Introduction
Atmospheric photochemistry produces a variety of radicals that exert a
substantial influence on the ultimate composition of the atmosphere. The OH
radical is the main oxidant in the atmosphere (Monks et al., 2009, and
references therein). Its actual concentration is determined by the balance
between its sources and sinks. While in many cases OH sources are well
understood, OH sinks are manifold and not completely characterised. OH
reactivity is a measure of the strength of the sinks for the OH radical. It
can be derived from the reaction rates of the reactants kOH+X and their
concentrations [X] (Kovacs et al., 2003):
kOH=∑kOH+VOCiVOCi+kOH+COCO+kOH+NONO+kOH+NO2NO2+kOH+SO2SO2+…
In situ measurements of OH reactivity have provided new insights into OH
loss chemistry and the oxidative ability of the atmosphere (e.g. Di
Carlo et al., 2004; Edwards et al., 2013; Hofzumahaus et al., 2009; Whalley et al., 2011; Yoshino et al., 2006). There are a
number of different techniques used for the direct measurement of OH
reactivity. The total OH loss rate measurement technique (TOHLM) was one of
the first techniques applied for determination of total OH reactivity based
on a single measurement (Ingham et al., 2009; Ren et al., 2003a; Shirley
et al., 2006). TOHLM is based on the measurement of the decay of
artificially produced OH following the introduction of reactants into an
ambient air sample within a flow tube. By varying the distance between the
OH injection point and the detector, the reaction time changes and provides
a series of relative decay rates (Kovacs et al., 2003; Kovacs and Brune,
2001). A similar approach is taken with the laser-induced pump and probe
technique, whereby decay in OH is detected by time-resolved laser-induced
fluorescence (Sadanaga et al., 2004). Another technique developed by
Sinha et al. (2008), called the comparative reactivity method (CRM), is based
on the measurement of a single reactant (most often pyrrole) which first
reacts with OH under clean air conditions and then under competitive
conditions with ambient air. The reaction takes place in a glass vessel and
is most commonly probed by PTR-MS. Recently,
Nölscher et al. (2012b) presented a GC-PID for the
detection of pyrrole for CRM.
These techniques enable comparison of directly measured OH reactivity to
calculated OH reactivity using Eq. (1) based on measurements of
individual compounds. The difference between the two is referred to
as missing OH reactivity. Reasons for an under-prediction of OH reactivity
maybe due to incomplete or inaccurate measurements of individual compounds
(Di Carlo et al., 2004; Kim et al., 2011; Kovacs and Brune, 2001).
Therefore, direct measurements of total OH reactivity can help to evaluate
the completeness of measured VOC budgets (Dolgorouky et al., 2012; Mao et
al., 2009; Mogensen et al., 2011).
In urban environments good agreement between measured and calculated OH
reactivity has been found. For example, no significant missing OH reactivity
was found in New York during summer (Ren et al., 2003b) and for both
Paris under clean marine air conditions (Dolgorouky et al., 2012) and
Tokyo (Yoshino et al., 2006) in the winter. Larger missing OH reactivity
of up to 30 % was found for all other seasons in Tokyo by Yoshino et
al. (2006), presumably owing to secondary reaction products, including semi-volatile oxygenated compounds, from atmospheric oxidation of VOCs. A similar
amount of missing OH reactivity was reported by Kovacs et al. (2003) for
urban measurements in Nashville. They suggest that non-measured short-lived
VOCs accounted for the missing reactivity. In Paris, a missing OH reactivity
of up to 75 % was found for continentally influenced air, which is also
attributed to highly oxidised compounds from photochemical processes during
transportation of these air masses (Dolgorouky et al., 2012). Similar
reasons were reported by Lou et al. (2010) to account for missing OH
reactivity measured in the highly populated Pearl River Delta.
Direct measurements of OH reactivity in rural areas generally tend to have
larger missing OH reactivity. Using PTR-MS and the CRM method in a boreal
forest in Finland during August 2008, Sinha et al. (2010) reported
missing OH reactivity of approximately 50 %. This site was revisited in
2010, when missing OH reactivity of 58 to 89 % was recorded
(Nölscher et al., 2012a). Similar results in a mixed deciduous
forest were obtained by Hansen et al. (2014), who reported missing OH
reactivity of 46 to 65 %. Both studies concluded that unmeasured
oxidation products were missing from the OH reactivity calculation.
Undetected biogenic emissions and transport of reactive compounds are also
cited as other reasons for missing OH reactivity. In contrast to those
findings, Ren et al. (2006) found no significant missing OH reactivity
on average during a summertime campaign in a deciduous forest in New York in
2002. They attributed this to differences in the composition of emitted
biogenic VOCs (BVOCs). Rainforests are a large sink for OH as they emit a huge
amount of VOCs. Measured OH reactivity in the rainforest of Borneo during
April 2008 yielded a missing OH reactivity of 70 % compared to calculated
reactivity from measurements of single compounds (Edwards et al.,
2013) and ∼ 53 % compared to modelled reactivity
(Whalley et al., 2011). Since isoprene makes up the biggest contribution
to OH reactivity, the effect of oxidation products of isoprene was discussed
(Edwards et al., 2013; Whalley et al., 2011).
While different possible explanations for missing OH reactivity have been
given, the wide range in reported missing OH reactivity suggests that many
reactants and processes remain unknown or cannot be measured at present.
Measurements of total non-methane organic carbon in the western Los Angeles
Basin (Chung et al., 2003) and results following the
application of a double-column (orthogonal) GC for urban air measurements
(Lewis et al., 2000) emphasise the large number of OH reactants that are
not measured with standard field equipment.
Measurements of non methane hydrocarbons (NMHCs) used for calculation of OH
reactivity are often performed with GC (Lou et al., 2010; Sadanaga et
al., 2005; Shirley et al., 2006) and therefore the time resolution of the
calculated OH reactivity is low due to long sampling and run times, e.g.
canister samples once every hour (Kovacs et al., 2003; Sadanaga et al.,
2005; Yoshino et al., 2006), when compared to measured total OH reactivity.
