SST Retrieval and Assimilation:An Integrated Approach

SST Retrieval and Assimilation:An Integrated Approach

•Why are we looking at this, and why now?

•What are the main issues?

•Possible solutions

•The way forward

•Why are we looking at this, and why now?

•What are the main issues?

•Possible solutions

•The way forward

Andy Harris, NOAA/NESDIS/ORA

With thanks to:Chris Merchant, U. EdinburghChelle Gentemann, Remote Sensing SystemsJean Thiébaux, NOAA/NCEPGary Wick, NOAA/ETLJo Murray, RAL

Andy Harris, NOAA/NESDIS/ORA

With thanks to:Chris Merchant, U. EdinburghChelle Gentemann, Remote Sensing SystemsJean Thiébaux, NOAA/NCEPGary Wick, NOAA/ETLJo Murray, RAL

The technology is in place:

• SST retrievals from polar-orbiting IR,geostationary IR and now MW

• Computing power is sufficient to permitoperational application of more sophisticatedretrieval methods

• Each data source has its strengths andweaknesses

• An optimally combined product would satisfymany users

The technology is in place:

• SST retrievals from polar-orbiting IR,geostationary IR and now MW

• Computing power is sufficient to permitoperational application of more sophisticatedretrieval methods

• Each data source has its strengths andweaknesses

• An optimally combined product would satisfymany users

“End user” communities that have requirements• Climate monitoring• Operational oceanography• Numerical weather prediction• Fisheries

“End user” communities that have requirements• Climate monitoring• Operational oceanography• Numerical weather prediction• Fisheries

The requirements are different, but are theymutually exclusive?

• Climate monitoring needs observation systemto be stable to better than 0.1 K/decade

• Timeliness for NWP, operationaloceanography. Less emphasis on absoluteaccuracy (±0.5 K)

• Fisheries. Often need fronts, and sometimesabsolute temperatures

The requirements are different, but are theymutually exclusive?

• Climate monitoring needs observation systemto be stable to better than 0.1 K/decade

• Timeliness for NWP, operationaloceanography. Less emphasis on absoluteaccuracy (±0.5 K)

• Fisheries. Often need fronts, and sometimesabsolute temperatures

What do we mean by optimal?

Minimize the “cost” of assimilating the data…

What do we mean by optimal?

Minimize the “cost” of assimilating the data…

( ) ( ) ( )( ) ( ) ( )( )J x x x x y x y xb b o o= − − + − + −− −1

2

1

2T T

B H E F H1 1

An equation of 2 halves: background (or 1st guess)error; observation and forward model error

So, we need to know how good our retrievals areand how good our background value is

SST retrieval has traditionally been performedby direct regression against in situ. Is this thebest way?

An equation of 2 halves: background (or 1st guess)error; observation and forward model error

So, we need to know how good our retrievals areand how good our background value is

SST retrieval has traditionally been performedby direct regression against in situ. Is this thebest way?

What is the main advantage of remote sensing?

• Provides data in remote regions where in situobservation are sparse or non-existent

To utilize remotely-sensed data to an optimumlevel, we need to be able to specify accuracy inthese remote regions

• This requires independent data in order togain the necessary confidence

Can retrieval accuracy be improved by the additionof other data sources?

• Inclusion of water vapor can probably only bedone at a rudimentary level using directregression Studies have demonstrated little

What is the main advantage of remote sensing?

• Provides data in remote regions where in situobservation are sparse or non-existent

To utilize remotely-sensed data to an optimumlevel, we need to be able to specify accuracy inthese remote regions

• This requires independent data in order togain the necessary confidence

Can retrieval accuracy be improved by the additionof other data sources?

• Inclusion of water vapor can probably only bedone at a rudimentary level using directregression Studies have demonstrated little

• The chief advantage of radiative transfer is thatit allows specification of the retrieval algorithmwithout bias towards the data-rich regions

• The in situ data can then act as a randomindependent sampling of the retrieval conditions.

• If the observed errors agree with the modeledones, then high confidence can be placed on themodeled errors in data-sparse regions

• Additional advantage is that other sources oferror can be accounted for explicitly, andexternal data (e.g. atmospheric profiles) can beincorporated

•This doesn’t mean it’s easy to do…

• The chief advantage of radiative transfer is thatit allows specification of the retrieval algorithmwithout bias towards the data-rich regions

• The in situ data can then act as a randomindependent sampling of the retrieval conditions.

