The speciation data quality was evaluated by examining the consistency between species concentrations measured by different methods: for example, gravimetric mass vs. mass from continuous monitor, gravimetric mass vs. reconstructed mass, SO42- vs. total S, K+ vs. total K, etc. Deming regression was applied in the examination using the scatter plot computer programme developed by Wu and Yu (2018), which is available at 10.5281/zenodo.832417 (last access: 9 December 2021). This technique is applied to consider the measurement uncertainties of both variables to be compared in the regression. Details of the evaluation are provided in Sect. S1 in the Supplement. In short, the evaluation shows that the speciation data are of adequate quality for the ensuing analyses.

The study-wide average concentration of PMcoarse is 14.9 ± 8.6 µg m-3 (± standard deviation), accounting for 45 % of ambient PM10 (32.9 ± 18.5 µg m-3). The daily concentrations range from 2.9 to 40.4 µg m-3. The contribution of geological material is estimated by assuming the crustal elements are in oxide forms, i.e. 1.89×[Al]+2.14×[Si]+1.2×[K]+1.4×[Ca]+1.67×[Ti]+1.43×[Fe]. This component has the largest contribution, making up 5.2 µg m-3 or 35 % of the PMcoarse mass. The next important component is nitrate (2.2 µg m-3, 15 %), followed by sea-salt-related ions (i.e. Na+, Mg2+, and Cl-) and organics, which represent 11 % and 8 %, respectively. The coarse organics were estimated by multiplying the measured OC by a factor of 2, assuming the organics are mainly associated with biological particles which are enriched in oxygenated compounds such as polyols and carboxylic acids (Edgerton et al., 2009). The composition forms a stark contrast with that of PM2.5 (18.0 ± 11.2 µg m-3), in which carbonaceous components (organics and EC, 41 %) and secondary ions (NH4+, NO3-, and SO42-, 38 %) are the major components. The difference is consistent with combustion and secondary aerosol formation processes being the major sources of fine particles, whereas coarse particles are primarily generated by mechanical processes. The organics here were approximated to be 1.6× [OC] considering typical features of urban aerosols with both primary and secondary contributions (Turpin and Lim, 2001).

The annual average concentrations of PMcoarse and selected major components measured in this study are compared with those in other locations in Table 1. Only studies that spanned at least 1 year or more and had all major species measured (i.e. elements, ions, OC, and EC) are considered. Our PMcoarse level is similar to those in other urban locations, more than 2 times higher than Milan in Italy, and ∼5 µg m-3 higher than Central Los Angeles, as well as only half of that in Casa Grande in Arizona and a 10th of that in Lahore in Pakistan. Our concentration is also comparable to two roadside studies carried out in Bern in Switzerland and in London and Birmingham in the UK. We note the PMcoarse concentration in a Hong Kong roadside study is ∼10 µg m-3 higher than the current study. Yet a straightforward urban vs. roadside comparison is not feasible given the roadside measurement was conducted more than 15 years ago. We also note that all the cited measurements were taken at least a decade ago. The lack of more recent measurements highlights the need for more PMcoarse speciation effort, considering the growing importance of PMcoarse in aerosol mass loading and health effect contributions as PM2.5 has been controlled effectively in many locations. Our PMcoarse concentration is also 3–4 times lower than that measured in the desert area in Arizona but one third higher than a desert-like area in Lancaster in Los Angeles.

Geological material is the single largest component in PMcoarse across all studies including ours, accounting for roughly 30 %–50 % (Lahore shows 74 %), underlining the importance in identifying fugitive dust sources (e.g. natural vs. anthropogenic) for effective mitigation of PMcoarse. We note that our nitrate concentration is the highest among all studies (except for the Lahore study, which is comparable to ours), constituting 2.2 µg m-3 or 15 % of the PMcoarse. Coarse-mode nitrate mainly forms by the uptake of HNO3 by pre-existing alkaline particles forming NaNO3 in reaction with sea salt and Ca(NO3)2 with soil dust (Bian et al., 2014). Our total carbon level of 0.7 µg C m-3 is among the lowest compared to other studies, with 86 % of it coming from OC. A quarter of PMcoarse mass is regarded as unidentified in this study. The percentage share is among those observed in other studies, which range between 8 % and 38 %. The nature of the unidentified mass will be discussed in Sect. 3.3.

