The measurement of elements in PM Xact is a registered trademark of Cooper Environmental Services, LLC.
Ambient particulate matter (PM) plays a major role in affecting human health and air quality. Trace elements represent a minor fraction of the atmospheric aerosol on a mass basis, but they can act as specific markers
for several emission sources. The short- or long-term exposure of ambient
particulate matter (PM) has significant negative effects on human health (Dao et al., 2012; Ancelet et al., 2012; Zhao and Hopke, 2004; Pope III and
Dockery, 2006; Dockery et al., 1993; Zhou et al., 2018). Cakmak et al. (2014) found significant association of acute changes in cardiovascular and
respiratory physiology with PM
Source quantification and characterisation is an important step in understanding the relationship between source emissions, ambient concentrations, and health and environmental effects. SA by receptor models has been widely used in recent years to identify and apportion the contributions of various sources to the airborne PM concentrations. Positive matrix factorisation (PMF) is one of the most widely used receptor models for SA of trace elements (Rahman et al., 2011; Ancelet et al., 2012; Cesari et al., 2014; Ducret-Stich et al., 2013; Kim et al., 2003; Rai et al., 2016; Zhang et al., 2013; Harrison et al., 2011; Hedberg et al., 2005). However, a limited number of studies are available for trace-element emission sources with high time resolution (hourly or sub-hourly; Visser et al., 2015; Crilley et al., 2016; Bukowiecki et al., 2010; Richard et al., 2011; Dall'Osto et al., 2013; Manousakas et al., 2015; Jeong et al., 2019; Wang et al., 2018, among others). Hourly elements data can be used to explore the diurnal patterns of emissions from traffic, biomass burning and industrial sources, thereby yielding more accurate and exposure-relevant SA results. Currently, there are very few offline instruments available for field sampling of elements with high time resolution, such as the rotating drum impactor (RDI; Bukowiecki et al., 2008), the streaker sampler (PIXE International Corporation; Lucarelli et al., 2011) and the semi-continuous elements in aerosol sampler (SEAS; Kidwell and Ondov, 2001). The large quantity of samples generated by these samplers requires a labour-intensive and time-consuming offline analysis. These offline analyses require high-precision and low-detection-limit techniques such as synchrotron-radiation-induced X-ray fluorescence spectrometry (SR-XRF) of aerosol samples collected with a RDI, particle-induced X-ray emission (PIXE) with the streaker sampler and graphite furnace atomic absorption spectrometry (GFAAS) with the SEAS. In practice, the offline samplers lead to undesirable trade-offs between time resolution and data coverage even for short duration field campaigns, whereas highly time-resolved long-term measurements are impractical. A recently introduced online high-time-resolution instrument can collect samples and perform analysis for elements simultaneously in a near-real-time scenario for long-term measurements without waiting for laboratory analysis. The XRF-based Xact 620, Xact 625 and the newer Xact 625i Ambient Metals monitors (Cooper Environmental Services, Tigard, Oregon, USA) have been developed in recent years and have been used in several field studies (Fang et al., 2015; Cooper et al., 2010; Furger et al., 2017; Park et al., 2014; Phillips-Smith et al., 2017; Tremper et al., 2018; Chang et al., 2018; Liu et al., 2019). However, only 10 studies included SA on Xact data (Park et al., 2014; Fang et al., 2015; Phillips-Smith et al., 2017; Chang et al., 2018; Liu et al., 2019; Ji et al., 2018; Sofowote et al., 2018; Jeong et al., 2019; Belis et al., 2019b; Cui et al., 2019).
The main focus of this work is the exploration of the use of the Xact for source apportionment in Europe, where the concentrations are considerably lower than in polluted areas in Asia. In the present study, we conducted SA using PMF to characterise the source contributions of highly time-resolved metals during a 3-week campaign at a traffic-influenced site in Härkingen, Switzerland. PMF was implemented through the multilinear engine-2 (ME-2) solver and controlled via SoFi, which allows for a comprehensive and systematic exploration of the solution space (Bozzetti et al., 2016; Canonaco et al., 2013). The rotational control available in ME-2 provides a means for treating extreme events such as fireworks within a PMF analysis. Such events are often excluded from the PMF input matrix to avoid modelling errors due to the pulling of a solution by outliers (Ducret-Stich et al., 2013; Norris et al., 2014). Although a few studies have already been carried out in the past at this location (Lanz et al., 2010; Hueglin et al., 2006; Furger et al., 2017), none of them have reported SA on elements.
