Department of Atmospheric, Oceanic, and Space Sciences, University of Michigan, Ann Arbor, Michigan, USA
Corresponding author: J. E. Penner, Department of Atmospheric, Oceanic, and Space Sciences, Space Research Building, University of Michigan, 2455 Hayward St., Ann Arbor, MI 48109-2143, USA. (firstname.lastname@example.org)
 Satellite-based estimates for the aerosol indirect forcing are consistently smaller than those from models due, in part, to the use of present-day results that do not capture the temporal changes between present day and pre-industrial conditions. Here, we use results from a coupled aerosol-climate model to pick regions and seasons that are sufficiently pristine to represent pre-industrial conditions. Then we use results from MODIS and CERES to estimate the forcing between present and pre-industrial conditions. The estimated forcing in the North Pacific Ocean region ranges from −1.8 to −2.2 Wm−2 from observations. This range is similar to our modeled forcing for this region, −2.65 Wm−2, but is smaller than the modeled forcing using the same methodology as that used with the satellite observations, −3.6 W/m2. Nevertheless, a previous estimate based on satellite observations was a factor of 10 smaller, −0.2 to −0.5 Wm−2. Results demonstrate that while the estimated forcing from models may be somewhat larger than estimates based on satellite data, a judicious choice of analysis methods, yields results that are much closer than previous satellite and model-based comparisons.
 The first aerosol indirect effect, or Twomey effect, is measured by the change in top of the atmosphere (TOA) solar fluxes caused by the change in cloud droplet number when cloud liquid water path and other morphological changes are held constant. Satellite measurements of cloud droplet number concentration (Nc) vs aerosol optical depth (τa) have been used to estimate the effect of changes in Nc due to anthropogenic aerosols and range from −0.2 Wm−2 to −0.5 Wm−2 [Quaas et al., 2006, 2008]. In contrast, model results that rely on mechanistic descriptions of the relationship between aerosols and cloud drop number concentrations range from −0.5 Wm−2 to −1.85 Wm−2 but can be even larger, depending on how aerosol nucleation is treated in the model and whether a lower limit for droplet number concentrations is specified [Lohmann and Feichter, 2005; Wang and Penner, 2009].
Penner et al. [2011a]used a model study to show that satellite-based estimates that rely on either the spatial variation of ln(Nc) vs ln(τa) or ln(AI), where AI is the Aerosol Index (Ångström exponent times τa), rather than temporal variations induced by changes between the present (PD) and pre-industrial (PI) aerosol concentrations can significantly underestimate the Twomey effect.
 Here, we use model estimates of the average droplet number concentration and the regression between ln(Nc) and ln(AI) for both present-day and preindustrial results to choose regions for the present day that can act as a surrogate for pre-industrial conditions. Then satellite data are used to estimate the difference between present-day and preindustrial TOA fluxes. Unfortunately, among 14 regions examined, only the South Pacific Ocean (SPO) can be considered as a “pristine” surrogate for a few polluted regions, and then only during June, July, August (JJA) and December, January, February (DJF). When this region is used as a surrogate for PI conditions in DJF and JJA in the North Pacific Ocean (NPO), the TOA forcing from the satellite analysis is much closer to that estimated from the model.
 We used the same 14 regions defined in Quaas et al.  as well as data from all 4 seasons to estimate the predicted Nc and slope of the relationship between ln(Nc) and ln(AI) from PD and PI simulations. Our goal was to use the model results to find pristine regions and seasons in the Southern Hemisphere that could be used to describe the preindustrial values in the Northern Hemisphere. The response of Nc to increases in AI is expected to be a better indicator of aerosol/cloud interactions than is the response of Nc to τa [Nakajima et al., 2001], so we focused on using these variables. Wang et al.  have previously compared the predicted aerosol concentrations to data and Wang and Penner  compared the model with droplet concentrations. Quaas et al.  compared the slope of ln Nc with ln τa from the model with satellite observations.
 All PI slopes for ln Nc vs ln AI and mean values of Nc differed significantly from PD values, except for the SPO region, where the difference in Nc between NPO PI DJF and JJA values and SPO PD JJA and DJF was less than 8%. The values of the slopes of ln Nc vs ln AI between the PD SPO and PI NPO were within 3% and a factor of 2 for NPO JJA and DJF, respectively (Table 1). Average τa and AI are also similar as well as the composition of the particles forming the cloud condensation nuclei concentrations (see auxiliary material). Therefore, we judged this area as “pristine” and used SPO as a surrogate for pre-industrial conditions in NPO in the following. We note however that the slopes of lnNc vs ln τa differ in these regions, so in this sense the SPO is not “pristine”.
Table 1. Modeled Average Value of Droplet Number Concentrations, Aerosol Optical Depth (τa) and Index (AI), and Slope of the Change in Droplet Number With Aerosol Optical Depth or Aerosol Index in NPO and SPOa
Values are given for June, July, August (JJA) and December, January, February (DJF).
The values in parentheses are standard deviation of the cloud droplet number concentration.
