From September 2015 to August 2016, a WIBS-4M was continuously sampling at the Cape Verde Atmospheric Observatory (CVAO), off the west coast of continental Africa (16∘51′49′′ N, 24∘52′02′′ W). The observatory is located 50 m from the coastline and is subject to persistent northeasterly trade winds. During this period, the observatory recorded a mean temperature of 23.3 ∘C and median wind speeds of 6.2 ms-1, which are similar to the long-term trends reported by Carpenter et al. (2010). Given the placement of the observatory on the island, as well as the lack of major coastal features, this site is well suited for observations of unpolluted marine air. Air was drawn vertically down a 10 m stainless steel pipe with an inner diameter of 1 in at a rate of 16.7 L min-1, from which an isokinetic inlet was used to draw a subsample of air at a rate of 0.3 L min-1. The sample line was heated to minimise condensation build up, and a Thermo PM10 inlet was used at the top of the pipe. Meteorological data were recorded atop a 30 m tower located at the observatory, with dominant wind direction recorded as one of 16 compass points. This was funded as part of the Ice in Cloud Experiment – Dust (ICE-D) campaign, which had the aim of taking measurements of Saharan dust concentrations to better understand aerosol–cloud interactions and reduce uncertainty in numerical models. Details of the ICE-D surface experiments and accompanying aircraft sampling campaigns are described by Liu et al. (2018). The WIBS-4M was used to observe long-term trends in bioaerosol concentrations and was regularly calibrated using National Institute of Standards and Technology (NIST) calibration polystyrene latex (PSL) spheres.

A 9σ threshold was applied when setting fluorescent baselines for each WIBS model to remove weakly fluorescent inorganic particles that could otherwise act as interferents. This is a substantial increase in the threshold used from previously published work but has been shown to significantly reduce the interference from mineral dust while maintaining the concentration of biological content (Savage et al., 2017). As such, all fluorescent particles observed should contain significant amounts of biological material and are subsequently considered to be a bioaerosol. It should be noted that these are not likely to all be single particles but could instead be agglomerates, including biological particles attached to non-biological ones. The minimum cut-off diameter at 50 % detection efficiency (Dp50) for this particular WIBS-4M is 0.8 µm. As such, only particles between 0.8 and 10 µm have been included in the analysis.

Agglomerative hierarchical cluster analysis was performed using Ward linkage with Euclidean distance for the fluorescent particles, using the method described in Crawford et al. (2015). This has been successfully tested in the past when applying a controlled set of fluorescent PSL spheres through the same instrument model (Crawford et al., 2015; Robinson et al., 2013) and has also been shown to accurately capture bioaerosol mixtures within a semi-arid forest (Gosselin et al., 2016). This process begins with individual data points being treated as their own cluster, which are sequentially combined into larger clusters based on the “distance” between data point values. Ward linkage identifies which pair of clusters would yield the minimum increase in total within-cluster variance after merging and consequently merges them next. A Calinski–Harabasz score is then calculated, suggesting an optimum number of clusters. This cluster analysis was performed separately for both the WIBS-4A and WIBS-4M datasets. The transmission efficiency of the inlet as a function of size has also been accounted when considering size distributions for particles up to 10 µm. Inlet penetration as a function of size has also been considered based on the Andersen 321A PM10 inlet (McFarland, 1984).

Analysis of the LAAP-ToF data acquired during the ICE-D intensive campaign has been presented by Marsden et al. (2019), in which the mineralogy and mixing state of north African dust transported to the Cabo Verde islands is described in detail with reference to dust samples collected on the ground from potential source areas in northern Africa. Using a combination of cluster analysis and subcompositional analysis, that work showed that differences in mineralogical composition of ground samples can be detected on a regional scale, whilst variations in internal mixing state are much more localised. In transported dust, however, the mixing state was much less varied, suggesting that local sources become well mixed during transport.

In the present study, we used the compositional classes already identified in the above work by cluster analysis and compared the temporal trends in particle number concentrations with that of the biological particles detected by the WIBS. In addition, we used mass spectra of the previously identified mineral dust class to examine internal mixing of dust with biological material using peak marker analysis. More specifically, biological markers CN- (m/z 26) and CNO- (m/z 42) were used to identify those mineral dust particles which had a significant marker of internal mixing with biological material and also compared the temporal trends in those particle number concentrations to the biological particles detected by WIBS.

