A 3-D evaluation of the MACC reanalysis dust product over Europe, northern Africa and Middle East using CALIOP/CALIPSO dust satellite observations

Abstract. The MACC reanalysis dust product is evaluated over Europe,
northern Africa and the Middle East using the EARLINET-optimized
CALIOP/CALIPSO pure dust satellite-based product LIVAS (2007–2012). MACC
dust optical depth at 550 nm (DOD550) data are compared against LIVAS
DOD532 observations. As only natural aerosol (dust and sea salt)
profiles are available in MACC, here we focus on layers above 1 km a.s.l.
to diminish the influence of sea salt particles that typically reside at low
heights. So, MACC natural aerosol extinction coefficient profiles at 550 nm
are compared against dust extinction coefficient profiles at 532 nm from
LIVAS, assuming that the MACC natural aerosol profile data can be similar to
the dust profile data, especially over pure continental regions. It is shown
that the reanalysis data are capable of capturing the major dust hot spots in
the area as the MACC DOD550 patterns are close to the LIVAS DOD532
patterns throughout the year. MACC overestimates DOD for regions with low
dust loadings and underestimates DOD for regions with high dust loadings
where DOD exceeds ∼ 0.3. The mean bias between the MACC and LIVAS DOD
is 0.025 (∼ 25 %) over the whole domain. Both MACC and LIVAS
capture the summer and spring high dust loadings, especially over northern
Africa and the Middle East, and exhibit similar monthly structures despite
the biases. In this study, dust extinction coefficient patterns are reported
at four layers (layer 1: 1200–3000 m a.s.l., layer 2:
3000–4800 m a.s.l., layer 3: 4800–6600 m a.s.l. and layer 4:
6600–8400 m a.s.l.). The MACC and LIVAS extinction coefficient patterns
are similar over areas characterized by high dust loadings for the first
three layers. Within layer 4, MACC overestimates extinction coefficients
consistently throughout the year over the whole domain. MACC overestimates
extinction coefficients compared to LIVAS over regions away from the major
dust sources while over regions close to the dust sources (the Sahara and
Middle East) it underestimates strongly only for heights below
∼ 3–5 km a.s.l. depending on the period of the year. In general, it
is shown that dust loadings appear over remote regions and at heights up to
9 km a.s.l. in MACC contrary to LIVAS. This could be due to the model
performance and parameterizations of emissions and other processes, due to
the assimilation of satellite aerosol measurements over dark surfaces only or
due to a possible enhancement of aerosols by the MACC assimilation system.


forecasts (optical depth and surface concentration) on a daily basis. The aerosol forecasts are produced using the same system that was operated for the production of the multiyear (2003-2012) MACC reanalysis. The MACC reanalysis was developed within the framework of GMES (Global Monitoring for Environment and Security) and a series of MACC projects funded by the European Union and coordinated by ECMWF (http://www.gmes-atmosphere.eu/about/project/). The MACC activities are now carried on under CAMS (Copernicus Atmosphere Monitoring Service) (Eskes et al., 2015). 5 Upon its release the MACC reanalysis aerosol product has been used in a many studies at a global and regional level. For example, it has been used in global estimates of the direct and indirect aerosol radiative effect (Bellouin et al., 2013), to study the sensitivity of clouds to aerosol loads and types over the oceans (Andersen et al., 2016), to constrain the influence of aerosols on cloud coverage (Gryspeerdt et al., 2016), to build regional climatologies in conjunction with satellite data (e.g. Nabat et al., 2013, Georgoulias et al., 2016a, as input for the production and evaluation of satellite-based surface solar 10 radiation products (Mueller et al., 2015;Alexandri et al., 2017), to support reports on the current state of the climate (Benedetti et al., 2014), etc.
