Apparent dust size discrepancy in aerosol reanalysis in north African dust after long-range transport

North African dust reaches the southeast United States every summer. Measurements taken in Miami, Florida indicate that more than one-half of the surface dust mass concentrations reside in particles with diameters less than 2.1 μm, while vertical profiles of micropulse lidar depolarization ratios show dust reaching above four km during pronounced events. These observations are compared to the representation of dust in the MERRA-2 aerosol reanalysis and closely-related GEOS5 Forward Processing (FP) aerosol product, both of which assimilate satellite-derived aerosol optical depths using a similar 5 protocol and inputs. These capture the day-to-day variability in aerosol optical depth well, in a comparison to an independent sun-photometer-derived aerosol optical depth dataset. Measured near-surface dust mass concentrations slightly exceed model values, with most of the modeled dust mass in diameters between 2-6 μm. Modeled-specified mass extinction efficiencies equate light extinction with approximately three times as much aerosol mass, in this size range, compared to the measured dust sizes. GEOS-5 FP surface-layer sea salt mass concentrations greatly exceed observed values, despite realistic winds and 10 relative humidities. In combination, these observations help explain, why, despite realistic total aerosol optical depths, 1) freetropospheric model volume extinction coefficients are lower than those retrieved from the micro-pulse lidar, suggesting too low model dust loadings, and 2) model dust mass concentrations near the surface are higher than those measured. The modeled vertical distribution of dust, when captured, is reasonable. Large, aspherical particles exceeding the modeled dust sizes are also occasionally present, but dust particles with diameters exceeding ten μm contribute little to the measured total dust mass 15 concentrations after such long-range transport. A further integrated assessment is needed to confirm this study’s interpretations.

overlap function were checked in 2016 using a mirror to send and receive the lidar laser beam through the atmosphere horizontally, on a day chosen for its clear, horizontally-homogeneous conditions. The correction for the near-field overlap occurring within the laser beam is most significant below one km (Delgadillo et al., 2018), and has less influence on measurements of free tropospheric aerosol. The depolarization ratio is not sensitive to the overlap correction. The lidar is kept in a temperaturecontrolled environment, which reduces fluctuations in the lidar backscatter-to-extinction ratio, and is located approximately 10 100 m away from where the filter measurements are made. Increases in the volume depolarization ratio from 2015 to 2016 coincide with the incorporation of a new laser diode pump in May, 2016. This will increase the laser beam intensity, increasing the beam's ability to penetrate into the atmosphere, increasing the signal to noise of the signal in both channels (cross-and co-polarization), and ultimately increasing the δ v value. As such, the volume depolarization ratios reported here serve more as a qualitative indicator for the presence of the dust, than a quantitative measure, although more intense dust events are typically 15 associated with higher volume depolarization ratios (at the same laser intensity).
The MPL measures a backscattered intensity, which can only be related to the more physically-relevant aerosol volume extinction coefficient through a retrieval that is under-constrained. A sun-photometer-derived aerosol optical depth provides a vertically-integrated constraint on the extinction retrieval (Delgadillo et al., 2018). The two sun photometers, located near the lidar on a rooftop of a three-story building, are part of the Aerosol Robotic Network (AERONET; Holben et al., 1998). The 20 Miami sun photometer version 2 data were calibrated and cloud-filtered (level 2) for 2015, and solely cloud-filtered during 2014 and 2016 (level 1.5). The retrieval also produces a column-average ratio of the backscattered intensity to extinction, known as the lidar ratio. This will also vary with particle size. Because the vertical column above the lidar also contains sea salt in the lower atmosphere, the column-average lidar ratio is not that of the dust in the free troposphere. Lidar ratios range from 40-60 sr for dust and 15-25 sr for sea salt (Burton et al., 2012); a column-average lidar ratio less than that appropriate 25 for dust is consistent with an overestimate in the lidar-retrieved extinction. The optical properties of dust are not thought to vary with relative humidity , but those for sea salt, a hydrophillic aerosol, will. We estimate the overall uncertainty in the retrieved lidar extinction conservatively with a factor of three. The lidar vertical resolution is 30 m and the time resolution is 15 seconds (done to facilitate detection of the small clouds common to Miami).

