Atmospheric Environment 36 (2002) 5607–5618

Application and comparison of two statistical trajectorytechniques for identification of source regions of

atmospheric aerosol species

Alexandru Lupu*, Willy Maenhaut

Ghent University, Department of Analytical Chemistry, Institute for Nuclear Sciences, Proeftuinstraat 86, Gent 9000, Belgium

Received 14 June 2002; received in revised form 12 August 2002; accepted 20 August 2002

Abstract

Two approaches for identification of source locations and preferred transport pathways of atmospheric particulate

trace elements and aerosol species are investigated, namely, versions of the potential source contribution function

method (PSCF) and the concentration field method (CF). Both methods are based on combining chemical data with

calculated air parcel backward trajectories. The two methods are applied to four multi-species multi-annual

concentration time series measured at sites in Finland, Norway, and Israel. A non-parametric bootstrap technique is

used to estimate the statistical significance of the calculated PSCF values. It is found that the methods agree well with

each other and correctly identify known emission sources. Examples of applying the methods are presented, mainly for

the site in Finland.

r 2002 Elsevier Science Ltd. All rights reserved.

Keywords: Receptor model; Potential source contribution function; Bootstrap; Aerosol; Long-range transport

1. Introduction

The residence time probability analysis of Ashbaugh

et al. (1985), also known as the ‘potential source

contribution function’ (PSCF) method, has been exten-

sively used in the identification of source locations and

preferred transport pathways of atmospheric trace

elements and particulate species, e.g., sulphate, nitrate,

ozone, black carbon, and mercury (Poirot and Wishin-

ski, 1986; Zeng and Hopke, 1988; Cheng et al., 1993;

Stohl and Kromp-Kolb, 1994; Polissar et al., 1999; Lin

et al., 2001; Polissar et al., 2001a). The method has also

been applied to locate the sources or source categories

identified by multivariate receptor models, e.g. principal

component analysis, positive matrix factorization of

Paatero and Tapper (1994) or Henry’s (1997) UNMIX

(Xie et al., 1999; Poirot et al., 2001; Polissar et al.,

2001b).

In a nutshell, PSCF is defined as the probability that

an air parcel with concentration higher than a specified

threshold arrives at the receptor site after having resided

in a certain grid cell of the spatial domain of interest. It

is calculated as the ratio between the grid cell number of

backward trajectory segment endpoints associated with

concentrations above the threshold and the total

number of trajectory segment endpoints for the specified

cell.

A modified version of this method, called ‘concentra-

tion field’ (CF) or ‘trajectory statistics’ (TS) method, has

been developed by Seibert et al. (1994) and refined by

Stohl (1996) into the redistributed CF method. The CF

method consists of assigning concentration values

measured at the receptor site to corresponding air parcel

backward trajectories. A mean or logarithmic mean

*Corresponding author. Tel.: +9-264-65-98; fax: +9-264-66-

99.

E-mail address: alexandru.lupu@rug.ac.be (A. Lupu).

1352-2310/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.

PII: S 1 3 5 2 - 2 3 1 0 ( 0 2 ) 0 0 6 9 7 - 0

concentration is calculated for each grid cell of the

spatial domain, by using as a weighting factor the time

spent by the air parcel over the cell. This approach has

been employed by Wotawa et al. (2000) to investigate

the transport of ozone in the Alps, and applied by

Virkkula et al. (1995) and Virkkula et al. (1998) to

identify the sources of non-sea-salt sulphate, ammo-

nium, sodium, and sulphur dioxide at a site in the

Finnish Lapland.

It is apparent that in both cases, a problem arises

because of grid cells crossed by a small number of

trajectories. Having poor counting statistics often results

in false positives if trajectories travelling over true source

areas extend beyond these sources. The problem is

‘fixed’ (Cheng et al., 1993), in the case of the PSCF

method, by multiplying the PSCF values with an

arbitrary weighting function which reduces the con-

tribution of the grid cells with low number of counts,

and, in the case of the CF method, either by simply

discarding the mean concentrations corresponding to

grid cells that have a number of trajectory segment

endpoints less than an arbitrary-set threshold, or by

smoothing the concentration field with a nine-point

filter, while imposing the constraint that the concentra-

tion values should be kept within their confidence

interval (Seibert et al., 1994). An objective method for

dealing with these ‘tailing effects’ (Cheng and Lin, 2001)

has been advocated by Hopke et al. (1995), who have

used a bootstrap technique to examine the statistical

significance of the PSCF values. To the same purpose,

Vasconcelos et al. (1996a,b) have employed two

statistical tests, one based on a bootstrap method, the

other, on the binomial distribution.

The performance of both methods has not been fully

discussed in the literature. Vasconcelos et al. (1996a)

have shown that the angular resolution of the PSCF

method is good, but that the radial resolution is poor,

since all trajectories converge towards the receptor site.

