2.1.1 Neural network aerosol classification algorithm

NATALI (Nicolae et al., 2018a) is an automated, optical-property-dependent aerosol layer classification algorithm. The typical aerosol profiles ($\mathrm{3}{b}_{\mathit{\lambda }\mathit{\alpha }}+\mathrm{2}{e}_{\mathit{\lambda }\mathit{\alpha }}+$ PLDR, optional; in NetCDF format) from the EARLINET database are used as inputs in order to retrieve the mean aerosol optical properties within the layer boundaries indicated by the gradient method (Belegante et al., 2014). The learning process of the ANN has been performed using a synthetic database developed by Koepke et al. (1997) along with the T-matrix numerical method (Waterman, 1971; Mishchenko et al., 1996) to iteratively compute the intensive optical properties of six pure aerosol types (the first six aerosol types in Table 1) presented to the artificial neural networks to perform the typing. The synthetic database is built for 350, 550, and 1000 nm wavelengths, which are then rescaled to the usual lidar wavelengths (i.e., 355, 532, and 1064 nm) using an Ångström exponent equal to 1. The mixtures are obtained through a linear combination of pure aerosol properties (Nicolae et al., 2018a).

Two classification schemes are used with different aerosol type (classification) resolutions when particle depolarization data are available. The first one is applied when all the high-quality aerosol optical parameters are provided (uncertainty of e_λα≤50 %, uncertainty of b_λα≤20  %, uncertainty of PLDR ≤30 %) and the aerosol typing is performed in the high-resolution (AH) mode. This means that the aerosol mixtures can be sufficiently resolved, providing the maximum number of output types (14 types, Table 1). In the second scheme, the values of the aerosol optical parameters have high uncertainty (uncertainty of e_λα>50 %, uncertainty of b_λα>20 %, uncertainty of the PLDR >30 %), and the typing is performed in the low-resolution (AL) mode. In this case, the number of output types is limited to six (first six aerosol types; Table 1). A “voting” procedure selects the most probable answer out of the two (possibly different) individual returns. The correct answer is selected based on a statistical approach considering two criteria: (i) which answer has a higher confidence and (ii) which answer is more stable over the uncertainty range (i.e., the percentage of agreement for values between error limits). Finally, there is the capability to perform the typing when the particle depolarization is not available, knowing that the mixtures cannot be resolved, so only the predominant aerosol type is retrieved (Nicolae et al., 2018a).

2.1.2 Mahalanobis distance aerosol classification algorithm

The automatic classification algorithm described in Papagiannopoulos et al. (2018) makes use of the Mahalanobis distance function (Mahalanobis, 1936) that relates an unclassified measurement to a predefined aerosol type. The method compares observations to model distributions that comprise EARLINET pre-classified data. Each Mahalanobis distance of an observation from a specific aerosol type is estimated, and the aerosol type is assigned for the minimum distance. Prior to the classification, two screening criteria are assumed to ensure correct classification. The algorithm is able to classify an observation to a maximum of eight (dust, volcanic, mixed dust, polluted dust, clean continental, mixed marine, polluted continental, and smoke) and minimum of four (dust + volcanic + mixed dust + polluted dust, mixed marine, smoke + polluted continental, and clean continental) aerosol classes depending on the lidar configuration.

In this study, we used four aerosol intensive properties: the backscatter-related Ångström exponent at 355 and 1064 nm, the aerosol lidar ratio at 532 nm, the color ratio of the lidar ratios, and the aerosol particle linear depolarization ratio at 532 nm, while the minimum accepted distance was set to 4.3 (Papagiannopoulos et al., 2018). As soon as the distance from a specific aerosol class is below the threshold and the remaining distances are higher than the threshold, the observation is assigned to that aerosol class. If more than one distance is below the threshold, the normalized probability of each class needs to be over 50 %.

This objective multidimensional classification scheme has found great applicability and has been used with lidar (e.g., Burton et al., 2012; Papagiannopoulos et al., 2018), space-based polarimetry (Russell et al., 2014), and spectral photometry (e.g., Hamill et al., 2020; Siomos et al., 2020) data. Voudouri et al. (2019) compared NATALI and the Mahalanobis distance classification algorithm for the EARLINET station in Thessaloniki. Their study used a 3b_λα+2e_λα lidar configuration and four aerosol classes (i.e., dust, maritime, polluted smoke, and clean continental) for each automatic algorithm. In general, they found fair agreement between MD and NATALI, and the differences were attributed to the class definition and the range of the class intensive properties.