GEOS-5 FP/MERRA-2 30
The GOES-5 model possesses 72 vertical layers, of which the mid-level of the lowest layer is at approximately 69 meters. Data from the lowest model layer are compared to the measured surface aerosol concentrations. The aerosol assimilation occurs eight times per day, and the output frequency of the full three-dimensional aerosol field is also every three hours. We primarily consider the dust and sea salt contributions. The GOCART dust emission parameterization depends on a source function, the near-surface wind speeds, and soil moisture (Ginoux et al., 2001). The dust sizes, prescribed at emission, can evolve thereafter during transport, with sedimentation the primary process capable of altering the dust size distribution. The parameterization of the dust mass distribution by size approximately follows that of Tegen and Lacis (1996) and contains five size categories, extending to a maximum diameter of 20 µm. These are indicated in Table 1 along with the density and mass extinction efficiency corresponding to each size category. Spheroidal particles are assumed, with more details on the dust parameterization available 5 in Colarco et al. (2010). The dust is assumed to not be hygroscopic.
The flux of sea salt off of the ocean surface is also parameterized into five size bins, but the size is allowed to vary with the surface wind speeds and sea surface temperature, and the sea salt can undergo hygroscopic growth as a function of the relative humidity. The sea salt parameterization is adapted from Gong (2003) and described further in Chin et al. (2002) and Bian et al. (2019). 10 Mass extinction efficiencies -the ratio of extinction to mass, specific for each diameter size range (see Table 1 for the dust mass extinction efficiencies) -connect the model aerosol mass mixing ratios to the assimilated aerosol optical depth. The clear-sky τ aer values at a wavelength of 550 nm are primarily assimilated from the Moderate Resolution Imaging Spectroradiometer (MODIS) and secondarily the Multi-Angle Imaging SpectroRadiometer (MISR) satellite instruments . Data from the surface-based AERONET sites are no longer assimilated , allowing the Miami 15 sun photometer data to independently validate the GEOS-5 FP τ aer s. The lowest model level is assumed to represent surface values.

HYSPLIT back trajectories
Daily average back-trajectories explore the differences in the integrated air flow for days with high and low dust mass concentrations. The calculations rely on the NOAA Air Resources Laboratory HYSPLIT model (Stein et al., 2015). These extend for 20 300 hours (12.5 days) and are driven by the NCEP Reanalysis Climate Diagnostics Center-1 product. Each day's trajectory is initialized at 00Z and at an altitude of 2000 m. This is a robust height for the presence of dust above Miami, as indicated later.
There is some evidence that the NCEP Reanalysis lower tropospheric winds may be weaker than observed (e.g., Adebiyi et al., 2015), which would increase the transport time, but this is not investigated further for this study.

25
A time series of the primary measurements displays the daily surface-based filter-derived dust mass concentrations and the corresponding vertical structure as captured by the lidar volume depolarization ratio, for each of the three dust seasons (Fig. 3).
The lidar measurements indicate that dust is often prevalent in the boundary layer, typically extends up to 2-3 km, and occasionally rises above 4 km but not above 5 km. The depolarization ratios indicate that when dust is present, it is typically also present within the boundary layer. This is corroborated by the filter measurements; only during mid-to-late August 2015 is no 30 dust detected within the atmospheric column by either the filters or the lidar. The synoptic and year-to-year variability is clear, with the highest dust mass concentrations occurring on just a few select days of each summer.
The back-trajectories indicate the crucial role of the north Atlantic subtropical high in guiding the atmospheric flow towards Miami ( Fig. 4(a-c)), explored further in Kramer et al. (2019). Many back-trajectories intersect the coastal region off of North Africa, with few continuing further eastward over the African continent. This may indicate a strong influence from the northerly coastal flow on the eastward side of the north Atlantic subtropical high. The HYSPLIT back-trajectories disregard explicit mixing with environmental air, but a back-trajectory with a large meridional component would imply more mixing of the 5 dusty air with air emanating from the northeast Atlantic and Europe. That this occurs is consistent with the co-existence of anthropogenic aerosols with dust ( Fig. 1). The separate influences of African versus Atlantic air on Miami and Caribbean thermodynamic profiles is also noted in Dunion (2011).
After twelve days, the integration confidence in the trajectories leading to the original location of air parcel is debatable, therefore we focus more on the trajectory characteristics over the Atlantic. Back-trajectories associated with days with dust  The GEOS-5 FP τ aer values slightly exceed those from the sun photometers at optical depths below 0.15, and are slightly less at τ aer > 0.15 (Fig. 5). The slight underestimation at higher τ aer s is also shown in Shi et al. (2019), in which it is attributed to missing emissions , and is also apparent in a comparison of MERRA-2 τ aer s to shipborne τ aer observations ( Fig. 9 of Randles et al., 2017). Surface albedo inhomogeneities at this coastal location could perhaps contribute to the overestimate at the lowest τ aer s, although another interpretation, shown next, is that an overestimate from the sea salt contri-25 bution becomes most apparent at the lowest τ aer s. An implication of the positive correlation, overall, is that the AERONET cloud screening is effective. Although not shown, the day-to-day variations in τ aer match well, as would be expected given the assimilation and availability of satellite-derived τ aer s.