Wotawa and Kr .oger (1999) have tested the ability of the

CF method of Seibert et al. (1994) and the redistributed

CF method of Stohl (1996) to reproduce emission

inventories in cases of negligible measurement and

transport errors, by applying them to a simulated

concentration data set generated with a Lagrangian air

quality model. They have concluded that the main

features of the emission inventory are successfully

replicated. Finally, Cheng and Lin (2001) have evaluated

the performance of the PSCF model using data observed

from the 1998 Central America smoke events. Their

findings indicate that the model is able to accurately

determine the smoke source locations and pollutant

transport pathways.

In what follows, we will apply versions of the two

methods described to four multi-species multi-annual

concentration time series measured at sites in Finland,

Norway and Israel, compare and discuss the results

obtained. A non-parametric bootstrap method will be

used for estimating the statistical significance of the

calculated PSCF values.

2. Methods

2.1. Potential source contribution function

Calculations are performed on a longitude–latitude

grid which covers the spatial domain of interest. We

assume that a species emitted within a grid cell is swept

into the air parcel and transported to the receptor site

without loss through diffusion, chemical transformation

or atmospheric scavenging (Cheng et al., 1993). Let nij

be the total number of trajectory segment endpoints

falling in the grid cell ði; jÞ over the period of study, and

let mij be the number of endpoints in ði; jÞ correspondingto trajectories associated with concentration values at

the receptor site exceeding a specified threshold. The

ratio Pij ¼ mij=nij (PSCF) is then the conditional

probability that an air parcel passing over the cell ði; jÞon its way to the receptor site arrives at the site with

concentration values above the threshold. Hence, high

values in the spatial distribution of Pij will pinpoint

geographical regions that are likely to produce high

concentration values at the receptor site if crossed by a

trajectory.

In order to identify the high PSCF values that might

have arisen purely by chance, it is necessary to test these

values against the null hypothesis that there is no

association between concentrations and trajectories

(Vasconcelos et al., 1996b). The statistical significance

of the spatial distribution of the PSCF values is

examined by making use of a non-parametric bootstrap

method (Wehrens et al., 2000). The method assumes that

the concentration values are independent and identically

distributed. We draw with replacement from the original

concentration data set, C ¼ fc1; c2;y; cNg; B random

subsamples of size equal to the length of the data set,

Cn ¼ fcn1 ; cn2 ;y; cnNg: We then calculate for each boot-

strapped sample k the corresponding PSCF spatial

distribution, Pnk;ij : Let Pn

ð1Þ;ijo?oPnðBÞ;ij be the ordered

values fPnk;ijg; k ¼ 1;y;B; and let a be the chosen

significance level. If

PijXPn

ððBþ1Þð1�a=2ÞÞ;ij ; ð1Þ

the null hypothesis is rejected at ð1� aÞ100% confidence

level. We decide to retain for our further analysis only

the PSCF values satisfying the above relation. Note that

if there are more than one trajectory assigned to a

concentration value, the simple bootstrap on the

concentration data set is equivalent to a blocked

bootstrap on the trajectory set.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–56185608

2.2. Concentration field method

In this method, a mean or logarithmic mean

concentration is calculated for each grid cell ði; jÞ: Theequation for the logarithmic mean is

log %cij ¼1

PMl¼1 t

lij

XM

l¼1

logðclÞtlij ; ð2Þ

where l is the index of the trajectory, M is the total

number of trajectories, cl is the concentration observed

on arrival of trajectory l; and tlij is the time spent in ði; jÞ

by trajectory l (Seibert et al., 1994; Stohl, 1998). A high

value of %cij implies that air parcels travelling over ði; jÞwill have, on average, high concentrations at the

receptor.

If the sampling time and/or the frequency of available

trajectories are not constant, we make, after reindexing,

the following adjustment to the weights in Eq. (2),

log %cij ¼1

PNs¼1

tsij

Ms

XN

s¼1

logðcsÞts

ij

Ms

; ð3Þ

where s is the index of the sample, Ms is the number of

trajectories arriving at the receptor site during the

sampling time of sample s;N is the total number of

samples, and tsij is the total time spent in ði; jÞ by the Ms

trajectories. If trajectory data endpoints are given at

equal time intervals, tsij can be replaced by the

corresponding number of trajectory segment endpoints,

nsij : When the sampling time is the same for all samples

and the frequency of trajectories is constant, Eq. (3)

reduces to Eq. (2).

3. Data

3.1. Chemical data

Four data sets have been used in this study. Aerosol

samples for the first set were collected at Sevettij.arvi

ð691350N; 281500E; 130 m above sea level (a.s.l.)), in

Finnish Lapland, from July 1992 until January 1996

(Fig. 1). The sampling device was a virtual impactor

which provides two particle size ranges: fine (aerody-

namic diameter, ADo2:5 mm) and coarse (AD 2.5–

15 mm). The collection time per sample was 48 h: Thesamples were weighed for particulate mass (PM),

analysed for black carbon (BC) by a light reflectance

technique, for major anions and cations by ion

chromatography, and for up to 46 elements by a

combination of instrumental neutron activation analysis

and particle-induced X-ray emission spectrometry

(Virkkula et al., 1999).