However, the sampling time during one GC cycle is often shorter than the
analysis time, and thus any high temporal variability in measured OH
reactivity is not easily captured when it is derived from GC data
(Nölscher et al., 2012a). When measured and calculated OH reactivity
are compared, high-time-resolution data are often averaged over intervals
that correspond to the GC cycle. The GC-FID systems for measurements of NMHCs
used by Dolgorouky et al. (2012), for example, were sampling over 10 min
and had an additional analysis run time of 20 min, resulting in a time
resolution of 30 min. Their GC-MS for measurements of OVOCs used a sampling
time of 30 min, resulting in an overall time resolution of 90 min,
corresponding to the frequency they calculated OH reactivity.
This work addresses the question of how temporal VOC concentration
variability is reflected with different sampling time resolutions.
Furthermore, the effect of averaging VOC data on calculated OH reactivity is
discussed alongside how this may affect the amount of “missing” OH
reactivity.
Relatively high time-resolved VOC data collected by PTR-ToF-MS are used to
calculate OH reactivity for selected compounds. Differing time resolutions
are analysed to explore the effects. Data from an urban winter campaign are
compared to measurements from a semi-rural summer campaign.
Experimental section
Two different sets of VOC mixing ratios measured with PTR-ToF-MS were used
for analysis. One was collected during the ClearfLo (Clean Air for London,
www.clearflo.ac.uk) (Bohnenstengel et al., 2015) winter
campaign in 2012 at an urban background site in London, UK. The second was
taken during the PARADE (PArticles and RAdicals: Diel observations of the
impact of urban and biogenic Emissions, http://parade2011.mpich.de/) campaign in late summer 2011 at a semi-rural
site located in the Taunus ridge, Germany.
Field data
ClearfLo
A PTR-ToF-MS (Series I; Kore Technology Ltd, UK) (see standard
PTR-MS apparatus in Barber et al., 2012;
Thalman et al., 2015) was
deployed at Sion–Manning School (51∘31′15′′ N, 0∘12′51′′ W) near the North Kensington urban background station in London
during the intensive observation periods of the ClearfLo project in 2012. A
general overview of the ClearfLo project and the measurement site is given
in Bohnenstengel et al. (2015). For background measurements a
hydrocarbon trap (activated carbon filter by Grace Alltech) was employed
once during the time period investigated here. Its efficiency was in the
range of 87–96 %. Calibration measurements were performed before
(acetone) and after (toluene and xylene) the campaign in the laboratory. For
the calibration of toluene and xylene a permeation tube was used, and
calibration of acetone was done by dilution of a gas standard with zero air.
The stability of the instrument during the campaign was monitored with a
bromobenzene internal standard. Based on these measurements no correction
needed to be applied. Of the two intensive observation periods (IOPs) (i.e.
winter: 6 January to 11 February; summer: 21 July to 23 August) data from
1 to 7 February 2012 were selected for analysis in this study. During this
period the measurement site was influenced by local sources, as well as by
air masses from other parts of the UK and the continent (Bohnenstengel et
al., 2015).
A dual-channel GC with flame ionisation detector (DC-GC-FID; Hopkins
et al., 2003) was deployed at the same site as the PTR-ToF-MS during the
ClearfLo IOPs. A wide range of VOCs, including alkanes, alkenes, dienes,
aromatic compounds and OVOCs, were measured (see Table 1). Stainless steel tubing heated to 80 ∘C was used as a sampling
line. This setup ensured the destruction of ozone present in the sample. The sampling time was
10 min,
while the analysis runtime was around 50 min, resulting in approximately one
measurement per hour.
Mean mixing ratios and concentrations with standard deviation, rate
coefficient and OH reactivity of the VOCs measured with DC-GC-FID during
ClearfLo from 1 to 7 February 2012.
Compound
VMR (ppbV)
Concentration
kOH
OH reactivity
(molecules cm-3)
(cm3 molecules-1 s-1)
(s-1)
Alkanes
Ethanea
12.91 ± 10.89
(3.14 ± 2.65) × 1011
2.40 × 10-13
0.075
Propanea
4.59 ± 3.35
(1.12 ± 0.81) × 1011
1.10 × 10-12
0.123
iso-Butaneb
1.42 ± 1.00
(3.45 ± 2.43) × 1010
2.12 × 10-12
0.073
n-Butaneb
2.35 ± 1.60
(5.71 ± 3.89) × 1010
2.36 × 10-12
0.135
Cyclopentaneb
0.10 ± 0.11
(2.50 ± 2.58) × 109
4.97 × 10-12
0.012
iso-Pentaneb
0.83 ± 0.62
(2.03 ± 1.50) × 1010
3.60 × 10-12
0.073
n-Pentaneb
0.42 ± 0.26
(1.02 ± 0.64) × 1010
3.80 × 10-12
0.039
2,3-Methylpentaneb∗
0.35 ± 0.29
(8.56 ± 6.93) × 109
3.10 × 10-11
0.265
n-Hexaneb
0.13 ± 0.09
(3.16 ± 2.29) × 109
5.20 × 10-12
0.016
n-Heptaneb
0.09 ± 0.07
(2.18 ± 1.58) × 109
6.76 × 10-12
0.015
2,2,4-TMPb
0.04 ± 0.02
(9.88 ± 5.22) × 108
3.34 × 10-12
0.003
n-Octaneb
0.03 ± 0.02
(6.75 ± 3.75) × 108
8.11 × 10-12
0.005
Alkenes
Ethenea
1.93 ± 1.04
(4.68 ± 2.52) × 1010
7.80 × 10-12
0.365
Propenea
0.43 ± 0.30
(1.05 ± 0.73) × 1010
2.90 × 10-11