2.1.3 Source Classification Analysis

SCAN is an automated aerosol layer classification process that is independent of optical properties and has been developed in the IDL programming language in the framework of this study. For each identified aerosol layer an X h HYSPLIT backward trajectory (Draxler et al., 2013) is used to calculate the amount of time traveled above predefined aerosol source regions before arriving over a lidar station at the specific date and height that the aerosol layer is observed. X is the number of hours of the backward trajectory, which can be decided by the user at the beginning of the process. SCAN assumes specific regions (Fig. 1, colored squares; from now on referred to as domains) in terms of aerosol sources (Penning de Vries et al., 2015). The polluted continental domains represent the regions with increased anthropogenic activity according to monthly means of tropospheric NO₂ from GOME-2 (Georgloulias et al., 2019).

Taking into account the information (latitude, longitude and height) from each HYSPLIT air-mass backward trajectory, SCAN implements a number of criteria: (i) if the geographical coordinates for the specific hour of the backward trajectory are within the boundaries of the marine domains and if the height of this trajectory over the domain is below 1 km (Wu et al., 2008; Ho et al., 2015), SCAN assigns this layer to the marine aerosol type. (ii) If the geographical coordinates for the specific hour of the backward trajectory are within the boundaries of the clean continental, polluted continental, or dust domains and if the height of this backward trajectory is below 2 km over the domain, SCAN assigns the specific hour to the clean continental, polluted continental, or dust aerosol type, respectively. (iii) For an hour to be assigned to the smoke aerosol type, except from the coordinates of the backward trajectory at this specific hour, which should be within the boundaries of the clean continental or polluted continental domains, the height of the trajectory at this specific hour should be below 3 km (Amiridis et al., 2010).

Figure 1SCAN's aerosol source type classification map. Different colored squares represent different aerosol sources: orange squares correspond to dust, blue squares to marine, brown squares to clean continental, and black squares to continental polluted aerosol sources.

SCAN draws fire information from the Fire Information for Resource Management System (FIRMS) (https://firms.modaps.eosdis.nasa.gov, last access: 1 August 2020) using data on actively burning fires, along with their location, time, and confidence value (in %) (Kaufman et al., 1998; Giglio et al., 2002; Davies et al., 2009; Justice et al., 2002) derived from the Moderate Resolution Imaging Spectroradiometer (MODIS). The selected time period is in accordance with HYSPLIT simulations (air-mass backward duration). From SCAN's point of view, a hotspot is assumed to be significant if the MODIS “confidence” value is higher than 80 % (Amiridis et al., 2010). In addition to the above criteria, the location of HYSPLIT air-mass backward trajectories at the specific hour must be a maximum distance of 8 km away from a hotspot of high confidence in order to be assigned as the smoke aerosol type.

SCAN performs the above classification process for all the HYSPLIT air-mass backward trajectories, and as a final step, it counts the hours that the air parcel spends above each geographical domain. If more than one domain is involved in following the backward trajectory's path, a mixture of more than one aerosol type is assumed. In case the aforementioned criteria (domain and height limitations) are not satisfied, the aerosol type is considered unknown.

Table 1Correspondence between the aerosol types, the shorthand used for this work, and the actual aerosol types defined with the NATALI, MD, and SCAN aerosol classification algorithms.

Download Print Version | Download XLSX

The maximum number of pure aerosol types that SCAN can assign to a layer is six, and the combination of them gives SCAN the capability to identify aerosols from different sources in a specific layer. In Table 1 one can find the pure aerosol types and the mixtures that SCAN has dealt with in this study. The whole classification procedure of SCAN is displayed in the flowchart in Fig. 2.

Figure 2Classification procedure of SCAN. This procedure is performed for each hour of the HYSPLIT back trajectory associated with each layer.