This approach is similar to that used by Zawadowicz et al. (2019), who examined the biological content of dust particles in North America by comparing temporal trends in particle number of a WIBS with SPMS, in that case the PALMS instrument. However, in our study, we have used different criteria for biomarkers in the mass spectra, using chlorine to normalise the biomarker signals rather than phosphorous. Chlorine is a ubiquitous signal in transported mineral dust and is easily ionised because of its high electro affinity. This reduces the temporal incoherence in biomarker ion signals caused by the matrix effect.

A clear seasonal trend for fluorescent and total particle concentrations can be seen in Fig. 1a, with both being highest during winter and lowest in summer. Monthly median bioaerosol concentrations as high as 45 L-1 were found with 95th percentile values exceeding 130 L-1 during strong dust events. Such peaks in concentrations are short lived and regular. Figure 1b suggests that bioaerosols are a minority particle type, with non-fluorescent concentrations consistently dominating particle contributions. The fraction of fluorescent particles compared to total particle concentrations was a mean value of 0.4 ± 0.2 %. The 99th percentile value was a fraction of 1.1 %, with some high-fraction events exceeding 1.5 % in October and May. These will be discussed in greater detail further on. Figure 1c highlights how larger particles are observed during winter. Note that there is no relationship between the total particle concentration and fraction of fluorescent particles (r2=0.02), but a strong correlation exists between fluorescent and non-fluorescent particle concentrations (r2=0.79).

Agglomerative hierarchical cluster analysis was performed using Ward linkage with Euclidean distance for all fluorescent particles. A four-cluster solution produced the greatest Calinski–Harabasz score, with information regarding their fluorescence in each channel, size and asymmetry shown in Table 1.

All fluorescent particles seen across the long-term campaign were included in Fig. 2. Fluorescence generally increases with size, but this trend is not equal across all channels and clusters. It is strongest for Clusters 1 and 4, which are also more fluorescent across most size ranges. Figure 2e shows the small size of observed particles, with Clusters 3 and 4 being much smaller than Clusters 1 and 2. Figure 2f shows the median fluorescence in each channel across all fluorescent particles combined.

The relative proportion that each cluster contributes to the total fluorescent concentration shows some variability but no clear trend. This would suggest that Cabo Verde is generally exposed to the same particles year round. Clusters 2 and 3 are clearly the two dominant clusters, accounting for a combined total of approximately 90 % of fluorescent particles. When compared to the average fluorescence values in Fig. 2, these two clusters notably have the weakest fluorescence.

The 120 h back trajectories using National Oceanic and Atmospheric Administration (NOAA)'s Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) merged with the Openair package (Carslaw, 2012) were calculated from the observatory, with a starting height set at 10 m. These trajectories were run every 3 h for all 11 months of the campaign, using monthly meteorological data files downloaded from NOAA. Figure 4 captures the enhanced fluorescent particle concentrations seen during periods of continental outflow as compared to Atlantic or coastal back trajectories. Furthermore, it appears that continental trajectories predominantly occur during the same months that the increase in fluorescent and total particles is seen (October–February). Back trajectories were also calculated for periods where the fluorescent fraction was greatest. Particularly high ratio events can be identified during October, May and August in Fig. 1b. These events all show coastal back trajectories, suggesting coastal sites have higher relative biological content.

Fluorescent particle properties were investigated as a function of trajectory direction. Three week-long periods were chosen, where the trajectories were consistently coming from a specific region. These regions were oceanic, coastal and continental. Oceanic regions were defined by trajectories that spent the previous 120 h passing across the Atlantic Ocean, never passing landmass. Coastal regions were defined by northeasterly trajectories passing along the edge of the African continent, and continental regions were described by more easterly trajectories passing through the Sahara. As it was important to have an entire week of trajectories coming from one source, each of these regions was picked out at different times of the year. The continental period took place from 25/01/2016 to 31/01/2016, the coastal from 20/06/2016 to 27/06/2016 and the marine from 20/04/2016 to 26/04/2016. Average fluorescence across the four clusters and three channels has been compared for particles from each region, as shown in Fig. 5. It does not appear that there are significant differences in the fluorescent properties of particles depending on their source region.

