Introduction
The atmospheric boundary layer (ABL) is the section of the atmosphere that
responds directly to the flows of mass, energy and momentum from the earth's
surface (Stull, 1988). Since most air pollutants are emitted or
chemically produced within this layer, its evolution plays an important role
in the transport, dispersion and deposition of air pollutants (Chen et al.,
2012; Fan et al., 2008; Whiteman et al., 2014; Platis et al., 2016; Wolf et al.,
2014; Wu et al., 2013). The ABL structure is determined by complex
interactions between atmosphere static stability and those mechanical
processes (such as wind shear from synoptic or terrain-induced flows)
(Stull, 1988). These processes can operate at a variety of different
heights and temporal scales, and their dominance may vary considerably with
height and time at any given location (Salmond and McKendry,
2005). This makes it very difficult to observe and predict the transport and
diffusion of air pollutants within the ABL (Chambers et al.,
2015b, a), particularly in regions of complex terrain
such as Beijing.
Beijing, the capital of China, is geographically located at the northwestern
border of the North China Plain. This city is surrounded by the Yan
Mountains to the north and the Taihang Mountains to the west, with the Bohai
Sea 160 km to the southeast (Fig. 1). Under favorable weather conditions
(e.g., stagnant weather), terrain-related circulations can be well developed
over Beijing and its surroundings, leading to a complex ABL structure, which
is thought to substantially affect Beijing's air quality (Hu et al.,
2014; Miao et al., 2017; Ye et al., 2016; Gao et al., 2016; Xu et al., 2016).
With high emissions of air pollutants from anthropogenic sources, Beijing is
suffering serious air pollution problems and the pollution can be even more
severe when southwesterly and southeasterly winds prevail within the ABL
(Chen et al., 2008; Ye et al., 2016; Zhang et al., 2014, 2012).
Several studies investigated the interactions between ABL meteorology and
air quality in Beijing using tower-based observations (Sun et al.,
2013, 2015; Guinot et al., 2006; Guo et al., 2014). However, the
results are not ideal because the tower-based observations have a low
observational height (325 m). In addition, numerous intensive ABL measures
were conducted using other approaches such as mooring boats, airplane and
ground-based remote sensing (Tang et al., 2015; Zhu et al., 2016; Zhang et
al., 2009; Hua et al., 2016). Since these approaches are complex, expensive
and labor-intensive, they are often restricted to the duration of specific
research campaigns and hence their results may be considered
to be unrepresentative. Overall, the existing knowledge of linkages between ABL
meteorology and air quality in Beijing is drawn largely from either low
observational height or short observational duration; therefore, the common
patterns of the influence of the changing ABL structures on Beijing's air
quality remains unclear and need to be further studied (Quan et al.,
2013; Miao et al., 2017; Guo et al., 2014).
Meanwhile, routine radiosondes are not being fully utilized to
investigate urban pollution issues. The advantage of using radiosondes over
other approaches seems to be their length, which usually spans several
decades. For a long time, it was challenging to reduce the wealth of
radiosonde data to characterize the ABL structure, and therefore the use of routine
radiosondes was very limited in case studies (Ji et al.,
2012; Zhao et al., 2013; Gao et al., 2016). Recently, self-organizing maps
(SOMs; a feature-extracting technique based on an unsupervised machine
learning algorithm) (Kohonen, 2001) were introduced to investigate the
ABL thermodynamic structure, indicating the capabilities of SOMs in feature
extraction from a large dataset of the ABL measurements (Katurji
et al., 2015). In fact, the application of the SOM has increased in atmospheric
and environmental sciences during the past several years (Jensen et al.,
2012; Jiang et al., 2017; Gibson et al., 2016; Pearce et al., 2014; Stauffer et
al., 2016), including a radiosonde-based application in South Africa
(Dyson, 2015). However, there has been no SOM application in air-pollution-related ABL structure research thus far. It is expected that such a new
analytical approach can tap the potential of routine radiosondes to reveal
the ABL mechanism of air pollution in Beijing.