The seasonal evolution of weather in Hong Kong is largely driven by the East Asian Monsoon system. Correspondingly, the atmospheric PM pollution in Hong Kong displays a distinct seasonal characteristic. In general, the PM loading in summer is mainly governed by local emissions due to the prevailing southerlies carrying clean marine air masses. In winter, the prevailing northerlies place Hong Kong immediately downwind of the continental region with intense industrial and agricultural activities. Under this situation, the PM loading is affected by both local and regional sources. The transient seasons – spring and fall – have more mixed wind directions. The seasonal contrast in precipitation frequency and ambient temperature, both being higher in summer and lower in winter, also contributes to the variation in PM concentration across different seasons (Louie et al., 2005; Yu et al., 2004).

The sampling period in this study is divided into four seasons based on the observed meteorological and weather patterns as analysed in Sect. S2. Table 2 lists the starting and ending dates of individual seasons, along with the seasonal averages of PM concentrations and several meteorological parameters. Note that the two winter periods at the beginning and the end of the sampling programme are regarded as two different winter periods considering the variability in weather conditions and that they span mostly different calendar months.

Figure 1 presents the PMcoarse concentration and composition by season. The PMcoarse exhibits a significant variation across different seasons, ranging from the lowest 8.1 µg m-3 in summer to the highest 24.8 µg m-3 in the second winter. Washout by precipitation plausibly plays a role in the seasonal contrast given that summer takes up 75 % of the rainfall for the whole study period (Table 2). Mixing layer height appears to play an insignificant role in controlling the variation in PMcoarse level. For example, although the mixing height in the first winter is the lowest among all seasons (509 ± 402 m) while that in the second winter is the highest (874 ± 408 m), the PMcoarse in the latter is more than twice higher than the former. The wind speed also shows small variation across the seasons, with a range of 1.9 to 2.3 m s-1. This range corresponds to a Beaufort scale number of 1–2, referring to the light wind condition. The meteorological data imply that the seasonal variation in PMcoarse levels is likely caused by changes in source intensity and/or air mass origin.

The composition information indicates that geological material is largely responsible for the variability in PMcoarse. This component takes up 22 %–43 % of the PMcoarse mass. The seasonal contrast in the contribution of this component could be attributed to enhanced wet deposition in the warmer season and elevated contribution from regional transport in the colder season. The unidentified mass also represents a major component in most seasons (except spring), accounting for 20 %–32 % of PMcoarse mass. Like geological material, this fraction has a significantly enhanced contribution in the colder season compared to the warmer season. As for other components, nitrate has the highest absolute contribution in spring and lowest in summer (3.2 vs. 1.2 µg m-3). Organics are the highest in the second winter and lowest in the first winter, showing an order of magnitude difference (2.5 vs. 0.2 µg m-3). The concentrations of sea-salt-related ions (i.e. Na+, Mg2+, and Cl-) are higher in the warmer season than that in the colder season, which is consistent with the enhanced influence of marine air mass in the warmer season.

Here the source origins of PMcoarse are discussed. The number of factors (or source categories) contributing to PMcoarse was determined in the PMF analysis. The PMF solution with four factors was selected for source interpretation after an examination of the physical interpretability of the resolved factor profiles for a series of PMF solutions with different factor numbers. The details are provided in Sect. S3. In brief, the three-factor solution was discarded as it gave poor modelling result for Cu, which is an important species in PM health effects associated with reactive oxygen species formation (Bates et al., 2019). The five-factor solution was not considered either because the fifth factor, which is a secondary nitrate factor, was assessed to be chemically inexplainable after examining the charge balance of the ionic composition. The stability of the four-factor solution has been tested against the bootstrapping and displacement functions embedded in the PMF software. The results show that the PMF solution is statistically robust for source analysis. Details of the uncertainty estimation are summarized in Table S2.