It can be seen in Fig. 6 that all of the dust samples dispersed through the WIBS-4M show relatively weak fluorescence across all three channels but are on average more fluorescent than Clusters 2 and 3. This is particularly true for Moroccan dust in Fl 1. Exposing these samples to gamma irradiation produced little effect, marginally reducing the fluorescent properties of Moroccan dust while increasing the fluorescence of the illite sample. The size and asymmetry values are quite similar across all samples and compare most closely with those of Cluster 2.

Analysis of the negative ion spectra revealed that sea-spray aerosol was the dominant aerosol detected at the site by the LAAP-ToF, accounting for approximately 87 % by number. The remainder was comprised of silicate mineral dust (∼ 5 %), calcium chloride (∼ 3 %) and secondary material (∼ 4 %). Cluster analysis of the WIBS-4A data produced a four-cluster solution which was similar to the long-term CVAO cluster solution, featuring two weakly fluorescent dominant clusters and two highly fluorescent, likely primary biological aerosol particle (PBAP) clusters. A number concentration time series of select LAAP-ToF and WIBS-4A data products are shown in Fig. 7, where similarities in the trends of silicate dust and WIBS-4A Cluster 3 can be observed during a dust event (10/08 onwards), suggesting that this cluster may represent a biodust mixture which will now be examined further.

Deeper analysis of the LAAP-ToF silicate dust class was performed to screen for potential biomarkers. Figure 8 shows example negative ion spectra from laboratory samples of bacteria (top panel) and c-means fuzzy cluster products from ICE-D representative of pure dust and dust containing biomarkers (middle and bottom panels, respectively). It can be seen that the biosilicate spectrogram contains silicate m/z peaks (e.g. SiO2–SiO3) and also bacterial biomarkers (e.g. CN-CNO-) which are not present in the pure dust spectra, suggesting that a subset of the observed dust is mixed with biological material.

The LAAP-ToF silicate dust is then filtered for biomarkers using the ratio of biomarkers to chlorine (CN + CNO / Cl).

A time series of the LAAP-ToF biosilicate and WIBS-4A Cluster 3 products are shown in Fig. 9. Here, it can be seen that the general trends are in good agreement (r2=0.63). This suggests that the WIBS is sensitive to dust particles which are mixed with bacteria, and that these particles can be segregated from the general fluorescent aerosol population and classified using the cluster analysis method described in Sect. 2.3. There is also a moderate correlation between the WIBS-4A Cluster 3 and pure silicate (r2=0.49) as expected, since the biosilicate will be a subset of the pure silicate. The correlation between Cluster 3 and sea spray is poor (r2=0.20), affirming that our observations of fluorescent particles are genuine and are not being skewed by known interferents.

The identified peaks in particle concentration are consistent with back-trajectory analysis showing the sources to be Saharan. Strong Aeolian generating mechanisms mean that potentially significant concentrations of soil bacteria could have attached themselves to dust particles, which are acting as an agent for their distribution and transport. More specifically, it is likely to be the Harmattan wind that is driving these events. This is a north-easterly trade wind that blows across Africa from approximately November to March, although high concentrations of dust are already present almost all year round (Prospero, 1999). Source regions in Tunisia and northern Algeria have been identified before, as well as dust that has been transported from the Chad Basin in the Bodélé Depression (Herrmann et al., 1999). Such Harmattan dust clouds have often been found to affect the coastal regions of the Gulf of Guinea. These clouds are a major annual event, noted for exacerbating cardiovascular health issues and increasing daily mortality by 8.4 % (Perez et al., 2008). Previous studies by Enete et al. (2012) found this dust to contain high quantities of silicon, which is consistent with the LAAP-ToF results presented here.