Specifically in dust oriented studies, MACC forecasts have been used in conjunction with measurements from dropsondes and lidars onboard aircrafts, ships and satellites (Cloud-Aerosol Lidar with Orthogonal Polarization onboard Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations -CALIOP/CALIPSO) to study the long-range transport of Saharan dust 15 across the Atlantic within the framework of the Saharan Aerosol Long-range Transport and Aerosol-Cloud-interaction Experiment (SALTRACE) campaign in spring and summer 2013 (Chouza et al., 2016;Ansmann et al., 2017). In these studies forecast fields were used instead of analyses (MACC reanalysis data stop in 2012) focusing on the total aerosol optical depth (AOD) and extinction coefficients rather than on dust. Cuevas et al. (2015) evaluated the MACC reanalysis dust product over Northern Africa and Middle East for two years (2007)(2008) using ground and satellite-based 20 measurements. Their comparisons focused on specific sites (AERONET sunphotometers, lidars and CALIOP/CALIPSO observations) while for spatial evaluations they utilized total AOD satellite data from passive sensors such as MODIS, MISR and OMI. Marinou et al. (2017) furthermore performed a first comparison of an optimized CALIOP/CALIPSO pure dust optical depth (DOD) patterns with MACC reanalysis DOD patterns; however, a detailed 3-dimensional spatiotemporal evaluation of the MACC dust product is still missing . 25 In this study we advance for the first time to a 3-D (optical depths and profiles) evaluation of the MACC reanalysis dust product over the Europe -Northern Africa -Middle East domain [13 o N-60 o N, 40 o W-70 o E] using the EARLINEToptimized CALIOP/CALIPSO pure dust satellite-based product for the time period [2007][2008][2009][2010][2011][2012]. It has to be highlighted that this is an independent observational product as it is not included in the MACC assimilation procedure. Details about the datasets used for the evaluation along with a description of the methodology followed are given in Sect. 2. The results from 30 the evaluation procedure are presented in Sect. 3, while in the end of the paper the main findings and conclusions of this research are summarized. Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2017-1238 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 26 January 2018 c Author(s) 2018. CC BY 4.0 License.

MACC reanalysis data
In this work, two MACC reanalysis datasets, one characterizing the columnar dust load and one indicative of the dust profile in the atmosphere, are evaluated for a six-years period spanning the period from 2007 to 2012. 3-hourly dust optical depth data at 550 nm (DOD 550 ) and natural aerosol (dust and sea salt) optical depth at 550 nm (AOD 550 ) profiles are available from 5 MACC. The spatial resolution of the data is ~ 78 km x 78 km with 60 vertical levels from the surface up to 0.1 hPa. It is obvious that the profile data as given by MACC are affected by a sea salt component; however, if used properly one can get an insight into the ability of MACC to simulate the dust profiles. In addition to the optical depth data, geopotential data (in m 2 /s 2 ) from MACC are also used in order to calculate the physical height of each model layer which is necessary for the comparison of the profile data with the satellite data as it is shown below. 10 The MACC reanalysis data used here are produced using the aerosol analysis and forecast system of ECMWF. This consists of a forward model (Morcrette et al., 2009) and a data-assimilation module (Benedetti et al., 2009). The MACC forecasting system assimilates, among other observational data (Eskes et al., 2014), AOD 550 measurements from the two MODIS sensors aboard Terra and Aqua through a 4D-Var assimilation algorithm to produce the aerosol analysis. The assimilation improves the representation of aerosols as shown in previous studies (see Benedetti et al., 2009;Mangold et al., 2011). The MACC 15 aerosol system accounts for a total of five aerosol species, mineral dust, sea salt, sulfates, black carbon and organic matter.
Three different size bins are used for mineral dust (0.03-0.55, 0.55-0.9 and 0.9-20 microns) and sea salt particles (0.03-0. 5, 0.5-5 and 5-20 microns). Black carbon and organic material are distributed to a hydrophilic and a hydrophobic mode. Sea salt emissions are given as a function of surface wind speed (Guelle et al., 2001;Schulz et al., 2004). Dust emissions are given as a function of surface wind speed, soil moisture, surface albedo and land cover following Ginoux et al. (2001). The 20 emissions of the other species are taken from inventories (e.g. SPEW, EDGAR) while a climatology is used for stratospheric aerosols.

LIVAS CALIOP/CALIPSO data
For the evaluation of the MACC reanalysis data, dust optical depth at 532 nm (DOD 532 ) and dust extinction coefficients at 532 nm (in km -1 ) from CALIOP/CALIPSO are used. The horizontal resolution of the data is 1 o x 1 o and the extinction 25 coefficient retrievals are available at 399 predefined heights which characterize a layer of ~60 m for altitudes below ~20 km and ~180 m for higher altitudes. As CALIPSO flies at a 705 km altitude sun synchronous polar orbit with a 16 day repeat cycle there are 1-3 measurements available per grid cell on a monthly basis. The satellite data utilized in this work have been produced using an EARLINET-optimized retrieval scheme that was developed within the framework of the LIVAS (LIdar climatology of Vertical Aerosol Structure for space-based lidar simulation studies) project (Amiridis et al., 2015). More 30 specifically the pure dust LIVAS product is used (see Amiridis et al., 2013;Marinou et al., 2017). This product is corrected Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2017-1238 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 26 January 2018 c Author(s) 2018. CC BY 4.0 License.