Surface sea salt mass concentrations
A first step within the assimilation scheme is the aerosol speciation, and how much of the aerosol is partitioned into sea salt 30 will also affect the amount available for dust. The GEOS-5 FP aerosol product overestimates the near-surface salt mass con-centration by an order of magnitude, with a model mean value of 61 µg m −3 versus a measured mean salt mass concentration of 7.7 µg m −3 estimated from the sodium measurements (Fig. 6). The latter is consistent with previous measurements (Savoie and Prospero, 1982;Prospero, 1999a). Further assessment of the GEOS-5 wind speeds and relative humidity using surface meteorological data indicate that these are realistic (not shown), suggesting that the underlying size and/or light extinction parameterizations may be the issue. This is consistent with the findings of Bian et al. (2019), in which an inconsistency between 5 an overpredicted salt mass concentration and underpredicted τ seasalt in GOES-5 simulations is reconciled through the model adoption of too-large sea salt sizes. As noted by Buchard et al. (2017), the assimilation of τ aer s that exceed the forecast τ aer s can further exacerbate the discrepany.

Surface dust mass concentrations
All else equal, an overestimate of the GEOS-5 FP sea salt mass concentrations might point to GEOS-5 FP dust mass con-10 centrations that are too low, for the same assimulated aerosol optical depth. Instead, the GEOS-5 FP near-surface dust mass concentrations also exceed observed values on almost two-thirds of the days, although the overestimation is not as pronounced as that for sea salt, with a mean bias of 3.46 µg m −3 (Fig. 7). A clear correlation is apparent, matching that reported in Buchard et al. (2017) for MERRA-2 based on Barbados measurements (which did not find a systematic bias). GEOS-5 FP approximately captures the monthly evolution from July to September, although the GEOS-5 FP overestimate is most pronounced for 15 June, and is enough to alter the perception of the monthly evolution (i.e. the GEOS-5 FP maximum in 2014 and 2016 occurs in June, in contrast to a July (August) maximum in the observations). The cause for this is unclear.
The model overestimate in the surface dust mass concentrations has consequences for inferences of the frequency of high and low dust loading days (Fig. 8). Measured dust mass concentrations exceeding 40 µg m −3 or less than 1 µg m −3 were infrequent, with dust present in the Miami boundary layer at concentrations between 1 to 10 µg m −3 over one-half of the time. 20 In contrast, GEOS-5 FP include more days with concentrations exceeding 10 µg m −3 than are observed, and correspondingly fewer days with concentrations between 1-10 µg m −3 . GEOS-5 FP overall places too much dust in the boundary layer. That dust is consistently present in the boundary layer is not controversial (e.g., Reid et al., 2002).
4.4 Does dust size explain the difference between the GEOS-5 FP/MERRA-2 and measured dust mass concentrations? 25 The use of a size distribution parameterization that permits a larger mass for the same visible extinction (as is thought to occur for sea salt; Bian et al. (2019)) is one explanation for why the MERRA-2 dust mass concentrations may exceed those observed, despite realistic assimilation-constrained aerosol optical depths. This idea is assessed using size-resolved dust mass concentrations from 17 days in 2016, collected in addition to the bulk filter samples. The 28 July to 9 August time period included one of the heaviest and longest dust mass concentration events from 2016, and data from another dust event spanning 30 1-2 September are also included. The trajectories for these days indicate the dust transport occurred directly over the Atlantic to Miami without passing over the Gulf of Mexico and recirculating back. By happenstance, the subtropical high was strongest and located furthest north in 2016 of the three years considered, based on the 1023 hPa sea level pressure contour (Fig. 4).