It is interesting that continental back trajectories during summer do not contain similar concentrations of fluorescent material, despite a similar fluorescent ratio and wind speeds to those seen in winter. This is likely due to the seasonal shift in the ITCZ, as outlined by Chiapello et al. (1995). During summer, the ITCZ is located between 6 and 10∘ N, and dust gets transported into the Saharan Air Layer (SAL) at a maximum altitude of 7000 m. At this height, there is little opportunity for the dust to mix with trade winds, and it is instead transported significant distances across the Atlantic. However, during winter, the ITCZ shifts south and the dust is no longer able to get taken into the SAL. This constrains the dust into lower altitudes, where they are subject to mixing with trade winds, including the Harmattan wind. Dust consequently gets deposited much sooner and often blankets Cabo Verde. The ecological consequences of this have been investigated by Korte et al. (2017), who placed sediment traps across the Atlantic to measure fluxes in biogenic constituents, including biogenic silica. They observed higher biogenic fluxes during winter, due to Saharan dust getting deposited during the previously discussed trade wind events. They also observed a seasonal maxima of biogenic silica towards the west, suggesting that the bioaerosols we observe are not as readily deposited into the ocean as some other particle types.

It is also interesting to consider potential differences in dust from the various source regions. This has been investigated by Patey et al. (2015), who also monitored dust concentrations at the CVAO. By looking at specific elemental ratios within dust samples, they were able to identify where the dust was thought to have come from, as well as what it was comprised of. They found during summer months that 92.5 % of samples contained a contribution from the Sahel, compared to just 52.3 % of samples collected during winter. These observations are supported by previous work from Prospero and Lamb (2003), who found that it was Sahelian dust that was predominantly transported across the Atlantic via the SAL during summer.

The relative contributions of each cluster in Fig. 3 highlight the importance of Clusters 2 and 3. Furthermore, it shows that the mixture of bioaerosols have no clear pattern, with each cluster present all year round. The intensive measurements from the LAAP-ToF and WIBS-4A have provided supporting evidence to suggest that most of the fluorescent particles observed are mixtures of biological and non-biological material, i.e. bacteria attached to dust. It is likely because of this that fluorescence intensity is correlated with particle size. Larger particles may carry greater numbers of bacteria due to increased surface area and as a result should carry more fluorophores for detection by the instrument. The relationship between fluorescence and particle size has been investigated extensively by Hill et al. (2015), who have shown a generally positive correlation.

Concluding that Clusters 2 and 3 are at least partially bacterial in nature is in contrast to findings by Savage et al. (2017), whose laboratory tests on Bacillus atrophaeus, Escherichia coli and Pseudomonas stutzeri have produced different fluorescent spectra to those observed here. In their tests, bacteria have been shown to fluoresce strongly in Fl 1 and weakly in Fl 2 and 3, while our results show little increase in Fl 1 concentrations during winter but substantial increases in Fl 2 and 3. However, it must be emphasised that they used “pure” bacterial samples, potentially unrepresentative of the mixed aggregates we observed.

Furthermore, they outline previous work by Agranovski et al. (2004), who found that an ultraviolet aerodynamic particle sizer (UV-APS) was effective in identifying fluorophores at similar wavelengths that a WIBS-4A was unsuccessful in doing so. They hypothesise reasons for this, including potential differences in gain voltages applied in the instrument, and weaker excitation intensity in Xe2 with respect to Xe1. With that considered, it must also be noted that there are high levels of agreement between the WIBS-4A used in the intensive measurements and the WIBS-4M for 11 months. As such, the more likely explanation may simply be differences between the samples used in the laboratory and those measured at the observatory. Without proper characterisation of a bacteria–dust aggregate in a controlled setting it is difficult to interpret how this could impact the emissions spectra. Internal reabsorption of fluorescence is known to occur when emitted fluorescent wavelengths overlap with a particle's absorption spectrum. In this instance, it is possible that the dust absorbs certain wavelengths emitted from the bacteria, skewing what is detected by the instrument. It must be emphasised that applying a 9σ threshold to the fluorescent baseline of the WIBS-4M means that our observations are unlikely to be interferents.

Clusters 1 and 4 are interesting because of their high fluorescence intensity. It is possible that these clusters represent “pure” bioaerosols, that have not mixed with non-fluorescent material or are relatively more exposed due to, for example, larger bacterial aggregates present. Cluster 1's spectral profile matches closely with laboratory experiments using bacteria (Savage et al., 2017), with a strong fluorescence in Fl 1 seen across all size ranges. However, Cluster 4 remains difficult to identify. It shows similarly high-intensity fluorescence to Cluster 1 but only in Fl 2 and 3. Its relative contribution to the total bioaerosol count is significant, yet given the location of the observatory seems unlikely to be other common terrestrial bioaerosols such as pollen fragments. Although it may therefore seem likely to have a marine source, its concentrations follow similar seasonal trends to the other clusters which are more easily explained by Aeolian events. With potential differences in their capacity to act as ice nucleators when compared to bacteria and dust aggregates, it is useful to quantify their concentrations.

