Accurate portrayal of the aerosol characteristics is crucial to determine aerosol contribution to the Earth's radiation budget. We employ novel satellite retrievals to make a new measurement‐based estimate of the shortwave direct radiative effect of aerosols (DREA), both over land and ocean. Global satellite measurements of aerosol optical depth, single‐scattering albedo (SSA), and phase function from PARASOL (Polarization and Anisotropy of Reflectances for Atmospheric Sciences coupled with Observations from a Lidar) are used in synergy with OMI (Ozone Monitoring Instrument) SSA. Aerosol information is combined with land‐surface bidirectional reflectance distribution function and cloud characteristics from MODIS (Moderate Resolution Imaging Spectroradiometer) satellite products. Eventual gaps in observations are filled with the state‐of‐the‐art global aerosol model ECHAM5‐HAM2. It is found that our estimate of DREA is largely insensitive to model choice. Radiative transfer calculations show that DREA at top‐of‐atmosphere is −4.6 ± 1.5 W/m 2 for cloud‐free and −2.1 ± 0.7 W/m 2 for all‐sky conditions, during year 2006. These fluxes are consistent with, albeit generally less negative over ocean than, former assessments. Unlike previous studies, our estimate is constrained by retrievals of global coverage SSA, which may justify different DREA values. Remarkable consistency is found in comparison with DREA based on CERES (Clouds and the Earth's Radiant Energy System) and MODIS observations.

1 Introduction Aerosol particles impact the Earth's radiation budget directly through scattering and absorption of radiation in the atmosphere. Scattering increases the planetary albedo, causing a negative forcing that cools the climate system. In contrast, absorption reduces the outgoing shortwave (SW) flux at top of atmosphere (TOA), leading to a positive forcing, which promotes temperature rise. The two effects are not in balance, and the net globally averaged radiative contribution by aerosols is negative [Intergovernmental Panel on Climate Change (IPCC), 2013]. Aerosol particles also influence the radiation balance indirectly through their interactions with clouds. These include perturbation of the albedo of clouds and their lifetime [Twomey, 1977; Albrecht, 1989], whose representation in climate models is challenging [e.g., Quaas et al., 2008; Lacagnina and Selten, 2014; Lacagnina et al., 2014]. Since some aerosols and aerosol precursor gases are emitted by human activities, part of the total aerosol effect is anthropogenic, referred to as forcing [Lohmann and Feichter, 2005; IPCC, 2013]. Estimates of both direct and indirect forcings are hampered by large uncertainties and make the largest contribution to the overall uncertainty in the radiative forcing causing climate change [Myhre et al., 2013]. In this study we focus on the direct radiative effect (natural + anthropogenic) of aerosols (DREA) in the SW spectral range. DREA depends on a number of factors, such as aerosol vertical placement with respect to clouds, surface albedo, and aerosol optical properties. Among the latter, the most important for radiative transfer calculations are aerosol optical depth (AOD), single scattering albedo (SSA), and, to a lesser extent, phase function. The columnar AOD quantifies the total extinction of a light beam interacting with aerosols and is related to their amount. The SSA is the ratio of the scattering to the total extinction and measures the degree of particle light absorption. Generally, purely scattering aerosols (e.g., sea salt) enhance the backscattering of solar radiation to space, resulting in negative DREA, i.e., a cooling effect. For absorbing aerosols (e.g., black carbon) a decrease of SSA can change the sign of DREA from negative (cooling) to positive (warming), depending on the albedo of the underlying surface [Liao and Seinfeld, 1998; Russell et al., 2002]. Finally, the phase function describes the angular distribution of scattered radiation, indicating how much energy is scattered backward or forward by particles. Two approaches are usually employed to estimate DREA. The first approach is model based: aerosol properties needed for radiative transfer calculations are given by global aerosol models' simulations [e.g., Schulz et al., 2006; Ma et al., 2012]. These estimates are biased by uncertainties in parameterizations of a variety of subgrid aerosol processes [Kinne et al., 2006] and by host model errors in simulating clouds, surface albedo, and radiation [Stier et al., 2013; Randles et al., 2013]. The second approach is measurement based: observations provide the basis for the assessment, with minor inputs from models. In many cases AOD is taken from satellite retrievals, while SSA and asymmetry parameter are from global aerosol models or sparse ground‐based measurements [e.g., Chung et al., 2005; Bellouin et al., 2008; Myhre, 2009]. Eventual observational gaps, due to algorithm limitations or scarcity of ground stations, are supplemented with model information. Uncertainties in this method arise from cloud contamination, instrument‐related errors, and retrieval assumptions [Yu et al., 2006]. DREA from both the model‐based and measurement‐based methods is associated with large uncertainties, because of inaccuracy of basic parameters. Among these, SSA accounts for much of this uncertainty [Myhre, 2009; Loeb and Su, 2010]. At present, two main satellite products provide quantitative global coverage SSA. One product [Hasekamp et al., 2011; Waquet et al., 2016] is based on multiangle multispectral photopolarimetric measurements from the POlarization and Directionality of Earth Reflectances (POLDER‐3) instrument onboard the PARASOL (Polarization and Anisotropy of Reflectances for Atmospheric Sciences coupled with Observations from a Lidar) microsatellite [Fougnie et al., 2007]. A second remote‐sensing data set is built on the Ozone Monitoring Instrument (OMI) on board NASA's Aura spacecraft [Torres et al., 2007, 2013]. The two sensors capture aerosol characteristics in different, but complementary, spectral bands: visible and near infrared in the former and ultraviolet (UV) in the latter. We rely on aerosol information from both instruments to make a new measurement‐based estimate of SW DREA over land and ocean. With respect to previous bottom‐up assessments [e.g., Yu et al., 2006; Myhre, 2009], our estimate is more constrained by observations, because aerosol phase function and SSA are taken from satellite retrievals, together with AOD. Moreover, land‐surface reflectance and cloud characteristics are from MODIS [MODerate resolution Imaging Spectroradiometer, Platnick et al., 2003] satellite products. Eventual gaps in observations (summarized in section 2) are filled with the global aerosol model ECHAM5‐HAM2 [Zhang et al., 2012], as described in section 3. Any (bottom‐up) measurement‐based method relies, to some extent, on some model simulations, and pure estimates from observational data are currently not possible [Myhre, 2009]. Here we take a few steps toward an increasingly measurement‐based approach. Resulting estimates of DREA and related uncertainties are discussed in section 5. Concluding remarks and summary of results are presented in section 6.