This may have decreased dust transport that year (Kramer et al., 2019), resulting in relatively fewer days with dust exceeding 20 µg m −3 for that year. The episodic-maximum dust mass concentration of 28 µg m −3 occurred on August 4, 2016 (Fig. 2).
Overall, the GEOS-5 FP size distribution is broader than that measured, with GEOS-5 FP placing 16% of the dust mass 5 concentration in diameters < 2 µm, with 46%, 32%, 5.4% and 0.05% in the subsequent larger size bins. In contrast, the measurements place 56% of the total dust mass concentrations in particles with diameters < 2.1 µm, with 18%, 21%, and 3.6% in the larger size bins. We note that the model size distribution places equal mass in the three size bins with diameters exceeding 2 µm, and that thereafter, the only mechanism that can alter the dust sizes is differential settling. This effect is evident in the GEOS-5 dust size distribution after long range transport, with most of the mass contained in the (2.0 -3.6 µm) 10 diameter range, and the mass distribution within the three larger size bins no longer equal. As has been previously noted by others (e.g., Kok, 2011), more of the measured mass is contained in particles with diameters > 6 µm, than in the model. Filters from 1 and 2 August indicate mass from particles with diameters > 10 µm, and these were interrogated further with scanning electron microscopy to verify the particle size (Fig. 10). The dust particle images independently corroborate the presence of larger sizes. Four of the five examples have a dimension exceeding 20 µm (see Table 2 for dimensions of each identified 15 particle), with the largest particle measuring 36.65 by 26.86 µm (Fig. 10b). The particles are highly aspherical, with aspect ratios ranging from 0.38 to 0.73. This could contribute to their survival, as aspherical particles will fall at slower terminal speeds than spheres of equivalent mass (e.g., Yang et al., 2013). These dust particles exceed the GEOS-5 upper limit of a 20 µm diameter. However, only 1 µg m −3 of the total measured dust mass concentration is contained in particles with diameters exceeding 10 µm, implying that the neglect of particles with diameters > 20 µm by GEOS-5 FP only results in a small error 20 in the total dust mass concentration after such long-range transport.
The more striking result is that the measurements place more than half of the mass in particles with diameters < 2 µm, or more than three times as much as does GEOS-5 FP. Particles with diameters < 2 µm possess larger mass extinction efficiencies than larger particles (Table 1), by at least a factor of three, implying that more mass is required to produce the same extinction for larger particles than for smaller particles. Fig. 9 indicates that one cause for the overestimated near-surace dust mass 25 concentrations in GEOS-5 FP above southern Florida may be a dust size parameterization that distributes most of the dust mass into a larger size, for which the mass extinction efficiency is lower.
We are not aware of other dust size distribution measurements gathered over coastal Florida against which to compare this study's measurements. Measurements made close to the African coast emphasize the presence of larger particles, with Ryder et al. (2013) reporting most of their measured dust mass in diameters exceeding 5 µm, consistent with Haywood et al. (2003).
Barbados  estimate most dust particles are 1.6-2.0 µm in diameter, which is more broadly consistent with the Miami impactor measurements. This is at first glance consistent with the in-situ dust mass concentrations reported in Jung et al. (2013), but these do not extend beyond 2.5 µm in diameter (most aircraft aerosol intake inlets cut off at 3 µm), thus do not resolve the larger sizes.

