1JANUARY 2006 FORSTER AND GREGORY 39 The Climate Sensitivity and Its Components Diagnosed from Earth Radiation Budget Data

One of the major uncertainties in the ability to predict future climate change, and hence its impacts, is the lack of knowledge of the earth’s climate sensitivity. Here, data are combined from the 1985–96 Earth Radiation Budget Experiment (ERBE) with surface temperature change information and estimates of radiative forcing to diagnose the climate sensitivity. Importantly, the estimate is completely independent of climate model results. A climate feedback parameter of 2.3 1.4 Wm 2 K 1 is found. This corresponds to a 1.0–4.1-K range for the equilibrium warming due to a doubling of carbon dioxide (assuming Gaussian errors in observable parameters, which is approximately equivalent to a uniform “prior” in feedback parameter). The uncertainty range is due to a combination of the short time period for the analysis as well as uncertainties in the surface temperature time series and radiative forcing time series, mostly the former. Radiative forcings may not all be fully accounted for; however, an argument is presented that the estimate of climate sensitivity is still likely to be representative of longer-term climate change. The methodology can be used to 1) retrieve shortwave and longwave components of climate feedback and 2) suggest clear-sky and cloud feedback terms. There is preliminary evidence of a neutral or even negative longwave feedback in the observations, suggesting that current climate models may not be representing some processes correctly if they give a net positive longwave feedback.


Climate sensitivity
In 1990 the Intergovernmental Panel on Climate Change (IPCC) suggested a range of 1.5-4.5 K for the global surface equilibrium temperature increase associated with a doubling of CO 2 (Houghton et al. 1990).Since then, despite a massive improvement in models and in our understanding of the mechanisms of climate change, the uncertainty in our projections of temperature change has stubbornly refused to narrow (Houghton et al. 2001).A significant part of this uncertainty is due to different cloud feedbacks between models, although differences in other feedbacks, such as water vapor, also play a role (Houghton et al. 1990;Cess et al. 1996;Colman 2003).Recently, there have been several efforts to try and narrow this range using a combination of observations and climate model data; however, the large uncertainties in climate sensitivity have, if anything, increased (Forest et al. 2002;Gregory et al. 2002;Harvey and Kaufmann 2002;Knutti et al. 2003).In particular, it has proved extremely difficult to rule out very high values of climate sensitivity using observations (Gregory et al. 2002).
There are many definitions of climate sensitivity in the literature.While the most quoted sensitivity is the equilibrium warming for 2 ϫ CO 2 this is not necessarily the most useful, as differences in the CO 2 radiative forcing can be confused with differences in climate response, and ideally one would like to know the climate response to any forcing mechanism.In this work, we use a standard linear definition of climate sensitivity, which we state here, and go on to show how it can be derived from observational data.
A linear climate feedback parameter (Y ) can be defined, where and where N is the net downward irradiance at the top-of-atmosphere (TOA) energy balance and ⌬T s is the surface temperature change.
Consider factors that directly affect the top-ofatmosphere flux (N ): where F is a surrogate parameter for external radiative forcing mechanisms and X represents all the other internally driven mechanisms that might affect N, which are not directly or indirectly related to surface temperature As we are concerned with linear changes in N, we can set N o ϭ 0 and ignore the nonlinear terms.The term (‫ץ‬N/‫⌬ץ‬F ) ⌬F is a radiative forcing term, which we chose to rewrite as Q: This equation includes a term [(‫ץ‬N/‫⌬ץ‬X ) ⌬X ] that allows N to vary independently of surface temperature.A similar equation was used by Gregory et al. (2004).If we regress (QϪN) against ⌬T s , we should be able to obtain a value for Y.
The X terms are likely to contaminate the result for short datasets, but provided the X terms are uncorrelated to ⌬T s , the regression should give the correct value for Y, if the dataset is long enough.
The concept of radiative forcing relates this climate feedback parameter to the expected equilibrium surface warming (⌬T eq ) caused by a constant radiative forcing (Q c ;Houghton et al. 2001): ͑4͒ If the radiative forcing (Q c ) is due to a doubling of carbon dioxide (approximately 3.7 W m Ϫ2 ), then ⌬T eq is termed the equilibrium climate sensitivity.A smaller value of Y NET leads to a larger equilibrium surface temperature change and corresponds to a larger equilibrium climate sensitivity.Models indicate that while most radiative forcing mechanisms have climate sensitivities that are within ϳ30% of that from carbon dioxide changes, some do not, such as absorbing aerosol and ozone changes (Hansen et al. 1997;Joshi et al. 2003).Also, for the same radiative forcing mechanism, a climate model's climate sensitivity can vary with time, and some climate feedbacks, such as certain cloud changes, may not be linear (Colman et al. 1997;Senior and Mitchell 2000;Gregory et al. 2004).Acknowledging these limitations, we search for a single value for Y NET in observations, and return to discussion of its applicability later.