for the dust LR value which is specific for the region we focus in this study, based on multi-year measurements performed by the ground-based lidar stations of the EARLINET (European Aerosol Research Lidar Network). These region-specific LRs are equal to 55±5 sr for the North Africa -Europe domain (Tesche et al., 2009(Tesche et al., , 2011Groß et al., 2011Groß et al., , 2015 and equal to 40±5 sr for Middle East and central Asia at longitudes further east than 30°E (Mamouri et al., 2013;Nisantzi et al., 2015;Hofer et al., 2017). The correction leads to an AOD 532 absolute bias of ~-0.03 compared to spatially and temporally 5 collocated AERONET observations above Europe and North Africa while the corresponding biases for the standard CALIPSO product are much higher (~-0.10) (Amiridis et al., 2013). The bias is lower (~-0.02) when compared against spatially and temporally collocated MODIS satellite data. In addition, the use of a new methodology for the calculation of the pure dust extinction from dust mixtures and an averaging scheme that includes zero extinction values for the non-dust aerosol types allow for further improvement of the LIVAS pure dust product (Amiridis et al., 2013). 10

Spatial and temporal collocation of the datasets
The DOD and profile datasets from MACC reanalysis have to be processed properly prior to the comparison with the LIVAS data. Generally, it is much more straightforward to evaluate the MACC columnar dataset. It has to be mentioned that while the MACC reanalysis data are available on a 3-hourly basis, the LIVAS data used here are available as monthly means.
However, the exact overpass date and time of the retrievals used for the calculation of the monthly data is given which 15 allows for the temporal collocation of the two datasets. The MACC DOD 550 data are first brought to the LIVAS 1 o x 1 o grid using bilinear interpolation and then only the MACC values closer to the to the LIVAS DOD 532 retrieval time are chosen.
Finally, the MACC DOD 550 data are averaged on a monthly basis and can be evaluated against the LIVAS data (see Fig. 1a for the whole procedure).
Much more effort is needed to bring the MACC reanalysis and the satellite-based profile data in a format suitable for 20 comparison (see Cuevas et al., 2015 andChouza et al., 2016 for previous efforts) prior to the horizontal and temporal collocation of the datasets. As the MACC reanalysis offers only natural (dust and sea salt) AOD 550 fields (unitless) for each one of the 60 MACC layers and the LIVAS data include extinction coefficients (in km -1 ) at 399 heights it is obvious that the two datasets are not directly comparable. Similar problems may emerge when evaluating simulations from other global or regional climate models. The method proposed here (see Fig. 1b for details) is a generic one and could be applied in future 25 model evaluation studies. First, the MACC reanalysis AOD 550 profiles are converted to extinction coefficients at 550 nm by dividing the given MACC geopotential fields with the gravity acceleration to obtain the physical layer heights. From the physical layer heights (upper layer minus lower layer physical height) the physical depth of each MACC layer is calculated. aerosol extinction coefficients for the 399 LIVAS levels are averaged on a monthly basis. To obtain more robust statistics, both the MACC and the LIVAS 399-level data are finally averaged vertically within a set of selected layers each one having a depth of 300 m (see also Cuevas et al. 2015). The evaluation procedure is implemented for 29 (300-meter) layers covering the troposphere from 300 m (first layer centered at 450 m) up to 9 km (last layer centered at 8.85 km). It needs to be reiterated that the profile data as given by MACC are contaminated with a sea salt component; however, if used properly one 5 can get an insight into the ability of MACC to simulate the dust profiles. Sea salt particles in the area are mostly accumulated within the marine boundary layer, generally at heights below 1 km (see Nabat et al., 2013); hence, sea salt is expected to have an impact only at the lower levels of the natural aerosol profiles. Therefore, in this work we focus on layers higher than 1 km. The extinction coefficient patterns (MACC, LIVAS and their difference) presented in this work are calculated by averaging vertically over four 1800-meter layers (layer 1: 1200-3000 m, layer 2: 3000-4800 m, layer 3: 4800-6600 m and 10 layer 4: 6600-8400 m). The first layer starts from 1200 m (> 1 km) in order to diminish the contamination of the extinction coefficients from sea salt particles as discussed above. Hence, we refer to MACC dust extinction coefficients hereafter and not natural aerosol extinction coefficients. One should keep in mind however that the dust extinction coefficients used here are still contaminated with a sea salt component at some degree especially over the sea and regions close to the coasts while our results should be considered more robust over pure continental regions. 15

Evaluation procedure
The MACC reanalysis dust product evaluation procedure comprises different steps.  Fig. 2d) it is shown that on average the bias becomes negative when DOD becomes higher than 0.28. It has to be 15 highlighted here that ~90% of the LIVAS DOD 532 values are below this critical value (see Fig. 2d), hence, the underestimations are connected to a large degree with source areas and episodic dust events.