• If the observed errors agree with the modeledones, then high confidence can be placed on themodeled errors in data-sparse regions

• Additional advantage is that other sources oferror can be accounted for explicitly, andexternal data (e.g. atmospheric profiles) can beincorporated

•This doesn’t mean it’s easy to do…

Modeling must be accurate:• Spectroscopy (mainly continuum)• Representative input data (atmospheric

profiles)• Noise characteristics of real data• Filter functions

Sensor calibration must be accurate

• However, RT does give rise to possibility ofcorrecting calibration in a more predictablefashion

Cloud masking, aerosols

Surface effects (skin vs. bulk)

Modeling must be accurate:• Spectroscopy (mainly continuum)• Representative input data (atmospheric

profiles)• Noise characteristics of real data• Filter functions

Sensor calibration must be accurate

• However, RT does give rise to possibility ofcorrecting calibration in a more predictablefashion

Cloud masking, aerosols

Surface effects (skin vs. bulk)

GOES-8 SST retrievals using radiative transferGOES-8 SST retrievals using radiative transfer

Note that operational retrievals are 0.73 K r.m.s. on same data with +ve bias

Diurnal thermocline is avery non-linear effect:

• Wind mixing energy isproportional to U*3

• The heat contained inthe upper layer is thesame, but the heatcapacity is different

• Skin effect isgenerally a moretractable problem

Diurnal thermocline is avery non-linear effect:

• Wind mixing energy isproportional to U*3

• The heat contained inthe upper layer is thesame, but the heatcapacity is different

• Skin effect isgenerally a moretractable problem

Diurnal cycle modelingcan now be done withreasonable confidence,but solar flux must beaccurate and have goodtime resolution.

Diurnal cycle modelingcan now be done withreasonable confidence,but solar flux must beaccurate and have goodtime resolution.

Other considerations

• Cloud masking – is it time to dispense withthresholds and serial application of tests?

• Should fast-forward RT models be used?

• What about archive data?

• Aerosols (stratospheric & tropospheric)

• ‘Blending’ of POES & GOES

Test the schemes – facilitate the feedback loopbetween validation and algorithm development

Other considerations

• Cloud masking – is it time to dispense withthresholds and serial application of tests?

• Should fast-forward RT models be used?

• What about archive data?

• Aerosols (stratospheric & tropospheric)

• ‘Blending’ of POES & GOES

Test the schemes – facilitate the feedback loopbetween validation and algorithm development

Microwave dataMicrowave dataNow have the capability to retrieve SSTs from TMI togood accuracy, except in precipitation and “near” land.AMSR will extend coverage

• Spatial resolution is of order 50 km, but oversampled

• Larger errors where emissivity modeling is moredifficult (high windspeed)

• Some calibration challenges

• Physically-based retrieval of SST and otherparameters (cloud liquid water, TCWV, windspeed)

• Immune to aerosol problems

Errors in IR and MW SST retrievals are essentially

Now have the capability to retrieve SSTs from TMI togood accuracy, except in precipitation and “near” land.AMSR will extend coverage

• Spatial resolution is of order 50 km, but oversampled

• Larger errors where emissivity modeling is moredifficult (high windspeed)

• Some calibration challenges

• Physically-based retrieval of SST and otherparameters (cloud liquid water, TCWV, windspeed)

• Immune to aerosol problems

Errors in IR and MW SST retrievals are essentially

Case study using GOESCase study using GOES

Relaxation of cloud masking thresholds has resulted inlower detection rateRelaxation of cloud masking thresholds has resulted inlower detection rate

Also some correlation with water vaporAlso some correlation with water vapor

Assimilation techniquesAssimilation techniques

Adaptive, anisotropic analysisstructure functions can bedeveloped from study of theobservation – backgroundincrement field

Where does the 1st guess(background) field come from?

• Previous analysis orpredictive model. Someusers will want to use theirown model

A predictive model can be usedt i il t ti

Adaptive, anisotropic analysisstructure functions can bedeveloped from study of theobservation – backgroundincrement field

Where does the 1st guess(background) field come from?

• Previous analysis orpredictive model. Someusers will want to use theirown model

A predictive model can be usedt i il t ti

Example of high resolution (9-km) multi-scale OIusing night-time AVHRR PathfinderExample of high resolution (9-km) multi-scale OIusing night-time AVHRR Pathfinder

SummarySummaryA single, optimal SST product (or methodology) thatmaximizes the strengths of each input dataset whilstminimizing the impact of the deficiencies requires:

• A common retrieval framework, with known errorcharacteristics (seasonally and geographicallyvarying)

• Modeling of surface effects (accurate fluxes)

• Assimilation methods to take account ofcharacteristics of input data (e.g. non-gaussian)

• This may require a predictive model withappropriate geophysical constraints & forcing

A l b t d bl t k!

A single, optimal SST product (or methodology) thatmaximizes the strengths of each input dataset whilstminimizing the impact of the deficiencies requires:

• A common retrieval framework, with known errorcharacteristics (seasonally and geographicallyvarying)

• Modeling of surface effects (accurate fluxes)

• Assimilation methods to take account ofcharacteristics of input data (e.g. non-gaussian)

• This may require a predictive model withappropriate geophysical constraints & forcing

A l b t d bl t k!