Other two data sets were obtained for Birkenes

ð581230N; 81150E; 190 m a:s:l:Þ and Skre(adalen ð581490N;

61430E; 465 m a:s:l:Þ in Southern Norway. The sam-

plings at the two sites were conducted in parallel,

according to a 2-2-3 day schedule, from January 1991

until March 1996. Gent PM10 stacked filter unit (SFU)

samplers were used for separating the size fractions of 2–

10 mm AD and of o2 mm AD, respectively. The filters

were analysed for PM, BC, and over 40 elements

(Maenhaut et al., 2000).

The fourth data set was measured at Sde Boker

(301510N; 341470E; 470 m a.s.l.), in the Negev Desert of

Israel. Sampling and measurement procedures are

similar to those applied at the Norwegian sites

(Maenhaut et al., 1996). The data set covers the period

from January 1995 to August 1998.

3.2. Trajectories

Three-dimensional ten-day backward trajectories ar-

riving at 300 m above ground level (a.g.l.) every 6 h have

been calculated for the four sites using the PC version

4.5 (March 2002 release) of the Hybrid Single-Particle

Lagrangian Integrated Trajectory Model (Draxler and

Hess, 1998; HYSPLIT4, 2002). We assume that the

chosen arrival height is within the mixed layer for all

four sites, so that there is effective coupling between

transport and surface measurements. NCEP/NCAR

reanalysis data have been used as meteorological data

input to the model. The resolution of the reanalysis data

archive is 6 h on a 2:51 2:51 global grid. The output ofthe trajectory model is hourly trajectory endpoints

indicating the geographical location and the height of

the air parcel. The position errors for ten-day trajec-

tories are rather large, of the order of 1000 km at the end

of the period (Stohl, 1998, for a review). It is supposed

that the use of the analysed wind field can significantly

reduce the trajectory uncertainty (Lin et al., 2001).

Moreover, the use of a large number of trajectories in

Fig. 1. Location of Sevettij.arvi, Skre(adalen, Birkenes, and Sde

Boker measurement sites.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–5618 5609

the statistics computations will minimize to a certain

degree the random errors generated by interpolation,

truncation and wind data.

Other trajectory data have been also available, but for

periods not fully covering the concentration data sets:

(1) three-dimensional 4-day back trajectories arriving at

950 hPa at Sevettij.arvi every 6–8 h; calculated by the

TRADOS model of the Finnish Meteorological Insti-

tute, (2) three-dimensional 5-day back trajectories

calculated once per day (1200 UT) and for up to 6

arrival levels (960, 900, 850, 800, 700, 500 hPa) for Sde

Boker by the German Weather Service, and (3) three-

dimensional 10-day back trajectories arriving at 300 m

a.g.l. at Birkenes and Skre(adalen every 3 h; computed

with the FLEXTRA model (Stohl et al., 1995) and using

ECMWF analysed wind fields.

4. Results and discussion

All computations have been performed on a 2:512:51 grid consistent with the meteorological grid. First,

we have applied the two methods to the fine sulphur

concentration time series measured at the four sites

(Table 1). The analysis of trajectories has revealed that

there is no predominant pathway for transport to

Sevettij.arvi; however, there is a slight preference for

transport from the Norwegian Sea. For the two sites in

Southern Norway, air masses arrive predominantly from

over the North Sea and the Atlantic Ocean, while at Sde

Boker, Israel, transport occurs mainly from South East

Europe. One would therefore expect, taking into

account the geographical distribution of the sources,

that fine sulphur concentrations measured at Sde Boker

should be higher than for the other three sites. Indeed,

Table 1 shows that the median concentration for fine

sulphur at Sde Boker is almost one order of magnitude

higher that at the three other sites.

4.1. Comparison and evaluation of the PSCF

and CF methods

Fig. 2 presents the fine sulphur concentration field as

computed with Eq. (3) (logarithmic mean) for all four

sites. The grid cells which have less than 5 sample counts

are not shown. Similar maps obtained by calculating the

mean concentration show good correspondence with the

logarithmic mean maps. Fig. 3 displays the correspond-

ing bootstrapped PSCF spatial distribution, calculated

at 95% confidence level with the 75th percentile of the

concentration distribution as threshold. PSCF and CF

calculations have also been performed for the other

trajectory data sets available, and a good agreement has

been found with the maps produced by making use of

HYSPLIT trajectories. The bootstrapped maps are

based on a population of 2000 samples.