3 Methodology Shortwave (SW, here considered from 0.25 to 4.00 μm) DREA at TOA may be expressed as the difference in net radiative flux F (downward minus upward) in presence (aer) and absence (noaer) of aerosols [Matus et al., 2015]: (1) (2) where the subscripts “cloud‐free” and “cloudy‐sky” indicate fluxes in case of absence of clouds and in case of completely cloudy sky, respectively. The sign convention is such that negative values of DREA indicate energy leaving the climate system due to aerosols, i.e., cooling effect. The DREA cloud‐free and cloudy‐sky is computed with a combination of POLDER and OMI‐derived aerosol parameters as input (see section 3.1). Outside the observational bands, SW wavelength dependence of aerosol properties is obtained from the ECHAM5‐HAM2 model (more details in section 3.1). Moreover, because of retrieval algorithm limitations, observational gaps may occur, e.g., under semipermanent cloudy conditions and snow and ice‐covered surfaces. In these situations ECHAM5‐HAM2 outputs are combined with our data set, following the procedure in section 3.2. 3.1 Spectral Variability As input to our flux calculations, we use the spectral AOD, SSA, and phase function. The latter is derived through aerosol microphysical properties, using the Mie/T‐matrix approach [Dubovik et al., 2006]. Both microphysical and optical parameters are provided by PARASOL products within the spectral range, i.e., 0.5 to 1 μm. Outside this range, wavelength variability of aerosol properties is obtained partly integrating OMI observations and partly relying on model guidance, as follows. We replace PARASOL SSA with OMI SSA in the spectral range where it is available, i.e., 0.35 to 0.50 μm. This approach has a major advantage: it reduces the uncertainties due to extending optical properties outside the PARASOL retrievals wavelength range (see section 2). When gaps in OMI data occur, PARASOL is used; when data are not available from both sensors, grid boxes are left empty. The distribution of the number of observations from PARASOL and OMI SSAs is shown in Figure 9. As far as AOD is concerned, we use PARASOL data only, because they are more accurate than OMI (see section 2). Figure 9 Open in figure viewerPowerPoint Number of observations of SSA from (a) PARASOL and (b) OMI for year 2006. Outside the satellite spectral bands, aerosol optical properties are extended to the remaining SW bands, using the variability supplied by ECHAM5‐HAM2, adjusted by a scaling factor. The scaling factor is obtained, dividing observed and modeled AOD, scattering AOD (SAOD), and absorption AOD (AAOD) at the last coincident wavelength (SAOD and AAOD are needed to derive SSA). The scaling factor is then multiplied by the appropriate modeled optical property at the wavelengths outside the observational spectral range. As a result, the spectral variability of AOD and SSA follows the model simulation, but the magnitude of these quantities is consistent with observations. The other aerosol property, which is wavelength dependent, is the refractive index, needed to compute the phase function. The procedure is similar to the one outlined in the previous paragraph, taking into account that refractive index from PARASOL is retrieved as constant over the satellite bands from 0.49 to 1.02 μm. Therefore, before using the modeled refractive index, this is spectrally averaged in the observational range 0.49 to 1.02 μm. Refractive index and the other microphysical properties are not directly available in ECHAM5‐HAM2 outputs, requiring additional postprocessing, described in Appendix A. 3.2 Filling Observational Gaps As mentioned above, gaps in observations may occur in unfavorable retrieval conditions. In these situations, it is a standard procedure to fill observational gaps with modeling outputs [e.g., Chung et al., 2005]. These offer continuous and consistent maps, which have the advantage of reducing sampling issues. Indeed, adjacent grid boxes can be sampled differently from satellites, because of observational gaps, leading to some regions being oversampled compared to others [e.g., Schutgens et al., 2016]. We use the ECHAM5‐HAM2 model to fill observational gaps of the following optical parameters: SAOD, AAOD (needed to derive SSA), AOD, and microphysical parameters: column number, effective radius, effective variance, refractive index, and fraction of spherical particles. These microphysical variables are needed to derive the phase function. Following Chung et al. [2005], at each 3° × 2° grid box scaling ratios are calculated by dividing monthly mean observational parameters over the appropriate modeling parameters. The resulting scaling ratios have grid boxes with missing values (gaps) as in the related observational maps. These gaps are filled by interpolating over the neighboring grid boxes with existing values, weighting their contribution according to the distance from the missing value. The new filled scaling ratios are multiplied by the ECHAM5‐HAM2 data. The resulting scaled values are used to replace the original observational data only where these contain gaps. Finally, the vertical distribution of aerosols is not available from the passive sensor products used in this study. Therefore, it is taken from ECHAM5‐HAM2, which is considered one of the best models in simulating profiles of aerosol extinction, compared to Cloud‐Aerosol Lidar with Orthogonal Polarization observations [Koffi et al., 2012]. At each model layer, fractional contributions of total column midvisible AOD are computed from ECHAM5‐HAM2. These fractional contributions are then multiplied by the total column midvisible AOD from PARASOL. The resulting aerosol altitude distribution is input to a radiative transfer model, described in section 4. 