Dust vertical structure 5
The GEOS-5 FP dust mass mixing profile combined with the mass extinction efficiencies specific to each size range (Table   1) can be assessed using extinctions retrieved from the lidar backscattered intensities (Delgadillo et al., 2018). This is similar to the strategy invoked within Liu et al. (2012)  The time series of the lidar-derived extinctions (Fig. 11c,f,i and l, all corresponding to nighttime, when the lidar signal is more robust) indicate dust extending up to 4 km at times, stratified into distinct layers supporting what appear to be gravity waves (Fig. 11). A peak at 2 km is specific to the 28-30 July dust event (Fig. 3), and a secondary peak at approximately 800 Corresponding model aerosol extinction profiles, resolved over the 3-12 UTC time frame encapsulated within the average of the cloud-free lidar extinction profiles, indicate that the GEOS-5 FP product does not distribute enough of the assimilated τ aer s above 1.5 km on any of the four days (Fig. 11b,e,h,k). Most of the model τ aer is confined to the boundary layer, where 25 the τ aer also contains a substantial sea salt contribution given relative humidities capable of exceeding 80% (RH not shown).
The lidar extinction values within the boundary layer are much lower than those from GEOS-5 FP, and, are lower than those in the free-troposphere, despite including a sea salt contribution. Overall this assessment suggests that GEOS-5 FP places too much dust within the boundary layer, and not enough in the free troposphere above the boundary layer, where the dust can be advected further more easily. Where GEOS-5 FPl does place dust within the free troposphere, its vertical distribution is 30 reasonable, however. Agreement between the model-derived and observed lidar extinctions, for the same dust mass concentrations, can in theory be produced through the application of higher mass extinction efficiencies. The model tendency to overestimate the mean particle size could also be one reason why the model may place too much of the dust at lower altitudes (Fig. 9). Since larger https://doi.org/10.5194/acp-2020-1 Preprint. Discussion started: 4 March 2020 c Author(s) 2020. CC BY 4.0 License. particles fall more quickly, they should be preferentially located at lower altitudes. Size-resolved model dust mass mixing ratios for 28 July 2016 indicate that larger particles do prefer lower altitudes (Fig. 12), and in particular much fewer particles with diameters > 6 µm occur above the boundary layer. In the two size ranges with the largest discrepancy between the model and measurements (2-3.6 µm versus 0.2-2 µm), the altitude difference between the vertical distribution of the two size-resolved dust mass mixing ratios is only on the order of 100-300 m, however.

Conclusions
Dust forecasts incorporating the assimilation of satellite-derived τ aer can circumvent a difficulty global aerosol models otherwise encounter, in which τ aer s tend to be too low further away from a source region (e.g., Kim et al., 2014;Evan et al., 2014;Ansmann et al., 2017). As such, the assimilation of observed τ aer holds the promise of more accurate depictions of the global aerosol distribution. While the GEOS-5 FP aerosol forecasts and MERRA-2 capture independently-measured τ aer s 10 and their variability well for an aerosol environment dominated by sea salt and dust transported far from its source region, the more challenging objectives of realistic aerosol vertical distributions and surface mass concentrations are less well met. A clear overestimate in the modeled sea salt loading, apparent in an in situ comparison (Fig. 6) and a comparison of model-derived extinctions to those from a lidar (Fig. 11), is corroborated by the independent findings of Bian et al. (2019), will foster an underestimate in the dust loading, for the same assimilated τ aer .

