Global Dust Optical Depth Climatology Derived from CALIOP and MODIS Aerosol Retrievals on Decadal Time Scales: Regional and Interannual Variability

Abstract. We present a satellite-derived global dust climatological record over the last two decades, including the monthly mean visible dust optical depth (DAOD) and vertical distribution of dust extinction coefficient at a 2º (latitude) × 5º (longitude) spatial resolution derived from CALIOP and MODIS. Dust is distinguished from non-dust aerosols based on particle shape information (e.g., lidar depolarization ratio) for CALIOP, and on dust size and absorption information (e.g., fine-mode fraction, Angstrom exponent, and single-scattering albedo) for MODIS, respectively. On multi-year average basis, the global (60° S–60° N) and annual mean DAOD is 0.029 and 0.063 derived from CALIOP and MODIS retrievals, respectively. In most dust active regions, CALIOP DAOD generally correlates well with the MODIS DAOD, with CALIOP DAOD being significantly smaller. CALIOP DAOD is 18 %, 34 %, 54 % and 31 % smaller than MODIS DAOD over Sahara Deserts, the tropical Atlantic Ocean, the Caribbean Sea, and the Arabian Sea, respectively. Over East Asia and the northwestern Pacific Ocean (NWP), however, the two datasets show weak correlation. Despite these discrepancies, CALIOP and MODIS show similar seasonal and interannual variations in regional DAOD. For dust aerosol over NWP, both CALIOP and MODIS show a declining trend of DAOD at a rate of about 2 % yr−1. This decreasing trend is consistent with the observed declining trend of DAOD in the southern Gobi Desert at a rate of −3 % yr−1 and −5 % yr−1 according to CALIOP and MODIS, respectively. The decreasing trend of DAOD in the southern Gobi Desert is in turn found to be significantly correlated with an increasing trend of vegetation and a decreasing trend of surface wind speed in the area.



Introduction
Mineral dust, referred to as dust for short, is one of the most abundant type of atmospheric aerosol in terms of dry mass (Textor et al. 2006;Yu et al. 2012;Kok et al. 2017). Dust aerosol directly interacts with both solar and thermal infrared radiation, known as the direct radiative effect, and thereby influences the Earth's radiative energy budget (Kok et al, 2017;Song et al., 45 2018;Di Biagio et al. 2020).Dust also influences the life cycle and properties of clouds by altering the thermal structure of the atmosphere (known as semi-direct effects) (Hansen et al., 1997) and by acting as cloud condensation nuclei (CCN) and ice nuclei (IN) (known as indirect effects) (Albrecht 1989;Rosenfeld and Lensky 1998;Twomey 1977). Dust storms and plumes can degrade air quality and generate adverse impacts on human health (Griffin, 2007;Querol et 50 al., 2019). Dust also contains a variety of nutrients and the deposition of dust during transport provides essential nutrients to marine and terrestrial ecosystems (Jickells et al. 2005;Yu et al., 2015b). The deposition of dust on snow reduces the snow albedo and promotes snow melting (Painter et al., 2007). All these impacts manifest the important role of mineral dust in the Earth systems (e.g. Evan et al., 2006;Lau & Kim, 2007;Miller & Tegen, 1998;Shao et al., 2011) 55 Dust production is sporadic in nature and it can be transported on intercontinental, hemispherical, and even global scales (Grousset et al. 2003;Uno et al. 2009;Yu et al. 2012Yu et al. , 2013. Thus, global and routine measurements of dust spanning over years or even decades are vital for studying dust transport and deposition, estimating the dust radiative effects, and 60 evaluating and constraining dust simulations in numerical weather and climate models. Satellite remote sensing is the only means to observe dust on regional to global scales. Satellite remote sensing techniques usually retrieve the optical depth or extinction profile for total aerosol in the https://doi.org/10.5194/acp-2021-1 Preprint. atmosphere with additional retrievals of particle size, shape, or absorption properties that are sensor specific. Passive sensors, such as the Total Ozone Mapping Spectrometer (TOMS) 65 (Prospero et al., 2002), Ozone Monitoring Instrument (OMI) (Chimot et al. 2017), Multiangle Imaging SpectroRadiometer (MISR) (Ge et al., 2014 andY. Yu et al. 2019), Moderate Resolution Imaging Spectroradiometer (MODIS) (Ginoux et al., 2010;Remer et al.,2005;Yu et al., 2009), multi-angular and polarimetric POLDER/PARASOL measurements (Chen et al. 2018) and IASI (Klüser et al., 2011;Clarisse et al. 2019) are used to detect dust sources and track dust 70 plumes at global scales. On one hand, these passive sensors provide global or quasi global coverage of column integrated properties of aerosol with satisfactory temporal resolution. On the other hand, they do not provide the vertical structure of aerosol that is critical for studying aerosol-cloud interactions and aerosol influences on the thermal structure of the atmosphere.
Space-borne lidar systems, such as the Cloud-Aerosol Lidar with Orthogonal Polarization 75 (CALIOP) onboard the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) spacecraft (Winker et al., 2010) and the Cloud-Aerosol Transport System (CATS) onboard the International Space Station (Yorks et al. 2015) are able to provide the vertical structure of aerosol and clouds, albeit with limited spatial coverage. All these passive and active remote sensing observations have been used extensively in studies of the spatial and temporal 80 evolution of aerosol over the past decade (e.g., Proestakis et al. 2018).