3.3 Cloud Effects The coexistence of aerosols and clouds in the same atmospheric column makes aerosol interactions with sunlight more complex. Compared to cloud‐free events, aerosols intercept less (more) solar radiation when they are below (above) cloud layers [e.g., Matus et al., 2015]. Therefore, clouds significantly modulate DREA, and their effects need to be taken into account in radiative transfer calculations. Here clouds are described by droplet effective radius, effective variance (a measure of the width of droplet size distribution [King et al., 1998]), and cloud optical thickness from MODIS. Wavelength‐dependent refractive index of water is taken from Segelstein [1981]. Single‐layer clouds are inserted into the atmospheric column of a radiative transfer model (described in section 4) at the level closest to the MODIS cloud top pressure. Clouds with tops warmer than 263 K are assumed to be liquid, as in Zelinka et al. [2012], and radiatively treated through Mie theory. Clouds with tops colder than 263 K are treated as ice clouds, following the parameterization of Fu [1996]. Since only monthly mean cloud properties are used to compute DREA, possible influences of the diurnal variability of cloud properties are neglected. This is a source of uncertainty for cloudy‐sky DREA calculations. Indeed, sensitivity studies by Min and Zhang [2014] show that this approximation leads to systematic biases in DREA estimates. A typical 10–20% of errors result when a constant cloud fraction is assumed. Quantification of radiative effects of aerosols in presence of clouds is considered when grid boxes contain both clouds and aerosols at monthly mean temporal scale. We do not use simultaneous colocated retrievals of both clouds and aerosols to estimate cloudy‐sky DREA, because aerosol properties used in this study are retrieved only in cloud‐free scenes. This represents a limitation and an additional source of uncertainty in computing cloudy‐sky DREA, compared to its cloud‐free counterpart. The detection and retrieval of aerosol properties above clouds is a recent line of research. Retrieval techniques applied to POLDER polarization measurements [Waquet et al., 2009] and radiometric observations at near UV [Torres et al., 2012] and visible [Jethva et al., 2013] wavelengths have been developed. These and other recent applications [de Graaf et al., 2012; Waquet et al., 2013; Peers et al., 2015; Feng and Christopher, 2015] document novel approaches to extract information of aerosols above clouds, offering the potential of reducing uncertainties in cloudy‐sky DREA in the near future.

4 Radiative Transfer Model We adapted the linearized vector radiative transfer (RT) model developed by Hasekamp and Landgraf [2005] to calculate SW fluxes over 30 bands, ranging from 0.25 to 4.00 μm. The radiation field is discretized in eight Gaussian streams, and the phase function is expanded in terms of Legendre moments, truncated with the delta‐M method [Wiscombe, 1977]. Molecular absorption parameters are from the HITRAN‐2012 database [Gordon et al., 2013] and are treated with the linear‐k method [Hasekamp and Butz, 2008]. The most important radiatively active gases are considered, such as CO 2 , CH 4 , and O 2 (set to constants) and water vapor and ozone. Monthly mean profiles of the latter two gases are taken from European Centre for Medium‐Range Weather Forecasts (ECMWF) ERA‐Interim reanalysis [Dee et al., 2011] for year 2006, along with monthly mean temperature and pressure profiles. Surface reflectance characterization is described in section 2.5. Radiative properties of aerosols are obtained as explained in section 3, and global means of AOD and SSA are reported in Table 2. Table 2. Global (60°S–60°N) Annual Mean (Weighted by Area) Aerosol Optical Depths and Single‐Scattering Albedos at 0.55 μm From PARASOL With Monthly Observational Gaps Filled 0.55 μm Global Ocean Land AOD 0.18 0.14 0.29 SSA 0.94 0.94 0.93 Finally, solar spectral irradiance is taken from Gueymard [2004], adjusted for variations in the eccentricity of the Earth's orbit. We restrict radiation calculations to the central day of every month. Consistently, the daily cycle is captured by using 24 different solar zenith angles (SZAs), according to latitude and the fifteenth day of the month. The radiation scheme is run with a horizontal resolution of 3° × 2° and 15 vertical layers, of which seven in the troposphere. Clouds are taken into account in radiative transfer calculations as follows. In order to compute cloud‐free DREA, the RT model is ran without clouds (cloud free, i.e., cloud optical thickness set to zero throughout the vertical layers), once in absence of aerosols and once in presence of aerosols (equation 1). The cloudy‐sky DREA needs two additional runs with a completely cloudy sky (i.e., cloud covers 100% of the area of a grid box) with and without aerosols (equation 2). The cloudy‐sky and the cloud‐free radiative effects are combined to form all‐sky DREA as follows: (3) where cf stands for monthly mean cloud fraction from MODIS. As far as evaluation of the RT model is concerned, results are examined against benchmark fluxes and DREA reported in a recent radiative transfer code intercomparison study by Randles et al. [2013]. We carried out experiments following the protocol laid out by Randles et