This study investigates the influence of ABL meteorology on Beijing's air
quality based on the SOM application to 5 years (2013–2017) of routine
meteorological radiosondes. First, we use the SOM technique to classify the
state of ABL through detecting topological relationships among the
radiosonde-based virtual potential temperature profiles (see Sect. 3.1).
Then, we provide a visual insight into near-surface pollutant variations
under various ABL types and discuss the potential physical mechanisms behind
their relationships (see Sect. 3.2–3.3). It is expected that such an
association between air quality and ABL type could provide local
policymakers with useful information for improving the predictions of urban
air quality.
Materials and methods
Data preparation and preprocessing
The recent 5-year (2013–2017) radiosonde data observed at the Beijing
Observatory (39.81∘ N, 116.48∘ E; WMO station number
54511) were collected from the University of Wyoming
(http://weather.uwyo.edu/, last access: 25 March 2018). The radiosondes were launched twice a day (08:00
and 20:00 LT, corresponding to the morning and evening,
respectively) and provided atmospheric sounding data (profiles of
temperature, relative humidity, wind speed, etc.) at the mandatory pressure
levels (e.g., surface, 1000, 925, 850, 700 hPa) and additional significant
levels. In addition, the hourly near-surface meteorological parameters
(including temperature, wind speed and relative humidity, etc.) in
2013–2016 were collected from the Beijing Meteorological Bureau.
We chose 2000 m above ground level as the upper limit of the ABL based
on a number studies investigating the ABL height over Beijing or North China
(Tang et al., 2016; Guo et al., 2016; Miao et al., 2017). This height
exceeded the top of the ABL in most cases, and therefore, most ABL processes
influencing the near-surface air quality were included in the analysis
herein. In our study period, the average number of data points in radiosonde
profiles was 3.7 below 500 m and 10.1 below 2000 m. Despite the coarse
resolution, there were enough profile shapes for the SOM technique to classify the
state of ABL. A previous radiosonde-based study indicated that surface
temperature inversions occur frequently in eastern China
(Li et al., 2012), suggesting that all of the radiosondes that measure twice daily mainly represent the nocturnal stable ABL. To record a whole
night, the daily ABL profiles were composited from the radiosondes at 20:00
and 08:00 of the next day.
The mass concentrations of atmospheric pollutants (including PM2.5,
O3, NO2, SO2 and CO) over Beijing during the period from
2013 to 2017 are obtained from the Ministry of Environmental Protection of
the People's Republic of China (http://datacenter.mep.gov.cn/, last access: 31 December 2017). In addition,
hourly PM2.5 data measured at the Beijing US Embassy
(http://www.stateair.net/, last access: 1 April 2018) are also used in this study. The PM2.5
values in the two datasets show a good consistence, with a mean correlation
coefficient of 0.94. The mean hourly standard error of PM2.5 across
sites changes little from 12.6 to 12.9 after the inclusion of the US Embassy data.
Hourly concentrations are calculated for the Beijing urban area by averaging
concentrations from nine urban sites (including Dongsi, Guanyuan, Tiantan,
Wanshouxigong, Aotizhongxin, Nongzhanguan, Gucheng, Haidianwanliu and US
Embassy). The daily pollutant concentration is then performed
afternoon to afternoon (15:00–15:00), in order to include one whole
night in each 24 h period.
Self-organizing map technique
The SOM is thought to be an ideal tool for feature extraction because the
input data are treated as a continuum without relying on correlation,
cluster or eigenfunction analysis (Liu et al., 2006). Since
Kohonen (1982) first proposed the SOM, it has been widely used for data
downscaling and visualization in various disciplines (Jensen et al.,
2012; Katurji et al., 2015; Dyson, 2015; Stauffer et al., 2016; Pearce et al.,
2014; Jiang et al., 2017). In this study, the SOM is introduced to classify
the ABL types through detecting topological relationships among the 5-year
(2013–2017) radiosonde-based virtual potential temperature profiles. Since
the SOM is sensitive to the virtual potential temperature value, the
deviation profiles, which are determined by subtracting the mean virtual
potential temperature of each profile from each level, are used as the SOM
input.