Over the whole domain (EU), the MB between the MACC and LIVAS DOD is 0.025 and the NMB is ~25% with an RMS error of 0.115. The correlation coefficient (R) of the linear regression (y=0.562x+0.068) between MACC and LIVAS DOD is 0.76 (for details see Table 1). Up to a DOD value of ~0.3 the MACC and LIVAS products are characterized by a strong 20 linear correlation with a slope close to 1 (y=0.986x+0.046 with R=0.76); however, there is no significant correlation for values higher than ~0.3, LIVAS exhibiting much higher DOD values than MACC (y=0.176x+0.269 with R=0.33). A similar situation is observed over sub-regions high DODs near the major dust sources such as CWSah, ESah and ME and less over the transitional (from dusty to clean conditions) sub-regions such as ATL, SWE, CM and EM. The sub-region with the highest MACC-LIVAS correlation (higher R value and slope closer to one) is ATL (y=0.713x+0.037 with R=0.87) and the 25 sub-region with the lowest MACC-LIVAS correlation is CE (y=0.300x+0.029 with R=0.40). The slope and the intercept of the linear regression line, the correlation coefficient R, the MACC-LIVAS MB, NMB and the RMS error, the MACC and LIVAS mean DOD levels for all the sub-regions of interest are given in Table 1.
It is concluded from the two paragraphs above that in general MACC overestimates DOD for regions with low dust loadings and underestimates DOD for regions with high DOD loadings. Similar results were shown in Amiridis et al. (2013)

and 30
Atmos. Chem. Phys. Discuss., https://doi.org /10.5194/acp-2017-1238 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 26 January 2018 c Author(s) 2018. CC BY 4.0 License. Tsikerdekis et al. (2017) where BSCDREAM8b and RegCM4 dust simulations were compared against CALIOP/CALIPSO satellite observations. Many reasons could be responsible for these overestimations/underestimations. First of all it might be related to the model itself (e.g. parameterization of dust emissions, the wind velocity, the distribution of dust particles in different bins, the dry and wet deposition, the convection scheme which is used, etc.). For example if the model overestimates the fine mode dust particles, the lifetime of dust in the air would increase leading to the transport of particles 5 away from the sources and at greater height levels. However, as discussed in Ansmann et al. (2017), the uncertainties stemming from the complex parameterizations used by the model make it difficult to reach a solid conclusion about the observed overestimations and underestimations.
Another reason for the underestimation of DOD close to the major dust sources in the area could be the assimilation of AOD 550 measurements only from the MODIS/Terra and MODIS/Aqua Dark Target (DT) product which does not include 10 observations over bright surfaces such as deserts, arid and semi-arid regions. The overestimation of dust away from the sources (MACC DOD never gets a zero value even over remote oceanic regions) might also be related to the assimilation procedure. The control variable of the assimilation is the total aerosol mixing ratio calculated by adding the mixing ratios of all species. The AOD 550 is calculated from the single species, summed, integrated and then compared to the observations. Through the 4D-Var assimilation algorithm increments in total aerosol mixing ratio are obtained. Those increments are 15 redistributed to all species proportionally to their fractional contribution to the total mass. It has to be noted that the model does not take into account ammonium nitrate aerosols which represents a large component over the greater European area (Giordano et al., 2015). As a result the model will most of the time underestimate AOD relative to the observations and hence the assimilation system will tend to increase the other aerosol components to give the correct AOD overall. Probably, the system allows the presence of dust even at tiny concentrations and so dust always receives a small contribution during 20 the assimilation even when there should be no dust in the atmosphere.
One more parameter contributing to the differences observed between the model and the observations is probably the limitation of CALIPSO to detect aerosol layers with signals lower that the satellite's signal-to-noise ratio (SNR) (Winker et al., 2013). In particular, in heights where the CALIPSO SNR is higher that the signal of the layer, the area is characterized as clear air, and a value of 0 km -1 is set. The detection thresholds are defined in terms of 532 nm scattering ratio and are 25 adjusted according to altitude, solar background illumination and averaging resolution (Vaughan et al., 2009). As higher thresholds are used during daytime than at night (because SNR is reduced by solar background illumination), weakly scattering layers which are detected at night may be missed during daytime. Typical values of CALIPSO layer thresholds in dust observations are 0.04±0.02 km -1 during daytime and 0.008±0.003 km -1 during nighttime. Indicatively, Kim et al. (2017) found a global mean undetected aerosol layer with an AOD of 0.031±0.052 after comparing 2 years of CALIPSO and 30 MODIS AODs. These undetected layers are expected to affect more the higher altitudes in the CALIPSO product.