al. [2013], as outlined in the caption of Table 3. The values published in their study are reported in Table 3, along with our results. It is shown that simulations from the RT model used here are in close agreement with the other RT scheme results. Discrepancies are mostly caused by differences in assumed spectral molecular absorption. Table 3. Radiative Fluxes (W/m2) for the SW Broadband for the Different Case Experiments Reported in Randles et al. [ 2013 a SZA Case I Case II Case III Randles et al. [ 2013 30° 204.7 ± 2.7 (210.1 ± 5.5) −8.2 ± 0.1 (−9.8 ± 1.4) 10.3 ± 0.2 (9.0 ± 1.4) 75° 75.2 ± 1.4 (77.8 ± 1.6) −18.0 ± 0.3 (−16.7 ± 1.5) −6.5 ± 0.4 (−5.7 ± 0.7) This study 30° 208.0 −7.5 11.7 75° 76.8 −16.9 −5.8 An additional case accounting for cloud radiation interactions is treated in the intercomparison study of Halthore et al. [2005]. They refer to this experiment as case IV. It consists of a cloudy layer of prescribed properties located in an aerosol‐free tropical standard atmosphere. Following the protocol laid out by Halthore et al. [2005], our experiments yield 745.1 W/m2 at SZA = 30° and 231.1 W/m2 at SZA = 75°, which are very close to the values reported in Halthore et al. [2005]: 746.7 ± 5.7 W/m2 and 231.0 ± 2.5 W/m2, respectively. Given the sufficient agreement with published results from other RT models, we regard the RT scheme used in this study as appropriate for DREA estimates.

5 Shortwave DREA Estimates Figure 10 shows our observational‐based estimate of the DREA at TOA for both cloud‐free (equation 1) and all‐sky (equation 3) conditions. Because of weak sunlight and high surface reflectance, PARASOL and OMI provide very few retrievals at high latitudes. As a consequence, the global assessments are confined to areas between 60°N and 60°S. Most of previous measurement‐based estimates were performed within these latitudes [Yu et al., 2006]. Figure 10 Open in figure viewerPowerPoint Annual mean distribution of various quantities for year 2006. DREA at TOA in (a) cloud‐free and (c) all‐sky conditions. Global (60°S–60°N) means (weighted by area) are displayed in parentheses. (b) AAOD at 0.55 μm from PARASOL with observational gaps filled with ECHAM5‐HAM2 and (d) total cloud fraction from MODIS. Figure 10a shows that aerosols enhance the planetary albedo through scattering the solar radiation in most of the globe, specifically over dark surfaces (e.g., ocean and vegetated land). The overall effect on the Earth's radiative balance is to cool the planet by −4.6 W/m2. Strongest negative values are found in regions of high aerosol load (Figure 3b), such as China, India, and central Africa. These areas are governed by a mix of aerosol types (e.g., dust, biomass burning, and pollution) [Duncan et al., 2003; Yang et al., 2013; Eck et al., 2005], with variable absorption properties (Figures 3a and 10b). Over the oceans, the central Atlantic experiences negative DREA (Figure 10a), due to a mixture of desert dust from Sahara and carbonaceous particles from biomass burning in the Sahel region [Kaufman et al., 2005]. High sea salt production in the ocean belt surrounding Antarctica is responsible for peaks of negative DREA at the high latitudes of the Southern Hemisphere. In contrast, other regions reflect less solar radiation to space: positive DREA can be found at high latitudes close to 60°N and over the Himalaya, most likely due to snow and its strong surface albedo. Geographical patterns of cloud‐free DREA (Figure 10a) are consistent with previous estimates [Yu et al., 2006; Kinne et al., 2013; Su et al., 2013; Matus et al., 2015]. However, discrepancies arise at regional scale. For instance, present estimate lacks of positive values over the Saharan region, compared to Kinne et al. [2013] and Matus et al. [2015], while negative values are shown in Su et al. [2013] and, to a lesser extent, Yu et al. [2006]. This may in part be attributed to diversities in aerosol absorption data sets. Indeed, Kinne et al. [2013] show much stronger AAOD (see their Figure 8) over Sahara compared to our data set (Figure 10b). A similar discrepancy, but of opposite sign, is found over central Africa. In this region, our estimate of cloud‐free DREA is more negative than in Kinne et al. [2013] and Matus et al. [2015], but it is similar to estimates by Su et al. [2013]. Compared to Kinne et al. [2013], aerosols in the present study exhibit more absorption in this area (Figure 10b). We stress that, unlike previous studies, here aerosol absorptivity is taken from global satellite retrievals, and it is not based on model assumptions. This can explain part of the discrepancies in regional DREA. Figure 10c shows DREA in all‐sky conditions. It is less negative than its cloud‐free counterpart (Figure 10a). This is because particles below clouds receive less sunlight than in cloud‐free conditions, weakening their effect on radiative fluxes. In addition, absorbing aerosols above cloud layers reduce the amount of light reflected by clouds, leading to positive effect. As a result, DREA shifts from negative to positive in some areas, and the globally averaged aerosol cooling reduces to −2.1 W/m2. It is worth noticing a hot spot of positive all‐sky DREA over Amazon (Figure 10c). This area exhibits high cloud fraction (Figure 10d), but it is usually not affected by strong events of aerosols above clouds [e.g., Waquet et al., 2013]. In these situations, all‐sky DREA is expected to become less negative than its cloud‐free counterpart, but it should not shift from negative to positive values. Positive