The training of the SOM is an unsupervised, iterative procedure, and the result
is a matrix of nodes (i.e., types) that represent the input data. The
following provides a simple introduction about the SOM algorithm, and the
details can be found in Kohonen (2001). To learn from the input data,
every SOM node has a parametric reference vector with which it is
associated, and these reference vectors are randomly generated. After
initialization of the reference vectors, a stochastic input vector is
compared to every reference vector, and the closest match, named the
best-matching unit, is determined by the smallest Euclidean distance. Each
reference vector is then updated so that the best-matching unit and its
neighbors become more like the input vector. Whether or not a reference
vector learns from the input vector is determined by the neighborhood
function. Only reference vectors that are topologically close enough to the
best-matching unit will be updated according to the SOM learning algorithm.
The first step of SOM training is to determine a matrix size of nodes for
initializing the reference vectors. This step is performed subjectively and
depends on the degree of generation required (Lennard and Hegerl,
2015). We test several SOM matrixes, and finally select a 3 × 3 matrix,
because it captured unique profiles without the profiles being too
general as with a smaller matrix, or being too similar as with a larger
matrix. In addition, the batch mode is chosen to execute the SOM algorithm,
because it is much more computationally efficient compared to the sequence
mode. The other user-defined settings in the SOM software are set at the
default, such as the hexagon topology or the Gaussian neighborhood function.
The SOM code used in this study is sourced from the MATLAB SOM Toolbox,
which is freely available from http://www.cis.hut.fi/projects/somtoolbox/ (last access date: 1 April 2018).
Measuring the discriminative power of SOM technique for pollution
assessment
The Kruskal–Wallis one-way analysis of variance is used as a nonparametric
method to test the difference of pollutant concentrations among the various
ABL types. A 1 % significance level is used and hereafter denoted as KW in
Sect. 3.2. Furthermore, the coefficients of variation (CVs) of pollutant means
across the various ABL types are also used to examine the discriminative
power of the SOM technique for pollution assessment.
Results and discussion
Self-organized boundary layer meteorology
We construct a 3 × 3 SOM matrix for daily virtual potential
temperature deviation profiles, and the self-organized output shown in Fig.
1 represents nine ABL types (i.e., SOM nodes). On the SOM plane, the most
notable feature is adjacency of like types (e.g., Nodes 1 and 2) and the
separation of contrasting types (e.g., Nodes 1 and 9). Such ordering is a
feature of the SOM algorithm (i.e., “self-organized”), which allows us to
distinguish the unique characteristics of nodes through the variation of
specific features across the SOM plane. According to the ordering feature,
the SOM nodes in the four corners (i.e., Nodes 1, 3, 7 and 9) can be thought
of as the typical types and the others can be considered as transitional
types. The four typical ABL types have a relatively higher occurrence
frequency (> 10 %), with the highest frequency associated to
Node 1 (23 %). Furthermore, the seasonal statistical results (Fig. 2)
reveal that these self-organized ABL types exhibit a strong seasonality. For
example, Node 3 occurs more frequently in winter and autumn, while Node 1
has a relatively higher occurrence in spring and summer.
The 3 × 3 SOM output for radiosonde-based virtual
potential temperature (θv) deviation profiles observed at the
Beijing Observatory. SOM nodes are shown in red, with the corresponding
individual profiles in grey. For reference, the overall average θv deviation profile and 25 and 75th percentile profiles are shown in
cyan. The top right of each panel shows the occurrence cases and frequency of each SOM
node.