The aerosol optical depth and aerosol index are separated by a comma. Standard deviations are in parentheses.
The values in parentheses are the correlation coefficient and the margin of error for the 95% confidence level for the regression slope.
0.13 (0.06), 0.12 (0.07)
0.40 (0.91, 0.06)
0.51 (0.92, 0.06)
0.11 (0.08), 0.08 (0.09)
0.33 (0.74, 0.09)
0.47 (0.95, 0.04)
0.07 (0.03), 0.03 (0.02)
−0.002 (−0.04, 0.04)
0.32 (0.89, 0.05)
0.07 (0.02), 0.02 (0.01)
−0.07 (−0.39, 0.05)
0.39 (0.94, 0.04)
0.07 (0.04), 0.04 (0.02)
0.15 (0.55, 0.07)
0.65 (0.98, 0.04)
0.06 (0.03), 0.02 (0.01)
−0.001 (−0.01, 0.03)
0.38 (0.84, 0.07)
Wang and Penner describe the off-line radiative transfer calculation usingNc, leading to changes in cloud albedo and forcing. We used saved values of Nc and aerosol concentrations from the coupled aerosol/climate model [Wang and Penner, 2010], rather than the previously described off-line calculation ofNc, since the effects of sedimentation, precipitation and coagulation are present in the satellite data, so should be present in the model as well. Here, as in Penner et al. [2011a], we used monthly average aerosol concentrations together with 4-hourly meteorological fields at a 2° × 2.5° resolution to estimate instantaneous aerosol optical depths and Ångström exponents as seen by satellites at the overpass time of 1:30 pm. Note that the use of hourly data does not alter our results significantly [Penner et al., 2011b]. Estimates of the cloud albedo change and forcing based on these inline quantities of Nc used PD cloud fields (water vapor, water path and cloud fraction) to eliminate any feedback to cloud liquid water path (LWP) or cloud fraction (f) caused by changes in aerosols in the coupled aerosol/general circulation model.
Figures 1a and 1bshow a plot of the all-sky albedo for JJA in NPO and DJF in SPO (a similar figure showing DJF in NPO and JJA in SPO is available in theauxiliary material), while Table 2shows average values for whole-sky, clear sky, and cloudy-sky albedos. The cloudy sky albedo is not directly available from satellite data, but can be estimated from
where α and αclrare the all-sky and clear sky albedos. There is clearly an increase in the all-sky NPO albedo compared to SPO, but this may be due to increases in cloud fraction, LWP, and aerosols in clear sky regions. Therefore, to avoid contamination by these factors, we first sorted the data using only those data withf > 99% (shown in Figures 1c and 1d), but also provide estimates for f > 50% (see Table 2). Then these albedo data were compiled into 10 LWP bins, evenly spaced in ln(LWP) intervals of 0.3 from 3 to 6 (with LWP in gm−2). A test using up to 30 bins gave results similar to those reported here. Figure 2 shows a plot of ln (albedo) vs LWP for f > 99% together with the standard deviation of albedo within each LWP bin for NPO in JJA and SPO in DJF (Figure 2a) as well as for NPO in DJF and SPO in JJA (Figure 2b). The relative increase (%) for each LWP bin is also shown. While albedo clearly increases with LWP (Figures 2a and 2b) there is not any systematic difference in the percentage change between NPO and SPO with albedo (Figures 2c and 2d). On average, we find a change in albedo of 1.5% between JJA NPO and DJF SPO, whereas the increase is 4.4% in NPO DJF. The albedo change is 1.6% in NPO JJA and 2.7% in DJF if we use all data with f > 50%.
Table 2. Satellite and Model Albedos for All-Sky, Clear Sky and Cloudy Sky and for Liquid Clouds Using f > 99% and f > 50%
Cloudy sky albedos are not directly available from satellite observations, but calculated from equation (1) summed over all grid boxes. Two values appear in the table. The first estimate was obtained by first averaging all values for a given spatial location over time, and then forming the spatial average, while the second was obtained by treating all values as equal when forming the average.
The percentage change is based on the average percentage change for the 2 methods of averaging described in footnote a.
 The observation-based forcing was calculated from the reflected solar fluxes in cloudy skies:
 Here, the first term is the all sky flux, while the second is the clear sky flux. The summation over x,y is taken over all grid boxes in a region. S0,x,y is the incoming solar radiation over each grid box, fx,y is the fraction of each grid box covered by clouds, αx,y is the observed all sky albedo in each grid box, αclr,x,y is the clear sky albedo, ΔαNPO-SPO is the average fractional change in albedo, and Ax,y is the area of the grid box.
Table 3summarizes the observation-based forcing calculated usingequation (2) from the satellite and the model. For model calculations based on equation (2), we replaced ΔαNPO-SPO with the average fractional change in cloudy sky albedo between PD and PI conditions in the NPO, but all other values were based on the observations. The estimated annual average forcing (average of JJA and DJF) for the observations ranges from −1.78 to −2.24 Wm−2 while that for the model ranges from −3.53 to −3.55 Wm−2. The JJA results from the model are a factor of 2.3 to 2.7 higher than the observations, but the DJF model results are a factor of 1.6 larger (for f > 50%) to 13% smaller (for f > 99%). The primary reason for the model overestimate in JJA is the much larger change in τa between PD and PI results than that in the observations between NPO and SPO (see discussion in auxiliary material and Table S2).