A quick visual inspection of Figs. 2 and 3 shows that

both methods identify the same source regions. This is

not unexpected, since both methods are applied to the

same concentration and trajectory data. The most likely

sources to affect Sevettij.arvi are located in the Volga and

Ural industrial regions, and, possibly, the oil and gas

complex in Western Kazakhstan. Sources in Central and

Eastern Europe are also producing high concentrations.

These various source regions agree with emission

inventories and global budgets for sulphur dioxide and

with information on major industrial areas in Russia

(NILU, 1984; Olivier et al., 1996; Vinogradova, 2000).

The same regions as for Sevettij.arvi account for high fine

sulphur concentrations at the two Norwegian sites. In

addition, these sites are affected to a higher degree by

sources in England, Belgium, Germany and Poland.

These countries are all important emitters of sulphur

dioxide (Olivier et al., 1996; Vestreng, 2001). The

measured concentration values for fine sulphur at the

two sites, which are about 100 km apart, were well

correlated, indicating that atmospheric levels of sulphur

in Southern Norway are mainly due to long-range

transport. This conclusion is supported by the general

patterns of the calculated maps for the two sites. It can

be noted that both the PSCF and the CF maps are fairly

similar, except that the concentrations calculated with

the CF method are systematically lower for Skre(adalen,

as are the measured concentrations (about 23% for the

median). It is assumed that topographical differences

between the sites may be responsible for this difference.

A somewhat different picture is obtained for Sde Boker.

For this site, the main sources are located in Eastern

Europe, the Ukraine (Donetsk), and North West Turkey

(Istanbul area).

Table 1

Percentile concentrations ðng=m3Þ for fine S at the study sites

Number of

Site samples Sampling period 100 95 90 75 50 25 0

Sevettij.arvi 577 7/1992–1/1996 2743.6 1394.9 1055.1 656.0 320.6 149.1 22.5

Skre(adalen 780 1/1991–3/1996 3694.0 1489.7 1118.0 643.3 290.5 141.9 11.2

Birkenes 749 1/1991–3/1996 4255.9 1841.9 1318.4 814.8 378.5 180.8 18.3

Sde Boker 492 1/1995–8/1998 5683.5 3646.4 3147.0 2494.7 1892.2 1303.1 163.9

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–56185610

The fine sulphur ‘hot spot’ over Bosnia and Herzego-

vina on the Sevettij.arvi map has to be considered with

caution, as it lies at the edge of the domain. A look at

the Birkenes map also reveals a possible source in the

region. To test whether we have a false positive coming

from the ‘smearing’ of a real source, a further check has

been performed by calculating the probability map at

Birkenes for an ideal source (Vasconcelos et al., 1996a,

count 1 if trajectory crosses ideal grid cell, 0 otherwise)

placed in the ‘Black Triangle,’ a known source of

anthropogenic sulphur, which is downwind of former

Yugoslavia. The map showed no correlation between

the ‘Black Triangle’ and the region of former Yugosla-

via. Tests for various locations of the ideal source and

for the receptor site at Sevettij.arvi have produced the

same result, except in the case when neighbouring

Hungary has been chosen as source. The EMEP

inventory confirms that Bosnia and Herzegovina are

an important emitter of sulphur dioxide (Vestreng,

2001). However, the calculated PSCF/CF maps cannot

distinguish between the sources in former Yugoslavia

and Hungary for both receptor sites. It is quite probable,

that the strength of the sources in the maps is

overestimated due to poor counting statistics.

Fig. 4 displays the fine sulphur probability maps

before bootstrapping corresponding to the plots in

Fig. 3. It can be seen that most of the edge cells with

high probability values due to poor counting statistics

are removed by the bootstrapping technique.

Fig. 5 shows the spatial distribution of P with and

without bootstrapping for fine arsenic at Sevettij.arvi and

for fine vanadium at Sde Boker. As can be seen, the cells

with low significance value being excluded, the boot-

strapped maps are easier to interpret in terms of known

Sevettijarvi

x

Sde Boker

x

Birkenes

x

Skreadalen

x

Fig. 2. Fine sulphur concentration field ðng=m3Þ as computed using CF method for Sevettij.arvi, Sde Boker, Birkenes, and Skre(adalen.

The location of the sites is indicated with an ‘x’.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–5618 5611

sources. High arsenic loads in samples at Sevettij.arvi are

associated with air transported from the copper–nickel

production facilities at Nikel/Zapolyarnyi on the Kola

Peninsula (Reimann et al., 2000). Other possible sources

are coal mining and combustion in the Pechora basin

and copper–nickel production, fossil fuel combustion

and steel and iron industry in the Ural region (Pacyna

et al., 1985). PSCF calculations performed by placing an

ideal source in the Nikel/Zapolyarnyi industrial region

showed that the source is smeared to the East and

eventually reaches the Pechora basin. The map obtained

was similar to the map in Fig. 5. This means that, while

it is certain that the Kola industrial region is a source of

arsenic, it is very probable that the maxima obtained

from the calculations with actual concentration data are

false positives. In other words, by examining the

(bootstrapped) PSCF map only, one is unable to tell

whether the Pechora basin or the Ural region are sources

of arsenic for Sevettij.arvi.

