Using Modern Satellite Data to Constrain Climate Models, and Baesian Climate Projection Stephen Lero, Yi Huang, Richard Good, James Anderson (Harvard Universit) International Detection and Attribution Group Boulder, Colorado Januar 8, 1
Outline New Satellite Data Using infrared spectral data Using infrared spectral data and GNSS radio occultation data Baesian Climate Projection Problems with normal multi-pattern regression: quick review Modifing IPCC model predictions using data Application to surface air temperature record, 197-1999 Summar Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection
Spectral Infrared Data Huang, Y., S. Lero, P.J. Gero, J. Dkema, and J. Anderson, 1: Separation of longwave climate feedbacks from spectral observations. J. Geophs. Res., In Press. Huang, Y., S.S. Lero, and J.G. Anderson, 1: Determining longwave forcing and feedback using infrared spectra and GNSS radio occultation. J. Climate, In Review. Radiance Trend [1-8 W cm - (cm -1 ) -1 ster -1 decade -1 ] Trend of Tropical IR spectrum in clear skies 5-5 -1-15 - Tropics, SRES A1B, clear NCAR CCSM3 GFDL CM. MPI/ECHAM5-OPYC UKMO HadCM3 5 1 15 Frequenc [cm -1 ] W m - Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 3
Spectral IR Signals CFMIP x CO experiment Partial Radiative Perturbation Use in optimal fingerprinting in information content stud All signals normalized b spectral integral Blue is mean signal form Red is signal uncertaint Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 4
Longwave Feedback Maps Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 5
Detected Feedback Error Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 6
Detected Feedback Error Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 7
Add GNSS RO Dr Pressure φ GPS v GPS p p ε φ LEO ( ln p N t λ Δν = v LEO cos φ LEO + v GPS cos φ GPS n(p) = 1 ε(p ) dp π p p ) h 1 H p ( h t ) p v LEO P (x, ) =P (x ) P () =P ( x) P (x) Radio occultation (RO) insensitive to clouds Refractive index mostl densit with some water vapor contribution in the lower troposphere Dr pressure nearl the same as pressure but with a precipitable water component in lower troposphere [Km] 3 5 15 1 5 x : : observation variables δln(pdr) prediction variables..1.1..3.4 all ta trop ta strat hus trop hus strat Σ Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 8
Detection Using IR Spectra and RO Dr Pressure Truth OD: IR OD: IR+RO Truth ta trop 1 IR OD ta trop IR+GPS OD ta trop Trop T 5 5 1 ta strat 1 ta strat ta strat Strat T.5.5 1 hus trop 1 hus trop hus trop Trop q 5 5 1 hus strat 1 hus strat hus strat Strat q.5.5 1 cld uppertrop 1 cld uppertrop cld uppertrop Upper Trop cld 5 5 1 Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 9
New Satellite Data: Summar Some longwave feedbacks are ambiguous in spectral domain. The can be resolved using radio occultation data, except low cloud and surface temperature signals. Earl detection of CO forcing and stratospheric trends likel. Detection times for other trends probabl similar to surface air temperature detection times. Stud under wa using infrared (AIRS, IASI) and GPS RO data (CHAMP, SAC-C, COSMIC). Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 1
Assumptions of Multi-pattern Regression Assumes a separable prior in signals shapes and signals amplitudes. Assumes prior knowledge of signals shapes but no prior knowledge of signals amplitudes. The opposite ma make more sense. Extrapolation of results to climate projection is nois at best, tenuous at worst. Lero, S.S., and J.G. Anderson, 1: Optimal detection of regional trends using gobal data. J. Climate, In Review. No need for first two assumptions. Derive new equations without them. P1: A climate model must be weighted according to how well it reproduces data. P: The data used must be relevant to the variable to be predicted. C: A simple application of Baes encapsulates these phenomena, ields a PDF for projection, and evaluates the overall agreement between climate model ensemble and data. Tebaldi et al., 5: Quantifing uncertaint in projections of regional climate change: A Baesian approach to the analsis of multimodel ensembles. J. Climate, 18, 154-154. Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 11
Baes (or just marginalization) P ( x, ) = P( x ) P( ) = P( x) P( x) x : observable variables : prediction variables Form joint PDF P(x,) using an ensemble of climate models. For each model, need (1) observation kernel to simulate data x from hindcast run, and () emissions scenario run to generate prediction variables. Natural variabilit in x and and uncertain phsics will both be accounted for. With data d, set x = d and P( x=d ) is the projection PDF with data incorporated. P(x) is a normalization constant that guarantees a unit integral of P( x) over. Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 1
Simple Solution Assume Gaussian statistics Covariance for (x,): Projection : σ ρσ σ x x Σ = ρσ xσ σ posterior = prediction σ + ρ σ x d Projection uncertaint: σ,posterior = σ (1 ρ ) x Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 13