0.306
trans-2-Buteneb
0.04 ± 0.03
(1.03 ± 0.81) × 109
6.40 × 10-11
0.066
1-Buteneb
0.08 ± 0.05
(1.90 ± 1.21) × 109
3.14 × 10-11
0.060
iso-Butenea
0.11 ± 0.07
(2.63 ± 1.77) × 109
5.10 × 10-11
0.134
cis-2-Buteneb
0.03 ± 0.02
(6.92 ± 5.72) × 108
5.64 × 10-11
0.039
trans-2-Penteneb
0.04 ± 0.03
(9.13 ± 7.37) × 108
6.70 × 10-11
0.061
1-Penteneb
0.03 ± 0.02
(7.32 ± 5.27) × 108
3.14 × 10-11
0.023
Acetylenea
1.43 ± 0.74
(3.47 ± 1.81) × 1010
7.50 × 10-13
0.026
Dienes
Propadieneb
0.02 ± 0.01
(4.40 ± 2.61) × 108
9.82 × 10-12
0.004
1,3-Butadieneb
0.05 ± 0.03
(1.14 ± 0.76) × 109
6.66 × 10-11
0.076
Isoprenea
0.02 ± 0.02
(5.37 ± 4.07) × 108
1.00 × 10-10
0.054
Aromatic compounds
Benzenea
0.41 ± 0.17
(9.88 ± 4.06) × 109
1.20 × 10-12
0.012
Toluenea
0.64 ± 0.48
(1.56 ± 1.17) × 1010
5.60 × 10-12
0.087
Ethylbenzeneb
0.14 ± 0.11
(3.48 ± 2.57) × 109
7.00 × 10-12
0.024
m+p-Xyleneb∗
0.18 ± 0.14
(4.28 ± 3.52) × 109
1.87 × 10-11
0.080
o-Xyleneb
0.17 ± 0.12
(4.02 ± 2.82) × 109
1.36 × 10-11
0.055
Oxygenated VOCs
Acetaldehydea
2.37 ± 1.38
(5.77 ± 3.35) × 1010
1.50 × 10-11
0.866
MACRb
0.16 ± 0.12
(3.89 ± 2.97) × 109
2.90 × 10-11
0.113
Methanola
1.44 ± 0.81
(3.50 ± 1.96) × 1010
9.00 × 10-13
0.031
Acetonea
1.11 ± 0.51
(2.69 ± 1.24) × 1010
1.80 × 10-13
0.005
MVKb
0.28 ± 0.15
(6.72 ± 3.61) × 109
2.00 × 10-11
0.134
Ethanola
5.48 ± 3.81
(1.33 ± 0.93) × 1011
3.20 × 10-12
0.426
Propanola
0.31 ± 0.21
(7.41 ± 5.15) × 109
5.80 × 10-12
0.043
Butanola
0.59 ± 0.33
(1.45 ± 0.80) × 1010
8.50 × 10-12
0.123
a IUPAC preferred value; b Atkinson and Arey (2003); * average of both.
PARADE
For comparison, data collected with a PTR-ToF-MS (Ionicon Analytik
GmbH, Austria) (described in Jordan et al., 2009) during the PARADE
field campaign were analysed. Measurements were taken between 15 August and
9 September 2011 at the Taunus observatory on the summit of Kleiner Feldberg
(50∘13′25′′ N, 8∘26′56′′ E) under various meteorological conditions. A detailed description of the
measurement site and measurements performed during PARADE can be found in
Crowley et al. (2010) and Bonn et al. (2014). The PTR-ToF-MS was
operated continuously with minor interruptions. Background measurements were
conducted regularly with zero air throughout the campaign and calibration
measurements were performed with a multicomponent gas standard before and
after the campaign in the laboratory. For this study, 2 weeks of data (21
to 27 August 2011 – Period 1; 1 to 6 September 2011 – Period 2) were
selected, each with approximately the same amount of data points as the
ClearfLo data set. Period 1 was mainly influenced by continental air masses
and only towards the end by air that travelled over the UK and the English
Channel (UK-marine). Period 2 was dominated by UK-marine air, but was also
influenced by air masses that travelled over the Atlantic (see Phillips et al., 2012).
Data
Time series of VOCs during ClearfLo (top) and PARADE Period 1
(bottom). The time resolution is 1 min.
![]()
While the ClearfLo data presented here were collected at an urban
background site with mainly anthropogenic emissions, the PARADE campaign
took place at a semi-rural site. Biogenic emissions were expected from the
direct vicinity, but some anthropogenic influence was apparent from the
proximity of the highly populated Rhein-Main area and Frankfurt.
Three mass channels were selected for the analysis corresponding to
acetone/propanal, toluene and ethylbenzene/xylene. For brevity, the combined signal of acetone and propanal is referred to as acetone, and
that of ethylbenzene and xylene is referred to as xylene. The compounds used for analysis represent different sources of VOCs.
Toluene and xylene are counted along anthropogenic VOCs, monoterpenes are of
biogenic origin, and the OVOCs (acetone and methanol) are emitted directly or
produced by photochemical oxidation in the atmosphere (Monks et al., 2009,
and references therein). They were selected because their volume
mixing ratios could be determined with low uncertainty for both instruments.
Aromatic compounds such as toluene and xylene are well suited for this
investigation because they often show short-term high variability. The
analysis of the PARADE data also includes methanol and the sum of
monoterpenes. The characteristic parameters of the measurements during
ClearfLo and PARADE are given in Table 2.
Characteristics of the two different PTR-ToF-MS deployed during
ClearfLo and PARADE. Given are the sensitivities based on normalised counts
per second (ncps), accuracy as error for the measurements, and the limit of
detection (LOD), which was calculated as 1σ for ClearfLo and
2.6σ for PARADE based on 1 min data.
Compound
Sensitivity
Accuracy
LOD (1σ)
(ncps ppbV-1)
(%)
(ppbV)
Acetone
9.89
18*
0.56
ClearfLo
Toluene
6.36
18* (22)
0.38
Xylene
9.00
18* (20)
0.41
Compound
Sensitivity
Accuracy
LOD (2.6σ)
(ncps ppbV-1)
(%)
(ppbV)
Acetone
37.0
16
0.08
Toluene
26.9
8
0.04
PARADE
Xylene
33.4
13
0.01
Methanol
12.7
17
0.24
Monoterpenes
14.1
10
0.02
Effects of isobaric overlap from fragmentation taken into account.