It is apparent that the bootstrap technique does not

overcome the problem of source ‘smearing’ due to the

spatial distribution of the trajectories which renders

emission cells and neighbouring ones indistinguishable

of each other. While the methods correctly identify the

direction where sources are located, they have limited

spatial resolution (Vasconcelos et al., 1996a). A

straightforward solution for improving the spatial

resolution would be to calculate maps of concentra-

tions/probabilities by making use of data from more

than one site—the more, the better—as trajectories

ending at different locations would cross source cells

from different directions. In our case, the sites are

either too close (Birkenes and Skre(adalen), or too far

apart (e.g., Sevettij.arvi and Sde Boker) to obtain an

Sevettijarvi

x

Sde Boker

x

Birkenes

x

Skreadalen

x

Fig. 3. PSCF bootstrapped spatial distribution of fine sulphur at 95% confidence level, with 75th percentile of the concentration

distribution as threshold, for Sevettij.arvi, Sde Boker, Birkenes, and Skre(adalen. The location of the sites is indicated with an ‘x’.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–56185612

unambiguous result. We should note, however, that the

multiplication cell by cell of the non-bootstrapped fine

sulphur maps calculated with the 50th percentile as

threshold for Sevettij.arvi and Sde Boker (Fig. 6), i.e., the

joint probability of source contributions to the two

receptor sites, shows that potential sources are better

spatially resolved: one can identify the Saint Petersburg

and Moscow areas, Middle Volga and Kama region

(Eastern Russia), Donetsk (Southern Ukraine) and North

West Turkey as sources affecting both receptor sites.

Stohl (1996) and Wotawa and Kr .oger (1999) have

advocated the use of the redistribution method. This

may partly remove the smearing problem. Stohl (1996)

applied the method to particulate sulphate measurement

data sets at 14 EMEP sites taken separately and

combined, and Wotawa and Kr .oger (1999) applied it

to a simulated concentration data set generated with a

Lagrangian air quality model. They could identify

potential source areas with higher resolution than with

the method of Seibert et al. (1994), i.e., the method was

able to reproduce the steep gradients between emission

areas and no-emission areas, though it also enhanced

unrealistic structures in regions with poor counting

statistics (Wotawa and Kr .oger, 1999). We have applied

the redistribution method to various combinations of

the four sites and separately to each of the sites. We have

also obtained higher concentration gradients when

applying the redistribution to one site only. However,

when combining more sites, the results have not been

conclusive, as there are not, statistically speaking, many

cells crossed from different directions by trajectories.

4.2. Application of PSCF and CF for source identification

The map for Sde Boker in Fig. 5 indicates that the

Arabian Gulf countries are a regional source for fine

Sevettijarvi

x

Sde Boker

x

Birkenes

x

Skreadalen

x

Fig. 4. PSCF spatial distribution of fine sulphur, with 75th percentile of the concentration distribution as threshold, for Sevettij.arvi,

Sde Boker, Birkenes, and Skre(adalen. The location of the sites is indicated with an ‘x’.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–5618 5613

vanadium. The map for fine nickel was quite similar to

that for fine vanadium. Furthermore, observed fine

nickel concentrations were well correlated with fine

vanadium concentrations. Both species had average

crustal enrichment factors of 24 relative to the average

crustal rock composition of Mason (1966), when using

aluminium as reference element, suggesting that they

have originated from other sources than mineral dust. It

is well known that both species are good indicators for

residual oil burning (Nriagu and Pacyna, 1988; Chow,

1995). Possible sources of fine vanadium and nickel

are at Saudi oil refineries on the Red Sea coast (mainly,

R%abigh) and in the Gulf region, or at desalination/

power plants. As in the case of arsenic at Sevettij.arvi,

there are cells with high P values for vanadium inside

the Arabian Peninsula that can be attributed to the

smearing of the real sources which are closer to the

receptor site.

Sevettijarvi

x

As V

Sde Boker

x

Sevettijarvi

x

As V

Sde Boker

x

Fig. 5. Spatial distributions of (top) PSCF and (bottom) bootstrapped PSCF with 75th percentile as threshold for fine arsenic, at

Sevettij.arvi, and for fine vanadium, at Sde Boker. The location of the sites is indicated with an ‘x’.

x

x

Fig. 6. Spatial distribution of joint probability of source

contributions to Sevettij.arvi and Sde Boker for fine sulphur.