all‐sky DREA in our estimate arises from model biases. Indeed, the ECHAM5‐HAM2 model here provides, among other things, vertical distribution of aerosols. Over Amazon, this model tends to simulate higher than observed aerosol extinction in midtroposphere [Koffi et al., 2012], leading to unrealistically high aerosol load above clouds. As a result, a considerable amount of absorptive aerosols (Figure 10b) is present above clouds over Amazon, leading all‐sky DREA to unexpected positive values. Solid lines in Figure 11a show the monthly variability of cloud‐free DREA, along with the relative contribution by Northern Hemisphere (NH) and Southern Hemisphere (SH). DREA is more prominent in NH, which experiences strongest cooling during March‐April‐May (MAM) and June‐July‐August (JJA) seasons, only partially compensated by peaks of less negative DREA in SH. Little differences between NH and SH occur during September‐October‐November (SON). This is the result of high (low) aerosol load in the NH (SH) (Figure 11b). Similar behavior, but smaller in magnitude, characterizes DREA in all‐sky conditions (dashed lines in Figure 11a). Comparison of Figures 11a and 11b shows that AOD is a leading factor modulating DREA, but the effects of aerosol absorption (seen through the SSA) partly counterbalance the effects of aerosol loading (seen through the AOD). For instance, AOD and coalbedo (1‐SSA) in SH are the lowest in May and increase during the other months, while SH DREA is almost constant from May to July. This is the result of high scattering in May associated with low AOD and lower scattering in July accompanied by higher aerosol load. In the NH, SSA is the lowest during December‐January‐February, due to biomass burning in India and Sahel [Duncan et al., 2003; Yu et al., 2010; Lacagnina et al., 2015], associated with low values of AOD. As a result, the DREA is less negative during boreal winter. Figure 11 Open in figure viewerPowerPoint (a) Monthly DREA at TOA for year 2006 in cloud‐free (solid lines) and all‐sky (dashed lines) conditions, over the Northern Hemisphere (NH, 0°N–60°N), Southern Hemisphere (SH, 0°S–60°S), and globally averaged (60°S–60°N). (b) Monthly AOD (solid lines) and SSA (dashed lines) at 0.55 μm for year 2006 from PARASOL. Averages are weighted by area. Global distribution maps of seasonal cloud‐free DREA (Figure 12) show that large aerosol cooling in JJA is associated with aerosols downwind of the Sahara and the Arabian peninsula over the ocean, that is primarily mineral dust leaving deserts [e.g., Kaufman et al., 2005]. These particles are characterized by high scattering at midvisible wavelengths [Kim et al., 2011]. During boreal spring and summer, Caribbean islands are known to be influenced by dust from the Sahara [Kaufman et al., 2005], confirmed by DREA as large as −10 W/m2 extending westward to Central America. During SON, the strongest cooling is detected over the Indonesian area, because of extensive fires contributing to much of the aerosol emissions [Gras et al., 1999]. Indonesian smoke is significantly less absorbing than smoke from Australia or Africa [Gras et al., 1999; Sayer et al., 2014; Lacagnina et al., 2015]. As a result DREA exceeds −20 W/m2 locally. During MAM, the North Pacific is affected by remote aerosol events (Figure 12b). In this season mineral dust mixes with the local anthropogenic pollution in Northeast Asia [Eck et al., 2005], and meteorological patterns extend aerosol influence to the adjacent ocean. Figure 12 Open in figure viewerPowerPoint Seasonally averaged DREA at TOA for year 2006 in cloud‐free conditions. Global (60°S–60°N) means (weighted by area) are displayed in parentheses. Boreal seasons are (a) winter, (b) spring, (c) summer, and (d) autumn. These findings, both in terms of spatial and seasonal patterns of DREA, are consistent with previous measurement‐based estimates. For instance, see for comparison Loeb and Manalo‐Smith [2005] and Patadia et al. [2008], who used completely different satellite observations to derive DREA over ocean and over land, respectively. Global (60°S–60°N) estimates from these studies and various others are summarized in Figure 13. Discrepancies among the assessments are partly related to interannual variability and differences in cloud screening methods [Yu et al., 2006]. Figure 13 Open in figure viewerPowerPoint Yu et al. [ 2006 Loeb and Manalo‐Smith, 2005 Patadia et al., 2008 Su et al., 2013 Kinne et al., 2013 Matus et al., 2015 Boucher and Tanré, 2000 Bellouin et al., 2003 Global (60°S–60°N) annual mean of DREA at TOA computed in this study (black boxes) compared with measurement‐based estimates from various assessments:] (red), CERES/MODIS (green) [], MISR/CERES/MODIS (blue) [], MODIS/MATCH (cyan) [], MAC‐v1 (purple) [], CloudSat/CALIPSO (yellow) [], POLDER (gray) [], and POLDER/AERONET (orange) []. All averages are weighted by area. The symbol (*) indicates averages between 90°S and 90°N. Yu et al. [2006] reported the median of several measurement‐based approaches; Loeb and Manalo‐Smith [2005] (Clouds and the Earth's Radiant Energy System (CERES)/MODIS) and Patadia et al. [2008] (Multiangle Imaging Spectroradiometer (MISR)/CERES/MODIS) used a combination of different satellite data; Su et al. [2013] (MODIS/(MATCH) Model for Atmospheric Transport and Chemistry) combined reanalysis model data with AOD from MODIS; Kinne et al. [2013] used an aerosol climatology (MAC‐v1) based on AERONET measurements merged onto multimodel ensemble maps; Matus et al. [2015] used active remote sensing data