Table 3. Estimates of Forcing (Wm−2) in the NPO Region Between PD and PI From Satellite Observations and Model Results
JJA, f > 50%
JJA, f > 99%
DJF, f > 50%
DJF, f > 99%
The model results in equation (2) use the regional average fractional change in albedo from the modeled PD and PI results for Nc together with the outgoing SW fluxes from the observations.
The model-based forcing from PD-PI refers to the results calculated with the off-line radiative transfer model using the inline calculationNc as in Penner et al. [2011a]. All cloud fractions are included.
 The model results based on equation (2) are on average larger than those from the satellite data, but they are much closer than the factor of 10 difference between the full model PD and PI estimates, −2.65 Wm−2 (shown as the last row in Table 3) and the values of −0.2 ∼ −0.5 Wm−2 estimated for this region from satellites using methods that do not try to calibrate to “pristine” regions for PI estimates [Quaas et al., 2008, Figure 5b]. Thus, we show that when applied in a judicious manner, satellite data give results for the first indirect forcing that are of the same order of magnitude as the results from a model.
 We have shown previously that the use of satellite data for present day without consideration of temporal variations is subject to large errors [Penner et al., 2011a, 2011b]. Here, we used data from the “pristine” South Pacific Ocean as a surrogate for pre-industrial conditions. When we do so, the satellite estimates of forcing for the NPO range from −1.8 to −2.2 Wm−2 while the corresponding results from the model using the same method of analysis range from −3.5 to −3.6 Wm−2. Perhaps surprisingly, results from the full PD-PI off-line calculated forcing are even closer to the satellite results, −2.65 Wm−2.
 To explain this result, we examined observed and modeled cloudy sky albedos and cloud fractions (Table S3 in the auxiliary material). The modeled cloud albedos are larger than the observed albedos, while the cloud fraction is much smaller. In JJA, the forcing based on the satellite observations is smaller than the modeled (PD-PI) forcing since the observed cloud fraction is only 7% larger, and the modeledτa difference is significantly larger than that in the observations (Table S2 in the auxiliary material). However, in DJF, the forcing from satellites is larger for the f > 99% case than the PD-PI modeled value, and similar for the f > 50% case, since the observed cloud fraction is 37% larger, even though theτa difference is similar (Table S3 in the auxiliary material). Thus, differences in LWP (as expressed in cloud albedos), cloud fraction, and aerosol optical depths lead to compensating differences that combine to produce a modeled PD and PI forcing that is close to that estimated from the observations.
 Our estimates for aerosol indirect forcing from observations are subject to several assumptions. First, we assumed that the SPO region could be used as a surrogate for PI conditions in the NPO region. Such an assumption has been made since the earliest attempts to gauge the effects of aerosols on clouds [e.g., Schwartz, 1988]. However, we tested the use of the SPO region by comparing model results for PI cloud and aerosol properties in the NPO with those from the PD SPO. Second, we assumed that ice clouds did not mask the change in outgoing shortwave radiation caused by changes in warm-phase clouds. This assumption can be tested by comparing the forcing in the model with and without ice-clouds included. When we do so, the estimated forcing is reduced by about 10%. Also, we assumed that the meteorological conditions leading to a change in albedo at a fixed LWP were similar in the NPO and SPO. Cloud resolving model studies have shown that the response of clouds to increasing aerosols depends in large part on the meteorological conditions or the specified large-scale forcing [e.g.,Ackerman et al., 2004; Guo et al., 2007]. But these model results include both the first indirect forcing (or albedo effect) as well as the response of LWP and cloud fraction to changes in aerosols. Our methodology avoids including these responses. Finally we assumed that the fractional increase in cloud albedo for liquid water clouds is the same for all cloud fractions. This is perhaps the most limiting assumption in our study. Our study selects conditions that only include clouds that cover at least 50% of a 1° × 1° grid, but estimates the forcing assuming that the fractional response of the cloud albedo for the first indirect effect for all cloud fractions is similar. This assumption should be approximately followed for the initial cloud droplet concentrations as long as the updrafts that determine the initial cloud drop change do not depend on the meteorological conditions that determine the cloud fraction, and as long as the mixing of cloud air with clear air is similar for all cloud fractions (i.e. the clouds are heterogeneously mixed to the same degree for all cloud fractions). Both assumptions can call our observational estimates into question. Indeed, since global model estimates of albedo effects often make these same assumptions, it is perhaps not surprising that the results from the model and that from observations using our method are similar. Still, it is comforting that if one adds such qualifications to the observational analysis, the results of the model and the observations are within a factor of two.
 The NCAR CISL provided computer time for this study. We thank both reviewers for suggestions that improved the manuscript.
 The Editor thanks two anonymous reviewers for their assistance in evaluating this paper.