PM
Map of the sampling location (NABEL site in Härkingen). The site is marked with the red Google pin. Map reproduced by permission of swisstopo (JA100119).
Sampling and analysis was conducted with an Xact 625 Ambient Metals Monitor
equipped with a PM
Positive matrix factorisation is one of the most common receptor models
based on a weighted least-squares fit (Paatero and Tapper, 1994). It is
used to describe the variability in a multivariate dataset as the linear
combination of a set of constant factor profiles and their corresponding
time series as shown in Eq. (1) in cell notation:
The PMF algorithm was solved using ME-2 (Paatero, 1999), which enables an efficient exploration of the solution space by introducing a priori
information to
The conditional bivariate probability function (CBPF) is a data analysis tool to identify the direction of source contributions and was applied to the PMF source factors. Polar plots are
used to present the CBPF analyses, where the number of events with a concentration greater than the
In our study, the PMF input consists of a data matrix and an error matrix of
hourly element measurements, where the rows represent the time series (456
points with 1 h steps) and the columns contain the elements (14 variables).
The input preparation of PMF was done by excluding some specific elements
for better source apportionment results. A common approach for the choice of
species to include in the PMF input depends on (1) the percentage of data
below the detection limit (Polissar et al., 1998) and (2) Xact data comparison with offline 24 h PM
An important step in the PMF analysis is the selection of the number of
factors by the user, as mathematical diagnostics alone are insufficient for
choosing the correct number of factors (Ulbrich et al., 2009; Canonaco et
al., 2013). The selection of factors is often based on an analysis of total
In a first step, we examined a range of solutions with 3–10 factors
at 10 seeds (number of PMF repeats) from unconstrained runs. The
unconstrained PMF solution resulted in mixed factors, such as sea salt mixed
with fireworks, in all factor solutions. We show an example of a mixed
nine-factor solution in Fig. S1 in the Supplement. This is likely because of the very high concentration and variation in composition of firework emissions during the firework period. Because the signal-to-noise ratio is very high, imperfections in the model description exert a strong influence on
The input dataset was divided into two parts: firework days (FDs; 31 July–4 August) and non-firework days (NFDs; all days except 31 July–4 August). To obtain a specific firework profile, we further selected only firework hours (FHs; 31 July 21:00 LT–1 August 07:00 LT (local time is
coordinated universal time
The unconstrained NFD PMF analysis resolved seven factors at all seeds, such
as sea salt, secondary sulfate, a non-exhaust traffic-related factor, road dust,
background dust, an industrial factor and a
The unconstrained PMF analysis of the complete dataset identified a
secondary sulfate factor (most of the S is apportioned in this factor, with
91 % of the factor mass) in the nine-factor solution (Fig. S1). The
identified secondary sulfate factor time series correlated very well with
ACSM sulfate (
We then performed the constrained PMF analysis on the FD and FH datasets.
Here we constrained the secondary sulfate factor profile (
In the FD and FH PMF analyses, the sea salt factor profile (
Note that the number of time points contained in the input matrices for the
data subsets (as opposed to the full dataset) are in some cases smaller than
typical recommendations for ambient PMF (Belis et al., 2019a). This is most
extreme in the case of the FH dataset, where only 11 time points are used.