Relative frequency of individual ABL types (i.e., SOM nodes) in
all four seasons: winter (DJF), spring (MAM), summer (JJA) and autumn (SON).
Figure 3 displays the average profiles of wind speed, relative humidity and
virtual potential temperature gradient according to the ABL types. Clearly,
each of the self-organized types features distinct dynamic and thermodynamic
conditions within the ABL, suggesting the SOM technique is feasible to
classify the boundary layer meteorology. Since the classification is based
on the twice-daily radiosondes, the resulting ABL types are dominated by
near neutral to strong stable conditions, and none of the types fall within
the unstable category (i.e., Δθv/Δz<0). While Node 3
features the strong static stability in the upper ABL (the
large Δθv/Δz values), Node 1 represents a near
neutral ABL condition, with the lowest Δθv/Δz values and the
highest wind shears in the lower ABL. In contrast, Node 7
corresponds to a moderate static stability in the lower ABL, and Node 9
relates to a strong static stability. Particularly, the virtual potential
temperature gradient in Node 9 remains at a high level (>0.7 ∘C / 100 m) from the surface
to approximately 800 m, indicating that a strong
and deep surface temperature inversion developed in this type. In addition,
due to the strong surface inversion, vertical mixing is suppressed,
resulting in a strong decreasing gradient in humidity profiles. Overall, the
SOM classification scheme reveals a significant coupling between dynamic and
thermal effects in the ABL, which is expected to considerably impact the
near-surface air quality.
Profiles of average wind speed (WS), relative humidity (RH) and
virtual potential temperature gradient (Δθv/Δz)
corresponding to individual ABL types (i.e., SOM nodes) at the Beijing
Observatory. The black, green and red labels of the horizontal axis
correspond to Δθv/Δz, WS and RH, respectively.
To detect the boundary layer development after sunrise, the daytime boundary
layer height (BLH) is estimated using the parcel method (Holzworth, 1964,
1967), i.e., intersecting each day's 08:00 radiosonde potential temperature
(θ) profile at Beijing Observatory with each hour's (from 09:00 to
15:00) surface θ values, which are calculated from surface air
temperature observations. As shown in Fig. 4, the BLH on the days following
a strong stable night (i.e., Node 1) is relatively flat, reflecting an
inadequate development of the daytime boundary layer. Similarly, Node 3 is also
followed by a flat daytime BLH variation. The maximum BLHs in these two types
are lower than 900 m, indicating a limited space for vertical mixing in the
day. In contrast, the afternoon BLH in Node 7 can reach up to 1100 m; this
mixing depth is conducive to the dilution of the pollutants accumulated in the
previous night. In Node 1, the convective boundary layer develops well, and
its maximum height on average exceeds 1500 m, far higher than the values in
other types.
Daytime boundary layer height (BLH) estimated for the four typical
ABL types (i.e., Nodes 1, 3, 7 and 9).
Implementing the SOM-based ABL classification scheme for urban air
quality assessment
In the previous section, it was seen that the SOM classification scheme is
an effective tool for delineation between various dynamic and thermodynamic
structures within the ABL. As a further evaluation, we implement the new
classification scheme to quantify changes in various urban pollutant
concentrations as a function of ABL types. Since the pollutant emissions
have a strong seasonality over Beijing and its surroundings, the analyses
are performed for winter (December to February), spring (March to May),
summer (June to August) and autumn (September to November), respectively.
Figure 5 shows the statistical distributions of daily pollutant
concentrations according to the nine ABL types, along with the results of
the Kruskal–Wallis test and the coefficients of variation of pollutant means
across the various types. Figure 6 displays the type-average pollutant
diurnal patterns composited for the four typical ABL types (i.e., Nodes 1,
3, 7 and 9).