Method
For our approach we use Eq. ( 3) to diagnose the climate feedback parameter (Y NET ) from observed time series of N and ⌬T s .We estimate corresponding values for Q and then regress QϪN against ⌬T s to find a value for Y NET from the slope of the regression line (also see Gregory et al. 2004).We use either seasonal data (with the annual cycle removed) or annual averages.Importantly, with this method, Y NET is only determined by changes in N, Q, and ⌬T s over the time period considered.Hence, we do not need to identify a base state of zero climate change.The methodology is applicable to: 1) an equilibrium nonforced situation, where Q would be constant and N is responding to internal variability in ⌬T s ; 2) transient forced climate change situations, where N and ⌬T s are responding to a radiative forcing (Q); and 3) a combination of internal variability and forced climate change.Sources of absolute error can safely be ignored as they only affect the intercept of the regression line and not its slope; N in particular might be expected to have large absolute errors, being difficult to measure.
The best estimate of the equilibrium climate sensitivity and its uncertainty range are derived from the linear regression analyses.The choice of linear regression model has an important influence on results, since the scatter is quite large.For our analyses, we adopted a robust fitting technique, based on ordinary least squares (OLS) regression of QϪN against ⌬T s (Feigelson and Babu 1992).Compared to other techniques this gave larger error estimates.For a number of reasons we believe that OLS methodology was the most appropriate choice of the regression model.Ordinary least squares regression will give a bias toward small Y values and high climate sensitivities; all the other linear regression models examined suggested less than a 2-K doubling for 2 ϫ CO 2 (see appendix).
For the uncertainty analysis we assume, like Gregory et al. (2002), that errors in the observable parameters (N, Q, and ⌬T s ) all have Gaussian distributions.In a Bayesian statistics framework, this is equivalent to assuming Gaussian observational errors and a uniform "prior" in each of the observables.Since the uncertainties in Q and N are much larger than in ⌬T s (a factor influencing our choice of regression model; see appendix), uncertainty in QϪN is linearly related to uncertainty in Y, so our assumption is also approximately equivalent to assuming a uniform prior in Y.Other studies (Forest et al. 2002;Knutti et al. 2003) have assumed a prior that is uniformly distributed in equilibrium climate sensitivity, which is proportional to 1/Y.Compared with our prior, theirs emphasizes the higher values of equilibrium climate sensitivity.We believe our choice is more appropriate for our analysis and input data, but it has the result that our range of climate sensitivity lies lower.Differences in priors should be taken into account when comparing quoted ranges of possible temperature change.

Data a. Net energy imbalance (N)
For this work we take N to be the net flux imbalance at the top of the atmosphere (positive downward) measured by instruments on the Earth Radiation Budget Satellite (ERBS); we use data from the 15 yr of widefield-of-view (WFOV) data and the shorter ϳ5 yr dataset from the scanning radiometer.These instruments measure radiation between 60°S and 60°N.The WFOV instrument uses two cavity radiometers to measure the absorbed solar radiation and the outgoing longwave radiation (OLR); the scanning instrument measures shortwave (SW) and longwave (LW) radiances with a thermistor bolometer (Wielicki et al. 2002).As a result of battery problems, there are several gaps in the data, especially during 1993, and for these reasons we do not use data from 1993 or after 1997 for computing annual averages.Our data are essentially the same as the datasets presented in Wielicki et al. (2002), except we have additionally applied a correction to the WFOV data for the decay in the satellite's orbit (T.M. Wong 2005, personal communication).This correction reduces the positive trend in OLR and increases the positive trend in absorbed solar radiation.The annual averaged anomalies used in this study from the WFOV instrument are shown in Fig. 1b.These data arguably represent the best observations we have for N and relative calibration in SW and LW fluxes are expected to be better than 0.5 W m Ϫ2 (for decadal timescale changes) and even better than this on interannual time scales (Wielicki et al. 2002).The changes we observe in the data are considerably larger than this (see Fig. 1b).

b. Surface temperature (⌬T s )
We use two global surface temperature anomaly datasets; one from the Goddard Institute for Space Studies (GISS; Hansen et al. 1999) and the other from the Hadley Centre and University of East Anglia's Climate Research Unit (HADCRU) (Jones et al. 1999).
Their anomalies are similar for the 1985-97 period (Fig. 1a).However, the two surface temperature time series shown give quite different changes between 1996 and 1997.As our approach is dependant on stable time series of robust data, 1997 was excluded from subsequent analysis.