However, there are two important differences between PMF analyses of these
sub-datasets and typical ambient PMF: (1) the sub-datasets are constructed
to maximise the variability in a factor or set of factors, and (2) we are concerned only with accurately characterising the profile(s) of these
selected major factors. These two points work together to greatly reduce the
number of time points required for the analysis (Fig. S6). A similar
approach has been successfully applied by Fröhlich et al. (2015), in which
short-duration spikes in organic aerosol concentration were combined into a
sub-dataset to determine an anchor profile related to local cigarette smoke,
and by Visser et al. (2015), in which a subset of a trace-element dataset with
high residuals was analysed separately to retrieve a factor profile related
to industrial emissions. In the final complete dataset PMF analysis, the
factor profiles of fireworks, secondary sulfate and sea salt were
constrained (
Residual analysis (
The statistical and rotational uncertainties were explored by the bootstrap
(BS) resampling strategy (Efron, 1979) and the exploration of the
We also performed separate random bootstrap analyses for 1000 times on the
correlation (
Source profiles of the finally retained eight-factor PMF results. The data and their corresponding uncertainty are given as box–whisker plot (bottom to top: p10–p25–p50–p75–p90) of good solutions from bootstrap runs. The left
During the non-firework period, uncertainties in the source apportionment
results are assessed by a bootstrap analysis as described above. However,
this approach cannot be used to assess uncertainties in the sea salt and
secondary sulfate factors during the firework period, as during this period
these factor time series are constrained with an artificially low
Secondary sulfate concentrations during the firework period were estimated
from the linear fit of the secondary sulfate factor to ACSM sulfate during
the entire non-firework period. Uncertainties of
As described above, the sea salt factor time series during the 4 d firework period was investigated to determine measurements that were affected or not affected by fireworks, where the measurements determined to be affected were replaced with a linear interpolation between the nearest good points. To determine the uncertainties of this approach, we applied this calculation to random segments of the non-firework data. Specifically, the 4 d long sequence of affected or non-affected time points determined during the firework period was applied to a randomly chosen segment of data, and the standard deviation of measurement data to the estimated values calculated by interpolation was determined. This analysis was repeated for 38 randomly selected locations through the non-firework data, and a mean standard deviation of
The solution that best represented the input data was the eight-factor
solution. The eight factors from the PMF results are as follows:
two firework factors with prominent relative contributions of a sea salt factor explaining a large fraction of a secondary sulfate factor mostly dominated by two dust factors, one dominated by Ca and showing traffic rush hours peaks and the other dominated by a non-exhaust traffic-related factor characterised by an industrial factor showing relatively high contributions of
In this section, the results of the PM
Mean diurnal patterns of the factors and of some corresponding external tracers with error bars (1 standard deviation).
CBPF analysis (at 90th percentile) of factors in terms of wind speed (m s
The separation of two dust factors is in line with Amato et al. (2009), where ME-2 yielded a road dust factor distinct from a mineral dust factor. The CBPF plot shows that higher concentrations of the road dust factor are associated with the southern wind sector, while the background dust factor is influenced by the south-western and north-eastern wind sectors (Figs. 5, S12). The diurnal pattern of the road dust factor shows morning rush hour traffic peaks similar to the
A source apportionment study of 14 elements in PM
The source apportionment model performance could possibly be additionally
improved by the inclusion of
The datasets are available upon request to the corresponding author.
The supplement related to this article is available online at:
PR performed SA analysis and wrote the paper. MF, RF and CH performed the measurement. MF and RF analysed the data for Xact and ACSM, respectively. FC and JGS provided expertise on software for SA analysis. MCM provided ICP-MS and ICP-AES analysis data. KP lent Xact 625 for measurement. UB, ASHP, MF, FC and JGS were involved with the supervision and conceptualisation. All authors commented on the paper and assisted in the interpretation of the results.
Krag Petterson is employed by Cooper Environmental Services, the manufacturer of Xact 625.
We thank René Richter and Roland Scheidegger of PSI for their technical support during the measurement campaign. We are grateful to Chris Koch and Varun Yadav of Cooper Environmental Services for instructions on the instrument and numerous technical discussions. We thank the operators of the NABEL station for providing all kinds of support during the measurements.
This study has been funded by the Swiss Federal Office for the Environment (FOEN). This research has been supported by the Swiss National Science Foundation (SNSF; grant nos. 200021_162448/1 and BSSGI0_155846). María Cruz Minguillón acknowledges the Ramón y Cajal Fellowship awarded by the Spanish Ministry.
This paper was edited by Willy Maenhaut and reviewed by four anonymous referees.