Download

3.1.1 NATALI versus MD

Figure 8a presents a comparison between the cases that MD classified as 1 type or a mixture against those classified as other by NATALI, while Fig. 8b shows the number of aerosol layers that MD classified as unknown and NATALI as 1 type. Finally, in Fig. 8c we present the number of aerosol layers that MD classified as unknown and NATALI as a mixture or other. The pie charts above each bar (Fig. 8b) and each stem (Fig. 8c) reveal the mean frequencies of each aerosol type as calculated by MD.

Figure 8(a) Comparison between the cases classified by MD as 1 type or mixture and as other by NATALI. (b) Number of aerosol layers classified by MD as unknown and as 1 type by NATALI. (c) Number of aerosol layers classified by MD as unknown and as mixture or other by NATALI. The pie charts above each bar and each stem reveal the mean frequencies of each aerosol type as calculated by MD.

Download

The inability of MD to classify the aerosol layers according to NATALI's classification (Fig. 8a) can be attributed to the characteristics of the aerosol layers not being well-modeled by the algorithm, which means that the intensive parameters are not within the accepted “borders” of the predefined classes of the algorithm. Concerning the 1 type aerosol layers according to NATALI (Fig. 8b), MD would have predicted them correctly if the labeling of these layers by MD was achieved by considering a higher percentage of aerosol types, as the pie charts above each bar indicate (Fig. 8b). This does not seem to be the case for the dust type labeled by NATALI, to which MD gave a 40 % probability to be continental polluted, 30 % smoke, and only 30 % dust and mixed dust (Fig. 8b). Concerning the mixture and other according to NATALI (Fig. 8c), it seems that MD found a high contribution of dust and mixtures of dust aerosols (approximately 50 %) inside these layers (yellow, orange, and dark red aerosol types according to the pie charts above the stems).

3.1.2 MD versus SCAN

Figure 9a presents a comparison between the cases that were classified by MD as 1 type and as a mixture by SCAN, while Fig. 9b shows the number of aerosol layers classified by MD as unknown and as 1 type by SCAN. Finally, Fig. 9c presents the number of aerosol layers classified by MD as unknown and as a mixture or other by SCAN. The pie charts above each bar (Fig. 9b) and each stem (Fig. 9c) reveal the mean frequencies of each aerosol type as calculated by MD.

Figure 9(a) Comparison between the cases that were classified by MD as 1 type and as mixture by SCAN. (b) Number of aerosol layers classified by MD as unknown and as 1 type by SCAN. (c) The number of aerosol layers classified by MD as unknown and as mixture or other by SCAN. The pie charts above each bar and each stem reveal the mean frequencies of each aerosol type as calculated by MD.

Download

The misclassification of an aerosol layer by MD compared to the classification as a continental polluted and smoke layer by SCAN (Fig. 9a) could again be attributed to the location of the observation compared to the location of the predefined aerosol types for the MD classification algorithm, which depend on the aerosol optical properties of the studied layers. Additionally, it seems that the aerosol optical properties of the mixture of continental polluted and smoke by SCAN are attributed to either clean continental or continental polluted aerosols by MD (Fig. 9a). From the cases that MD classified as unknown, eight are classified as continental polluted (cp) by SCAN (Fig. 9b), while another six cases are classified as continental polluted and smoke (cp + s) (Fig. 9c). Finally, from these unknown cases by MD, there are another 11 cases that SCAN classified as clean continental (two cases), dust (one case), marine (one case; Fig. 9b), clean continental and marine (one case), continental polluted and clean continental (one case), continental polluted and marine (two cases), continental polluted, smoke, and marine (one case), or continental polluted, clean continental, and marine (two cases; Fig. 9c).

3.1.3 SCAN versus NATALI

Figure 10 presents a comparison between the cases that (a) SCAN classified as a mixture and NATALI as 1 type, (b) SCAN classified as 1 type and NATALI as other, and (c) SCAN classified as a mixture and NATALI as other.

Figure 10Comparison between the cases classified by (a) SCAN as mixture and as 1 type by NATALI, (b) SCAN as 1 type and as other by NATALI, and (c) SCAN as mixture and as other by NATALI.