A significant hurdle of applying satellite remote sensing measurements for dust studies is how to distinguish dust from other aerosol types in a quantitative way. While many studies have used total aerosol retrievals by focusing on regions and seasons where dust dominates, some 85 studies have developed sensor-specific methods of partitioning total aerosol into dust and non-https://doi.org/10.5194/acp-2021-1 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. dust components with varying assumptions (Kaufman et al., 2005;Kalashnikova et al. 2005;Dubovik et al. 2006;Ginoux et al., 2010;Yu et al., 2009Yu et al., , 2013Yu et al., , 2015a. In general, the dust separation methods are based on dust physical and optical properties such as their large size, their irregular or nonspherical shape, and absorption characteristics. For example, CALIOP dust 90 classification is mainly based on the fact that dust aerosols are nonspherical in shape and their lidar depolarization ratio is significantly larger than those spherical aerosols. In contrast, the wide spectral coverage of MODIS measurements enables the retrieval of aerosol particle size information, such as effective radius, fine-mode fraction (FMF), and aerosol extinction Angstrom exponent, as well as spectral gradient of absorption (decreasing of absorption from 95 UV to red) (Remer et al., 2005). The combinations of these retrievals provide the basis for dust separation and DAOD retrievals from MODIS. Some recent studies have also characterized dust distribution through integrating satellite measurements with other data sources and model simulations. For example, Voss and Evan (2020) Voss and Evan (2020) determined these characteristic FMFs from AERONET measurements. Voss and Evan (2020) also extended the MODIS-based method to AVHRR over-ocean retrievals with some assumptions and produced the long-term (1981-2018) record of dust optical depth. Gkikas et al. 105 (2020) developed a global fine resolution (0.1º x 0.1º) DAOD dataset for the period 2006-2017 by scaling MODIS retrieved AOD with the DAOD-to-AOD ratios provided by MERRA-2 (Modern-Era Retrospective analysis for Research and Applications, Version 2) reanalysis (Gelaro et al., 2017). Given that MODIS and other remote sensing measurements (e.g., MISR https://doi.org/10.5194/acp-2021-1 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. and AERONET) have been assimilated in the MERRA-2 reanalysis to constrain the aerosol 110 optical depth, the DAOD-to-AOD ratio reported by MERRA-2 is the same as that from the underlying GOCART aerosol transport model in the MERRA-2 reanalysis system.
In this study, we focus on the dust optical depth derived from CALIOP and MODIS with two major objectives. First, we produce a decadal (2007-2019) record of global DAOD and dust 115 vertical extinction coefficient profile climatology from the CALIOP observations, which represents an extension of the trans-Atlantic dust transport and deposition studies by Yu et al. (2015aYu et al. ( , 2015b, both in terms of spatial and temporal coverages. Second, we compare the CALIOP DAOD climatology with the MODIS DAOD over both land and ocean (Yu et al. 2020;Pu and Ginoux, 2018) to identify and understand their differences in terms of global dust 120 distribution and interannual variabilities including decadal trend in key dust regions. Our analysis goes beyond broad dust-laden regions by zooming into potential dust source areas, which provides important insights into local dust activities. A systematic comparison and better understanding of DAOD from the two sensors based on distinct retrieval algorithms is critical for applying satellite measurements to evaluate global dust modeling . In 125 comparison to some most recent studies (Voss and Evan, 2020;Gkikas et al. , 2020), our dust climatology is derived by using the satellite observations in a self-consistent way without blending in other measurements (e.g., AERONET) or models (e.g., MERRA-2) (see section 2 for details). As discussed in Yu et al. (2009), the self-consistent use of MODIS data could minimize the introduction of additional biases due to discrepancies in FMF between MODIS and 130 AERONET. Furthermore, we use the latest version 4.2 CALIOP products and version 6.1 MODIS products in this study to characterize the three-dimensional distributions of dust. rest of the paper is organized as follows. Section 2 provides a description of the methodology of deriving dust climatology from CALIOP and an overview of MODIS dust retrieval algorithm.
Section 3 provides the main results including analysis of CALIOP dust climatology data and its 135 comparison against MODIS dust data. Section 4 discusses the uncertainties in CALIOP as well as MODIS DAOD retrievals. Section 5 provides a summary of the study along with the main conclusions.