*First column does not include effect of isobaric overlap from aromatic fragmentation; second column includes estimation of isobaric overlap.
Figure 1 shows the time series of the VOCs for
ClearfLo (top) and PARADE Period 1 (bottom). The range of mixing ratios for
ClearfLo is much wider and higher mixing ratios are reached. Acetone shows
values up to a factor of 1.8 higher in ClearfLo compared to PARADE, while
the aromatic compounds are 2 orders of magnitude higher. This emphasises
the diversity of the two field sites. In the box plots, presented in
Fig. 2, some interesting patterns are apparent.
For ClearfLo all three compounds exhibit a similar interquartile range (0.60
to 0.86 ppbV) but also very high maximum values. For PARADE a different
distribution is depicted. Acetone has a wider interquartile range of 1.83 ppbV
and has a higher mean value than toluene and xylene. The aromatic
compounds have a much smaller range compared to ClearfLo (0.03 to 0.08 ppbV). Methanol has a wider range than acetone and the monoterpenes look
similar to the aromatic compounds. Both periods of PARADE show the same
pattern. The ranges of the mixing ratios during the campaigns are summarised
in Table 3. Values below the limit of detection
(LOD) as well as negative values are not disregarded in this analysis to
preserve the full range of the data in order that they can be compared to a
randomly generated data set.
Overview of the range of VOC mixing ratios in ppbV during ClearfLo
and PARADE (PAR).
Minimum
Maximum
Mean
Interquart. range
Max–Min
ClearfLo
Acetone
-0.294
9.816
1.459
0.864
10.110
Toluene
0.058
13.982
1.162
0.862
13.924
Xylene
0.038
13.519
0.861
0.601
13.482
PAR 1
Acetone
0.426
5.447
2.400
1.833
5.021
Toluene
-0.030
0.592
0.076
0.078
0.622
Xylene
-0.008
0.277
0.041
0.030
0.285
Methanol
0.851
10.775
4.438
3.858
9.923
Monoterp.
-0.008
0.801
0.124
0.116
0.809
PAR 2
Acetone
0.544
4.873
1.987
1.797
4.329
Toluene
-0.017
0.646
0.078
0.073
0.663
Xylene
-0.004
0.358
0.046
0.039
0.362
Methanol
0.781
10.649
3.776
2.739
9.869
Monoterp.
-0.009
0.692
0.075
0.086
0.701
Box plots for ClearfLo (left) and PARADE Period 1 (right) showing
the minimum, maximum, mean (∎), median, interquartile range (box)
and percentiles at 1 and 99 % (×).
![]()
OH reactivity relating to the VOCs under study is calculated from the first
term of Eq. (1). Reaction rates for acetone (1.8×10-13 cm3 molecule-1 s-1), toluene (5.6×10-12 cm3 molecule-1 s-1),
methanol (9.0×10-13 cm3 molecule-1 s-1) and α-pinene (5.3×10-11 cm3 molecule-1 s-1)
are taken from http://iupac.pole-ether.fr/index.html. The exact composition of the
monoterpene signal is not known; thus, only the reaction rate of α-pinene is used. For xylene the average of the reaction rates of
ethylbenzene and o-, m- and p-xylene (14.5×10-12 cm3 molecule-1 s-1) (Atkinson and Arey, 2003) was applied.
Table 4 summarises the minimum, maximum and mean
reactivity calculated from these VOCs as described.
The time resolution of PTR-ToF-MS is only limited by the signal-to-noise
ratio and resulting detection limit. Both instruments were operated with a 1 min time resolution. Volume mixing ratios of VOCs were averaged over
different intervals and standard deviations were derived. An average was
only included for analysis if its data recovery was at least 50 %. The OH
reactivity for each VOC was calculated and summed as required. Only the
standard deviations were propagated as errors of reactivity, as the focus of
this work is on investigating VOC variability.
For clarity throughout this paper the notation R=ROH for reactivity
regarding the OH radical replaces kOH (see Eq. 1; see also
Nölscher et al., 2012a). Indices denote the origin of the data (PTR (PTR-ToF-MS) or GC (DC-GC-FID) and CL (ClearfLo) or PAR (PARADE)).
Numbers indicate the averaging time in minutes. If only some VOCs are taken
into account for calculating the reactivity, this will be indicated, e.g.
RPTR,CLOVOC,5is the OH reactivity calculated from the 5 min mean
concentration of acetone, measured with the PTR-ToF-MS during ClearfLo.
Distribution of OH reactivity of measured VOCs
Minimum, maximum and mean OH reactivity and standard deviation
calculated from the VOCs under study for ClearfLo and PARADE.
OH reactivity (s-1)
Minimum
Maximum
Mean
SD
CF
ATX
0.036
4.864
0.463
0.289
PAR1
ATX
0.000
0.191
0.035
0.028
ATX+M+MT
0.026
1.296
0.292
0.205
PAR2
ATX
0.001
0.222
0.036
0.024
ATX+M+MT
0.044
0.987
0.215
0.119
ATX: acetone, toluene and xylene. ATX+M+MT: acetone, toluene, xylene, methanol and monoterpenes
.
For a more general view of the factors that drive variation in OH reactivity
of VOCs, its frequency distribution was investigated. GC data from the winter
(9 January–9 February 2012) and summer IOP (18 July–19 August 2012)
during ClearfLo were applied. The OH reactivity, RGC,CLVOCi,
was calculated for each measured VOCi and ranged from 0.003 to 0.822 s-1
in winter and from 0.001 to 1.568 s-1 in summer with a total
OH reactivity RGC,CLTVOC of 4.010 s-1 and 3.862 s-1,
respectively. The majority of RGC,CLVOCi
values lies below 0.1 s-1, as can be seen from the frequency distribution plotted in
Fig. 3, where more than 70 % of the winter and
80 % of the summer data are in the first interval from 0 to 0.1 s-1.