c. Radiative forcing (Q)
Radiative forcing time series were obtained from various sources (Sato et al. 1993;Myhre et al. 2001;Hansen et al. 2002).We found that volcanic and wellmixed greenhouse gas radiative forcings played the most important role in determining Y NET ; changes in total solar irradiance also had some effect, but these values could be derived directly from the ERBS instruments, which largely eliminated them as a source of uncertainty.Time series of the important datasets are shown in Fig. 1c.The tropospheric aerosol direct and indirect radiative forcings were found not to impact our regression.Despite potentially large absolute errors in these forcings, their impact on our analysis is likely to be small, as the tropospheric aerosol forcing in the datasets analyzed changed very little over 1985-96 (Myhre et al. 2001).However, we acknowledge that any undiagnosed change in aerosol radiative forcing could have significant effects on our shortwave results.Over this relatively short time period any undiagnosed forcing changes would likely be uncorrelated with global mean surface temperature changes; these forcings would therefore increase the uncertainty, or noise, in the analysis, rather than lead to systematic errors in Y.The forcing that had by far the largest impact on the regression was volcanism and, in particular, the eruption of Mt.Pinatubo in 1991.For this study, we used SW and LW volcanic forcing estimates from several sources (derived from data outlined in Myhre et al. 2001;Sato et al. 1993).This volcanic forcing may be in error by as much as 30% (Houghton et al. 2001); the impact of this uncertainty on the results is discussed in the results (and in Table 3).
As the ERBS instruments measured SW and LW components of the energy budget separately, the shortwave climate feedback parameter (Y SW ) and longwave climate feedback parameter (Y LW ) could also be found.
To do this, we separately considered SW or LW components of Q and N in Eq. ( 1) for the regression analysis.For the scanning instrument, we were able to further subdivide N into clear-and cloudy-sky components and use these to find corresponding Y values.For the regression of QϪN versus ⌬T s an appropriate Q, representative of the N being measured, needed to be estimated.Therefore, to derive Y SW and Y LW we had to split radiative forcing into SW and LW components.For the cloudy case where N cloud is taken to be NϪ N clear we set Q to zero.This assumes that there is no difference between the clear-sky and all-sky radiative forcing terms.In practice, over the short 5-yr time period where the scanner data were available, Q uncertainties had little impact on the derived Y values.

Results
For several time periods, we present values for the total climate feedback parameter (Y NET ) and its components.For the ERBS WFOV instrument, we divide the record into the pre-Pinatubo period , the Pinatubo years , and the 1985-96 period; 1993 is missing in the ERBE record.Table 1 gives the results of the regression analyses, using both annual averages and seasonal data (with the annual cycle removed); it also compares the results from using the two surface temperature datasets.Corresponding figures to several of these regressions are shown in Figs. 2 and 3. Figure 4 summarizes the overall findings for the annual results of the WFOV instrument.Consistent with the lower correlations, seasonal climate feedback parameters are slightly smaller in magnitude than the feedback parameters derived from annual averages, although they agree to within their respective uncertainty estimates.Results derived from annual averages typically have the highest correlations.Table 2 and Fig. 5 show the results of the scanning instrument analyses, which are summarized in Fig. 6.
The results show that the linear model of climate sensitivity adopted works well in the longwave, and we are able to quantify Y NET quite well and Y LW values better.For Y SW the simple model does not work as well, but we are still able to usually determine a sign.Figure 4 shows that over both the pre-Pinatubo period and the 1985-96 record we find that a Y LW of ϳ4 W m Ϫ2 K Ϫ1 is partly offset by a likely negative Y SW , to give a Y NET of ϳ2.0 W m Ϫ2 K Ϫ1 .Several climate models exhibit a similar "NET" positive feedback (Houghton et al. 2001).The NET climate feedback derived from the Pinatubo response is similar, but its LW and SW components are very different.Uncertainties in the radiative forcing (see below) cannot account for these differences, so these results indicate that the Pinatubo year's climate response is not the same as the longerterm response.The 1985-89 results from the scanning instrument agree well with those from the WFOV instrument (Table 2).Although errors in the regression are much larger, the scanning instrument also hints at the clear-and cloudy-sky feedbacks.