The Kruskal–Wallis test demonstrates that the self-organized ABL types are
able to distinguish between high and low loadings of air pollutants, with
KW < 1 % in all seasons (except for summertime SO2 with
a KW value of 1.5 %). Furthermore, it is found that the SOM technique has a
stronger discriminative power for SO2, PM2.5 and CO assessments,
which is supported by relatively higher CV values (CV > 0.30).
According to the seasonal CV values, this discriminative power shows the
following seasonal ordering: winter > autumn > spring > summer. Particularly, the wintertime CV value in PM2.5
assessment reaches the maximum (0.56), indicating an extremely strong
dependence of PM2.5 air quality on the changing ABL meteorology in
winter. In summer, the stable nocturnal ABL develops later due to the longer
day (Li et al., 2012), and hence avoids the larger daytime
pollutant emissions, particularly the traffic peak emissions. In the absence
of larger sources, the nocturnal stable layers exert a limited influence on
near-surface air quality; therefore, the classified ABL types have
relatively weakened discriminative power for summertime pollution
assessments. In addition, wet deposition (more precipitation in summer)
plays an important role in modulating summertime air quality, and to some
degree disrupts the linkages between the ABL type and air quality.
Daily concentrations of (a) PM2.5, (b) O3, (c) NO2,
(d) SO2 and (e) CO in Beijing for all nine ABL types separately in (1)
winter, (2) spring, (3) summer and (4) autumn. The solid dots denote the
mean. The box and whisker plot presents the median, the first and third
quartiles and the 5 and 95th percentiles, respectively. The numbers
above the plots denote the results of the Kruskal–Wallis test (KW) and the coefficients of
variation (CV) of pollutant means across the various types.
In the case of PM2.5, NO2 and CO, the most stable atmospheric
conditions (i.e., Node 9) are associated with dramatically increased
near-surface pollutant concentrations in all seasons except summer. The
wintertime average concentrations of PM2.5, NO2 and CO in Node 9
reach up to 197.2, 100.2 and 3.6 µg m-3,
respectively. These values are 3–8 times higher than that in Node 1 (i.e.,
near neutral condition), with the highest increasing amplitude (a factor of
7.3) related to PM2.5. As mentioned, Node 9 corresponds to the
strongest nighttime stability in the lower ABL and the lowest daytime BLH.
All of these ABL characteristics are extremely conducive to the accumulation
of air pollutants emitted near the ground. For Node 3, the concentrations of
PM2.5, NO2 and CO are the second-highest compared to those of the
other types. This ABL type features the strongest stability in the upper
ABL, suggesting that processes operating at the different heights throughout
the ABL may have a significant impact on near-surface pollutant
concentrations.
The diurnal cycles of PM2.5, NO2 and CO are extremely pronounced
under the strong stable conditions, although they are very reduced on the days with
near neutral night. On average, the wintertime diurnal range of PM2.5
increases from 18.2 µg m-3 in Node 1 to 95.4 µg m-3 in
Node 9. The corresponding diurnal range increase for NO2 is 18.9 to
33.6 µg m-3, and for CO 0.2 to 1.7 µg m-3. In Node 1, the
diurnal variations are characterized by a weak two-peak pattern, following
the traffic rush hours, suggesting that traffic is the primary driver of
these pollutants' diurnal cycles in Beijing (Liu et al., 2012).
However, the diurnal effects of traffic emissions are significantly
amplified by the stable ABL dynamics. It is clear that the more stable
the conditions near the ground, the higher the peak concentrations that are observed. In
winter, the stable ABL conditions exert a more important influence on the
evening traffic emissions, resulting in a broad evening peak. In contrast,
the morning peak signature is much lower since the morning emission is
counteracted by the destabilization of the ABL. However, as human activities
begin earlier during the warm season, maximum concentrations in spring and
summer are typically observed during the morning rush hours.
