnone University of Southern Mississippi, Stennis Space Center MS

Jean-François Cayula Qinetiq North America, Stennis Space Center MS

May 1, 2013

Improvements of Satellite SST Retrievals at Full Swath

2

A FRESH PERSPECTIVE

Search for Clues on how to Improve SST RetrievalsSST algorithms Used to Create CluesScatter Plots Used to Identify CluesModel Run Results to Discern Clues

Putting the Clues Together

Interpretation of Results

Conclusions

3

MCMCMCMCMC cTTScTTcTcSST 41211312112111

NLNLfield

NLNLNL cTTScTTTcTcSST 41211312112111

fieldTfieldTfieldTfieldTfieldTfield

Tfield TccTTScTTcTcSST 541211312112111

SST ALGORITHMS

MC SST

NL SST

Tfield SST

(same as MC SST, except for addition of Tfield as separate predictor)

1sec zenithS

First guess temperature field

4

BUOY DATA SET

75.1282.1475.2009.1 4321 MCMCMCMC cccc

MCMCMCMCMC cTTScTTcTcTSST 4121131211211111 1

YoffsetTTSccTSST MCMCMC 12113211

NAVOCEANO buoy data set for the month of June 2012: 115,036 daytime points After routine NAVOCEANO filtering: 61,782 points from 0° to 53° zenith angle (53.70%) 97,496 points from 0° to 70° zenith angle (84.75%)


Linear regression coefficients using MC SST:

In the original spirit behind the MC SST formulation:

MCMC cTcYoffset 4111 1 where

75.1009.0 11 TYoffset


1sec zenithS

MCSST

Pinch effect

No scatter

Filtering Parameters |Tfield-Buoy| < 2.0° |Zenith| < 53 deg Ddistance < 25 km Dtime < 4 hours

SSTMC-T11 vs. T11-T12


1sec zenithS

MCSST

Pinch effect

Data displays scatter

No scatter

No Pinch effect

Filtering Parameters |Tfield-Buoy| < 2.0° |Zenith| < 53 deg Ddistance < 25 km Dtime < 4 hours Goal is to capture the

features and structure of the distribution.

BUOY

a

b

BUOY-T11 vs. T11-T12 SSTMC-T11 vs. T11-T12

MCMCMCMCMC cTTScTTcTcSST 41211312112111 NLNL

fieldNLNL

NL cTTScTTTcTcSST 41211312112111

1sec zenithS

NLSSTMCSST

Pinch effect

Data displays scatter

No scatter Signs of scatter

Pinch effect

No Pinch effect

Filtering Parameters |Tfield-Buoy| < 2.0° |Zenith| < 53 deg Ddistance < 25 km Dtime < 4 hours Goal is to capture the

features and structure of the distribution.

BUOY

a

b c

BUOY-T11 vs. T11-T12 SSTNL-T11 vs. T11-T12

fieldTccTTScTTcTcSST 541211312112111

ClimK100K10

|Tfield-Buoy| < 0.1 0.021 0.057 0.026 5.921 0.978 0.999 0.07175|Tfield-Buoy| < 0.2 0.074 0.194 0.088 20.33 0.927 1.001 0.12447|Tfield-Buoy| < 0.5 0.245 0.643 0.261 66.82 0.759 1.004 0.25600|Tfield-Buoy| < 1.0 0.471 1.218 0.535 128.5 0.541 1.012 0.36675|Tfield-Buoy| < 2.0 0.649 1.663 0.736 177.0 0.364 1.013 0.45077|Tfield-Buoy| < 5.0 0.752 1.914 0.853 204.9 0.259 1.011 0.49914

rmserror

Clim

c1 + c5


rmserror

K100

c1 + c5


rmserror

K10

c1 + c5

Filtering

Table1: Effects of pre-filtering and increased spatial resolution of Tfield.

Tfield AS SEPARATE PREDICTOR

Blue Red COLOR PLOTTING Red Blue COLOR PLOTTING

a

c

b

d

DATA

MC SST

a b

c d

Blue Red COLOR PLOTTING Red Blue COLOR PLOTTING

DATA

Tfield SST with K10

From GUI results

Using global coefficients 97507 0.41055 3.66e-12 0.294 0.610 0.388 80.10 0.719 1.013 no K10 0.65727 1.09e-11 1.034 2.204 1.363 281.8 0.000 1.034

+60 to +70 21213 0.41834 -8.37e-13 0.230 0.292 0.296 62.7 0.794 1.024+50 to +60 19667 0.39780 -2.03e-13 0.296 0.621 0.349 80.9 0.721 1.017+40 to +50 15353 0.39463 2.83e-13 0.343 0.814 0.447 93.4 0.665 1.008+30 to +40 12415 0.38942 -9.51e-13 0.407 1.004 0.742 110.8 0.594 1.001+20 to +30 10231 0.38372 -2.92e-12 0.469 1.264 -0.011 127.7 0.533 1.002+10 to +20 9477 0.37942 7.47e-13 0.521 1.325 1.087 141.9 0.482 1.003 0 to +10 9151 0.37119 -6.32e-13 0.493 1.292 0.922 134.3 0.506 0.999

|Zenith| Band #pts rmserror bias c1 c2 c3 c4 c5 (c1+c5)

|Zenith| Band #pts rmserror bias rmserror bias

+60 to +70 21213 0.44234 -0.04142 0.86911 -0.16031+50 to +60 19667 0.39847 -0.00778 0.63142 0.01057+40 to +50 15353 0.39828 0.00676 0.58992 0.05601+30 to +40 12415 0.40274 0.02223 0.58093 0.06392+20 to +30 10231 0.40334 0.01332 0.53541 0.05429+10 to +20 9477 0.40560 0.03403 0.50201 0.06380 0 to +10 9151 0.39493 0.02127 0.52607 0.04228