Dust optical depth seasonal variability
In this section, the seasonal variability of the MACC, the LIVAS DODs and their differences are discussed. The seasonal patterns of MACC DOD 550 and LIVAS DOD 532 along with the MACC-LIVAS MB patterns are presented in Fig. 3 while the corresponding monthly variabilities per region of interest are shown in Fig. 4. A number of studies using passive and active satellite-based observations have revealed the spatiotemporal variability of dust and its pathways over the greater 5 Mediterranean area during the last two decades (e.g. Moulin et al., 1998;Prospero et al., 2002;Barnaba and Gobbi, 2004;Antoine and Nobileau, 2006;Gkikas et al., 2009Gkikas et al., , 2013Gkikas et al., , 2016Israelevich et al., 2012;Ginoux et al., 2012;Pey et al., 2013;Varga et al., 2014;Georgoulias et al., 2016a,b;Tsikerdekis et al., 2017;Marinou et al., 2017). It is well known today that over western Mediterranean dust peaks in summer, over eastern Mediterranean in spring while central Mediterranean is a transitional region with high dust loadings throughout summer and spring. Dust concentrations over western Europe peak in 10 summer while over central and eastern Europe are higher during spring and summer than during the rest of the year. The seasonal variability of dust depends mostly on the seasonality of the emissions over the source areas and the dominating wind patterns. Over the western Sahara the dust emissions peak in summer, over the eastern part of the desert in spring and over the Middle East dust activity peaks in late spring and summer (for details see in the studies given above and the references therein). As shown in Fig. 3 and autumn and to a lesser extent in winter (Figs. 4a-f). Over those sub-regions, DOD is slightly underestimated by MACC only in one case (in February over EM). The overestimation is stronger over the regions of CE and EE which are far away from the dust sources in the South. In addition, the overestimation is stronger from late spring to early autumn when MACC shows enhanced DOD values contrary to LIVAS. Over SWE, CM, EM and ATL the monthly variability of MACC DOD is closer to the LIVAS one compared to CE and EE; however, significant biases are observed for spring, summer and autumn. 5 We see here that both MACC and LIVAS depict clearly the difference in the peak period between the Western (summer peak), the Central (transitional region) and the Eastern (spring peak) part of the Mediterranean Basin. Over CWSah MACC underestimates DOD from February to September when dust loadings peak (Fig. 4g) while over ESah MACC overestimates DOD consistently throughout the year (Fig. 4h). Over ME MACC and LIVAS DODs are very close from February to July while MACC overestimates DOD during the rest of the year (Fig. 4i). In line with the discussion in the previous paragraph 10 DOD peaks in summer over CWSah, in spring over ESah and during spring and summer over ME.

Annual dust profiles
In this section, the evaluation of the annual MACC reanalysis profiles is presented taking advantage of the unique ability of CALIOP/CALIPSO to retrieve dust extinction coefficient profiles. As discussed in Sect. 2.3, the extinction coefficient 15 patterns presented in this work are reported at four 1800-meter layers that cover the first ~9 kilometres of the troposphere (from 1200 to 8400 m above the sea level -a.s.l.). In accordance to Fig. 2, the MACC and LIVAS dust extinction coefficient patterns presented in Fig. 5 are similar over areas characterized by high dust loadings, especially for the first 3 layers (1200-6600 m). The fourth layer (layer 4: 6600-8400 m a.s.l.) is characterized by zero or near zero LIVAS extinction coefficients everywhere while this is not the case for MACC. MACC overestimates extinction coefficients consistently over the whole 20 EU domain within layer 4 showing that small amounts of dust are always present in MACC even at altitudes up to ~9 km and also over remote oceanic regions. As shown in Fig. 5, within the other 3 layers the overestimations and underestimations from MACC compared to LIVAS are observed at the same areas where MACC overestimates or underestimates DOD (Fig.   2). The absolute MB values are higher in the lowest layer (layer 1: 1200-3000 m a.s.l.) decreasing in layer 2 (3000-4800 m a.s.l.) and layer 3 (4800-6600 m a.s.l.) which is expected taking into account that dust mostly resides within the first 5 km of 25 the atmosphere in this area (see also Tsikerdekis et al. 2017;Marinou et al., 2017).
In line with Fig. 2, strong overestimation is found over the region situated on the west of the Caspian Sea, the region of eastern Sahara and the Arabian Sea for the first three layers but also over the north-eastern Atlantic Ocean within layer 1.