CALIOP Dust Detection and AOD Partition
CALIPSO is in a sun-synchronous polar orbit with an equator crossing time of around 13:30 local time and 98 orbit inclination. CALIOP is a two-wavelength (532nm and 1064nm) polarization-sensitive lidar onboard CALIPSO. CALIPSO orbit track repeats every 16 days, CALIOP sensor never provides global coverage due to its small footprint. At Earth's surface, the 145 diameter of CALIOP footprint is around 70m, with spacing distance of 333m between two adjacent footprints along the orbit track. CALIOP utilizes three receiver channels (one measuring the 1064nm backscatter intensity and two measuring orthogonally polarized components of the 532nm backscatter) to provide high vertical resolution 30-60m of aerosol and cloud structure profiles (Winker et al., 2009). 150 Aerosol subtype classification and a priori assumption of lidar ratio for specific aerosol type are critical for CALIOP aerosol retrievals. CALIOP Level 2 product has been validated by comparing with ground-based measurements. The comparison between aerosol subtypes in CALIOP level 2 V2.01 and NASA Aerosol Robotic Network (AERONET) aerosol types shows that 70% of the CALIOP and AERONET aerosol types are in agreement. Best agreement is 155 achieved for dust and polluted dust (Mielonen et al. 2009). Schuster et al. (2012) compared CALIOP AOD to the collocated AERONET AOD measurements and found a CALIPSO bias of −13%, corresponding to an absolute bias of −0.029 relative to AERONET AOD on global average. Further comparison between CALIPSO AOD measurements and the collocated AERONET AOD measurements for the columns that contain the dust subtype exclusively 160 showed a larger bias (i.e., −29% and corresponding absolute bias of −0.1), although they show a relatively high correlation of R=0.58; this indicates that the assumed lidar ratio (40 sr) for the CALIPSO dust retrievals is too low. Omar et al. 2013 showed that CALIOP AOD are lower than AERONET AOD especially for low AOD. Furthermore, they found that the median of relative AOD difference between CALIOP and AERONET (500nm) is 25% of AERONET AOD for 165 AOD > 0.1. CALIOP observations have been used widely in previous studies of the spatial and temporal evolution of dust aerosols over the past decade (Huang et al. 2007(Huang et al. , 2008Yang et al. 2012;Xu et al. 2016;Kim et al., 2019). It is important to note that these studies are regional in scope and they use the standard CALIPSO product and aerosol subtype classification algorithm (Omar et al. 170 2009). In the standard CALIPSO product, each detected aerosol layer is classified as one of the six subtypes: dust, polluted dust, polluted continental, smoke, clean marine and clean continental.
In the latest CALIOP version, another sub-type "marine-dust" is introduced (Kim et al. 2018). In these studies, the "dust" subtype or a combination of "dust" and "polluted dust" subtypes is categorized as dust. While the former assumption leads to an underestimate of dust due to 175 neglecting dust component in the "polluted-dust" subtype, the latter assumption results in an overestimate of dust because of accounting for non-dust component in the "polluted-dust" subtype. In order to better distinguish dust component from each CALIOP detected aerosol https://doi.org/10.5194/acp-2021-1 Preprint. layers, Yu et al. (2015a) developed an algorithm independent of the standard aerosol subtype classification to distinguish dust from non-dust aerosol by using their respective thresholds of 180 particulate depolarization ratio. They further used the derived three-dimensional distribution of dust extinction to quantify the trans-Atlantic dust transport and deposition and its implications for Amazon rainforest (Yu et al., 2015b. In this study, we use the methodology in Yu et al. (2015a) to derive the monthly mean dust 185 extinction profile under clear-sky conditions from the latest V4.20 CALIOP products on a global scale from 2007 to 2019. First, we select the clear-sky profiles based on the operational CALIOP vertical feature mask and cloud layer product. In order to increase the sampling, we define clearsky cases in this study either as columns that are completely cloud-free or with the presence of optically thin (cloud optical depth < 0.2) and high-level (cloud base > 7km) clouds. This is 190 justified that the presence of high-level optically thin clouds does not significantly affect the retrieval of aerosol layers below the clouds (Yu et al. 2015a). After clear-sky screening, we use the operational 5 km level 2 CALIOP aerosol profile product that contains aerosol depolarization, backscatter and extinction profiles over a global scale (Young et al. 2018) to derive the dust extinction profile. The depolarization ratio from CALIOP is a key variable for 195 detecting and distinguishing dust from non-dust aerosol. Backscatter by spherical particle largely retains the polarization of the incident light, resulting in a depolarization ratio of nearly zero. In contrast, dust particles are generally non-spherical in shape and large in size, which gives them non-zero depolarization ratio that is significantly larger than other types of aerosol. The cloudaerosol discrimination (CAD) score in the products gauges the level of confidence for a feature 200 being classified as aerosol or cloud. In this study, in order to screen out low-confidence aerosol https://doi.org/10.5194/acp-2021-1 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. and cloud discrimination, we select layers with CAD scores between −90 and −100 (high level of confidence for aerosol feature) by following . Aerosol profile product also provides extinction quality control flag (Ext_QC) to indicate problematic retrievals. This study only uses layers with Ext_QC values of 0, 1, 18, and 16 . Only nighttime 205 data are used to avoid sunlight interference in aerosol signals.
For each backscatter coefficient profile, we derive the fraction of dust backscatter to total backscatter ( ) at each altitude from the following equation where is CALIOP observed particulate depolarization ratio, and is a priori knowledge of depolarization ratios of dust and non-dust aerosols respectively. Clearly, the calculations of 210 in Eq. (1) rely on the a priori depolarization ratios of dust and non-dust aerosols (i.e., and ). To account for various types of non-dust aerosols with different depolarization ratio, we follow Yu et al. 2015a and assume 0.02 and 0.07 as lower and upper bounds for (Burton et al., 2012;Fiebig et al., 2002;Sakai et al., 2010). Dust aerosols have significantly larger depolarization ratio compared to non-dust aerosols. In order to account for the variability of dust 215 shape and size, we use 0.2 and 0.3 as lower and upper bounds for (Ansmann et al., 2012;Esselborn et al., 2009;Sakai et al., 2010). Given an observed dust depolarization ratio , the based on Eq.