The location of the sites is indicated with an ‘x’.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–56185614

Fig. 7 presents the concentration fields obtained by

making use of Eq. (3) for various species in the fine size

fraction measured at Sevettij.arvi. Similar plots have

been obtained by employing the PSCF method, and a

good correspondence has been found. The major sources

of fine lead are located in the Volga region in Russia, in

x x

x x

x x

Fig. 7. CF at Sevettij.arvi for fine PM, BC ðmg=m3Þ; Zn, NHþ4 ; Pb, and NO�

3 ðng=m3Þ: The location of the site is indicated with an ‘x’.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–5618 5615

agreement with the location of Pacyna et al. (1995). The

same area is the main potential source of fine PM, BC,

Zn, and NHþ4 : A principal component analysis (PCA)

on the concentration data set (De Ridder et al., manu-

script in preparation), indicated that these four vari-

ables, together with fine S, K, V, and Mn, were all highly

loaded on a single component, which was a general

pollution component. The CF maps for fine S, K, V, and

Mn resembled those for fine PM, BC, Zn, and NHþ4 (the

map for fine S was shown in Fig. 2), indicating that the

various species which were highly loaded on the same

pollution component originated from roughly the same

source area. That fine PM was highly correlated with the

pollution component, and thus also with fine S, is quite

logical, as fine sulphate was an important contributor to

the fine PM, accounting for nearly 40% on average. It is

noteworthy that ammonium correlated well with sul-

phur, but not with nitrate. According to Seinfeld and

Pandis (1998), ammonia is first neutralized by (hydro-

gen)sulphate and afterwards by nitrate. The average

molar ratio of fine ammonium to fine non-sea-salt

sulphur at Sevettij.arvi was 0.76 (De Ridder et al.,

manuscript in preparation), thus indicating that sulphate

was only partially neutralized to ammonium hydro-

gensulphate. From the CF maps for fine S and NHþ4 it

appears that their source areas and those of their

precursor gases (SO2 and NH3) are roughly the same.

Fine nitrate was not correlated at all with fine

ammonium; it formed a unique component in the PCA

(De Ridder et al., manuscript in preparation) and its CF

map in Fig. 7 indicates that it originates from entirely

different source regions than the species that are

associated with the general pollution component. The

major source regions of fine nitrate are in Western

Europe, with maxima at the Benelux countries, Italy and

Spain. The maximum in Spain has to be treated with

caution, as it is at the edge of the domain. Those at the

Benelux countries and in Italy agree with maxima for

gridded emission data for NO2 (a precursor of

particulate nitrate) (Vestreng, 2001). Furthermore, they

agree fairly well with maxima in the European aerosol

nitrate concentration field that was derived from quality

analysed data by Schaap et al. (2002).

Similar maps as shown in Figs. 2–7 have been

produced for other species and each of the four sites.

More specific studies will be presented in forthcoming

publications, including an analysis of the effect of

changing the air mass arrival height and, in the case of

PSCF method, the threshold value. Seasonal variations

will also be investigated.

5. Conclusions

Two approaches for identification of source locations

and preferred transport pathways of atmospheric

particulate trace elements and aerosol species have been

investigated. Both approaches are based on combining

chemical data with calculated air parcel backward

trajectories. The two methods have been applied to four

multi-species multi-annual concentration time series

measured at sites in Finland, Norway, and Israel. It

has been found that the methods agree well with each

other and correctly identify known emission sources. We

have shown that the use of an ‘ideal source’ test can

provide supplementary information about real and

‘ghost’ sources arising from highly correlated probabil-

ity/concentration values. A non-parametric bootstrap

test has been developed in order to estimate the

statistical significance of the probability maps.

As for the relative merits of the PSCF and CF

methods, we feel that they complement each other. The

PSCF method works with probabilities, the CF method

with concentrations. CF maps may be more readily

understood and easier to interpret, in particular by

workers outside the field. However, when one looks for

common source regions for two or more sites, the PSCF

method has the advantage that working with probabil-

ities provides the same comparison base for all sites.

Another advantageous feature of the PSCF method is

that it allows one to examine the smearing of a real

source by means of an ideal source test. We may

conclude that application of trajectory statistics methods

can provide in many cases acceptable qualitative

information about source regions.

Acknowledgements

We gratefully acknowledge the financial support from

the ‘‘Belgian Office for Scientific, Technical, and

Cultural Affairs’’ (OSTC) and the ‘‘Fonds voor We-

tenschappelijk Onderzoek–Vlaanderen.’’

References

Ashbaugh, L.L., Malm, W.C., Sadeh, W.Z., 1985. A residence

time probability analysis of sulfur concentrations at Grand

Canyon National Park. Atmospheric Environment 19,

1263–1270.

Cheng, M.-D., Lin, C.-J., 2001. Receptor modeling for smoke

of 1998 biomass burning in Central America. Journal of

Geophysical Research 106, 22871–22886.