The strong overestimation by MACC within layer 1 over the north-eastern Atlantic Ocean is probably due to the presence of sea salt aerosols despite the fact that this layer is expected to be higher than the oceanic boundary layer as discussed in detail 30 in Sect. 2.3. On the other hand, MACC underestimates extinction coefficients significantly over the region of central and western Sahara and over the largest part of the Middle East for the first three layers as in the case of DOD (see Fig. 2). In general, over EU, the MB between the MACC and LIVAS extinction coefficients is 0.006 km -1 in layer 1, 0.003 km -1 in layer 2 and 3 and 0.002 km -1 in layer 4. The corresponding NMB values are 23%, 22%, 105% and 1866%. The correlation coefficient (R) of the linear regression between MACC and LIVAS extinction coefficients is 0.62, 0.75, 0.66 and 0.13 respectively for the four layers. These values along with the mean MACC and LIVAS extinction coefficients, the RMS error, the slope and the intercept of the MACC-LIVAS regression line for each layer for EU and the 9 sub-regions of interest can 5 be found in Table 2. In general, layer 2 exhibits the best correlation between MACC and LIVAS, layers 3 and 1 follow while the correlation is pretty low in layer 4. Fig. 6 shows the 300-m extinction coefficient profiles from MACC and LIVAS and the corresponding biases for the whole EU domain and the nine sub-regions of interest. As discussed above we focus on altitudes higher than 1 km a.s.l. to avoid as much as possible the interference of sea salt aerosols and hence assume that the MACC natural aerosol extinction 10 coefficients can be similar to dust extinction coefficients. In general, over EU, MACC overestimates extinction coefficients consistently from 1 up to 9 km (Fig. 6j). The overestimation is stable for heights below ~2 km a.s.l. (~0.006 km -1 ) decreases gradually up to ~4 km a.s.l. (~0.002 km -1 ), increases again moderately up to 6 km a.s.l. (~0.003 km -1 ) and finally decreases again, the bias being equal to ~0.002 km -1 for heights above 7-8 km a.s.l.. Over CE, EE, SWE and ATL the overestimation by MACC gradually decreases with height until it gets a value of 0.002 km -1 at heights above ~7 km a.s.l. (Figs. 6a, b, c and  15 f) while over CM and EM it increases first up to ~2 km a.s.l. and then decreases gradually (Figs. 6d and e). Over CWSah extinction coefficients are consistently underestimated for heights up to ~5.5 km a.s.l. and overestimated thereafter (Fig. 6g).
Over ESah MACC overestimates significantly extinction coefficients within the first 4 km a.s.l. with a bias peak at ~2 km a.s.l.. Above ~4 km a.s.l. MACC overestimates extinction coefficients less strongly with a bias peak at ~6 km a.s.l. (Fig. 6h).
The appearance of non-zero extinction coefficients at heights well above 5 km a.s.l. in the MACC aerosol product, in contrast to ground or satellite-based observations, can be spotted in figures of previous studies (e.g. Fig. 9 in Cuevas et al., 2015 andFig. 5 in Ansmann et al., 2017). However, there has been no effort to understand the reasons for this situation.
According to the discussion in Sect. 2.1.1, this might be due to the way the model deals with the dust distribution in different 25 size bins and dust deposition, vertical transport and mixing. An overestimation of the fine mode dust particles, an underestimation of the dry or wet deposition or a model parameterization that enhances the vertical transport of dust in the atmosphere would justify the existence of particles at heights up to 9 km a.s.l. over the whole EU domain (see . In addition, the appearance of non-zero dust extinction coefficients at remote oceanic areas and areas away from the sources could be related to the MACC assimilation procedure. As discussed in Sect. 2.1.1 in detail, an underestimation of 30 the modeled total AOD at each time step over the greater European domain relative to MODIS DT data is possible as ammonium nitrate aerosols, which can affect the AOD directly and indirectly through the absorption of water (Karydis et al., 2016), are not included in the model. In this case, during the assimilation procedure the concentrations of the various aerosol components and consequently dust will be enhanced to match the observed AOD. Undetected aerosol layers by CALIPSO (see Sect 3.1.1) may also play some role. These factors or a combination of them could be responsible for the consistent MACC overestimation at great heights even over regions such as CWSah and ME where dust extinction coefficients are underestimated by MACC for heights below ~6 km a.s.l. and ~2 km a.s.l., respectively.

Seasonal biases between MACC and LIVAS dust profiles
In this section, the seasonal variability of the bias between MACC and LIVAS dust extinction coefficient profiles is 5 discussed. The seasonal patterns of the biases between the MACC and LIVAS extinction coefficients for the four reference layers are presented in Fig. 7. In accordance to Fig. 5 the absolute MB values are higher in layer 1, decreasing in layer 2 and layer 3. In layer 4 MB is consistently positive during all the seasons. MACC overestimates extinction coefficients strongly (MB values higher than 0.02 km -1 ) over the region confined by the Caspian Sea, Kazakhstan, Uzbekistan and Turkmenistan in spring, summer and autumn and less in winter. The overestimation continues up to layer 4, particularly in spring and 10 summer. Over the eastern Sahara MACC overestimates extinction coefficients mostly in layer 1 throughout the year. Over the Arabian Sea MACC overestimates extinction strongly in layer 1 particularly in winter and autumn. The stronger overestimation in layer 2 appears in spring, summer and autumn while a strong overestimation still appears in layer 3 in summer. Over the north-eastern Atlantic Ocean MACC overestimates extinction strongly within layer 1 throughout the year, particularly in winter, spring and autumn, the overestimation being much lower in the next 3 layers. On the other hand, 15 MACC strongly underestimates extinction coefficients (MB values lower than -0.02 km -1 ) over western Sahara during summer and spring and to a lesser extent during winter and autumn. Strong underestimation is also seen in spring and summer in layer 2 and in summer in layer 3. Over the Middle East strong underestimations by MACC are seen in spring and summer mostly in layers 1 and 2.