from CloudSat and CALIPSO. Moreover, Boucher and Tanré [2000] (POLDER) and Bellouin et al. [2003] (POLDER/AERONET) estimated DREA based on AOD from POLDER. In these studies, they used a similar sensor to POLDER‐3 on board the PARASOL satellite, but the retrieval algorithm was not able to provide aerosol absorption properties. As a consequence, SSA and asymmetry parameter were provided by models [Boucher and Tanré, 2000] or sparsely distributed AERONET stations [Bellouin et al., 2003], whereas here these quantities are taken directly from global satellite observations. Figure 13 shows that DREA computed in the present study is comparable with previous assessments. Remarkable consistency is found between our estimates and the values calculated by Loeb and Manalo‐Smith [2005] (CERES/MODIS), for both cloud‐free and all‐sky ocean conditions. Despite some regional discrepancies, especially over areas characterized by small aerosol load, global patterns, seasonal, and zonal averages compare well with Loeb and Manalo‐Smith's [2005] results (Figures 10-12). They used completely different observations and methods. The CERES/MODIS approach bypasses the need of SSA and most aerosol characteristics by using CERES (Clouds and the Earth's Radiant Energy System) measurements of radiative fluxes in conjunction with MODIS retrievals. Error bars in their estimates arise from calculations using two different cloud masks. Compared to previous assessments, the present study estimates of DREA are generally smaller over ocean, likely due to the fact that PARASOL often observes maritime SSA smaller than models [Lacagnina et al., 2015]. More absorbing aerosols can lead to less negative oceanic DREA than anticipated by previous works. Over land our estimate is smaller than in MODIS/MATCH, whereas it is higher than in Yu et al. [2006], MAC‐v1, and CloudSat/CALIPSO and within error bar of the top‐down estimates in MISR/CERES/MODIS. Continental DREAs based on the MAC‐v1 and CloudSat/CALIPSO data sets are the least negative, compared to the other estimates. This is partly because MAC‐v1 and CloudSat/CALIPSO averages include polar regions, characterized by high surface reflectivity, leading to weak negative or even positive DREA. In addition, compared to our geographical maps (Figure 10), their estimate is more positive over Sahara and less negative in central Africa, contributing to weaker aerosol cooling. Normalizing DREA by AOD might give more insights. Radiative efficiency (E τ ), defined as DREA/AOD 550 [Anderson et al., 2005], is mainly governed by aerosol SSA and phase function, surface reflectivity, and solar irradiance. Globally (60°S–60°N) averaged cloud‐free E τ from our estimates is −30 W m−2τ−1 over ocean. This value is at the lower edge of the E τ range derived from other groups, i.e., −29 to −52 W m−2τ−1 [Anderson et al., 2005; Loeb and Manalo‐Smith, 2005]. This suggests that discrepancies in SSA can cause weaker aerosol cooling over ocean, compared to the other estimates (Figure 13). Over land, cloud‐free E τ is −20 W m−2τ−1, which is less negative than in former assessments (e.g., MISR/CERES/MODIS: −28 W m−2τ−1). This would lead to weaker aerosol cooling over land, while we calculate stronger cooling compared to the previous estimates (Figure 13), suggesting that PARASOL may retrieve higher continental AOD. However, given the uncertainties in observations, more accurate satellite‐derived aerosol characteristics are needed to draw more definite conclusions. 5.1 Uncertainty in DREA Estimates Uncertainties in our estimate of DREA may arise from different sources. These include uncertainties in the sets of cloud and aerosol retrievals and how these are ingested in the RT scheme. We estimate the uncertainties from the latter by carrying out several sensitivity experiments. The RT scheme was run for March only, varying specific inputs, as reported in Table 4. One possible variation to the reference experiment (described in section 3) consists in using wavelength‐dependent microphysical properties from PARASOL/ECHAM5‐HAM2 to compute AOD and SSA outside the satellite wavelength range, instead of extrapolating AOD and SSA in that spectral range (experiment “paraomi‐optmic”). Another option is not using OMI data, but relying exclusively on PARASOL retrievals. In this case two additional experiments are conducted: one is similar to the reference experiment, but OMI SSA is not assimilated (“para‐opt”); a second one is similar to experiment “paraomi‐optmic,” but OMI SSA is not assimilated (“para‐optmic”). We also conducted an experiment employing the TM5‐v3 model [van Noije et al., 2014], instead of ECHAM5‐HAM2, to provide aerosol vertical distribution, wavelength variability of aerosol parameters, and to fill observational gaps (“paraomi‐opt‐TM5”). Both models were constrained by similar meteorological conditions and aerosol precursors emissions, complying with the protocol of the AeroCom phase II intercomparison project experiment A2‐CTRL‐06 [Schulz et al., 2006]. An additional experiment consists in filling observational gaps without any model guidance, but only by interpolating satellite‐derived values for grid boxes with missing data (“paraomi‐opt‐NOMODEL”). In this case, ECHAM5‐HAM2 provides only aerosol vertical distribution and wavelength variability of aerosol parameters. Table 4. List of Sensitivity Experiments Label Experiment Description Reference Detailed description in sections 3–4 Paraomi‐optmic The same as the reference experiment, but aerosol microphysical retrievals are used