Cheng, M.-D., Hopke, P.K., Zeng, Y., 1993. A receptor-

oriented methodology for determining source regions of

particulate sulfate at Dorset, Ontario. Journal of Geophy-

sical Research 98, 16839–16849.

Chow, J.C., 1995. Measurement methods to determine com-

pliance with ambient air quality standards for suspended

particles. Journal of Air and Waste Management Associa-

tion 45, 320–382.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–56185616

Draxler, R., Hess, G.D., 1998. An overview of the HYSPLIT 4

modeling system for trajectories, dispersion and deposition.

Australian Meteorological Magazine 47, 295–308.

Henry, R.C., 1997. History and fundamentals of multivariate

air quality receptor models. Chemometrics and Intelligent

Laboratory Systems 37, 525–530.

Hopke, P.K., Li, C.L., Ciszek, W., Landsberger, S., 1995.

The use of bootstrapping to estimate conditional proba-

bility fields for source locations of airborne pollutants.

Chemometrics and Intelligent Laboratory Systems 30,

69–79.

HYSPLIT4, 2002. Hybrid single-particle lagrangian integrated

trajectory model. NOAA Air Resources Laboratory, Silver

Spring, MD, Web address: http://www.arl.noaa.gov/ready/

hysplit4.html.

Lin, C.-J., Cheng, M.-D., Schroeder, W.H., 2001. Transport

patterns and potential sources of total gaseous mercury

measured in Canadian high Arctic in 1995. Atmospheric

Environment 35, 1141–1154.

Maenhaut, W., Salomonovic, R., Cafmeyer, J., Ichoku, C.,

Karnieli, A., Andreae, M.O., 1996. Anthropogenic and

natural radiatively active aerosol types at Sede Boker, Israel.

Journal of Aerosol Science 27, S47–S48.

Maenhaut, W., Fran-cois, F., Ptasinski, J., Mertens, S.F.,

Hanssen, J.E., 2000. Five-year study of the atmospheric

aerosol composition, sources and chemical mass closure at

two sites in southern Norway. Journal of Aerosol Science

31, S180–S181.

Mason, B., 1966. Principles of Geochemistry, 3rd Edition.

Wiley, New York.

NILU, 1984. Emission sources in the Soviet Union. Technical

Report 4/84, The Norwegian Institute for Air Research,

Lillestr�m.

Nriagu, J.O., Pacyna, J.M., 1988. Quantitative assessment of

worldwide contamination of air, water and soils by trace

elements. Nature 333, 134–139.

Olivier, J.G.J., Bouwman, A.F., Van der Maas, C.W.M.,

Berdowski, J.J.M., Veldt, C., Bloss, J.P.J., Vesschedijk,

A.J.H., Zandveld, P.Y.J., Haverlag, J.L., Dec. 1996.

Description of EDGAR Version 2.0: A set of global

emission inventories of greenhouse gases and ozone-deplet-

ing substances for all anthropogenic and most natural

sources on a per country basis and on 11 11 grid. RIVM/

TNO Report 771060 022, Rijkinstituut voor Volksgezond-

heid en Milieu, Bilthoven, Netherlands.

Paatero, P., Tapper, U., 1994. Positive matrix factorization: a

non-negative factor model with optimal utilization of error

estimates of data values. Environmetrics 5, 111–126.

Pacyna, J.M., Ottar, B., Tomza, U., Maenhaut, W., 1985.

Long-range transport of trace elements to Ny (Alesund,

Spitsbergen. Atmospheric Environment 19, 857–865.

Pacyna, J.M., Scholtz, M.T., Li, Y.-F., 1995. Global budgets of

trace metal sources. Environmental Reviews 3, 145–159.

Poirot, R.L., Wishinski, P.R., 1986. Visibility, sulfate and

air mass history associated with the summertime aerosol

in northern Vermont. Atmospheric Environment 20,

1457–1469.

Poirot, R.L., Wishinski, P.R., Hopke, P.K., Polissar, A.V.,

2001. Comparative application of multiple receptor meth-

ods to identify aerosol sources in northern Vermont.

Environmental Science and Technology 35, 4622–4636.

Polissar, A.V., Hopke, P.K., Paatero, P., Kaufmann, Y.J., Hall,

D.K., Bodhaine, B.A., Dutton, E.G., Harris, J.M., 1999.

The aerosol at Barrow, Alaska: long-term trends and source

locations. Atmospheric Environment 33, 2441–2458.

Polissar, A.V., Hopke, P.K., Harris, J.M., 2001a. Source

regions for atmospheric aerosol measured at Barrow,

Alaska. Environmental Science and Technology 35,

4214–4226.

Polissar, A.V., Hopke, P.K., Poirot, R.L., 2001b. Atmospheric

aerosol over Vermont: chemical composition and sources.

Environmental Science and Technology 35, 4604–4621.