Seasonal differences in OH reactivity emission rates have previously been
described by Nölscher et al. (2013) for measurements at a Norway
spruce between spring and early autumn. Although the composition of VOCs
during ClearfLo changed from winter to summer, no seasonal dependency could
be found in the shape of the frequency distribution of RGC,CLVOCi. In both cases RGC,CLTVOC is dominated by the sum of
low-reactivity contributions and less by single compounds with high reactivity.
OH reactivity, RGC,CLVOCi, frequency distribution for
the ClearfLo campaign in winter (left) and summer (right). Bin size is 0.1 s-1 for both plots.
![]()
Generation of a randomised data set
To differentiate between pure statistical effects and measurement related
characteristics, a randomised data set was produced and analysed in the same
way as the PTR-ToF-MS data. The distribution of OH reactivity is skewed
towards smaller values and only positive values of OH reactivity are
expected; hence, it is better described by a log-normal distribution compared
to a normal distribution (Limpert et al., 2001). The data set of random
numbers was generated by simulating an OH reactivity distribution comparable to
the ClearfLo data set. The sample mean m=0.463 s-1 and standard
deviation SD = 0.289 s-1 from the ClearfLo 1 min data set were used to
define the parameters μ (Eq. 2) and σ (Eq. 3) for
the log-normal distribution of random numbers.
μ=logm1+SD2m2σ=log1+SD2m2
A log-normal distribution of a total of 8040 random numbers was generated
using the dlnorm (#, μ, σ) function in R. This provides a
set of data comparable to 134 h of OH reactivity measurements with a
time resolution of 1 min. Figure 4 shows the random
data set as a time series together with the hourly mean containing 60 data
points. On inspection of Fig. 4, it becomes
obvious that the range of the hourly average is very small, with a standard
deviation of 0.034 s-1.
Time series of randomly generated log-normal data set containing
8040 numbers. The “1 min” data are shown in black. The average over 60 data
points is plotted in red.
![]()
Results and discussion
Initially the OH reactivity RPTRVOC was calculated from the PTR
data for ClearfLo and PARADE. For both campaigns the signals of acetone,
toluene and xylene (referred to generally as VOCs) were used. The effects of
differing sampling intervals on the derived reactivity were explored. For
each campaign and VOC data set a correlation of the average values of
RPTRVOC,t for different intervals (t=5, 10, 20 and 30 min)
against the 60 min average RPTRVOC,60 was calculated. The intervals
are chosen to be the first t minutes of each hour to simulate the initiation
of a GC sequence; thus, the 10 min average also covers the 5 min averaging
period and so on.
Figure 5 shows the linear correlation of the 5 min
average RPTR,CLVOC,5 versus the 60 min value RPTR,CLVOC,60
for the ClearfLo winter campaign. Data were fitted with a bivariate
regression line with an intercept (bvf) and forced through the origin
(bvfo). The deviation from the slope of the linear regression to a unity
gradient mres=mR<60/R60-1 is taken as a
measure of how well the value of hourly OH reactivity is represented by the
shorter interval average and is further referred to as the residual slope.
Effects of different sampling intervals
The slopes of both fits in Fig. 5 are below 1.0,
indicating an under-prediction of the reactivity during ClearfLo by the
value calculated from the first 5 min of each hour. In this case there is
only a small deviation (1.3 %) from a unity gradient (see
Fig. 6). For all averaging intervals the slope is
equal to 1 in the range of the uncertainties of the fit.
Standard deviation of ΔR and results from Gaussian fits of
the ratio of OH reactivity RPTR,CLVOC,t<60/RPTR,CLVOC,t=60 calculated from shorter
interval averages to 60 min average for ClearfLo.
ΔR
Gaussian fit
Notation
Time interval
SD
Centre
FWHM
(min)
(s-1)
RPTR,CLVOC,5
5
0.12
0.998 ± 0.011
0.337 ± 0.025
RPTR,CLVOC,10
10
0.12
0.997 ± 0.008
0.244 ± 0.020
RPTR,CLVOC,20
20
0.10
0.988 ± 0.006
0.246 ± 0.013
RPTR,CLVOC,30
30
0.06
0.992 ± 0.004
0.198 ± 0.009
For the different averaging intervals the difference to the hourly average
(ΔR=RPTR,CLVOC,t<60-RPTR,CLVOC,t=60) was
calculated, and their standard deviations are given in Table 5
as a measure of variance. ΔR generally decreases with increasing
averaging time. Also presented in Table 5 are the
results from fitted Gaussian functions to the frequency distribution of the
ratio of the shorter interval averages to the 60 min average. Bins of 0.1
were chosen for the frequency distributions. Ideally, the centre of the
Gaussian fit is 1, while the full width at half maximum (FWHM) describes the
spread of the distribution around its centre. Both the standard deviation of
ΔR and the FWHM decrease when averages are calculated for
longer intervals. The centres of all Gaussian fits achieve 0.99.
Linear correlation with bivariate fit (bvf: fit with intercept;
bvfo: fit forced through the origin) of the OH reactivity calculated from
the signals of acetone, toluene and xylene for average intervals of 5 min
and 60 min for ClearfLo. The standard deviations of the 5 min means are
plotted as error bars.
![]()
Development of the slope of the correlation of OH reactivity
RPTR,CLVOC,t depending on the sampling interval for ClearfLo.
Slopes for bvfo (black) and bvf (blue) with their lower and upper limits are
shown.
![]()
For Period 1 of the PARADE data (PAR1) the results show a slope greater than
1 (Fig. 7). The high variability in the data is
reflected by a higher divergence of the slopes of 1.13 for bvf fit and 1.05
for the bvfo fit based on 5 min averaged data. The small standard deviations
of ΔR given in Table 6 highlight the narrow
range of calculated OH reactivity RPTR,PAR1VOC However, the high
variability in the data is reflected by the FWHM of the frequency
distributions of the ratios, which is higher for each interval when compared
to ClearfLo.