Simple Solution Assume Gaussian statistics Covariance for (x,): Projection : σ ρσ σ x x Σ = ρσ xσ σ posterior = prediction σ + ρ σ x d σ ρσ Projection uncertaint: σ,posterior = σ (1 ρ ) σ x x Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 14
Simple Solution Assume Gaussian statistics Covariance for (x,): Projection : σ ρσ σ x x Σ = ρσ xσ σ posterior = prediction σ + ρ σ x d σ σ d ρσ Projection uncertaint: σ,posterior = σ (1 ρ ) σ x data x Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 15
Multivariate Solution Σ Σ Σ xx x Correlation of = Σx Σ prediction and data Optimization -1 posterior = prior + ΣxΣxx( d xprior ) Σ d = Σ Σ x Σ -1 xx Σ x Ensemble uncertaint Information provided b data Posterior uncertaint is alwas less than the prior uncertaint. Matrix inversion required onl in the space of the observables. Indeterminate number of prediction variables allowed. Matrix pseudo-inversion, meaning EOF truncation, is necessar for small ensembles. Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 16
Evidence for Ensemble N / 1/ 1 T -1 P( d E) = (π ) Σxx exp ( d xprior ) Σxx( d xprior ) Precision of ensemble Accurac of ensemble Detection and attribution: One ensemble E 1 is run with forcing/emissions scenario Another ensemble E is run without forcing Compare the evidences P(d E 1 ) and P(d E ). No need for projection. Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 17
Application HadCRUT3, trends 197-1999. Eight regions: North Pacific, Warm Pool, Tropical East Pacific, North America, Equatorial Atlantic, North Atlantic, Northern Europe, Entire Globe. Use grid cells where data is continuousl valid to construct regional average trends. Subset CMIP3 th centur runs in exactl the same wa (where HadCRUT3 data values are valid). Predict a regional surface air temperature change from until ears 1,, 3, 4, 5, 6, 7, 8 Average over entire region Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 18
HadCRUT3 Trends, 197-1999 HadCRUT3 Temperature Trend -1. -.5..5 1. 1.5..5 3. 197- Temperature Change [K] Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 19
CMIP3 Trends, 197-1999 Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection
CMIP3 Trends, 197-1999 Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 1
Regional Trends, 197-1999 HadCRUT3 trends 1 Δ Temperature, 197-1999 [K] CMIP3 model trends (+) Natural (internal) variabilit included in scatter of models. Entire Globe North Pacific North America North Atlantic Equatorial Atlantic Warm Pool Tropical East Pacific North Europe Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection
Data-Prediction Correlation CMIP3 models (+) Data ( ) Δ Temperature, North America, -4 [K] 3 1 3 1 Entire Globe Equatorial Atlantic 1 North Pacific North America North Atlantic Warm Pool Tropical East Pacific North Europe 1 1 1 Δ Temperature, 197-1999 [K] Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 3
Application Result 5 5 North America ΔT [K] 4 3 1 neofs = 1 χ =.5567 North America ΔT [K] 4 3 1 neofs = χ =.17386-1 4 6 8 Year -1 4 6 8 Year 5 5 North America ΔT [K] 4 3 1 neofs = 3 χ = 1.7 North America ΔT [K] 4 3 1 neofs = 4 χ = 1.7945-1 4 6 8 Year -1 4 6 8 Year χ T 1 = ( d x) Σxx( d x) Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 4
Application Result 5 5 North America ΔT [K] 4 3 1 neofs = 5 χ = 1.844 North America ΔT [K] 4 3 1 neofs = 6 χ = 14.1853-1 4 6 8 Year -1 4 6 8 Year 5 5 North America ΔT [K] 4 3 1 neofs = 7 χ = 4.7733 North America ΔT [K] 4 3 1 neofs = 8 χ = 4.7779 χ -1 4 6 8 Year T 1 = ( d x) Σxx( d x) -1 4 6 8 Year Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 5
How to Truncate? Objective testing Comparing P(d E) for different ensembles E requires a test that allows for large χ, because large χ informs on structural problems with a perturbed phsics ensemble. Attribution is a special case of application of relative P(d E), one ensemble subjected to external forcing and the other not. Objective testing demands a truncation algorithm independent of values of χ. Probabl North criterion. No need for Σ x. Climate projection Little interest in data components that fall outside the bounds of an ensemble. Points toward an F-test, constraint on χ. Definite need for Σ x. Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 6
Summar: Baesian Projection Method that assumes less than multi-pattern regression; fuses model and data to produce a projection; onl need joint PDF for data and prediction P(x,). Clearl links detection & attribution, testing climate models, climate projection. Gaussian joint PDF has simple solution for projection and its uncertaint. Looks like optimal detection. Method of EOF truncation depends upon application, objective testing or climate projection. Projecting 1 st centur trends using CMIP3 and 3 ears of HadCRUT3 produces little improvement over CMIP3 prediction. Needs a dense joint PDF, use cpdn or equivalent. Better use of HadCRUT3 data, 19- piecewise. Use other data, precipitation, sonde. Januar 8, 1 Lero, Huang et al.: Satellite Data and Baesian Projection 7