Handling Cloud-Affected Infrared Radiances in the GSI

Will McCartyGSFC/Global Modeling and Assimilation Office

JCSDA Workshop10 October 2012

Introduction• In GMAO forward processing, infrared radiances are

assimilated from IASI, AIRS, and HIRS• Heritage “multi”-spectral sounders like HIRS (~ 18

channels) and the GOES Sounder are being phased out– The US HIRS instruments replaced by CrIS from NPP

onward (hyperspectral – 1297 ch total, 399 for DA)– The final European HIRS launched on MetOp-B. MetOp-

C will only fly IASI (hyperspectral – 8461 ch, 616 for DA) – No Sounder in US GEO beginning w/ GOES-R– Hyperspectral sounding potentially in GEO in a number

of future longitudes

3

Number of observations considered for assimilation

Number of observations used for assimilation

Observation volumeJanuary 1977 to present

Reduction of observations heavily due to presence of clouds in observations

Observationsprocessed per 6h

1979 − 2011

Observationsused per 6h 1979 − 2011 AIRS

IASI

AIRS

IASI

After thinning, QC

Before thinning, QC

How are Clouds Handled in GSI

• Cloud screening is a two-step process1. Retrieve a cloud height

• This is done via a minimum residual method (Eyre and Menzel 1989)

2. Compare cloud height against transmittance profile

• If layer-to-top of atmosphere transmittance of a channel at the retrieved cloud height is greater than 2% reject the channel

• For channels most-sensitive to the surface, this rejects ~80% of these data.

Further Exploiting IR Data

• To further exploit IR data, the next step is to include some characterization of clouds in the analysis

Clouds in the Infrared

• Quick lession in radiative transfer:– All forward radiative transfer in the GSI is done via the

CRTM (in Europe, they use RTTOVS) as the observation operator

– In a very basic sense, consider CLEAR-SKY radiative transfer equation as:

Clear IR Measurement = Surface + (Atmospheric Layers)

Clouds in the Infrared

• If you consider the signature of a very, very dense cloud in the IR, we can make some assumptions and then define CLOUDY SKY radiative transfer equation as:

Cloudy IR Measurement = Cloud Top + (Atmospheric Layers above cloud)

..this is known as a blackbody assumption, as the cloud top is considered “black”

Clear IR Measurement = Surface + (Atmospheric Layers)

Cloudy IR Measurement = Cloud Top + (Atmospheric Layers above cloud)

Retrieved Cloud Height

• In the cloud height retrieval, a cloud fraction, N, is also solved

• Under the graybody assumption, the partially cloudy observation can then be considered for a single, fractional cloud as:

• In the GSI, we can then restructure the H operator to include the Cloud Height and Cloud fraction to allow for a partially cloudy forward operator (and also partially cloudy Jacobians)

Clouds in the Infrared

Partially Cloudy IR Measurement = N * Cloudy IR Measurement +

(1 – N) * Clear IR Measurement

10

Obs minus Forecast (clear) Obs minus Forecast (cloudy)

• Considering the O-Fs versus cloud fraction, it is seen that the O-Fs are closer, but the cold bias is, as expected, amplified for higher (colder) clouds

• The accuracy of the calculated cloudy radiance is fundamentally dependent on the accurate retrieval of cloud height and fraction

11

Cloudy Infrared Radiance Assimilation within the GSI

• Jacobians are adjusted to move sensitivity from below cloud to cloud surface

• Single footprint assimilation shows that the system is drawing to the retrieved cloud top• Magnitude is inflated due to

low observation errors.• Error in CTP will result in an

erroneous O-F, which then can negatively impact the analysis

• To compensate, CTP is allowed to vary in the minimization as a control variable

UncontaminatedIncluding Cloud

Cloud Top

Observation-Centered Control Variables

• Current GSI implementations consider control variable only in terms of grids (2D & 3D) and channel-by-channel bias predictors

• Bias prediction coefficients are of the dimension [5,number of channels]– each satellite channel on each instrument has its own set of

predictiors (i.e. MetOp_AMSU-A channel 8 will have the same set of five coefficients across every footprint globally

• Observation-Centered control variables – consider a control variable at a footprint location over all

channels measured at that point – Dimension dynamic -> any number of observation-centered

control variables can be appended to the control vector 12

Observation-Centered Control Variables

• Once developed, the functionality was expanded to CTP– Cloud Fraction still considered constant and set as the

retrieved value• Jacobians

– In addition to modified TB/T(p), TB/qv(p), etc., the minimization now incorporates the CTP Jacobian, TB/pcld.