The factor profiles resolved in the four-factor solution are shown in Fig. 2. The four factors can be broadly classified into the sea salt category consisting of the first and second factors and the dust category consisting of the third and fourth factors. The first factor is marked by the high loading of Cl- with the additional presence of Na+ and Mg2+, which are strong indicators for fresh sea salt. The molar equivalent of Cl- is balanced by that of Na+ and Mg2+, and it has an anion-to-cation equivalence ratio of 0.99, adding credence to the validity of this factor. The second factor is loaded with a substantial fraction of Na+ and Mg2+, which are markers for sea salt. The absence of Cl- and presence of NO3- indicate this factor specifically represents aged sea salt given that Cl- in sea salt is actively depleted by gaseous HNO3 forming non-volatile NaNO3 (Bian et al., 2014). This factor is termed aged sea salt mixed with secondary inorganic aerosols. The third and fourth factors are clearly associated with fugitive dust, as indicated by the high abundance of crustal elements (e.g. Al, Si, Ca, Ti, and Fe). However, the chemical fingerprints in these two factor profiles only provide limited information for pinpointing the more specific sources responsible for the aerosol burden.

To better understand the sources behind the PMF-resolved factors, the association between air mass origins and source contributions was investigated through backward air mass trajectory analysis. The back-trajectories were computed by the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model using meteorological data from the 1∘ horizontal resolution Global Data Assimilation System (Stein et al., 2015). Past 48 h back-trajectories of air mass reaching Hong Kong at 300 m height at the end of each sampling event at midnight were computed. The trajectories were clustered based on similarity between the trajectory end points. Four trajectory clusters are resolved, and the means for each cluster are displayed in Fig. 3a. The average source composition associated with each cluster is shown in Fig. 3b.

As shown in Fig. 3, the source compositions exhibit features that agree with the travelled source areas of the corresponding air masses. For example, both fresh and aged sea salt contributions are higher when the monitoring site is under the influence of a marine air mass (clusters 2 to 4) than under the influence of a continental air mass from the north (cluster 1). It is also noted that the contribution of aged sea salt is higher in clusters 2 and 3. By examining the individual trajectories in these clusters, it can be seen that cluster 2 is mostly composed of air masses passing through the coastal areas, whereas cluster 3 consists of a mix of marine air masses from the east and short-distance continental air masses from the northeast direction (see Fig. S6). The higher aged sea salt contribution could possibly be explained by the observation that clusters 2 and 3 have more mixed contributions from sea salt and HNO3, whereas for the other two clusters either there is a deficiency in the availability of sea salt as in cluster 1 or deficiency in HNO3 as in cluster 4.

The magnitude of total dust contribution exhibits a descending order from clusters 1 to 4, corresponding to a transition from a continental air mass from the north to a coastal air mass from the northeast/east and to an oceanic air mass from the south (Fig. 3). The results suggest the inner continental region to the north of Hong Kong could be a significant dust-emitting area, with source intensity strong enough to influence the dust aerosol burden in Hong Kong through regional transport. Recent studies on anthropogenic air pollutant emissions in Guangdong Province (a larger geographical territory of GBA) based on emission inventory developments showed that the dust source and the industrial process source are the main contributors of PM10 loading (Bian et al., 2019; Huang et al., 2021). For those emission inventories, the dust source mainly consists of road dust and construction dust emissions, whereas the industrial process source includes emissions from the manufacturing processes of a list of raw materials, including paper, rubber, steel, ceramics, cement, etc. Analysing the hourly field measurement data for elemental species with PMF approach, Zhou et al. (2018) resolved two dust-related source categories responsible for the atmospheric PMcoarse loading in the city of Foshan in Guangdong Province. The first category is road dust with brake and tire wear, while the second is construction dust. Being one of the most important industrial hubs in the GBA, Foshan could represent one of the source areas responsible for the degraded air quality in Hong Kong resulting from regional transport of air pollutants.