Reimann, C., Banks, D., de Caritat, P., 2000. Impacts of

airborne contamination on regional soil and water quality:

the Kola Peninsula, Russia. Environmental Science and

Technology 34, 2727–2732.

Schaap, M., Muller, K., ter Brink, H.M., 2002. Constructing the

European aerosol nitrate concentration field from quality

analyzed data. Atmospheric Environment 36, 1323–1335.

Seibert, P., Kromp-Kolb, H., Baltensperger, U., Jost, D.T.,

Schwikowski, M., Kasper, A., Puxbaum, H., 1994. Trajec-

tory analysis of aerosol measurements at high alpine sites.

In: Borrell, P.M., Borrell, P., Cvita$s, T. (Eds.), Transport

and Transformation of Pollutants in the Troposphere.

Academic Publishing, Den Haag, pp. 689–693.

Seinfeld, J.H., Pandis, S.N., 1998. Atmospheric Chemistry and

Physics. From Air Pollution to Climate Change. Wiley,

New York.

Stohl, A., 1996. Trajectory statistics—a new method to

establish source–receptor relationships of air pollutants

and its application to the transport of particulate sulfate in

Europe. Atmospheric Environment 30, 579–587.

Stohl, A., 1998. Computation, accuracy and applications

of trajectories—a review and bibliography. Atmospheric

Environment 32, 947–966.

Stohl, A., Kromp-Kolb, H., 1994. Origin of ozone in Vienna

and surroundings, Austria. Atmospheric Environment 28,

1255–1266.

Stohl, A., Wotawa, G., Seibert, P., Kromp-Kolb, H., 1995.

Interpolation errors in wind fields as a function of spatial

and temporal resolution and their impact on different types

of kinematic trajectories. Journal of Applied Meteorology

34, 2149–2165.

Vasconcelos, L.A.P., Kahl, J.D.W., Liu, D., Macias, E.S.,

White, W.H., 1996a. Spatial resolution of a transport

inversion technique. Journal of Geophysical Research 101,

19337–19342.

Vasconcelos, L.A.P., Kahl, J.D.W., Liu, D., Macias, E.S.,

White, W.H., 1996b. A tracer calibration of back trajectory

analysis at the Grand Canyon. Journal of Geophysical

Research 101, 19329–19335.

Vestreng, V., 2001. Emission data reported to UNECE/EMEP:

Evaluation of the spatial distribution of emissions. EMEP/

MSC-W Note 1/01, The Norwegian Meteorological

Institute, Oslo, July.

Vinogradova, A.A., 2000. Anthropogenic pollutants in the

Russian Arctic atmosphere: sources and sinks in spring and

summer. Atmospheric Environment 34, 5151–5160.

Virkkula, A., M.akinen, M., Hillamo, R., Stohl, A., 1995.

Atmospheric aerosol in the Finnish Arctic: particle number

concentrations, chemical characteristics, and source analy-

sis. Water, Air and Soil Pollution 85, 1997–2002.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–5618 5617

Virkkula, A., Hillamo, R.E., Kerminen, V.-M., Stohl, A., 1998.

The influence of Kola Peninsula, continental European and

marine sources on the number concentrations and scattering

coefficients of the atmospheric aerosol in Finnish Lapland.

Boreal Environment Research 2, 317–336.

Virkkula, A., Aurela, M., Hillamo, R., M.akel.a, T., Pakkanen,

T., Kerminen, V.-M., Maenhaut, W., Fran-cois, F., Caf-

meyer, J., 1999. Chemical composition of atmospheric

aerosol in the European subarctic: contribution of the Kola

Peninsula smelter areas, central Europe, and the Arctic

Ocean. Journal of Geophysical Research 104, 23681–23696.

Wehrens, R., Putter, H., Buydens, L.M.C., 2000. The boot-

strap: a tutorial. Chemometrics and Intelligent Laboratory

Systems 54, 35–52.

Wotawa, G., Kr .oger, H., 1999. Testing the ability of trajectory

statistics to reproduce emission inventories of air pollutants

in cases of negligible measurement and transport errors.

Atmospheric Environment 33, 3037–3043.

Wotawa, G., Kr .oger, H., Stohl, A., 2000. Transport of ozone

towards the Alps—results from trajectory analyses and

photochemical model studies. Atmospheric Environment

34, 1367–1377.

Xie, Y.-L., Hopke, P.K., Paatero, P., Barrie, L.A., Li, S.-M.,

1999. Locations and preferred pathways of possible sources

of Arctic aerosol. Atmospheric Environment 33, 2229–2239.

Zeng, Y., Hopke, P.K., 1988. A study of the sources of acid

precipitation in Ontario, Canada. Atmospheric Environ-

ment 23, 1499–1509.

A. Lupu, W. Maenhaut / Atmospheric Environment 36 (2002) 5607–56185618