Download

From Fig. 10 it can be concluded that of the 12 cases that SCAN classified as continental polluted and smoke (cp + s), NATALI classified 6 of them as clean continental (cc), 4 as continental polluted (cp) (Fig. 10a), and 2 as continental polluted and marine/cloud-contaminated (cp + m/cc) (Fig. 10c). Moreover, of the six cases that SCAN classified as continental polluted (cp), three of them were classified as continental polluted and marine or cloud-contaminated (cp + m/cc), two as marine/cloud-contaminated (m/cc), and only one as continental polluted, smoke, and marine or cloud-contaminated (cp + s + m/cc) by NATALI (Fig. 10b). It seems that the aerosol optical properties of this mixture are attributed to either the clean continental or continental polluted aerosol types based on the NATALI classification.

3.2.1 Clean continental (cc) aerosols

Aerosol layers classified as clean continental by both NATALI and MD present medium LR₃₅₅ values (45–46 ± 5 SR), medium to low LR₅₃₂ values (37–39 ± 5 SR), medium $A_{b\mathrm{355}/\mathrm{532}}$ and $A_{b\mathrm{532}/\mathrm{1064}}$ values (1.0 ± 0.3), high $A_{e\mathit{\lambda }\mathrm{1}/\mathit{\lambda }\mathrm{2}}$ values (2.0 ± 0.3), and low PLDR values at 532 nm (3 ± 1 %). These values are in accordance with others reported in previous studies concerning cc aerosols (Ansmann et al., 2001; Omar et al., 2009; Giannakaki et al., 2010).

Table 4Mean values and standard deviations of aerosol optical properties according to each classification method.

Download Print Version | Download XLSX

3.2.2 Continental polluted (cp) aerosols

Aerosol layers classified as continental polluted by both NATALI and MD present medium LR₃₅₅ nm (57 ± 6 SR), slightly higher LR₅₃₂ values (62 ± 7 SR), medium $A_{b\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{532}/\mathrm{1064}}$, and $A_{e\mathrm{355}/\mathrm{532}}$ values (1.1–1.4 ± 0.3), and low PLDR values at 532 nm (3 ± 1 %). On the other hand, the aerosol layers similarly classified by SCAN present medium LR₃₅₅ values (50 ± 6 SR) and LR₅₃₂ values (49 ± 5 SR), medium $A_{b\mathrm{355}/\mathrm{532},}$ $A_{b\mathrm{532}/\mathrm{1064}}$, and $A_{e\mathrm{355}/\mathrm{532}}$ values (1.0–1.5 ± 0.3), and low PLDR values at 532 nm (3 ± 1 %). These values are also in accordance with those reported in previous studies concerning this type of aerosol (Müller et al., 2007; Giannakaki et al., 2010; Gross et al., 2013; Burton et al., 2013; Mattis et al., 2008). The similarity of these values to those of the clean continental aerosol type is the reason why it remains difficult to distinguish between these two aerosol types.

3.2.3 Smoke (s) aerosols

Smoke aerosol layers according to SCAN show medium LR₃₅₅ values (50 ± 5 SR), medium to small LR₅₃₂ values (37 ± 4 SR), medium $A_{b\mathrm{355}/\mathrm{532}}$ and $A_{b\mathrm{532}/\mathrm{1064}}$ values (0.9–1.3 ± 0.3), medium to high $A_{e\mathrm{355}/\mathrm{532}}$ values (1.6 ± 0.3), and low PLDR values at 532 nm (3 ± 1 %). These values are in accordance with those reported in previous studies concerning this type of aerosol (Wandinger et al., 2002; Müller et al., 2005, 2007; Burton et al., 2013; Baars et al., 2012; Balis, et al., 2003; Papanikolaou, et al., 2020). Again, the similarity of these values to those of the clean continental and continental polluted aerosol types is the reason why it remains difficult to distinguish between these three aerosol types.

3.2.4 Marine/cloud-contaminated (m/cc) aerosols

Aerosol layers classified as marine/cloud-contaminated by NATALI showed low values of LR₃₅₅ (35 ± 4 SR), even lower LR₅₃₂ values (27 ± 3 SR), small to medium $A_{b\mathrm{355}/\mathrm{532}}$ and $A_{b\mathrm{532}/\mathrm{1064}}$ values (0.9-1.0 ± 0.3), increased $A_{e\mathrm{355}/\mathrm{532}}$ values (1.7 ± 0.3), and low PLDR₅₃₂ values (4 ± 2 %). These values are in accordance with those reported by Cattrall et al. (2005), Burton et al. (2012, 2013), and Dawson, et al. (2015) concerning marine aerosols.