below 0.35 μm and above 1.02 μm Para‐opt The same as the reference experiment, but no OMI data are used Para‐optmic The same as the “para‐opt” experiment, but aerosol microphysical retrievals are used below 0.49 μm and above 1.02 μm Paraomi‐opt‐TM5 The same as the reference experiment, but TM5‐v3 model [van Noije et al., 2014 instead of ECHAM5‐HAM2 Paraomi‐opt‐NOMODEL The same as the reference experiment, but no model is used to fill observational gaps. Aerosol vertical profile is given by ECHAM5‐HAM2 Results from these runs are shown in Figure 14. The various input choices to the RT scheme have little impact on DREA estimates. The largest differences arise toward the south pole when another model (cyan line) or no model at all (purple line) is used. At those latitudes satellite observations are less frequent (Figure 9) and model guidance becomes more relevant. Moreover, more pronounced diversity arises at regional scale (not shown), and consideration of different temporal scales (e.g., daily) may result in larger variability. However, it is interesting to note that using completely different aerosol global models to fill observational gaps and to provide wavelength variability of observed aerosols, as well as their vertical distribution, does not affect significantly computations of DREA (black and cyan lines). This indicates that results presented here have little dependency on model preferences. Given the small impact that different approaches have on global mean DREA estimates, the standard deviation arising from the sample of sensitivity experiments is small: ±0.2 W/m2 for cloud‐free and ±0.1 W/m2 for all‐sky conditions. Figure 14 Open in figure viewerPowerPoint Zonal means of DREA at TOA for cloud free (solid lines) and all sky (dashed lines) from the sensitivity experiments reported in Table 4 . Values are for March 2006. Other sources of satellite uncertainties are less straightforward to estimate, compared to model‐based approaches [Bellouin et al., 2005]. Here we give the upper bound of these uncertainties, following the method outlined below. Additional sensitivity experiments are carried out by perturbing the values of AOD or/and SSA of the reference experiment. This is done in such a way that one experiment is ran with uniformly increased AOD, one with decreased AOD, other two with increased and decreased SSA, and other four run with a combination of increased/decreased AOD and SSA varied together. Perturbations are ±0.05 for AOD and ±0.03 for SSA. These are the ranges where the majority of points from PARASOL and OMI falls within the absolute difference with AERONET (Figure 1 and Figures 7a and 7c). Running the RT model for March 2006 yields a standard deviation in DREA fluxes of ±1.5 W/m2 in cloud‐free and of ±0.7 W/m2 in all‐sky conditions. These values dominate the total uncertainty.

6 Summary and Discussion We presented the first global estimate of SW DREA at TOA over land and ocean, based on new global satellite observations of SSA, phase function, and AOD from PARASOL, in synergy with OMI SSA retrievals. Aerosol information from these two sensors is combined with land‐surface BRDF and cloud properties from MODIS. The satellite‐based approach is completed by filling observational gaps with outputs from the aerosol global model ECHAM5‐HAM2. The methodology devised to integrate observational and modeling data sets of aerosol microphysical and optical properties has been described in detail. With respect to previous bottom‐up assessments of DREA [e.g., Myhre, 2009; Su et al., 2013], this study is different for a number of reasons. First, our estimate is more constrained by measurements, because aerosol phase function and SSA are taken from global satellite retrievals, together with AOD and surface albedo. Second, PARASOL and OMI, performing measurements in different but complementary wavelength ranges, are used in synergy. This study also provides an overview of the reliability of the PARASOL AODs and SSAs, which are available over land for the first time. Based on these inputs, we have produced monthly mean DREA maps during 2006 with radiative transfer calculations. The estimated global mean SW DREA is −4.6 ± 1.5 W/m2 for cloud‐free and −2.1 ± 0.7 W/m2 for all‐sky conditions. All‐sky DREA is less negative than its cloud‐free counterpart, because of enhanced planetary albedo by clouds and cloud masking effects on aerosol radiation interactions. Consistent with previous assessments, different mixtures of scattering and absorbing aerosols lead to diverse DREA values at the regional scale. Largest effects are in Northeast Asia, India area, and north central Africa, where DREA can exceed −20 W/m2 locally. Geographical distribution of DREA is asymmetric between the two hemispheres, with the Northern Hemisphere experiencing the strongest aerosol cooling. At seasonal scale, DREA is more negative during the boreal summer. This is mostly the result of high AOD and SSA, due to extensive mineral dust outbreaks downwind of the Sahara, the Arabian peninsula, and Mongolian deserts and pollution emissions, larger in the Northern Hemisphere. Estimates of DREA are affected by uncertainties, which have been discussed. Those include possible different approaches to process input parameters for the RT scheme. We carried out several sensitivity experiments varying certain input choices and found that these have little impact on global mean DREA. Moreover, using two different aerosol global models to complete the observational‐based approach, we found that our estimate