(1) has the minimum value when = 0.30 and = 0.07 and the maximum value when = 0.20 and = 0.02. In order to account for this variability, the final is based on the mean of the lowest (i.e., = 0.30 and = 0.07) and the highest (i.e., = 220 0.20 and = 0.02) dust scenario. The DAOD is also calculated for low dust and high dust scenarios for uncertainty study in section 4. Dust backscatter coefficient profiles are derived by multiplying CALIOP total backscatter coefficient with the calculated from Eq. 1. In order to derive dust extinction coefficient from 225 dust backscatter coefficient, we assume dust lidar ratio (LR), i.e., extinction to backscatter ratio, of 44 sr at 532nm, consistent with CALIOP Version 4.20 operational retrieval (Kim et al., 2018).
The use of globally uniform LR could also induce uncertainty to the derived regional DAOD, which is discussed in section 4.

MODIS Dust Detection and AOD Partition
As described above, the CALIOP-based DAOD derivation mainly makes use of dust nonsphericity in shape to separate dust aerosol from others. Another important difference of dust aerosol from other types of aerosols is their relatively large size. This difference provides the basis for the dust separation and DAOD derivation scheme based on the Moderate Resolution 235 Imaging Spectroradiometer (MODIS) retrievals that is introduced in this section. this consideration, we choose to use the nighttime CALIOP product that is free of solar noise, in hoping that the better data quality would outweigh the diurnal difference between nighttime CALIOP product and daytime MODIS retrievals.

250
MODIS aerosol retrievals employ two complementary algorithms to achieve the global coverage. The Dark Target (DT) algorithm is applicable for the retrieval of aerosol loading and properties over dark surfaces, including ocean-water and vegetated land. The MODIS aerosol AOD retrievals over ocean are found within the retrieval errors of Δ = ±0.03 ± 0.05 relative to AERONET AOD measurements (Remer et al. 2005). An approach was developed in 255 previous studies to separate DAOD from other types of aerosol by using aerosol optical depth and fine mode fraction retrieved from MODIS DT retrieval over ocean (details can be found in Kaufman et al., 2005;Yu et al., 2009Yu et al., , 2020. Over land, MODIS aerosol properties including AOD, Angstrom exponent, SSA are retrieved from the Deep Blue (DB) algorithm (Hsu et al. 2004(Hsu et al. , 2013. The MODIS aerosol AOD retrievals over land are found within the retrieval errors 260 of Δ = ±0.05 ± 0.15 relative to AERONET AOD measurements (Remer et al. 2005). To separate dust from scattering aerosols, it is required that the single-scattering albedo at 470nm to be less than 0.99. Then a continuous function relating the Angstrom exponent to fine-mode AOD is used to separate dust from fine particles (more details can be found in Pu and Ginoux, 2018).
Overall, multi-wavelength observations from MODIS contains aerosol size information such 265 as fine-mode fraction and Angstrom exponent in the observed reflectance spectral pattern, which was used to separate dust aerosol from others in MODIS dust retrieval over ocean and land. In this study, the latest retrieved aerosol properties from MODIS Collection 6.1 are used. We use data from Aqua MODIS only, because Terra MODIS retrievals may generate spurious dust trend (Yu et al. 2020). In order to minimize cloud contamination and avoid the infrequent sampling to 270 bias DAOD in MODIS dust retrieval over ocean, we screen the data by requiring a minimum of 10 DAOD retrievals in a month.

Global Dust Climatology
Based on the dust detection and separation schemes of two sensors described above, we derived the following two datasets: 275 1. The monthly mean CALIOP-based total aerosol optical depth (TAOD) and DAOD, as well as the vertical extinction profile on a 2º (latitude) ×5º (longitude) spatial resolution grids for the period of 2007 -2019. This relatively coarse resolution is limited by CALIOP's sampling.
2. We combine the monthly mean Aqua MODIS over-ocean (Yu et al., 2020) and over-land (Pu and Ginoux, 2018) TAOD and DAOD on a 1º ×1º spatial resolution grids to get the monthly 280 mean MODIS-based TAOD and DAOD from 2003 to 2019. In order to compare with CALIOP-based dust climatology data, we aggregate the 1º ×1º MODIS-based data to 2º ×5º resolution grids.
3. For evaluation and comparison purpose (see section 3.1), we also produce a seasonal global distribution of conditionally sampled DAOD from CALIOP. Different from the 285 climatological DAOD introduced above, where we include all cloud-free cases in the average of dust extinction and DAOD regardless of the presence of dust or not. In other words, DAOD and dust extinction are assumed to be zero when no dust is detected. In the conditionally sampled DAOD calculation, we only average those cases where dust is detected (i.e., DAOD and dust extinction are non-zero). Therefore, the conditionally sampled 290 DAOD is directly related to the intensity of the detected dust events, whereas the climatological DAOD is determined by a number of factors including not only the intensity https://doi.org/10.5194/acp-2021-1 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. of the detected dust events but also the frequency of the dust events as well as the capability of the instrument to sample the dust events.