3.2.5 Dust and marine aerosols (d&m)

Concerning the dust and marine mixture according to the MD algorithm classification, these aerosols showed medium LR₃₅₅ values (43 ± 4 SR), low LR₅₃₂ values (46 ± 5 SR), small $A_{b\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{532}/\mathrm{1064}}$ (0.0–0.7 ± 0.2), $A_{e\mathrm{355}/\mathrm{532}}$ values (−0.2 ± 0.2), and medium PLDR₅₃₂ values (15 ± 5  %). These values indicate large and depolarizing aerosol particles, confirming the type of these particles as a mixture of dust and marine ones, according to Burton et al. (2012) and Papagiannopoulos et al. (2016).

3.2.6 Continental polluted and smoke (cp&s) aerosols

The continental polluted and smoke mixed aerosols classified according to SCAN showed medium LR₃₅₅ values (52 ± 8 SR), medium LR₅₃₂ values (47 ± 7 SR), medium $A_{b\mathrm{355}/\mathrm{532}}$ and $A_{b\mathrm{532}/\mathrm{1064}}$ values (0.8–1.3 ± 0.4), high values of $A_{e\mathrm{355}/\mathrm{532}}$ (1.5 ± 0.4), and low values of PLDR₅₃₂ (4 ± 2 %). The medium LR₃₅₅ and LR₅₃₂ values indicate continental polluted aerosols (Müller et al., 2007; Giannakaki et al., 2010; Gross et al., 2013; Burton et al., 2013; Papanikolaou, et al., 2020), while the high $A_{e\mathrm{355}/\mathrm{532}}$ values indicate smoke aerosols (Wandinger et al., 2002; Müller et al., 2005).

3.2.7 Continental polluted and marine (cp&m) aerosols

The continental polluted and marine mixture according to NATALI showed a large difference between the LR₃₅₅ and LR₅₃₂ values, with the latter being smaller (LR₃₅₅=69 ± 11, LR₅₃₂=31 ± 5.3 SR). The $A_{e\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{355}/\mathrm{532}}$, and PLDR at 532 nm showed low values (0.9 ± 0.4, −1.2 ± 0.4, 2.7 ± 1.6, respectively), while the $A_{b\mathrm{532}/\mathrm{1064}}$ showed large values (1.3 ± 0.2). SCAN showed medium LR₃₅₅ values (45 ± 6 SR), increased LR₅₃₂ values (55 ± 7 SR), low $A_{b\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{532}/\mathrm{1064}}$, and $A_{e\mathrm{355}/\mathrm{532}}$ values (0.3–0.7 ± 0.3), and low PLDR₅₃₂ values (8 ± 4 %). The low PLDR values are indicative of non-depolarizing aerosols such as continental polluted (Müller et al., 2007; Giannakaki et al., 2010; Gross et al., 2013; Burton et al., 2013) and marine aerosols (Gross et al., 2011; Burton et al., 2012, 2013; Gross et al., 2013). Additionally, the low $A_{b\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{532}/\mathrm{1064}}$, and $A_{e\mathrm{355}/\mathrm{532}}$ values are indicative of coarse-mode aerosols such as marine, while the increased LR₅₃₂ values are more indicative of continental polluted aerosols rather than marine ones.

3.2.8 Continental polluted, dust, and marine or clean continental (cp&d&m/cc) aerosols

Finally, the continental polluted, dust, and marine or cloud-contaminated aerosol mixture classified by NATALI showed medium LR₃₅₅ values (41 ± 4 SR), medium LR₅₃₂ values (43 ± 5 SR), low $A_{b\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{532}/\mathrm{1064}}$, and $A_{e\mathrm{355}/\mathrm{532}}$ values (−0.1–0.7 ± 0.3), and medium PLDR₅₃₂ values (13 ± 4 %). Here, the medium PLDR values are indicative of dust mixtures (Gross et al., 2011, 2016; Burton et al., 2013), while the low $A_{b\mathrm{355}/\mathrm{532}}$, $A_{b\mathrm{532}/\mathrm{1064}}$, and $A_{e\mathrm{355}/\mathrm{532}}$ values are indicative of coarse-mode aerosols, such as dust and marine.