Standard deviation of ΔR and results from Gaussian fits of
the ratio of OH reactivity RPTR,PAR1VOC,t<60/RPTR,PAR1VOC,t=60 calculated from
shorter interval averages to 60 min average for PARADE – Period 1.
ΔR
Gaussian fit
Notation
Time interval
SD
Centre
FWHM
(min)
(s-1)
RPTR,PAR1VOC,5
5
0.016
0.980 ± 0.011
0.379 ± 0.027
RPTR,PAR1VOC,10
10
0.015
0.976 ± 0.012
0.353 ± 0.026
RPTR,PAR1VOC,20
20
0.012
0.997 ± 0.013
0.310 ± 0.030
RPTR,PAR1VOC,30
30
0.008
0.995 ± 0.009
0.273 ± 0.020
Development of the slope of the correlation of OH reactivity
RPTR,PAR1VOC,t depending on the averaging time for PARADE – Period
1.
![]()
For Period 2 (PAR2), an over-prediction of the OH reactivity
RPTR,PAR2VOC can be observed again (Fig. 8), but with an even greater slope of 1.26. In both periods of PARADE the
slope approaches a value of 1 as increasing averaging time takes more of
the variability within 1 h into account. Standard deviations of ΔR and FWHM values are similar to Period 1 of the PARADE data, while the
centres of the Gaussians are closer to 1 (Table 7).
Development of the slope of the correlation of OH reactivity
RPTR,PAR2VOC,t depending on the averaging time for PARADE – Period
2.
![]()
Standard deviation of ΔR and results from Gaussian fits of
the ratio of OH reactivity RPTR,PAR2VOC,t<60/RPTR,PAR2VOC,t=60 calculated from
shorter interval averages to 60 min average for PARADE –
Period 2.
ΔR
Gaussian fit
Notation
Time interval
SD
Centre
FWHM
(min)
(s-1)
RPTR,PAR2VOC,5
5
0.013
0.996 ± 0.014
0.352 ± 0.034
RPTR,PAR2VOC,10
10
0.009
0.994 ± 0.013
0.296 ± 0.031
RPTR,PAR2VOC,20
20
0.008
0.992 ± 0.008
0.238 ± 0.019
RPTR,PAR2VOC,30
30
0.006
1.010 ± 0.004
0.238 ± 0.010
When OH reactivity is calculated from GC measurements of VOCs, some of the
variability in the data is not captured, because air sampling alternates
with the GC run itself (Hopkins et al., 2003). In this manner, the
analytes are collected for a short duration which is then used to represent
the whole measurement cycle. This work suggests that a discrepancy between
60 min averages and shorter intervals can be caused owing to the variable
nature of atmospheric VOCs. A sampling time of only 5 min can cause a
deviation of more than 25 %. Accordingly, this would then artificially contribute to a deviation in OH reactivity, causing a positive or negative bias. Thereby, it is an additional error source when comparing
measured total OH reactivity to OH reactivity calculated from GC data.
The deviation is greater for the semi-rural measurements in the Taunus
during PARADE compared to the urban measurements in London. Although the
range of the analysed OH reactivity of VOCs is smaller during PARADE, the
highly frequent fluctuations cause a greater variability in OH reactivity
for the investigated intervals.
The distribution of residual slopes across consecutive 5 min
intervals
In the previous section, only reactivity calculated from the average of the
first 5, 10, 20 and 30 min was compared to the hourly mean. Naturally, these
averages have different values, depending on the point at which they are
selected from the hour under study. They may over- or under-predict the
hourly mean as can be seen from Fig. 9, where
residual bvf slopes between RPTRVOC,5 and RPTRVOC,60
(see Fig. 5) are plotted for consecutive 5 min averaging periods within the
hour. Depending on the selected 5 min interval the bvf resulted in a
divergence of -0.1 to 44 % for ClearfLo, 1 to 13 % for PAR1, -3 to 26 % for PAR2 and -2 to 10 % for the randomised data. A
tendency towards an over-prediction of OH reactivity was observed for both
campaigns (ClearfLo – top left; PARADE – bottom) and also for the
randomised data set (top right). For the randomised data set bvfo was used – a bvf has a much higher slope as the data are clustered together within a
small range. On average the residuals are nearly 10 % with a standard
deviation of 0.1 % or less (8.6 % ± 0.1 % for ClearfLo;
8.85 % ± 0.03 % for PARADE 1; 9.5 % ± 0.1 % for PARADE
2; 4 % ± 4 % for the randomised data).
Residual slopes of the correlation of all 5 min means to the hourly
mean.
![]()
For linear regression the standard least-squares fit is widely used. This
method is less adequate when errors in both y and x are assumed or when the
assignment of the independent variable is not clear (Isobe et al.,
1990). Other methods for bivariate fitting in natural sciences have been
discussed in the literature (Cantrell, 2008; Isobe et al., 1990; Warton
et al., 2006). Cantrell (2008) found that a bivariate fit is less
sensitive to outliers compared to an ordinary least-squares (ols) fit.
Warton et al. (2006) described the major axis (ma) and standard or
reduced major axis regression (sma/rma). These methods are preferred when
the agreement between two measurement techniques is investigated. For
equally important deviations from the regression line in the x and y
directions ma is used, while sma can be used when the scales in x and y are
not comparable. These two functions are implemented in the smatr package in R.
The ma function is used to produce the bivariate regression line (bvf) and
the bivariate regression forced through the origin (bvfo) in this work. In
the work of Isobe et al. (1990) the ordinary least-squares regression,
major axis and reduced major axis regression, and additionally ols bisector
(ols-bis) regression, are compared. They point out that different slopes are
to be expected for all the bivariate fits (ma, sma, ols-bis). For ma they
find large uncertainties for the slope. To carry out a symmetrical analysis
they recommend using the ols-bis regression.