Table 2: Coefficients Dependence on Zenith

DEPENDENCE ON ZENITH ANGLE

From GUI results

Using global coefficients 61795 0.3894 3.05e-12 0.416 1.039 0.515 113.4 0.589 1.005 no K10 0.5305 6.62e-12 1.006 2.550 1.174 274.1 0.000 1.006

+50 to +70 7112 0.36861 -1.38e-12 0.375 0.980 0.314 102.2 0.626 1.001+30 to +50 17404 0.39976 5.79e-13 0.513 1.191 0.626 139.8 0.497 1.010+10 to +30 12910 0.3532 -6.44e-13 0.279 0.672 0.339 76.3 0.731 1.010-10 to +10 4856 0.3521 -1.66e-13 0.242 0.795 0.240 63.9 0.672 0.914-30 to -10 8947 0.34733 -1.81e-13 0.432 1.080 0.554 117.9 0.580 1.012-50 to -30 9098 0.38322 -6.67e-13 0.514 1.054 0.612 140.0 0.501 1.015-70 to -50 1468 0.36599 -2.07e-14 0.359 0.777 0.359 98.1 0.673 1.032

Latitude Band #pts rmserror bias c1 c2 c3 c4 c5 (c1+c5)

Latitude Band #pts rmserror bias rmserror bias

+50 to +70 7112 0.37013 -0.02832 0.5056 -0.07924+30 to +50 17404 0.40451 0.12566 0.51462 0.11358+10 to +30 12910 0.36430 -0.01473 0.57448 -0.11061-10 to +10 4856 0.38496 -0.11888 0.58587 -0.11294-30 to -10 8947 0.34863 -0.07716 0.47300 0.00595-50 to -30 9098 0.38908 -0.06359 0.49011 0.05780-70 to -50 1468 0.37408 0.03549 0.47590 -0.00664

Table 3: Coefficients Dependence on Latitude, Zenith 0 to 53

DEPENDENCE ON LATITUDE

Table 4: Percent reduction in rms error

Zenith MCSST Tfield = Clim Tfield = K100 Tfield = K10

0o to 53o

53o to 70o

0o to 70o

0.53052 0.45077 15.0% 0.42551 19.8% 0.38937 26.6%

0.73712 0.51667 29.9% 0.45622 38.1% 0.41298 43.9%

0.65627 0.49989 23.8% 0.45507 30.7% 0.41053 37.4%

RMS ERROR REDUCTION

c1 c2 c3 c4 c5 rmserrorMCSST

TfieldSST

1.006 2.541 1.173 274.10 ---- 0.5058

0.251 0.617 0.312 68.42 0.752 0.2869

1.000 2.467 1.243 272.58 ----

TfieldTfieldi cc 1/

c1 c2 c3 c4 c5 rmserrorMCSST

TfieldSST

1.034 2.20 1.360 281.99 ---- 0.6382

0.163 0.326 0.221 44.48 0.844 0.2987

1.000 2.00 1.356 272.88 ----

TfieldTfieldi cc 1/

Zenith angles from 0 to 53 degrees |Tfield-Tbuoy|=0.7°

Zenith angles from 0 to 70 degrees |Tfield-Tbuoy|=0.7°

ANOTHER CLUE!

MCi

TfieldTfieldi ccc 1Clue 2:

15

fieldTfield

MCTfield

Tfield TcSSTcSST 11 1

fieldMCTfield

fieldTfield TSSTcTSST 1

MCi

TfieldTfieldi ccc 1

fieldTfieldMCMCMCMCTfield

Tfield TccTTScTTcTccSST 5412113121121111

151 TfieldTfield cc

22

122

12 1 Tfield

TfieldSST

TfieldSST cc

MCTfield

22

2

1

MCSSTTfield

TfieldTfieldc

22

2

511MC

MC

SSTTfield

SSTTfieldTfield cc

PUTTING IT ALTOGETHER

Clue 2:Clue 1:

fieldTfieldTfieldTfieldTfieldTfield

Tfield TccTTScTTcTcSST 541211312112111

If the errors in Tfield and SSTMC are uncorrelated, then the variances:

Substituting:

Simple Linear Weighting

and

where

Past Present - Past

or

16

SCATTER PLOT OF ERRORS

R2 < 0.04

Zenith angles from 0 to 70 degrees

Blue Red Color Plotting Red Blue Color Plotting

PRACTICALLY NO CORRELATION!