295
In this section, we compare shape-based CALIOP global dust retrieval against size-based MODIS dust retrieval, more specifically MODIS ocean dust retrieval from Yu et al. (2009Yu et al. ( , 2020 and land dust retrieval from Pu and Ginoux (2018), we analyze the similarities and differences between two dust climatology data and furthermore study seasonal cycle and decadal trend of dust aerosols based on these datasets. 300

Comparison between CALIOP and MODIS dust Climatology
The DAOD climatology datasets derived from the CALIOP and MODIS observations, as described in the last section, have two major sources of uncertainty: 1) The uncertainty associated with the AOD retrieval. The primary uncertainty sources in MODIS AOD retrieval include instrument calibration errors, cloud-masking errors, inappropriate 305 assumption of surface reflectance and aerosol model selection (Remer et al. 2005;Levy et al. 2013Levy et al. , 2018. Uncertainty sources in CALIOP aerosol retrieval include instrument calibration errors, errors in discriminating cloud from aerosol, uncertainties associated with the a priori assumption of lidar ratios, and the under detection of tenuous aerosol layers, and overestimation of the elevation height of heavy aerosol plume base (Winker et al. 2009;Yu et al., 2010;310 Schuster et al., 2012;Thorsen and Fu, 2015;Rajapakshe et al. 2017).
2) The uncertainty associated with dust detection and separation. As explained in section 2, CALIOP-and MODIS-based dust detection and separation methods are based on different characteristics of dust aerosols in comparison with other types of aerosols, as summarized in Table 1. CALIOP-based method makes use of the fact that depolarization ratio of dust aerosols is 315 much higher than other types of aerosols, primarily because of irregular non-spherical shape and also to a lesser extent because of coarse size of dust particles. MODIS-based method is largely based on the characteristics of coarse particle size. Over ocean, DAOD is derived from total aerosol AOD (TAOD) and fine mode fraction (FMF) with a priori characteristic FMF for individual aerosol types. Over land, DAOD is derived using spectral dependence of aerosol 320 extinction (i.e., Angstrom exponent) and single scattering albedo. In other words, MODIS retrieves DAOD based on dust size supplemented by absorption characteristics.
Given these retrieval uncertainties and methodological differences, some discrepancies between the two DAOD climatology datasets are expected. In this section, we will compare the two datasets to identify and understand their similarities and differences. Since the mechanisms 325 of dust generation, dust transport and dust removal processes all have a seasonal cycle Parrington et al. 1983), we first present and discuss dust spatial distributions for each season in this section. Table 2 summarizes the seasonal and annual mean DAOD and TAOD values averaged over ocean, land and the globe (all limited to 60° S-60° N), respectively, based on MODIS and CALIOP dust retrievals from 2007 to 2019. On multi-year average basis, the 330 global, annual mean DAOD (TAOD) is 0.029 (0.112) and 0.063 (0.167) according to CALIOP and MODIS retrieval, respectively. Generally, DAOD from two retrievals differ by a factor of about 3 over ocean and less than 2 over land, while TAOD differ by a factor of less than 2 over both ocean and land. The ratio of DAOD over land to that over ocean is about 2 and 3 for MODIS and CALIOP, respectively. For TAOD, the land to ocean ratio is about 2 for both 335 products. Overall, the difference in TAOD between two retrievals is less than their difference in DAOD. On a global average, both MODIS and CALIOP-based DAOD peaks in boreal summer These regions do stand out in MODIS DAOD maps (i.e., the second column in Figure 1). 350 Interestingly, these regions indeed show up in the DAOD to TAOD ratio plot based on both two sensors (i.e., the last two columns in Figure 1). One of possible reasons for this is that dust activities in those regions are more intermittent and CALIOP's narrow swath results in more frequent miss of detection than MODIS does. To test this hypothesis, we compare the seasonal climatological DAOD and conditional DAOD product. The second column of Figure 2 shows the 355 seasonal climatological DAOD which is the average dust load over a geographical domain and time interval. It contains information of both the intensity and frequency of dust activities. On the other hand, the seasonal conditionally sampled DAOD shown in the first column of Figure 2 eliminates the impacts from dust frequency by excluding dust-free cases in the average. It is mainly related to the intensity of observed dust events. Therefore, the comparison between 360 climatological and conditionally sampled DAOD sheds a light on the frequency and intensity of https://doi.org/10.5194/acp-2021-1 Preprint. Discussion started: 17 February 2021 c Author(s) 2021. CC BY 4.0 License. dust events. For example, the third column in Figure 2 shows the relative difference between conditionally sampled DAOD and climatological DAOD with respect to the climatological DAOD. In 'dust belt' regions, especially in Sahara Desert and Middle East where dust activities are persistent, climatological DAOD is very close to conditional DAOD. In Australia, Southwest 365 United State, South America and South Africa, however, the conditional DAOD (column 1 in Figure 2) and the difference (column 3 in Figure 2) are relatively high. This suggests that dust activities in those regions are highly episodic and/or occur in relatively small scales. As a result, the dust events in those regions are prone to be missed by CALIOP due to its once-a-day sampling over limited spatial coverage. Even if the episodic dust events are sampled by CALIOP, 370 the monthly averaging would diminish the sparse daily DAOD retrievals in those regions. Indeed, Prospero (1999) reported that dust signals were shown in the daily TOMS aerosol index (AI) product in those regions but were not captured in TOMS monthly-mean AI product. The difference also is very large in open oceans, suggesting that dust aerosols are present at a very low frequency. 375 Having analyzed the conditionally sampled DAOD from CALIOP, we now return to climatological DAOD and comparison between CALIOP and MODIS. Hereafter, all AOD values are climatological without otherwise explicit statement. Figure 3 shows the difference in seasonal mean TAOD, DAOD and the percentage of DAOD in TAOD between MODIS 380 retrievals and CALIOP retrievals. We note in Figure 3 that CALIOP-based DAOD is generally smaller than MODIS-based DAOD over Northeast Asia and Asian dust outflow region (Northwest Pacific-NWP). There could be several reasons for this. First, this region is a major outflow region of Asian pollution (Yu et al., 2020). It is possible that the internal mixing of dust aerosols with industrial pollution in this region changes the dust morphology making it less non-385 spherical (Li and Shao 2009) but larger in size, which leads to smaller depolarization ratio and smaller fine-mode fraction. As a result, CALIOP shape-based DAOD derivation method could not capture the dust particles contained in the mixture, while those dust particles can be captured by MODIS size-based method. Another potential reason could be associated with that dust plumes in this region are vertically dispersed (Yu et al., 2010;Su and Toon, 2011). These 390 tenuous dust layers are likely to go undetected by CALIOP because of its relatively low sensitivity. However, MODIS retrieves aerosol from the columnal integrated reflectance which is not dependent on the vertical distribution of aerosol.