However, increasing atmospheric stability has the opposite effect on
near-surface O3 concentrations. Since aerosols can absorb and reflect
solar radiation and thereby inhibit the photochemical production of O3 (Gao et al., 2016;
Kaufman et al., 2002), the lowest average O3
concentration is observed in Node 9. In addition, considering that ozone is
mainly produced in the upper ABL, near-surface O3 should be strongly
modulated by downward mixing processes (Tang et al., 2017b, a). In light of this, the varying daytime O3 peaks across the
ABL types can be partly attributed to the various magnitudes of vertical
mixing. This is supported by the daytime BLH. As is shown in Fig. 4, the
daytime BLH is highest in Node 1, followed by Node 7 and Node 3, and it is the lowest
in Node 9. Such ordering is generally consistent with the daytime O3
peaks in these types. Due to the persistent downward mixing caused by strong
wind shears, the near-surface O3 remains a relatively high nocturnal
concentration (e.g. about 45 µg m-3 in winter) in Node 1. In
contrast, the stable nocturnal conditions (e.g., Nodes 9, 7 and 3) are
commonly associated with low O3 concentration (e.g. about 16 µg m-3 in winter) due to the lack of vertical mixing, as well as the
strong chemical titration by NO emitted from vehicles.
Mean hourly composites of (a) PM2.5, (b) O3, (c) NO2,
(d) SO2 and (e) CO in Beijing for the four typical ABL types
separately in (1) winter, (2) spring, (3) summer and (4) autumn.
The highest average SO2 concentrations are typically observed in Node
3, but occasionally in Node 9. Over the North China Plain, high
stacks emit a significantly larger amount of SO2 compared to
small stacks (Zhao et al., 2012). The frequent surface
temperature inversions, together with the large SO2 emissions from
higher stacks, favor the formation of elevated SO2 pollution layers over
Beijing (Chen et al., 2009). If sufficiently strong, the
surface temperature inversion can even isolate near-surface observations
from the influence of elevated pollution layers (Salmond and
McKendry, 2005). This explains the commonly lower near-surface SO2
concentration in Node 9 than that in Node 3. However, after a stable night,
the burst of turbulent activity in the morning coincides with a rapid
increase in the near-surface SO2 concentration, resulting in a pre-noon
peak. Since there is no significant increase in SO2 emission at the
surface at that time, the result strongly suggests that the SO2 peaks
are due to the downward mixing from the elevated SO2 pollution
layers. Regarding the physical mechanism of the noontime-peak SO2 pattern,
Xu et al. (2014) gave a detailed explanation in a previous
study. Nevertheless, the wintertime SO2 concentration signature does
not always show a distinct pre-noon peak (e.g., Node 7). This may be
attributed to the increased SO2 emissions from household heating in winter
(Liao et al., 2017). Like other primary pollutants, the local SO2
emissions become trapped close to the surface under stable nocturnal
conditions, resulting in a much higher nighttime peak compared to the
pre-noon peak.
Quantifying the contribution of ABL anomaly to typical-month PM2.5 air quality
To improve air quality, the Chinese government promulgated the “Air Pollution
Prevention and Control Action Plan” in 2013. As a consequence, observed
annual mean PM2.5 concentrations decreased by about 37 % over Beijing
during 2013–2017. However, severe wintertime PM2.5 pollution events
still frequently wreak havoc across Beijing and its surroundings, which
result in severe damage to the environment and human health (Gao et
al., 2017, 2015). It is therefore a pressing issue to understand the
factors affecting the occurrence of such serious PM2.5 pollution.
Previous studies highlighted the potential importance of atmospheric conditions
to the wintertime PM2.5 air quality (Cai et al., 2017). Since the
fraction of time for which the different atmospheric conditions dominate can
vary from year to year, elucidation of the meteorological roles in those
serious pollution periods is of significant importance. In this section, we
evaluate the contribution of the ABL anomaly to elevated PM2.5
concentration in three typical pollution months, i.e., January 2013,
December 2015 and December 2016.
