1. Climate Change: Moonshine, Millions of Models, & Billions of Data New Ways to Sort Fact from Fiction Bruce Wielicki March 21, 2007 University of Miami Lecture

3. Shows consistent calibration stability at < 0.3 Wm -2 per decade (95% conf) Unfortunately only works for tropical mean ocean (nband vs bband issues) Regional trends differ by +2 to -5 Wm -2 /decade SeaWiFS vs CERES Loeb et al. 2007 J. Climate 0.21 Wm -2

4. Using CERES to Determine Length of Climate Data Record Needed to Constrain Cloud Feedback Given climate variability, 15 to 20 years is required to first detect climate trends at cloud feedback level with 90% confidence, and 18 to 25 years to constrain to +/- 25% in climate sensitivity Half of Anthrop Forcing of 0.6 Wm -2 /decade Loeb et al. 2007 J. Climate

5. Annual Mean Global SW TOA Flux Anomaly (Earthshine versus CERES: 2000 to 2004) Loeb et al. 2007 GRL Earthshine data implies large change of 6 Wm -2 in global reflected SW flux: is the Earth's albedo changing? (Palle et al., Science, 2004) CERES shows an order of magnitude less variability than Earthshine Earthshine approach is incapable of capturing changes in global albedo at climate accuracy.

6. CERES Shortwave TOA Reflected Flux Changes: Ties to Changing Cloud Fraction Unscrambling climate signal cause and effect requires complete parameter set at climate accuracy, e.g. for forcing/response energetics: radiation, aerosol, cloud, land, snow/ice, temperature, humidity, precipitation Tropics drive global albedo variations: global is in phase with tropics and 1/2 the magnitude Cloud fraction variations are the cause (not optical depth)

7. Early Cloud Feedback Signals in the Arctic from CERES data Seiji Kato and the CERES Science Team CERES: Derived from MODIS radiances by the CERES cloud algorithm. Radar: Derived from a ARM cloud radar. Lasers: Derived from a micro-pulse lidar and a Vaisala ceilometer Error bars and dashed lines indicate max. and min. during 4 years. Snow/sea ice fraction changed at a rate of 0.064 per decade (significant at an 80% confidence level) Cloud fraction changed at a rate of 0.047 per decade (significant at an 80% confidence level) Albedo change is insignificant at an 80% confidence level. From Kato, S., N. G. Loeb, P. Minnis, J. A. Francis, T. P. Charlock, D. A. Rutan, E. E. Clothiaux, S. Sun-Mack, 2006: Seasonal and Interannual Variations of Top-of-Atmosphere Irradiance and Cloud Cover over Polar Regions Derived from the CERES Data Set, Geophys. Res. Lett., 33, L19804, doi:10.1029/2006GL026685. Mean Cloud Fraction at Barrow AKTrends derived from Terra and Aqua Data over the Arctic Terra Aqua Linear Fit to Terra Missing days

8. Ocean Heat Content and Net Radiation A case study in the need for independent observations & analysis Possible Causes of 2004/5 Drop in Ocean Heat Storage: - transition from XBT to ARGO ocean in-situ data and sampling? - cooling upper 750m of the ocean, but larger heating deeper? - unmeasured heating under sea ice? - The answer: warm bias in XBTs (dominate pre-2002) cold bias in ARGO (dominate post 2002): no cooling in 2004/5 when bias is corrected. mystery solved. Ocean Cooling? Lyman et al., Science 2006 Net Radiation: no Altimeter Sea Level: no GRACE Ice Sheet: no 1992 to 2003 data from Wong et al. J. Climate 2006

9. "Global Dimming": is it real? What about new CERES fusion satellite surface fluxes? ARM/BSRN/CMDL/Surfrad Surface Radiation Sites

10. Surface SW Flux Validation Noise Spatial mismatch of surface point to satellite area Error decreases as simple 1/sqrt(N) random noise: but takes 20 sites for 1 year to reach 1 Wm -2 : 10,000 samples. (Wielicki, AMS Radiation Conference, 2006)

11. CERES Surface Fluxes vs Surface Sites: Interannual Anomalies Consistent at 0.2% or 0.3 Wm -2 Global satellite sampling of radiation fields remains key: regional variability (climate noise) is very large: 10 times the global forcing of 0.6 Wm -2 /decade: even averaging 40 disperse surface sites. Result from GEWEX Radiative Flux Assessment (in progress)

12. How well can we pull climate records from meteorological satellite data like ISCCP from geostationary? Loeb et al., 2007 J. Climate Geo calibration & sampling errors dominate inter- annual signals Uncertainty in Geo trends are a factor of 10 larger than climate goal: can we learn how to improve past data sets?

13. Trend in All-sky Downward SW flux at the Surface (2000-2004) ISCCP vs CERES CERES (SRBAVG_GEO) ISCCP minus CERES -ISCCP trends show systematic regional patterns that coincide with the area of coverage by the individual GEO instruments. -Artifacts in the GEO data are removed in CERES processing by a normalization procedure that corrects for GEO calibration, narrow-to-broadband, and radiance-to-flux coversion errors, so that fluxes from each GEO instrument are consistent with CERES.

14. Stainforth et al., 2005, Nature

15. Neural Net Structure Climate OSSEs Input Variables Input Variables Planet “I” - Planet “J” Planet “I” - Planet “J” base state CO 2 climate base state CO 2 climate TOA SW Flux TOA LW Flux Total Cloud Fraction Conv. Cloud Fraction Total Precipitation Large Scale Snowfall Large Scale Rainfall Surface Latent Ht Flux Surface Net SW Flux Surface Net LW Flux Surface Net Radiation Neural Network Output Variables Output Variables Planet “I” - Planet “J” 2xCO 2 minus 1xCO 2 2xCO 2 minus 1xCO 2 Surface Temperature Summer U.S. Precip Sea Level etc... AddObservationError Bias,  Difference in neural net performance with and without observation errors Isolates effect of observation error on constraining climate uncertainty

16. Climate OSSE's & Perturbed Physics Ensembles Early Conclusions: - Uses 2500 climateprediction.net mixed layer CO 2, 2 x CO 2 runs - Trains neural net on 5% of 6 million climate run "pairs", tests rest -Use of base climate state model differences can predict to 0.4C (1  ) the doubled CO 2 sensitivity differences over range of 2 to 12C. - Of 33 global mean climate variables, most information is in 11 (radiative fluxes, cloud cover, precipitation, snowfall, latent heat) - Use of base state climate metrics is highly nonlinear: linear regression factor of 2.5 lower accuracy. -Use of base state climate metrics from climateprediction.net simulations fails to predict accurately for IPCC mixed layer runs. - Use of climate change metrics (e.g. decadal change) are much more accurate than base state, are more linear, and show modest loss of accuracy when applied to IPCC runs. - Adding observation errors seriously degrades accuracy of neural net predictions if observational error exceeds 25% of climate change - Writing up for BAMS, next steps are coupled ocean/atmosphere models and more complete climate metric tests.