Uncertainties in deriving Y values
The 95% uncertainties for Y, presented in the tables, are purely statistical and are from the straight-line fits.However, we also analyzed the effects of using different temperature and/or radiative forcing datasets (Table 1  and Table 2), and compared results from annually averaged data with those from seasonal data (Table 1).
In the LW, QϪN values are well correlated to ⌬T s values, especially for the 1985-90 period, where correlations are larger than 0.95 for the annual averages and the regression gives a near-perfect straight line for Y LW .For all the time periods analyzed, both Y LW and Y NET TABLE 1.For different time periods the table shows the Y SW , Y LW , and Y NET values derived from the WFOV instrument.A 95% uncertainty and the correlation of QϪN vs ⌬T s are also presented.Data are shown for the two surface temperature anomalies (HADCRU and GISS).These data use the GISS well-mixed greenhouse gas forcing.Results are shown for seasonal and annual anomalies, with the seasonal cycle removed from the seasonal data.are reasonably well constrained by the regression analysis and correlations of annually averaged data are almost always above 0.6.The regression analysis does not work as well for deriving Y SW ; uncertainties are larger and correlations smaller, compared with those for Y LW and Y NET .Although its correlations are similar, the HADCRU surface temperature dataset gives higher Y LW values than the GISS dataset.However, again, these differences are no larger than the 95% errors in the regression analysis.Annual Y NET values for 1985-96 calculated with the two datasets differed by 0.2 W m Ϫ2 K Ϫ1 .For the summary in Figs. 4 and 6 the results from the analysis using the GISS surface temperature and forcing datasets are used as these often had higher correlations and smaller errors than found when using other analyses.
Table 3 analyzes how uncertainty in the radiative forcing term affects the derived Y NET values.The volcanic forcing and the well-mixed greenhouse gas forcing (WMGHG) both play a role in determining Y values, and results become largely meaningless if either forcing is ignored.However, the two estimates of WMGHG forcing are very similar, so uncertainty in this forcing has a small effect on Y (0.1 W m Ϫ2 K Ϫ1 ).In addition, we also examined the effect of including a 10% uncertainty estimate in the WMGHG forcing (Houghton et al. 2001); this also added 0.1 W m Ϫ2 K Ϫ1 to the overall uncertainty estimate.Varying the magni- tude of the volcanic radiative forcing by its suggested uncertainty of 30% (Houghton et al. 2001) changed Y NET values by 0.5 W m Ϫ2 K Ϫ1 .Over the 1991-92 time frame of the major Pinatubo climate response, the forcing term plays a larger role; but, even here, uncertainty in the forcing only varied Y SW by 1.0 W m Ϫ2 K Ϫ1 and Y LW by 0.5 W m Ϫ2 K Ϫ1 .We can therefore conclude that, provided we are not ignoring an important forcing term, uncertainties in radiative forcing ought not to have had a major impact on the derived Y values.If there are undiagnosed aerosol changes, they would impact our analysis.However, as the temperature changes over this time period are far from being a monotonic increase (Fig. 1), it would hopefully be unlikely that any undiagnosed forcing would correlate with ⌬T s .If these forcings were uncorrelated with ⌬T s , aerosol would be unlikely to alter the slope of the regression; rather they would make the regression noisier.Some of the noise in the Y SW regression could easily be the result of this undiagnosed forcing.However, experiments with the Third Hadley Centre Coupled Ocean-Atmosphere GCM (HadCM3), in which all the forcings were known (Gregory et al. 2004), also found a noisier Y SW .It is likely that shortwave cloud feedback processes also account for at least part of this noise (also see section 7).If, however, aerosol was correlated with ⌬T s (such as could conceivably be the case with forest fires), it could have a major impact, especially on our estimate of Y SW.
To estimate the complete error in our analysis of Y we also need to include the likely error introduced by uncertainties in the ERBE time series.Over 1985-96 the total change in QϪN is about 2-3 W m Ϫ2 (Fig. 1), which can be compared to a suggested interdecadal calibration error of less than 0.5 W m Ϫ2 (Wielicki et al. 2002).A change in the N of 0.5 W m Ϫ2 would significantly alter the results.However, these errors would not necessarily correlate with ⌬T s , so the fact that the regression gives any meaningful results at all, and sometimes very good straight line fits, indicates that the ERBE data may be reliable.Similar analyses with other available datasets for N produced little or no meaningful result.The N values were also estimated using 1981-87 Nimbus-7 data (Jacobowitz et al. 1984(Jacobowitz et al. ), 1950(Jacobowitz et al. -2000 ocean heat uptake data (Levitus et al. 2000), and a 1983-2000 dataset derived from several satellite instruments (Zhang et al. 2004).Several time periods were considered.For the Nimbus-7 analysis, although uncertainties are much larger, correlations were found to be around 0.5 for Y NET and Y LW , and the derived Y values agreed well with those from the ERBE data.For the other datasets, QϪN and ⌬T s were always uncorrelated and no significant relationships were found.Using the WFOV data without their latest orbital decay corrections also gave an indication of how uncertainty in N affected our results.For the 1985-96 time period, the uncorrected data, compared to the results in Table 1, gave a much higher value for Y LW (but the correlations were ϳ0.2 smaller).This resulted in a Y NET of around 4.0 W m Ϫ2 K Ϫ1 , which would correspond to only a 1-K warming for a doubling of carbon dioxide.
The errors in ERBE that would have the largest affect on our derived Y values are those caused by longterm calibration drift; in these cases a correlation would remain.The data used in this study have had latest correction data applied; using the uncorrected data would increase Y NET by over 50%.For the summarized  Y values and their uncertainties, presented in Figs. 4 and 6 and in the conclusions, we assume that the corrected ERBE time series are perfect apart from an unknown constant offset.