The dust samples dispersed into the WIBS-4M also support our argument that bacteria–dust aggregates are our dominant bioaerosol type. Many of the sampled particles shared similar spectral profiles to those for Clusters 2 and 3, being weakly fluorescent in all three channels. Each sample also predominantly consisted of non-fluorescent particles, sharing a similarly low fluorescent particle ratio to our campaign observations.

Although the 9σ threshold we have used should eliminate weakly fluorescent non-biological particles, the potential for more highly fluorescent particles to act as interferents should be discussed. Soot is one example, with previous studies having observed higher fluorescence than is typically seen for non-biological particles. Despite this, there are multiple reasons that we do not believe interferents are contributing to particle concentrations. Firstly, studies that found soot to fluoresce above their thresholds had usually only done so when using 3σ thresholding. Toprak and Schnaiter (2013) found propane flame soot to only weakly fluoresce in Fl1 at this threshold, and so we would not expect it to be fluorescent at a more conservative 9σ thresholding. Secondly, the sizes of the observed fluorescent particles are larger than we would expect for soot. Toprak and Schnaiter (2013) found generated soot to only be 0.8 µm after significant coagulation time in the NAUA chamber, while Savage et al. (2017) used a mechanically dispersed dry diesel soot powder to investigate potential interferent aerosol fluorescence. They noted that this powder fluoresced above a conservative 9σ threshold, but this sample aerosol was much larger than soot typically observed in the atmosphere when aerosolised (∼ 1.1 µm). Savage et al. (2017) also acknowledged that fluorescence intensity is a strong function of particle size due to surface area/volume effects and that this test soot was likely to be significantly more fluorescent than ambient diesel soot as a result. Furthermore, Savage and Huffman (2018) acknowledge that more highly fluorescent soot is representative of freshly generated soot close to source and is not representative of aged or processed soot. Ambient soot at CVAO should not be fluorescent at 9σ. While it is possible that soot could have internally mixed with dust and therefore become larger, this would still represent aged soot and would be less fluorescent.

We also acknowledge the fraction identified as biological is small (< 1 %) and that concentrations would consequently be significantly affected by even minor errors in the classification of particle types. However, if a fraction of non-biological particles were “bleeding” through and influencing our concentrations, their mass spectral signatures would differ from our “biosilicate” class. As there is a close correlation between the biosilicate particle counts and our fluorescent fraction, we do not believe that bleeding is significantly changing our observations. More studies comparing such a technique may elucidate the degree to which bleeding occurs, but we believe our study provides a good first estimate of bioaerosol concentrations in this region. As discussed by Savage et al. (2017), UV-LIF results should be considered uniquely in all situations with appreciation of possible influences. We are confident that many common interferent particles such as soot can be further discounted when evaluating properties such as particle size, as well as an appreciation for modelled back trajectories and identified source regions.

The results presented here are broadly consistent with previous studies demonstrating that mineral dust is often observed to be mixed with biological material, e.g. Yamaguchi et al. (2012), Griffin et al. (2001) and Maki et al. (2018). Similar to the approach used here, a recent study by Zawadowicz et al. (2019) assessed the prevalence of biological material over the continental United States using single particle mass spectrometry and UV-LIF, finding that 30 % to 80 % of biological particles were mixed with mineral dust. While they provide evidence for biomineral mixtures in their study, the fluorescent number concentration derived from the WIBS is likely an overestimate of the true bioaerosol concentration due to the choice of 3σ thresholding including pure mineral dust interferents in their assumed bioaerosol population, as a small but significant subset of mineral dust naturally exhibits weak fluorescence (Huffman et al., 2019; Savage et al., 2017; Crawford et al., 2016; Pöhlker et al., 2012). The conservative 9σ threshold used in our study excludes these non-biological interferents from the presented PBAP classes (Savage et al., 2017). Furthermore, the use of only negative ion spectra makes resolving biominerals from pure minerals challenging. The use of both positive and negative ion spectra in our study provides greater particle information and consequently improves our ability to classify biological and pure minerals as distinct groups (Shen et al., 2018).