The monthly variability of the bias between MACC and LIVAS 300-m dust extinction coefficient profiles over EU and over 20 the 9 sub-regions of interest is shown in Fig. 8. As it was previously suggested (Sects. 2.3 and 3.2.1), we should focus here on altitudes higher than 1 km a.s.l. to avoid as much as possible the interference of sea salt aerosols in MACC profiles. Over the whole EU domain (Fig. 8j) and for heights lower than ~5 km a.s.l. MACC overestimates extinction coefficients from September to April and underestimates extinction coefficients from May to August. Over ~5 km a.s.l. MACC consistently overestimates extinction coefficients (MB of ~0.002) with a bias peak in summer (MB of ~0.004) for heights of 6-7 km a.s.l. 25 (Fig. 8a). Over CE, EE and SWE MACC overestimates extinction coefficients consistently throughout the year (Figs. 8a, b and c), the MACC-LIVAS MB decreasing gradually with height. As shown in Fig. 8d, over CM MACC overestimates extinction coefficients at all heights except for the months from June to September at heights lower than ~2 km a.s.l. (for July a small underestimation is seen at a layer located at around 4 km a.s.l.). Over EM we see a similar situation but here the underestimation period for heights below ~2 km a.s.l. spans from April to September (Fig. 8e). Over ATL MACC 30 overestimates extinction coefficients consistently throughout the year at all the heights except for the summer months when we see a small underestimation at the layer from ~3 to ~5 km a.s.l.. The overestimation is strong at heights below 2-3 km Atmos. Chem. Phys. Discuss., https://doi.org /10.5194/acp-2017-1238 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 26 January 2018 c Author(s) 2018. CC BY 4.0 License. a.s.l. and gradually decreases (Fig. 8f). Over CWSah MACC underestimates extinction coefficients strongly for heights below 4-5 km a.s.l. from January to October. For the same period MACC overestimates extinction coefficients for heights above 4-5 km a.s.l. while during November and December an overestimation is seen at all height levels (Fig. 8g). Over ESah MACC underestimates extinction coefficients only for heights below ~2 km a.s.l. from January to May and for heights between 3 and 5 km a.s.l. from March to June. A strong overestimation is seen at heights below ~3 km a.s.l. in November 5 and December. Finally, over ME MACC underestimates extinction coefficients strongly for heights below 2-3 km a.s.l. from January to October while overestimating in November and December. Over ~3 km a.s.l. MACC consistently overestimates throughout the year. Overall, we find that MACC generally overestimates extinction coefficients compared to LIVAS over all the sub-regions except for those who are close to the major dust sources (CWSah, ESah and ME) where MACC underestimates strongly for heights below ~3-5 km a.s.l. depending on the period of the year. 10

Conclusions
In this work, the MACC reanalysis dust product is evaluated over Europe, Northern Africa and Middle East (EU domain) using CALIOP/CALIPSO satellite observations for the period 2007-2012. Specifically, MACC dust optical depth (DOD) data and MACC natural aerosol (dust and sea salt) extinction coefficient profiles at 550 nm are evaluated against DOD and dust extinction coefficient profiles at 532 nm respectively from the LIVAS pure dust product (Amiridis et al., 2013). As 15 MACC reports only natural extinction coefficients and not dust extinction coefficients a direct evaluation is unfortunately impossible. By focusing on heights above 1 km a.s.l. the influence of sea salt particles (that typically reside at low heights) is diminished and hence it can be assumed that the MACC natural aerosol profile data can be similar to the dust profile data especially over pure continental regions while our results should be considered less robust over the sea and regions close to the coasts. The main findings of this study are summarized in the following: 20 -The annual MACC DOD 550 patterns are close to the LIVAS DOD 532 ones showing that the reanalysis data are capable of capturing the major dust hot spots in the area. However, MACC overestimates DOD over continental Europe, parts of Turkey and Iran and over the sea (Atlantic Ocean, Mediterranean and Arabian Sea). The overestimation is relatively high over the region situated on the west of the Caspian Sea and over the eastern Sahara. MACC underestimates DOD 25 significantly over of central and western Sahara and over the largest part of the Middle East. In general MACC overestimates DOD for regions with low dust loadings and underestimates DOD for regions with high dust loadings (DOD exceeds ~0.3).