Correlation
To obtain the overall uncertainty estimate presented in the abstract and conclusions, we took the central best estimate, with the GISS dataset and used its regression uncertainty range.We then added surface temperature and radiative forcing uncertainties (discussed above), assuming they are independent of each other.This gave an overall uncertainty estimate for Y NET of 2.3 Ϯ 1.4 W m Ϫ2 K Ϫ1 .This uncertainty range does not account for uncertainty in the ERBE time series (N ).We used the GISS central estimate and its regression uncertainty, as these data had the highest correlations.Reasons for using the highest correlation to select the central value are that it may suggest that this dataset is more reliable, and the regression will have the smallest bias in the OLS slope.

Applicability of results for long-term climate change
Here we ask two questions about the applicability of the results for determining future climate change.change (Hansen et al. 1997;Joshi et al. 2003).If our Y NET value is derived from tropical temperature variability and El Niño, it could be unrepresentative of long-term climate change (Soden 1997;Houghton et al. 2001).Furthermore, it has also been suggested that subannual changes in temperature are often confined to the surface layer and that using such changes to diagnose climate feedback leads to an underestimate of the water vapor feedback (Hall and Manabe 1999).In addition, the ERBE instruments do not measure radiation poleward of 60°, so they may not adequately detect certain climate feedbacks, particularly the surface albedo feedback.
The years 1995 and 1985 span the range of global ⌬T s anomalies seen in the time series.Figure 7 shows that the change in surface temperature between these two years has a pattern that is a little like that of both the observed long-term surface temperature trends (e.g., Houghton et al. 2001, their Fig. 2.9d).Furthermore, Fig. 1 shows that the 1985-97 temperature changes are occurring throughout the troposphere, in a similar manner to that expected from long-term climate change.Therefore, we argue that the Y values we derive from the regression are likely to be representative of those due to longer-term climate change.In addition, using shorter time periods where the temperature changes may not be representative of longer-term climate change (such as 1991-92) gives distinctly different answers (Fig. 4).However, as it is likely that the observed temperature changes are a combination of variability (on different time scales) and forced climate change, with contributions from different forcing agents, there is still the possibility that such variability biases our result.For example, this variability could involve circulation responses that differ from those expected in a long-term climate change situation.Thus, our diagnosed climate sensitivity may also be a signature of this variability.
Integrations from HadCM3 were also employed to test the regression technique.We analyzed data from a HadCM3 integration of 1860-2100 with historical greenhouse gas concentrations and future concentrations following the IFCC Scenario 1992a (IS92a;Legget et al. 1992).The result from the regression was Y ϭ 1.11 Ϯ 0.02 W m Ϫ2 K Ϫ1 , which is much more precise than from the observations because the future radiative forcing and the resulting climate change are large, so the signal/noise is improved.Gregory et al. (2004) used the regression method of diagnosing climate sensitivity to compare its value in the coupled model (HadCM3) to that in the slab version of the same model.(The "slab" model uses a mixed layer ocean of 50-m depth with prescribed heat convergences rather than a 3D ocean GCM.Because the slab ocean heat capacity is relatively small, this kind of model is frequently used to diagnose equilibrium climate sensitivity.)They found the CO 2 climate sensitivity of the slab model to be slightly higher than in the first century of coupled model integrations, the latter being in agreement with our value above from the IS92a integration.If equilibrium and transient integrations give different Y values, the latter  is practically more relevant for projections of climate change in coming decades.
We then analyzed the years 1985-96 in an ensemble of four HadCM3 simulations of the twentieth century forced by both natural and anthropogenic factors (from Stott et al. 2000).The mean of the four Y values was 1.6 W m Ϫ2 K Ϫ1 and the standard deviation was 1.2 W m Ϫ2 K Ϫ1 , implying a standard error of estimate of 0.6 W m Ϫ2 K Ϫ1 on Y from the ensemble.The "true" value of 1.1 W m Ϫ2 K Ϫ1 lies within this interval of 1.6 Ϯ 0.6 W m Ϫ2 K Ϫ1 , suggesting that the observed data should similarly give an estimate of the climate sensitivity relevant to projections in the real world.Moreover, the ensemble standard deviation of 1.2 W m Ϫ2 K Ϫ1 was similar to the typical uncertainty on the slope estimated from the regression procedure in the individual integrations; the mean of the regression uncertainties was 1.1 W m Ϫ2 K Ϫ1 , suggesting that the regression uncertainty from the observed data should give a good estimate of the accuracy of our Y value.In the model, excluding data poleward of 60°from calculations of global radiation made only a small difference to the Y value (reducing it to 1.5 Ϯ 0.6 W m Ϫ2 K Ϫ1 ), because the surface albedo term is not a dominant component of the total climate feedback (Colman 2003), and in the model much of the surface albedo feedback term comes from seasonal snow cover on the northern continents south of 60°N.Further discussion of differences between model and observations is outside the scope of this study.