The source nature of our dust factors is inferred by comparing our factor profiles with those in Zhou et al. (2018). We noted that the chemical profile of their PMF factor containing road dust is similar to that of our third factor, both accounting for over half of the coarse Al, Si, K, Ca, Ti, and Fe by mass, inferring contributions from road dust. The elemental ratios of our third factor are also close to that of the local paved road dust reported by Ho et al. (2003): for example, 0.39 in our study vs. 0.39 in Ho et al. (2003) for Al/Si, 0.30 vs. 0.46 for Ca/Si, and 0.23 vs. 0.26 for Fe/Si. Aside from road dust, industrial emissions and coal combustion could also be the contributors due to the presence of Zn and Pb. High loadings of both Zn and Pb are also seen in the industrial coal combustion factor by Zhou et al. (2018). Tire wear could also be a potential source of Zn (Pant and Harrison, 2013; Zhou et al., 2018). The carbonaceous components in this factor can be attributed to deposition of combustion emissions on aerosol dust and emission of biological aerosols. Taken together, this dust factor is named “soil dust/industrial and coal combustion”. The term soil dust is used instead of road dust because soil dust is broader, covering both road dust and desert dust potentially contributed by desert or loose soil dust from the inner continental region to the north of Hong Kong; it is inferred from Fig. 3 that cluster 1 samples have the highest dust contribution.

Both the construction dust factor of Zhou et al. (2018) and our fourth factor differ from the first dust factor by a higher abundance of Ca than Si, while they are depleted in Al, Si, and K. The enrichment in Ca can be regarded as an indication of construction activity. This element is enriched in construction dust because of the use of cementitious materials. A point to note is that the fourth factor contains a high loading of Cu. Common sources of Cu in PMcoarse include brake wear generated from abrasion of brake lining material and brake discs (Pant and Harrison, 2013) and industrial emissions (Taiwo et al., 2014). However, no coarse-mode Cu was reported in the PMF factor profiles by Zhou et al. (2018), and hence it remains uncertain to what extent the fourth factor resolved in this study is similar to the construction dust factor resolved by Zhou et al. (2018). Considering the additional presence of the characteristic Cu peak, the fourth factor is termed “construction dust/copper-rich emissions”.

Figure 4 presents the absolute and relative source contributions by season in ascending order of PMcoarse concentration. The secondary nitrate represents the nitrate from all factors to better characterize the contribution by nitrogen oxides (NOx) emission. The two nitrate-free sea salt factors are aggregated into one sea salt factor. During summer when oceanic wind from the south prevails and ambient PMcoarse is the lowest in concentration, sea salt contributes nearly half of the PMcoarse, representing the most important contributor in this season (47 % or 3.7 µg m-3). Note that the source contributions are based on the PMF-apportioned mass, and thus the contributions include residue LWC, which is mainly associated with enhanced uptake of water by aged sea salt aerosols. The soil dust/industrial and coal combustion factor accounts for 24 % (1.8 µg m-3) of the PMcoarse, followed by secondary nitrate (16 % or 1.2 µg m-3) and the construction dust/Cu-rich emission factors (13 % or 1.0 µg m-3). The source composition changes slightly in the first winter and spring periods when PMcoarse is higher, with sea salt and soil dust/industrial and coal combustion similarly contributing to one third of the PMcoarse in both periods. The rest is evenly shared by the other two sources in both periods, i.e. 18 %–19 % for secondary nitrate and 14 %–17 % for construction dust/ Cu-rich emissions.