15
A comparison to dust mass concentrations measured at the surface and to lidar profiles of retrieved extinction indicates that GEOS-5 FP often distributes dust too low in the atmosphere, with too much mass placed in particle sizes that are larger than observed (except at the largest sizes). An overestimate of the amount of dust in the boundary layer has implications for modeldeduced ocean fertilization by the soluble iron (e.g. Colarco et al., 2003), and for cloud nucleation. Size-resolved measurements place most of the mass in diameters smaller than 2 µm, while GEOS-5 FP/MERRA-2 places most of the dust mass in diameters 20 between 2 to 3.6 µm. The prescribed dust mass extinction efficiency, by which the dust portion of the assimilated aerosol optical depth is cast as a dust mass mixing ratio, is more than a factor of 3 larger for the smaller size (Table 1). Thus, a model dust size distribution that places more of the dust in the smaller sizes, but otherwise using the same prescribed size-resolved mass extinction efficiencies, will place less dust mass within the boundary layer. A relative increase in the amount of dust within the smaller sizes in the free troposphere may be important for the overall large-scale spatial distribution of the dust, although the 25 vertical distribution of dust differs little for the smaller sizes (Fig. 12).
The overestimate in the model dust size after long-range transport is opposite to that documented for most global aerosol models, in which the number of small particles can be overestimated relative to the large particles (e.g., Kok, 2011). We recognize a recent emphasis on the presence of very large dust particles within the Saharan Air Layer (e.g., Ryder et al., 2019), and their ability to be transported for long distances. Large aspherical dust particles are also detected at Miami after a transport 30 of at least 20,000 km, exceeding the amount modeled, but their contribution to the overall mass is negligible at 1 µg m −3 .
Their presence is consistent with other observations that perceive little mixing of African dust with other air masses en route for select cases (Karyampudi et al., 1999), and find larger dust particles closer to the top of the Saharan air layer (Jung et al., 2013;Yang et al., 2013;Gasteiger et al., 2017). The existence of the large particles may be more typical of the dust events that advect more directly to Miami and undergo little precipitation. The neglect of particles exceeding the maximum GEOS-5-specified particle diameter of 20 µm introduces little error in the total GEOS-5 dust mass concentrations, but may be more important for the direct aerosol radiative effect, which is not addressed here. Overall global aerosol models with and without aerosol assimilation benefit from a more realistic modeling of dust particle size and its evolution with transport and age after 5 emission (Adebiyi et al., 2019).
Our dust size measurements, at a location further away from the dust source than Puerto Rico (Reid et al., 2002Maring et al., 2003b, a) and Barbados (Jung et al., 2013;Weinzierl et al., 2016), are difficult to compare in a consistent manner to those reported in previous studies. In-situ measurements made during the Puerto Rico Experiment suggest most of the dust mass resides in diameters between 5-7 µm, while recent lidar measurements tend to perceive larger concentrations of aerosols 10 with diameters < 2 µm (Haarig et al., 2019). Reasons for such discrepancies are not entirely understood. Neither the impactor measurements nor the GOCART dust size parameterization discriminates for aerosol diameters < 2 µm, and published dust size distributions are not always directly comparable. Although critiques can be made of each individual measurement presented within this study, in their totality a consistent interpretation emerges based on retrieved lidar extinctions, and near-surface dust and sea salt mass concentrations, suggesting the model mean dust sizes, by mass, are too large, with relatively too much dust 15 mass placed in the boundary layer. The total dust loading is too low within GEOS-5 FP and MERRA-2, possibly because too much of the assimilated aerosol optical depth is speciated into sea salt. We consider the current analysis a pilot study, however, and recommend a further dedicated assessment with a multi-wavelength depolarization extinction lidar as well as a more complete set of in-situ measurements, both at the ground and in the free troposphere. This will better anchor ideas for changes to the GEOS-5 sea salt and dust size parameterizations. Author contributions. PZ designed the study and led the writing of the manuscript. SK carried out the data collection and initial analysis at

25
Miami and drafted the initial manuscript. The GEOS-5 FP and MERRA2 values were provided by RG. RD provided the lidar analysis. CA contributed to the Miami data collection, with LC providing input on the methodology and additional mentorship of SK. AB provided the SEM analysis. PC provided substantial insight into the GEOS-5 and MERRA2 methodology. All authors commented on the manuscript.

Competing interests. No competing interests are present
Prosprero into the dust filter sampling program at U of Miami, its analysis and intellectual interpretation.  Table 2. Measured size of imaged particles in Fig. 9.     Figure 9. a) Daily-mean dust mass concentration over each diameter bin for each of the 17 sample days in measurements and b) GEOS-5 FP (black). c) 17-day-average dust mass concentration for each diameter bin from measurements (blue) and GEOS-5 FP (black). Note that GEOS-5 FP and the measurements fall within separate diameter bins (see Table 1).