b. How good is the linear model of climate sensitivity?
As well as data errors, poor correlations could also arise if different responses from different forcings were being observed in the same dataset.This is certainly the case for 1985-96, as we include the Pinatubo response, knowing its climate feedback to be different (Fig. 4).However, excluding 1991-92 only decreased uncertainties in the 1985-96 derived Y values by ϳ20% (not shown), which indicates that this issue is unlikely to be a major source of uncertainty in the regression.The linear model appears to work particularly well in the LW, but not as well in the SW (Figs. 2 and 3).This is also found to be the case in models, where the poorer fit in the SW has been attributed to nonlinear cloud effects (Colman et al. 1997).The scanner data support this hypothesis, as they reveal a better straight-line fit for clear-sky Y SW values, compared to cloudy Y SW values.Another possibility is that internal variability in the climate system produces fluctuations in cloud cover, causing cloud forcings, especially in the shortwave, that will be seen in N and produce a lagged response in ⌬T s and cloud feedbacks.This mixture of forcing and feedback will degrade the correlation.Analysis of the HadCM3 control integration, which has no externally forced changes, supports this possibility.There is internally generated variability in TOA fluxes and ⌬T s ; the clearsky Y values are statistically consistent with the response to greenhouse gases, but the cloudy-sky response shows poor correlation.

Climate feedbacks
The Y values derived can be interpreted in terms of climate feedbacks.With no feedback, the earth behaves like a blackbody and we expect a Y NET ϭ Y LW response of roughly 3.3 W m Ϫ2 K Ϫ1 (Colman 2003).This Y value can be increased through negative feedbacks that make the climate response smaller [through Eq. ( 2)] or decreased by positive feedbacks that would make the climate response larger.For the long-term changes, a substantial positive SW feedback (Y SW Ͻ 0) and a small negative LW feedback combine to decrease Y NET and give an overall positive feedback of ϳ1.3 W m Ϫ2 K Ϫ1 .
Over the short 1985-90 time period, we appear to find a near neutral LW feedback (Y LW is close to 3.3 W m Ϫ2 K Ϫ1 ).This result is particularly interesting as most climate models exhibit a positive longwave feedback (Houghton et al. 2001).If our results are accurate, it could mean either that there is little or no positive water vapor feedback and a neutral cloud feedback or it could imply that the longwave cloud feedback is negative, offsetting the positive water vapor feedback.In contrast, during the Pinatubo years the longwave feedback is positive (ϳ2.3 W m Ϫ2 K Ϫ1 ).This value agrees both with estimates of the water vapor feedback diagnosed from observations over the same period and with that diagnosed from climate models (Soden et al. 2002;Forster and Collins 2004).Therefore, if the water vapor feedback is similar over the entire time period (which appears to be the case; see below), one might speculate that during the Pinatubo cooling the longwave cloud feedback is near to neutral, but over the full period it is negative.Differences in global feedbacks (and hence in Y ) for different forcing agents could plausibly be caused by the differing geographical and seasonal distribution of the forcings concerned.Volcanic forcing, for instance, is principally in the shortwave and hence greater when insolation is high, whereas CO 2 forcing is in the longwave.
To try to answer some of these questions we can extend this analysis to other climate feedbacks with the scanner data for the 1985-90 period.Note that this discussion is somewhat speculative as errors in the regressions for the scanner data are large and the conclusions are particularly dependent on the ability of the ERBE algorithms to unambiguously separate clear and cloudy scenes.This feedback analysis uses the cloud radiative forcing concept (Cess et al. 1996;Houghton et al. 2001).As a result of cloud masking of the noncloud feedback terms, feedbacks analyzed using cloud radiative forcing will be systematically different than those from partial radiative perturbations (Soden et al. 2004).Soden et al. found that in one climate model the cloud radiative forcing feedback was 0.3-0.4W m Ϫ2 K Ϫ1 more negative than the actual cloud feedback.Likewise the clear-sky feedback analyzed employing cloud radiative forcing would be more positive than the actual clear-sky feedback.