The difference may also be caused by uncertainties in MODIS aerosol retrievals. The West Pacific Ocean is cloudy almost all year long (see the last column in Figure 3), which makes 395 MODIS aerosol retrievals being susceptible to cloud contamination. The cloud contamination can lead to an overestimation of TAOD but underestimation of FMF. Although the MODIS retrieval algorithm neither assume coarse particles are exclusively from dust aerosols nor assume dust particles are all coarse particles (Yu et al., 2020), coarse mode aerosols are primarily dust.
Thus, the overestimation of TAOD and underestimation of FMF will lead to an overestimation in 400 DAOD. An exception occurs during winter when cloud fraction is large in NWP. The MODISbased DAOD is smaller than CALIOP-based DAOD, even though MODIS TAOD is larger than CALIOP TAOD. Similarly, over the southeastern Atlantic Ocean, CALIOP-based DAOD is also generally smaller than MODIS-based DAOD. On one hand, cloud contamination may have biased the MODIS dust retrieval high. On the other hand, CALIOP clear-sky sampling is not 405 large enough to capture some dust events in this region. In southern part of Sahel and India, MODIS-based DAOD is generally smaller than CALIOP-based DAOD.
We further compare DAOD ( Figure 5) and TAOD ( Figure S1 in the supplementary) retrievals from CALIOP and MODIS over major dust laden regions (as shown in Figure 4), including three source regions on land (i.e., Sahara Desert, Middle East and Eastern Asia) and 410 six oceanic outflow regions (i.e., the tropical Atlantic Ocean -TAT, the Caribbean Basin -CRB, the Mediterranean Sea -MED, the northwest Pacific Ocean -NWP, the Arabian Sea -ARB as well as the tropical Indian Ocean and the Bay of Bengal -IND). Each data point in the scatter plot represents a monthly mean DAOD (or TAOD) in a 2º × 5º grid. The density of data is represented by different color. To avoid our analysis being biased by some extreme and rare 415 cases, we exclude those data points within the lowest 5% of data density (grey points in Figure 5).
Overall, the DAOD from the two instruments correlate well (R 2 > 0.5) and on average CALIOPbased DAOD is 18%, 34%, 54% and 31% lower than MODIS-based DAOD over the Sahara  (Figure 5h) regions, respectively. Over the Sahara Desert, the good agreement in DAOD between the two sensors (bias of 18% and R 2 = 420 0.61) suggests that over the Sahara Desert dust particles can be adequately characterized by both irregular non-spherical shape and coarse size. As a result, both CALIOP-and MODIS-based methods are able to detect and separate the dust from other types of aerosols. In TAT and ARB regions, two instruments correlate well (R 2 > 0.7) in both DAOD and TAOD. For TAOD, CALIOP is smaller than MODIS by 2% in TAT and larger than MODIS by 15% in ARB. By 425 comparison differences in DAOD are larger, with CALIOP DAOD lower than the MODIS DAOD by 34% and 31% in TAT and ARB, respectively. This suggests that the differences in DAOD from the two instruments are mainly resulted from differences in the dust separation method. In East Asia and NWP, on contrast, both TAOD and DAOD show poor correlation between the two methods (Figure 5c, 5g, S1(c) and S1(g)). As discussed earlier, the poor 430 correlation between the two methods may be contributed by many factors. For example, the total TAOD retrievals from MODIS are subject to larger uncertainties due to cloud contamination, or the DAOD retrieval from CALIOP may miss spherical dust particles that are coated by large combustion emissions from East Asia. shows a peak in late spring or even summer months for some years. This could have resulted from cloud contamination in MODIS retrievals due to the large cloud fraction in summer [Yu et 445 al., 2020]. In addition, a secondary maximum of dust activity with high elevation plume in summer over the Taklamakan desert (Ginoux et al., 2001) may also contribute to the seasonality trend captured by MODIS over NWP.