Figure 10 shows the residual slopes between
RPTR,CLVOC,5 and RPTR,CLVOC,60 for consecutive 5 min
intervals of the ClearfLo data using the different regression methods (ols,
ols-bis, bvfo, bvf). The mean of the residual ratios (i.e. the average
ratio minus 1) of RPTR,CLVOC,5 to RPTR,CLVOC,60 is also
shown in Fig. 10. The bvfo puts more weight onto
low OH reactivity values compared to bvf and produces a line that much better matches the
majority of the data. Therefore smaller residuals are observed
compared to the bvf. The very small residual of the average ratio also
emphasise that deviation from the ideal slope of 1 is mainly driven by
outliers. The ols-bis regression shows a negative residual for all 5 min
intervals. Mean deviations and ranges for all regression methods based on
consecutive 5 min averaging periods are summarised in
Table 8, where it can clearly be seen that ols and the ratio have a negligible deviation once
averaged across 12 intervals.
On average the ols shows the smallest deviation from the ideal slope of 1,
but in terms of stability across all 5 min intervals the ols-bis performs
better. This analysis shows that the extent of under- or over-predicting OH
reactivity by short sampling intervals is a matter of how the data are
compared to each other.
Summary of the statistics of the residual slopes and ratios from
the comparisons of the 5 min means to their 60 min means for the ClearfLo
data.
Method
Min
Max
Range
Mean
SD
ols
-0.129
0.250
0.379
-0.008
0.112
ols-bis
-0.119
-0.024
0.094
-0.080
0.036
bvfo
-0.033
0.116
0.149
0.014
0.043
bvf
-0.019
0.443
0.462
0.861
0.130
Ratio
-0.030
0.031
0.061
-0.006
0.017
Residual slopes from different linear regression methods and the
mean residual ratio for all 5 min intervals. Error bars depict the standard
error of the slope for the ols fit, the square root of the variance for the
ols-bis, and the lower and upper limit for the bvfo and bvf fits and the
standard deviation of the ratios.
![]()
The same analysis was performed with an extended data set that included 10
times the number of data points of a randomised log-normal distribution to
test for any artefacts relating to the limited sample size of the PTR-ToF-MS
data. No appreciable difference was obtained when compared to the smaller
data set. Hence, we conclude that the observed bias to an overestimation for
the bivariate fits and an underestimation for the ols-bis regression on
average is real and not an artefact caused by computing a shorter time
series.
At what sampling interval can the hourly mean be represented with a smaller
subsample?
The question being further investigated here is: how many data points are
needed to calculate an average value that represents the hourly mean within
its standard deviation? The ClearfLo data set of OH reactivity, based on
acetone, toluene and xylene, was used to calculate 60 min means of
consecutive 1 min data. Small gaps in the time series were skipped such that
60 contiguous data points were computed. However, data was discarded if it
included larger gaps, e.g. 1 h or more. The set of 60 data points was
further subdivided into smaller intervals to calculate means of OH
reactivity RPTR,CLVOC,t<60 of 2, 3, 5, 10, 15, 20 and 30 min.
Residual reactivities for these averages were calculated by subtracting the
hourly mean RPTR,CLVOC,60 before being plotted against the number of data
points n, which in this case corresponds to minutes
(Fig. 11). Corresponding standard deviations were calculated for each 60 min
mean, but only the minimum and maximum values are plotted in Fig. 11 (dashed and
solid grey lines, respectively).
Additionally, two models are plotted, describing the course of the functions
f1(1/n) (light blue) and f2 (1/n) (dark blue) starting
at the maximum and minimum value (both marked as red dots). The positive
range of residual OH reactivity is much wider than the negative range and is
capped by the 1/n function. The negative values show a slower approach to the mean.
The 20 min averages all lie within the maximum standard deviation, but even
when averaging over 30 min, the range is much wider than the minimum standard
deviation of OH reactivity.
Dependency of the deviation in OH reactivity from the hourly mean
on the number of data points for the entire ClearfLo data set.
![]()
The 2, 3, 5, 10, 20 and 30 min averages are now compared directly to their
hourly mean and standard deviation to summarise the findings from
Fig. 11. As can be seen in
Table 9 at 20 min, 2.78 % of the ClearfLo
data still exceed their hourly mean. At 30 min all data lie within the range of
the standard deviation. Therefore, a sampling time greater than 20 min would
be required to represent the hourly mean. The random data reach a comparable
level of data exceeding the hourly mean by 2.80 % for averaging over 5 min
only. Here, sampling for only 10 min would be sufficient for
representing an hour's worth of data. The required sampling times mentioned
here correspond to the VOC variability in the analysed data sets. Likewise,
longer sampling times could be necessary for representing hourly OH
reactivity in other environments such as measurements closer to industrial
sources. For example, Gilman et al. (2009) have shown that a much
broader range of OH reactivity of VOCs with a high degree of variability can
be found in the proximity of heavily industrialised areas like the Houston
and Galveston Bay area in Texas, USA.
Bivariate fit results between 5 min averaged to 60 min averaged
reactivity. Slopes are plotted for ClearfLo (blue shaded area, left) and
PARADE (Period 1 – grey shaded; Period 2 – green shaded, right).
Correlations were analysed separately for OVOCs (acetone for ClearfLo and
acetone and methanol for PARADE), BVOCs (monoterpenes) and aromatic compounds
(toluene and xylene). The error bars indicate lower and upper limits of the
fitted slopes.
![]()
Comparison to the hourly averages and their standard deviation for
ClearfLo and a data set of log-normal distributed randomised numbers. Listed
are the number and the percentage of data that exceed the SD of the
hourly mean for different n. n refers to the number of minutes that were
averaged in each case.