a. Water vapor and lapse rate
Models and observations suggest a positive feedback of around 1.7 Ϯ 0.4 W m Ϫ2 K Ϫ1 from water vapor and lapse rate changes (Colman 2003).When this number is subtracted from the blackbody Y LW value, a clear-sky Y LW of 1.6 W m Ϫ2 K Ϫ1 would result.Although it is one of the most uncertain of our Y retrievals, a similar number is seen for our clear-sky Y LW results (Fig. 6).

b. Surface albedo
Models have a positive clear-sky albedo feedback (Y SW Ͻ 0) of 0.4 Ϯ 0.2 W m Ϫ2 K Ϫ1 (Colman 2003).However, our clear-sky SW results indicate a negative clear-sky albedo feedback.Applying the Soden et al. (2004) correction to our clear-sky feedback could mean that the actual clear-sky feedback is more negative still.The negative feedback we find may be due to the ERBE instruments being unresponsive to changes in N caused by the cryosphere, as they do not measure latitudes higher than 60°.However, climate model results indicated that this was unlikely to be the problem (see section 5b).Examining the pattern of N variations revealed no obvious large-scale regional pattern associated with a possible mechanism.However, it is also possible that feedback from thin clouds (for example) is being misinterpreted as a clear-sky feedback by the ERBE algorithms.

c. Clouds
The NET cloud feedback appears close to neutral in the scanner data; it is made up of components due to a positive SW feedback and negative LW feedback of around 2.0 W m Ϫ2 K Ϫ1 each (Fig. 6).Soden et al. (2004) point out that this feedback is not representative of the actual cloud feedback term.However, we are still able to compare it to similar diagnostics from other models.One model out of the 10 models presented in Fig. 7.2 of the 2001 Intergovernmental Panel on Climate Change (IPCC) report had a similar cloud radiative forcing feedback (Houghton et al. 2001); other models behaved quite differently.As the clear-sky Y values from the scanner data (Fig. 6) resemble the total Y values for the Pinatubo response (Fig. 4), it is interesting to speculate that this may indicate that Pinatubo caused few cloud changes.