The MB between the MACC and LIVAS DOD is 0.025 over the whole EU, the normalized mean bias (NMB) is ~25%, the root mean squared error (RMS error) is 0.115 and the correlation coefficient (R) of the linear regression (y=0.562x+0.068) between MACC and LIVAS DOD is 0.76. For DODs lower than ~0.3, the MACC and LIVAS products are characterized by 30 a strong linear correlation with a slope close to 1. 22%, 105% and 1866%. R is 0.62, 0.75, 0.66 and 0.13 respectively for the four layers. In general, layer 2 exhibits the best MACC-LIVAS correlation. Layers 3 and 1 follow while the correlation is low in layer 4. Overall, over EU, MACC overestimates extinction coefficients consistently from 1 up to 9 km a.s.l.. The overestimation is stable for heights below ~2 km a.s.l. (~0.006 km -1 ) decreases gradually up to ~4 km a.s.l. (~0.002 km -1 ), increases again moderately up to 6 km a.s.l.
(~0.003 km -1 ) and finally decreases again, the bias being equal to ~0.002 km -1 for heights above 7-8 km a.s.l.. 25 -The absolute MACC-LIVAS MB values are higher in layer 1, decreasing in layer 2 and layer 3. In layer 4 MB is consistently positive during all the seasons. Over the whole EU domain and for heights lower than ~5 km a.s.l. MACC overestimates extinction coefficients from September to April and underestimates extinction coefficients from May to August. Over ~5 km a.s.l. MACC consistently overestimates extinction coefficients (MB of ~0.002) with a bias peak in 30 summer (MB of ~0.004) for heights of 6-7 km a.s.l.. MACC generally overestimates extinction coefficients compared to LIVAS over the sub-regions away from the major dust sources. On the contrary, over CWSah, ESah and ME MACC underestimates strongly for heights below ~3-5 km a.s.l. depending on the period of the year. Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2017-1238 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 26 January 2018 c Author(s) 2018. CC BY 4.0 License.
Overall, it is shown in this work that MACC overestimates DOD for regions with low dust loadings and underestimates DOD for regions with high DOD loadings. Non-zero MACC DODs appear over remote areas (away from the source areas in the South) where LIVAS returns zero DODs. In contrast to LIVAS, non-zero MACC dust extinction coefficients can be spotted over the whole EU for heights up to 9 km a.s.l. throughout the year. As discussed above, this could be due to the model performance and its parameterizations of emissions and other processes, and/or due to the assimilation of AOD 550 5 measurements only over dark surfaces (omitting this way the regions where dust is produced) and/or due to a possible enhancement of all the aerosol components (including dust) by the MACC assimilation system over the greater European area. By including a pure dust product such as LIVAS in the assimilation procedure, part of the observed biases would have probably been addressed. Apart from being a potent assimilation tool it is shown here that LIVAS constitutes an ideal observational dataset for the evaluation of climate model simulations and reanalysis datasets, and the need for more studies 10 towards this direction is acknowledged. It is suggested that dust products from more recent reanalysis projects such as the CAMS interim reanalysis (CAMSiRA) (Flemming et al., 2017), the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) (Gelaro et al., 2017) and the Japanese Reanalysis for Aerosol (JRAero) should be evaluated in a similar way (2-D or 3-D evaluation depending on the availability of dust profile data).

Figure 4:
Monthly variability of the MACC DOD 550 (orange color) and the LIVAS DOD 532 (black color) and their mean bias (red color) over the nine sub-regions of interest (a-i) and over the whole Europe -North Africa -Middle East domain (j). Different scale is used for each sub-region so that the differences between the two datasets per sub-region are depicted more efficiently. The monthly variability of DOD for all the sub-regions of interest together for MACC (k) and LIVAS (l) is also presented here to get an insight into the differences in 5 DOD levels over different sub-regions.

Figure 6
: 300-m resolution profiles of the MACC dust extinction coefficient at 550 nm (in km -1 ) (orange color), the LIVAS dust extinction coefficient at 532 nm (in km -1 ) (black color) and their mean bias (red color) over the nine sub-regions of interest (a-i) and over the whole Europe -North Africa -Middle East domain (j). Different scale is used for each sub-region so that the differences between the two datasets per sub-region are depicted more efficiently. The profiles for all the sub-regions of interest together for MACC (k) and 5 LIVAS (l) are also presented here to get an insight into the dust profile differences over different sub-regions.

Figure 7:
Seasonal patterns (DJF: column 1, MAM: column 2, JJA: column 3 and SON: column 4) of the mean bias between the MACC average dust extinction coefficient at 550 nm (in km -1 ) and the LIVAS average dust extinction coefficient at 532 nm (in km -1 ) over the Europe -North Africa -Middle East domain for layer 1 (a, b, c, d), layer 2 (e, f, g, h), layer 3 (i, j, k, l) and layer 4 (m, n, o, p).