– TB/pcld can be directly differentiated from the radiative transfer equation (i.e. the appendix of Li et al. 2001)

• Background error for CTP

13

Background error for CTP

• Background error for CTP (BCTP) was

considered first in a single-footprint case:– Initial CTP – 624 hPa– Initial N – 0.968

• Consider behavior of three values of BCTP

compared to clear-sky observations only and a static CTP (no variational CTP)– B

CTP = 50 hPa, 10 hPa, and 5 hPa

14

Background error for CTP

15

Clear

Cloudy Static CTP

Cloudy varCTP

Background error for CTP

16

Clear

Cloudy Static CTP

Cloudy varCTP

Background error for CTP

17

Clear

Cloudy Static CTP

Cloudy varCTP

Background error for CTP• Variational CTP acts as a “sink”, as a function of

BCTP

– As the bkg error is increases, the cloud signal is absorbed into the CTP variable

– the solution approaches clear-only result– As bkg error is decreased, result approaches static CTP

• Expected as CTP is tightly constrained to retrieved guess

• This is only for a single footprint. How does the analysis respond to a full suite of observations– Since only CTP is varying, only consider cloudy IR if 1.0

> N > 0.9 -> higher confidence in cloud height for opaque clouds

18

CTP Increments

BCTP = 5 hPa

19

CTP Increments

BCTP = 50 hPa

20

Background error for CTP• The CTP control variable increment variance is a

function of:– Latitude– Height

• Convergence is degraded as a function of background error– A 5 hPa error aloft is much different in terms of cloud

temperature than near the sfc– Also seen at ECMWF in similar McNally 2009

implementation• A simplified error model, as a f(CTP) implemented,

ranging from 5 hPa aloft to 13 hPa below – convergence on par w/ fixed error 21

Differences of Analyses

22

Std. Dev. of (A(5) – A(FullOS))

• The Variance of the difference between a CTL analysis (all obs) and a Cloudy AIRS analysis (CTL + AIRS Cld) at 0000 UTC, 850hPa

• Satellite Track of Aqua satellite evident

23

• The Variance of the difference between a CTL analysis (all obs) and a Cloudy AIRS analysis (CTL + AIRS Cld) at 0000 UTC, 850hPa

• Satellite Track of Aqua satellite evident

Differences of AnalysesStd. Dev. of (A(5) – A(FullOS))

Differences of Analyses

24

Std. Dev. of (A(5) – A(FullOS))

• In regions of expected persistent cloudiness, there are more changes to the analysis (month-averaged CTP, top, and cloud fraction, bottom)

Future Efforts• This work is still ongoing• Cycling runs w/ advanced CTP bkg error

model are under investigation• There are a number of issues to consider:

– Inclusion of a 2nd outer loop showed bias w.r.t. CTP solution --- GSI expanded to interpolate between levels

– GSI calls setuprad each outer loop, and this work is thus re-initialized w/ a second retrieved CTP

• CTP is not re-linearized about the solution from the first outer loop

– CTP was only considered for 15-11 μm 25

Future Efforts

• Expansion of observation-centered predictors to other studies– ECMWF has had this infrastructure, and it’s my

thought that this has a lot of possibilities for future work

– A logical potential expansion of this study is to cloud fraction

– With various assumptions, this capability can be expanded to SST, sfc emissivity, Cloud-Water Path, etc.

– With a little cleaning up, it will be as simple as setting an initial value, background error, and Jacobians of any predictor 26

27

Differences of Analyses

28

Std. Dev. of (A(50) – A(FullOS)) Std. Dev. of (A(5) – A(FullOS))

• Changes away from tracks for BCTP = 50 hPa

are a result of decreased convergence in the minimization