Conclusions
Data from the ERBS instruments appear sufficiently well calibrated to estimate the climate feedback parameter.Using the best estimate of radiative forcing and the GISS surface temperature dataset, a value for the climate feedback parameter is particularly well determined in the LW for 1985-90, where Y LW is 3.7 Ϯ 0.9 W m Ϫ2 K Ϫ1 .Overall, for 1985-96 the ERBE data suggest that Y NET is 2.3 Ϯ 1.3 W m Ϫ2 K Ϫ1 (Fig. 4).This indicates a net positive feedback, compared to the blackbody response.Uncertainties in specifying surface temperature and radiative forcing changes are examined and found to be a smaller source of error than quoted uncertainties in the regression analyses.Forcing uncertainty is estimated to vary Y NET by no more than ϳ0.5 W m Ϫ2 K Ϫ1 , and surface temperature change uncertainty varies Y NET by ϳ0.2 W m Ϫ2 K Ϫ1 .Our data gave an overall uncertainty range for Y NET of 2.3 Ϯ 1.4 W m Ϫ2 K Ϫ1 (not accounting for uncertainty in the ERBE time series, N ).
Our results show a net positive feedback relative to the blackbody response, which is consistent with climate models; they also hint at the possibility of nearneutral longwave feedback, which is inconsistent with most (if not all) climate models (Houghton et al. 2001;Colman 2003).However, as our longwave results are only well constrained for the 1985-90 period, it is probably premature to conclude that climate models are incorrectly modeling longwave feedbacks.
These results indicate a 1.0-4.1-Krange warming for a doubling of CO 2 (assuming Gaussian-distributed errors in the observable parameters and a radiative forcing for a doubling of CO 2 of 3.7 W m Ϫ2 ).Several stateof-the-art climate models exhibit climate sensitivities closer to the upper bound (Houghton et al. 2001), which would imply a Y NET ϭ 1.0 W m Ϫ2 K Ϫ1 .Simply inspecting the range of variations in N and ⌬T s in the figures shows that this value is extremely unlikely.It was noticed that climate models underestimate the ERBE decadal variability (Wielicki et al. 2002) in N.Even though this conclusion was based on data from the uncorrected ERBE analysis, our results suggest that this underestimate in variability could be because climate models tend to overestimate the equilibrium climate sensitivity.A climate model with a smaller sensi-tivity would produce larger variability in N, for the same variability in ⌬T s .
Patterns of surface and vertical temperature change and Hadley Centre model results indicate that the Y values found for the ERBE record are likely to be representative of that associated with longer-term climate change scenarios and the equilibrium climate sensitivity.The climate feedback from Mt. Pinatubo time period appears to be different.This means that the Y values derived may not necessarily be applicable to all forcing mechanisms, but it could also mean that short time-scale variability has different feedback properties.
The ERBE scanner record, although it only covers 5 yr, gives some intriguing, but somewhat uncertain, insights into the different components of the climate feedback parameter.In particular, it suggests a nearneutral cloud feedback (by offsetting a negative LW cloud feedback against a positive SW cloud feedback).It also suggests a positive surface albedo feedback.The scanning data show the potential benefits that could arise from 10 or more years of well-calibrated data of this kind.
We believe we have illustrated a powerful technique for determining linear components of climate feedback.Until now, it has been very difficult assess how realistic a given climate model's feedback is, as we lacked measures of distinguishing a "good" model from a "bad."Comparing these ERBE analyses with those from climate models would facilitate such judgments.compared to our 1.0-3.6-Krange quoted in the conclusions of the paper.

FIG. 1 .
FIG. 1.The 1985-97 globally averaged seasonal anomalies in (a) surface temperature (HADCRU, GISS) and midtropospheric temperature (MSU-TMT), (b) ERBE WFOV N data, and (c) well-mixed greenhouse gas and volcanic radiative forcing.The seasonal cycle has been removed from these datasets.ERBE N values are all measured positive downward, and 1993 data are not shown due to missing data in the ERBE record.

FIG. 2 .
FIG. 2. Data used to derive Y values from WFOV data.Shown are the SW, LW, and NET flux anomalies and the surface temperature anomalies (multiplied by 2 for scaling).GISS surface temperature and well-mixed greenhouse gas forcing datasets are employed.Data are for (top) 1985-90 annual data, (middle) 1985-96 annual data, and (bottom) and 1991-92 seasonal data.

FIG. 3 .
FIG. 3. Robust OLS linear regression of QϪN vs ⌬T s from the data shown in Fig. 2. A Y value, with 95% uncertainty, and the correlation of QϪN vs ⌬T s are also shown.GISS surface temperature and well-mixed greenhouse gas forcing datasets are employed.Regressions are for (top) 1985-90 annual data, (middle) 1985-96 annual data, and (bottom) 1991-92 seasonal data.
FIG.6.Same as in Fig.5, but from ERBE scanner data.Results are also shown for the clear-sky and cloudy values.Seasonal values derived for the GISS surface temperature datasets are used.

TABLE 2 .
The seasonally averaged Y SW , Y LW , and Y NET values derived from scanner and WFOV data for 1985-89.The table also shows the 95% uncertainty in Y and the correlation of QϪN vs ⌬T s .For the scanner data, Y is also separated into clear and cloudy scenes.The GISS surface temperature anomaly and well-mixed greenhouse gas forcing are employed.The results with the HADCRU surface temperature dataset and/or GISS radiative forcing dataset are very similar.
Values of Y SW , Y LW , and Y NET and their 95% uncertainties (vertical lines) from ERBE WFOV data.Annual values derived for the GISS surface temperature dataset are used.

TABLE 3 .
[Hansen et al. (2002)from WFOV data, using different radiative forcings.Two WMGHG forcings are used[Hansen et al. (2002): GISS and Myhre et al. (2001): MYHRE].The volcanic forcing from Sato et al. (1993; SATO) is also varied.The table also shows the 95% uncertainty in Y NET and the correlation of QϪN vs ⌬T s .The GISS surface temperature anomaly is used.Y NET values are shown as they are affected more by Q than either Y SW or Y LW , this is because Y NET is essentially as residual between a negative Y SW and positive Y LW .