The significantly elevated PMcoarse concentration observed in fall and the second winter is driven by the increase in dust contribution, which can be attributed to the prevailing northerlies from the continental region. The two dust-related factors cause a disproportionate impact on the ambient PMcoarse loading during these two seasons, contributing three quarters (72 %–79 %) of the PMcoarse mass in total. Compared to summer, the total contribution is 5 times higher or more in fall and the second winter, being 16.0 and 19.8 µg m-3, respectively. The seasonal difference is consistent with an earlier source apportionment study by Yuan et al. (2013), in which an 11-year-long (1998–2008) speciation data set obtained from the Hong Kong PM10 network was analysed by receptor modelling approach. Specifically, the study reported a 3 times higher crustal soil and dust contribution to PM10 in winter than in summer (9.7 vs. 3.2 µg m-3). Moreover, it showed the contributions of this source category are spatially and temporally similar across different monitoring stations in Hong Kong, implying the regional nature of this source.

The Foshan source apportionment study of Zhou et al. (2018) mentioned earlier was conducted in October–December 2014. Assuming the difference between PM10 and PM2.5 contributions by their motor vehicles/road dust factor is due to road dust, their road dust and construction dust sources contributed to 17.7 and 9.4 µg m-3 of PMcoarse, respectively. These contributions are higher than the 12.9–15.2 and 3.2–4.6 µg m-3 levels estimated for the soil dust/industrial and coal combustion and construction dust/Cu-rich emission factors in fall and the second winter of this study. This spatial gradient lends support to the dust contributions in Hong Kong being associated with regional transport. Once entrained into the atmosphere, the lifetime of mineral dust can be up to several days, and therefore it can be transported over long distances (over thousands of kilometres), and the concentration would decrease with transport distance away from the source regions.

The potential source areas are mapped by coupling the PMF-derived source contributions at the receptor with the associated backward air mass trajectory. In this analysis, the geographical domain of interest is divided and represented by a grid cell matrix. By coupling the trajectory end points in the grid cells with the concentrations at the receptor, each grid cell will receive a value representing the potential source strength in the corresponding area. The concentration-weighted trajectory (CWT) method is applied for the analysis (Hsu et al., 2003). In this method, each grid cell receives a weighted concentration value obtained by averaging the sample concentration that has associated trajectories crossing the corresponding grid cell, weighted by the residence time of air mass in that grid cell. The weighted concentration value (or CWT value) is expressed by Eq. (1): 1CWTij=∑l=1LClτijl∑l=1Lτijl, where Cl is the concentration at the receptor site associated with back-trajectory l, τijl is the number of end points of trajectory l falling into grid cell i,j (i.e. the residence time of the trajectory in the grid cell), and L is the total number of trajectories over a time period. To improve the robustness of the CWT analysis, the input trajectory information was augmented by considering all the trajectories calculated every 3 h for each sampling day and assuming the same concentrations over the day (Petit et al., 2017). The geographical domain was defined based on the spatial range of the trajectories travelled, with the dimension of grid cells set to be 0.5∘×0.5∘. A weighting function was applied to down-weight grid cells with an insufficient number of end points following the software guidelines. The CWT analysis was performed using the Zefir programme (Petit et al., 2017). The analysis was performed by season to account for the potential variability in source strength and meteorological conditions.

Figure 5 presents the CWT results for summer and the second winter and indicates the potential source areas. The results for other seasons are displayed in Fig. S7. It can be seen that for the two dust-related factors, the elevated contributions are associated with continental air masses originating from the north, whereas the sea-salt-related contributions are associated with marine and coastal air masses. These results are consistent with the general understanding of source origins of these categories of sources. An important finding revealed from this analysis is that the GBA or the greater part of southern China is shown to have significant fugitive-dust-related emission sources and that these dust sources are implicated in causing days of high ambient PMcoarse loading in Hong Kong.