Compared to the MODIS dust retrieval, CALIOP has a unique capability of detecting dust aerosol vertical distribution. summertime dust aerosol has the highest DAOD and reaches to the highest altitude extending from surface up to 6km in altitude. 455 The analysis above has been performed over the broad dust-laden regions. Here we focus on MODIS and CALIOP comparison in major potential source areas (PSAs) for dust in North Africa, namely NAF-1 to NAF-6 as illustrated in Figure 8 (adapted from Fig. 1 in Formenti et al., 2011). Among all dust source regions around the globe, the Sahara Desert and its margins in North Africa are the largest dust emitter. Within this region, prominent dust sources are often 460 associated with topographical lows and foothills of mountains (Prospero et al. 2002). Figure 9 shows scatterplots of CALIOP DAOD against MODIS DAOD over the six PSAs (corresponding scatterplots for TAOD are shown in Figure S2 (Prospero et al. 2002), and dust activity in the region occurs with a high frequency during all seasons except fall . However, CALIOP TAOD and DAOD are much smaller than MODIS retrievals in this region. In terms of dust seasonality (Figure 10 In summary, MODIS and CALIOP DAOD show largest differences under the following conditions: (1) highly cloudy oceanic regions and (2) dust-pollution internal mixtures with high relative humidity. Their differences can be explained as follows.
1. Over cloudy ocean, cloud screening is critical to the quality of aerosol retrievals. As an 480 active sensor, CALIOP is more reliable in discriminating clouds and aerosols than passive imager MODIS. In addition, active sensor is able to avoid impact from cloud side scattering. Therefore, MODIS is subject to more cloud contamination than CALIOP.
Large cloud contamination in MODIS results in overestimation in TAOD and underestimation in FMF, introducing a high bias in DAOD over ocean cloudy regions 485 (e.g., NWP).

2.
Pure dust particles are hydrophobic and will not absorb water vapor. However, for dust aerosols coated by other types of aerosols (such as the deliquescent dust-nitrate Ca(NO3)2) and saline mineral dust particles emitted from saline topsoil in arid and semiarid areas (Tang et al. 2019), those types of dust particles will take up water vapor 490 and grow to be larger in size and more spherical in shape (Wu et al. 2020). This phenomenon is most prominent for dust aerosols in polluted region (e.g., EAS) as well as with relatively high relative humidity. While such coarse spherical dust particles will not be accounted as dust in CALIOP shape-based method, they are categorized as dust in the MODIS size-based method. 495

DAOD Inter-annual variation from CALIOP and MODIS observations
In this section we examine the inter-annual variation of DAOD captured by two sensors over some major dust source and outflow regions. Figure 11   Dust over NWP comes mainly from East Asian dust sources. The broad East Asian region (ESA defined in Figure 4) show statistically significant DAOD decreasing trends (Figure 12c) which is consistent with the DAOD decreasing trend in NWP. It is also imperative to further 525 examine which of six major PSAs in East Asia (ESA-1 to ESA-6 in Figure 7) contribute to the decreasing trend of DAOD. As shown in Figure 13, among the six PSAs, the satellite data show consistent interannual declining trend of DAOD in EAS-5 (Southern Gobi Desert) at a rate of −4.8% yr -1 and −2.8% yr -1 for MODIS and CALIOP, respectively. In spring, DAOD in EAS-5 shows a significantly declining trend at a rate of −5.6% yr -1 and −3.3% yr -1 for MODIS and 530 CALIOP ( Figure S4). Figure 14 assesses the correlation between DAOD in EAS-5 and DOAD in NWP based on MODIS and CALIOP, respectively. For annual mean DAOD from 2007 to 2019, both sensors show a good correlation between EAS-5 and NWP with 2 ≈ 0.4 ( = 0.02). In spring, the correlation of DAOD from two regions is slightly reduced based on CALIOP ( 2 = 0.36 , = 0.03), while a much weaker correlation ( 2 = 0.28, = 0.07) was found based on 535 MODIS. We further examine potential factors contribute to the declining trend of DAOD in ESA-5. The first row in Figure 15 shows the springtime trend of MODIS enhance vegetation index (EVI), MERRA2 near-surface (at 10 m) wind speed and precipitation in EAS-5 region.
While EVI shows a significantly increasing trend with R 2 = 0.71 (p<0.05), the surface wind speed shows a decreasing trend with R 2 = 0.36 (p<0.05). There is no significant trend for 540 precipitation. The second and third row in Figure 15 shows the correlations of the three factors with MODIS DAOD and CALIOP DAOD, respectively. Clearly, EVI is anti-correlated with both MODIS and CALIOP DAOD with R 2 > 0.42 and p<0.05. While the surface wind speed is correlated with MODIS DAOD with R 2 = 0.53 and p<0.05, its correlation with CALIOP DAOD is weaker (R 2 = 0.29 and p=0.06). Note that EVI and surface wind speed are not independent 545 https://doi.org/10.5194/acp-2021-1 Preprint. variables that affect dust emissions. An increase of EVI or vegetation cover could reduce the surface wind speed. However, given the relatively coarse resolution of MERRA2, the surface wind speed trend may largely reflect the change in atmospheric circulations other than local wind decrease induced by more vegetation. The precipitation shows no statistically significant correlation with MODIS and CALIOP DAOD. 550 As discussed earlier, our results suggest that the decrease of NWP DAOD is likely a result of the decreasing dust events in Asian deserts (i.e., EAS-5 Gobi) in turn likely due to change of vegetation. This is also reported in several recent studies. Sternberg et al. (2015) found that Gobi

Uncertainty Analysis
The uncertainty of CALIOP DAOD retrieval come from several sources: One is some technical uncertainty such as instrument calibration errors, errors in discriminating cloud from aerosol and failure to detect aerosol layers (including tenuous aerosol layer and the lower part of 585 heavy dust layer. For example, Thorsen and Fu (2015) estimated that CALIOP may have underestimated 30%-50% in the magnitude of aerosol direct radiative effect due to its low sensitivity to tenuous layer), which is likely to translate into low bias in DAOD. In heavy aerosol conditions (e.g., strong dust storms in source regions and outflow regions), CALIOP laser cannot penetrate to the bottom of aerosol layer due to the laser attenuation (Chamara et al., 2017). As a 590 result, CALIOP AOD is biased low. DAOD is also subject to uncertainty due to the assumption https://doi.org/10.5194/acp-2021-1 Preprint. of dust lidar ratio (extinction to backscatter ratio). Different desserts produce dust with different minerology, thus different lidar ratio. Voss et al., (2001) measures LR for African dust as 41 ± 8 sr using a micropulse lidar and Liu et al. (2002) measures LR for Asian dust as 42-55 sr.