is virtually independent of model choice. However, more pronounced diversity may arise at different temporal and regional scales. Global annual mean DREA values assessed in the present study are comparable with previous works, relying on different data sets and approaches. Remarkable consistency is found between our estimate and the values calculated by Loeb and Manalo‐Smith [2005], based on completely different observations, namely, flux measurements from CERES and retrievals from MODIS. In general, our estimates are less negative over ocean and often more negative over land, compared to former observational‐based bottom‐up assessments. Less negative values are likely due to the fact that PARASOL often observes smaller oceanic SSA than models over many regions [Lacagnina et al., 2015]. More absorbing aerosols can lead to less negative DREA than anticipated by previous assessments. However, given the uncertainties in observations, more accurate satellite‐derived SSA and aerosol characteristics (e.g., refractive index and Mie size parameters) are needed to draw more definite conclusions. Accurate portrayal of the aerosol properties is crucial to determine aerosol contribution to the Earth's radiation budget. Since each sensor has its own strengths and limitations when retrieving aerosols, the best strategy lies on integrating measurements from multiple platforms [Stephens et al., 2002; Schoeberl et al., 2006]. The present study has taken some steps in this direction, showing very encouraging results from combining PARASOL and OMI retrievals. Further refinement of these products, integrated with additional observations from other sensors, will be beneficial to future estimates of DREA and its anthropogenic component.

Acknowledgments This work has been supported by the User Support Space Research program of the Netherlands Organization for Scientific Research (NWO) through project ALW‐GO/13‐38. The authors are grateful to the anonymous reviewers for their constructive comments that have helped for the improvement of this paper. We thank the AERONET principal investigators and their staff for establishing and maintaining the sites used in this study. The AERONET data set can be obtained from http://aeronet.gsfc.nasa.gov/. We thank NASA for the online availability of the OMI (http://disc.sci.gsfc.nasa.gov/Aura/data-holdings/OMI/omaeruv_v003) and MODIS products (https://ladsweb.nascom.nasa.gov/data/search.html and https://lpdaac.usgs.gov/dataset_discovery/modis/modis_products_table/mcd43b3). The authors are grateful to Kai Zhang for providing ECHAM5‐HAM2 model's data and to Twan van Noije for making available TM5 model's data. These modeling data can be accessed following the instructions at the AeroCom portal http://aerocom.met.no/data.html. The PARASOL data set, the results, and the code used in this study are available at ftp://ftp.sron.nl/open-access-data/carlol/doi-jgr-2016.

Appendix A: Details About Methodology to Combine Observational and Modeling Data Sets In order to combine the observational and modeling data sets, these need to be made compatible. Indeed, the two data sets describe aerosols in different ways, requiring further postprocessing. The PARASOL retrieval algorithm accounts for two aerosol modes (fine and coarse), characterized by columnar‐retrieved properties. In contrast, ECHAM5‐HAM2 accounts for one soluble nucleation mode plus three pairs of soluble and insoluble modes (Aitken, accumulation, and coarse), whose characteristics are given at each vertical level. We consider only the last two pairs of modes from the model, because of the negligible radiative contribution from the remaining modes, characterized by much smaller particle sizes. Hereafter, we refer to the model accumulation mode as the fine mode. A number of steps are required to make the modeling data set compatible with the description used in the retrieval algorithm. As a first step, the volume size lognormal distribution of each ECHAM5‐HAM2 mode is computed. This requires converting mass mixing ratios of the various chemical species to volume and then adding the volumes of the components contributing to a specific mode. At each grid box and vertical level, the total volume (V i ) of each mode (i) is used to compute the median radius ( ) of the lognormal distribution, following Seinfeld [1986]: (A1) N i and σ i are the concentration number and the geometric standard deviation of mode i, respectively, as given by ECHAM5‐HAM2. At this point, volume size lognormal distribution of mode i can be computed and then vertically integrated. whereandare the concentration number and the geometric standard deviation of mode, respectively, as given by ECHAM5‐HAM2. At this point, volume size lognormal distribution of modecan be computed and then vertically integrated. As a second step, the soluble volume size lognormal distribution, describing the fine (coarse) mode, is added to its insoluble counterpart. Mie‐scattering size parameters are derived through nonlinear fitting of the new fine (soluble + insoluble) mode distribution and of the new coarse (soluble + insoluble) mode distribution. Finally, model refractive indexes of the new fine (soluble + insoluble) and coarse (soluble + insoluble) modes are computed by averaging volume‐weighted refractive indexes of the appropriate chemical species. At this point, the set of aerosol parameters extracted from ECHAM5‐HAM2 is compatible with the correspondent set of PARASOL observations. The satellite and modeling data sets are now ready to be assimilated.