ClearfLo
Randomised data set
n
No. data
No. data > SD
% of data > SD
No. data
No. data > SD
% of data > SD
2
3960
838
21.16
4020
534
13.28
3
2640
457
17.31
2680
199
7.43
5
1584
225
14.20
1608
45
2.80
10
792
80
10.10
804
0
0
15
528
38
7.20
536
0
0
20
396
11
2.78
402
0
0
30
264
0
0
268
0
0
Effect of different VOC classes on OH reactivity
Many different atmospheric VOCs have been identified (Goldstein and
Galbally, 2007), all of which contribute to OH reactivity. Based on their
chemical characteristics they are often divided into different classes. In
order to identify how the variation in individual components contributes to
the observed deviation of RPTRVOC,5 from RPTRVOC,60, correlations between 5 min and hourly mean reactivities were analysed for
different VOC classes separately. The results are shown in
Fig. 12 for ClearfLo (blue area) and PARADE (grey
and green areas), where OVOC contains the data from acetone for ClearfLo
(RPTR,CLOVOC,5) and acetone and methanol for PARADE
(RPTR,PAROVOC,5). The aromatics are calculated from toluene and
xylene and BVOC refers to the sum of the monoterpenes, which were only
available for PARADE. Again, a greater deviation from 1 is observed for the
PARADE data. The OVOCs show no significant deviation from 1 for both
campaigns, and while the aromatics are close to 1 for ClearfLo, they show a
significantly different value for PARADE with a deviation of up to 31 %.
Finally, BVOCs deviate from a perfect correlation by 21 % for the second
period of PARADE.
These results are in line with observations from Williams et al. (2000),
who investigated the variability–lifetime relationship of VOCs measured in an
unpolluted region of Surinam based on the standard deviation of the natural
logarithm of their concentration. They found a higher variability for
toluene compared to acetone and methanol. Compounds with a lifetime below 2 days did not seem to fit into this relationship.
Scaling the effect to the share of OH reactivity of VOCs during
ClearfLo
The observed deviations of the slopes from the ideal slope of 1 in
Fig. 12 for ClearfLo were scaled by their share
to determine the overall effect on total OH reactivity of VOC RCLTVOC. Data from the same week as the PTR-ToF-MS data were used to calculate the
influence of VOC speciation on OH reactivity. Over the period of 1 to 7 February 2012 the total OH reactivity of these compounds is
RGC,CLTVOC=4.05 s-1. Based on
Table 1, OVOCs contribute most to reactivity at
43 %, followed by alkenes at 26 % and alkanes at 21 % of
RGC,CLTVOC. The aromatic compounds have a share of 6 %, and
dienes, including isoprene, account for 3 %. Finally, the contribution of
the only measured alkyne is less than 1 %.
The extent of which different VOC classes' variability effects
RGC,CLTVOC was calculated by weighting the deviation derived from
the correlations for the different classes (i.e. the deviation of the slope
between RPTR,CLclass,5 and RPTR,CLclass,60 from 1) by the
proportion that each class contributes to the total reactivity (calculated
from Table 1). Here, it is assumed that deviations
derived from measurements of only a few compounds is representative of each
class of VOCs under study.
Percentage deviation in OH reactivity for different sampling
intervals owing to VOC variability for the ClearfLo data. Results are
plotted separately for aromatic compounds (black and grey squares) and OVOCs
(green and teal triangles) for both bivariate fits without (bvfo) and with an
intercept (bvf). The deviations are based on the share of the VOC's class to
total OH reactivity for each investigated averaging interval. The error bars
indicate lower and upper limits of the fitted slopes.
![]()
Based on Fig. 13, 5 min averages over-predict OH reactivity by up to 2.6 % due to
variability in OVOC concentrations. This value decreases for increasing
averaging time but shows a maximum of 3.4 % for the 30 min mean. There is
no significant contribution of the aromatic compounds to a deviation from
the hourly mean OH reactivity for any averaging interval.
A similar behaviour could be expected for other classes of VOCs such as the
alkenes and alkanes, the second and third most important classes in
Table 1. However, this could not be tested in the
present study using PTR-ToF-MS data. However, this study shows how the effect of
using short sampling intervals could account for a missing or over-predicted
OH reactivity in the range of 10 % or more.
Lidster et al. (2014) investigated the potential increase in OH
reactivity owing to higher substituted aromatic compounds, which are
normally not measured in field campaigns. They state that they can
contribute to up to 0.9 s-1 in OH reactivity. This would increase the
share of aromatic compounds by more than a factor of 3; however, based on the
results in Fig. 13, the effect on OH reactivity would still be in the range of less
than 1 %, while the contribution of the OVOCs would only be altered
slightly.
Conclusions
The effect of using short sampling intervals for VOC measurements on
resulting OH reactivity was investigated using two different monitoring
campaigns as case studies. OH reactivity was found to be both under- and
over-predicted due to missing variability in VOC data. The divergence between OH
reactivity calculated from 5 min sampling intervals and hourly values was
found to be around 2–26 % and 0–44 % for the PARADE and CleafLo
campaigns, respectively, owing to the variability in the VOC concentrations.
These discrepancies may contribute to missing OH reactivity when compared to
direct measurements. Results from the urban and the semi-rural site show on
average similar effects when comparing reactivity averaged over 5 min
intervals to the hourly mean.
Comparison to a randomised data set with a similar distribution as the
CleafLo data showed that the variability in the VOC concentrations with time
is the main reason for deviating results from shorter sampling intervals. For
the randomised data a sampling time of less then 10 min is sufficient so
that all data points are within the range of the hourly standard deviation,
while for the ClearfLo data it takes more than 20 min.
The effect of short sampling times of VOC concentrations on calculated OH
reactivity is differently pronounced for each VOC class. When comparing OH
reactivity calculated from VOCs sampled over a 5 min period to the hourly
mean, a larger divergence was found for the OVOCs than aromatic compounds
during ClearfLo. The opposite trend was observed for the PARADE campaign,
while the effect of OVOCs is almost negligible. Biogenic VOCs, with the
monoterpenes as representatives, were added for analysis. They show a
similar behaviour to the OVOCs, but with a slightly greater divergence.
The bigger proportion of measured OVOCs, compared to the aromatic compounds,
at the urban site during ClearfLo contributes to a higher deviation in
calculated OH reactivity when using short sampling intervals. Taking the
results from Lidster et al. (2014) into account, the effect of aromatic
VOCs increases but is still small.