Globally observed lidar ratios are summarized in Müller et al., (2007) and Baars et al., (2016). 595 Typical lidar ratio values for desert dust aerosols range from 35sr to 55sr. This study assumes dust lidar ratio to be 44 sr at 532nm, which is the value used in the CALIOP V4 product (M.-H. Kim et al. 2018) and is comparable to previous studies and nevertheless induce potential uncertainties to DAOD. When separating dust from non-dust aerosol, the choice of depolarization ratio for dust aerosols and non-dust aerosols also introduces uncertainty in DAOD. 600 To quantify the uncertainty caused by DPR selection, we also calculated DAOD in the lowest ( = 0.30 and = 0.07) and the highest ( = 0.20 and = 0.02) dust fraction scenarios.
We estimated that the uncertainty in monthly DAOD is 35%-47% in regions with DAOD larger than 0.06 and up to 80% in regions with very low DAOD.

605
MODIS dust detection is also subject to some uncertainties for both over ocean and over land retrievals. Over ocean, the persistent presence of clouds in some regions (e.g., North Pacific Ocean, southeastern Atlantic Ocean) pose a challenge to MODIS aerosol retrievals, probably causing a high AOD bias, and low FMF bias, and thereby a high DAOD bias. In addition, DAOD was calculated from the MODIS-retrieved AOD ( ) and FMF ( ) with appropriate 610 parameterizations of marine aerosol AOD ( ), FMF of dust ( ), combustion ( ) and marine ( ) aerosols. All the parameterizations could also introduce uncertainty in the derived DAOD, in particular on a regional basis (see details in Yu et al. 2020). Over land, the derived MODIS DAOD represents the coarse-mode fraction of dust only. The exclusion of fine mode of dust aerosol at emission could induce a less than 10% underestimation of the emitted mass (Kok et al. 615 2017). The comparison of Aqua MODIS DAOD retrievals against AERONET coarse-mode AOD shows that Aqua MODIS DAOD values are underestimated with an error of 0.08+0.52DAOD (Pu and Ginoux, 2018).
A rigorous way to evaluate these uncertainties and validate the two dust detection methods is to compare with an independent measurement of DAOD. AERONET measurements 620 have been considered as ground truth and often used to evaluate satellite aerosol optical depth retrievals. However, so far there is not a valid method to derive DAOD from AERONET AOD measurements to compare our results with. Some studies use coarse-mode AOD from AERONET measurements as a proxy for DAOD (Pu and Ginoux, 2018), while CALIOP-based DAOD retrieval and MODIS-based oceanic DAOD retrieval do not assume dust aerosols are 625 exclusively coarse particles. Therefore, AERONET measurements could not be used to validate DAOD retrievals in this study. frequency by excluding dust-free cases in the average. The comparison between DAOD climatology data and conditional DAOD data suggests that dust activities in those regions are highly episodic. As a result, the dust events in those regions may be missed by CALIOP which 640 has a very limited spatial sampling coverage.

Summary and Conclusion
CALIOP distinguishes dust aerosols based on its non-spherical shape, whereas MODIS separates dust aerosols from others based on its large size characteristics. The discrepancy in dust retrieval based on two instruments are expected due to the uncertainty associated with their TAOD retrieval and the uncertainty associated with their different mechanism in dust detection 645 and separation. The comparison between CALIOP dust retrieval and MODIS dust retrieval facilitate a better understanding of advantages and limitations of each dust product and also provide some insights on dust morphology and dust size. Through the comparison, we found generally CALIOP-based DAOD correlates well with MODIS-based DAOD over dust-laden regions such as Sahara, TAT, CRB and ARB, but with CALIOP-based DAOD 18%, 34%, 54% 650 and 31% lower than MODIS-based DAOD over those regions respectively. This result is consistent with the different treatment of the dust-pollution mixtures in the dust separation approaches of two instruments. The better agreement (k=0.82) and correlation (R 2 =0.61) in Sahara Desert suggest that dust aerosols are irregular non-spherical and at the same time large in size in this region. In some regions such as NWP, the DAOD correlation between two sensors is 655 quite low. There could be many reasons for this, for example, the total TAOD retrievals from MODIS have larger uncertainty due to cloud contamination, or the DAOD retrieval from CALIOP may miss coarse spherical dust-pollution mixtures.
The interannual variability of DAOD over dust-laden regions show no clear trend except the NWP region at a rate of −1.6% −1 and −1.7% −1 based on CALIOP and MODIS 660 respectively, this trend is mainly attributed to the decreasing trend in spring with a rate of