17

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

rms error Tfield

rm

s er

ror

SS

TM

C

Desired rms error = 0.1




RMS ERRORS RELATIONSHIPS

222

111

MCTfield SSTTfieldSST

Guarantees that will always be less than or !TfieldSST

MCSST Tfield

If the errors in Tfield and SSTMC are uncorrelated, then the variancesare related as:

Resistors in Parallel Analogy

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

rms error Tfield

rm

s er

ror

SS

TM

C





RMS ERRORS RELATIONSHIPS

Interestingly, we have control of Tfield rms errorthrough more aggressive filtering. But at what price?

Plot of rms error and % filtered points vs.|Tfield – Tbuoy| Zenith from 0 to 53

0 1 2 3 4 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

|Tfield - Tbuoy|

rm

s er

ror

0 < Zenith Angles < 53

No Tfield

Tfield=ClimTfield=K100

Tfield=K10

0 1 2 3 4 50

10

20

30

40

50

60

70

80

90

100

|Tfield - Tbuoy|

% b

uo

y d

ata

(so

lid

)

&

% T

fiel

d c

on

trib

uti

on

(d

ash

)


No Tfield


Tfield=K10

TRADE-OFF: rms error vs. # buoy data points

SSTMC weakly affected by aggressive filtering

Stability

Low Dropoff

0 1 2 3 4 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

|Tfield - Tbuoy|

rm

s er

ror


No Tfield


Tfield=K10

0 1 2 3 4 50

10

20

30

40

50

60

70

80

90

100

|Tfield - Tbuoy|

% b

uo

y d

ata

(so

lid

)

&

% T

fiel

d c

on

trib

uti

on

(d

ash

)


No Tfield


Tfield=K10




Stability Low Dropoff

0 1 2 3 4 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

|Tfield - Tbuoy|

rm

s er

ror


No Tfield


Tfield=K10

0 1 2 3 4 50

10

20

30

40

50

60

70

80

90

100

|Tfield - Tbuoy|

% b

uo

y d

ata

(so

lid

)

&

% T

fiel

d c

on

trib

uti

on

(d

ash

)


No Tfield


Tfield=K10




Stability

Low Dropoff

Zenith from 0 to 53 |Tfield-Tbuoy| = 0.7

MCSST

TfieldSST with K10


EVIDENCE OF REAL RMS ERROR REDUCTION


MCSST

TfieldSST with K10




MCSST

TfieldSST with K10



25

HELP ME UNDERSTAND, LENA!

SSTMC

MCSST

26

SSTMC Tfield

+

MCSSTTfield

uncorrelated noise


27

SSTTfield SSTMC Tfield

=

22

2

1

MCSSTTfield

TfieldTfieldc

22

2

5

MC

MC

SSTTfield

SSTTfieldc

MCSSTTfield

MCTfield SSTTfield

SST c 1

fieldTfield

MCTfield

Tfield TcSSTcSST 51

+

Always clearer image!


28

A FRESH PERSPECTIVE TWO-PRONGED IMPROVEMENT EFFORTS

fieldTfield

ANYTfield

Tfield TcSSTcSST 51

Using Satellite Data Only!

NAVOCEANOK100, K10, K2

IN SITUBUOY DATA

HIGH ZENITH ANGLES(emissivity, sea roughness)

HARDWARE (2 or more looks) CLOUD MASK

29

CONCLUSIONSAlthough Tfield is used in NLSST, its additional information is tamed due to its appearance as a multiplier of T11-T12.

Use of Tfield as a separate predictor results in a significant increase in accuracy forall existing SST algorithms, daytime and nighttime, due to resulting variance always being less.

Tfield now is on equal footing with existing SST algorithms predicitions and improvements in its accuracy will benefit all combinations of existing SST algorithms and Tfield.

TfieldSST algorithm was found to be extremely stable. Reduced to a formulation that leads to specific relationships between variances of MCSST and Tfield.

NAVOCEANO’s Tfield characterizations, K100 and K10, only use previous satellite data. NAVOCEANO is currently working on K2, at 2km resolution.

TfieldSST algorithm allows for rms error under 0.3K over the full swath (0 to 70), while sacrificing a very modest number of buoy data points (from original 84.75%):

75% of buoy data points left from 0 to 70 with TfieldSST versus 55% of buoy data points left from from 0 to 53 PRESENTLY

30

QUESTIONS ?

Adding Tdiffoffset to NLSST

Tdiffoffset = 0 rmserror = 0.52546 Tdiffoffset = -1 rmserror = 0.62069

Tdiffoffset = +10 rmserror = 0.44501 Tdiffoffset = +5 rmserror = 0.45413

Pinch effectmoves to +1 Pinch effect

Data

• Pinch effect disappears• Scatter is more pronounced• rmserror diminishes

a

c

b

d

INCREASING Tfield IMPORTANCE

NLNLfield

NLNLNL ctTdiffoffseTTSctTdiffoffseTTTcTcSST 41211312112111

Documents

Walton McBride U.S. Naval Research Lab, Stennis Space Center MS Robert Arnone University of Southern Mississippi, Stennis Space Center MS Jean-François