Heavy PM2.5 pollution episodes occurred frequently in January
2013, December 2015 and December 2016, resulting in anomalously high
monthly averaged PM2.5 concentrations in the Beijing urban area (180.1, 151.8 and 151.2 µg m-3,
respectively). Figure 7 shows the hourly PM2.5variations in the
three selected months, along with daily ABL types. In general, the pollution
episodes were associated with Nodes 3 and 9, and the clean episodes
corresponded to Node 1. For example, the severe pollution episode that
occurred from 9 to 14 January 2013 was due to the alternate control of Nodes 3
and 9, and the pollution episode from 15 to 21 December 2016 was related to
the persistency of Node 9. Conversely, multi-day control of Node 1 caused a
clean episode from 14 to 16 December 2015. The linkages between PM2.5 air
quality and ABL type are consistent with the previous long-term
analyses, indicating that the changing ABL type is one of the primary
drivers of day-to-day variations in wintertime PM2.5 air quality over
Beijing.
Time series of hourly PM2.5 concentrations in (a) January
2013, (b) December 2015 and (c) December 2016. The daily ABL types
(i.e., SOM nodes) are shown at the top of each plot (red numbers).
Figure 8 illustrates a comparison of the occurrence frequency of the nine
ABL types between the three selected months and the whole 5-year winter
period. Compared with the winter-averaged frequency, notable differences are
that the stable ABL conditions increased and the near neutral nights
decreased during the three polluted months. For example, Node 9 occurrence
was nearly trebled in December 2016, and total occurrence of Nodes 3 and
9 doubled in January 2013. These results highlight the potential
contribution of the ABL anomaly to the elevated PM2.5 concentrations in
these pollution months.
Occurrence frequency of the ABL types (i.e., SOM nodes) during
(a) January 2013, (b) December 2015 and (c) December 2016. The winter-averaged
frequency during the 5-year (2013–2017) period is repeated on each plot for
comparison.
Estimated contribution of ABL anomaly to elevated PM2.5
concentration in January 2013, December 2015 and December 2016.
Pollution
Monthly averaged PM2.5
Total anomaly C′
ABL-driven anomaly
Contribution ratio of
month
concentration (µg m-3)
(µg m-3)
CABL' (µg m-3)
ABL-driven anomaly (%)
January 2013
180.1
76.1
44.4
58.3
December 2015
151.8
47.8
22.2
46.4
December 2016
151.2
47.2
34.6
73.3
Assuming the linkages between ABL type and PM2.5 loading are constant
in different years, the contribution of the anomalous ABL meteorology to
PM2.5 air quality can be estimated through a
meteorology–environment method, which is a revised version of the
circulation–environment method proposed by Zhang et al. (2012). For
each selected month, we define the deviation in PM2.5 from the 5-year
winter-averaged concentration (CWIN) as the total anomaly (C′), which is
due to the combined effects of emission and meteorology. The anomaly
calculated from the mean PM2.5 loadings for nine ABL types and their
occurrence frequencies during each month can be considered as the
“ABL-driven” anomaly, which represents the PM2.5 change caused by
the anomalous ABL meteorology. The ABL-driven anomaly (CABL') is
calculated through ∑iFi⋅Ci-CWIN, where
Fi is the occurrence frequency of type i ABL during a specific month and
Ci is the corresponding PM2.5 loading featuring that type. The
ratio of CABL′ to C′ is then used to assess the relative contribution of the
ABL anomaly to the total anomaly. The calculated results (Table 1) show
that the ABL-driven PM2.5 changes are 44.4 µg m-3 in January
2013, 22.2 µg m-3 in December 2015 and 34.6 µg m-3 in
December 2016, which explain 58.3, 46.4 and 73.3 % of the total
anomaly in respective months. These quantitative estimations demonstrate
that the elevated PM2.5 concentrations during the three polluted months
can be largely attributed to anomalous ABL conditions.
Summary
The influence of ABL meteorology on Beijing's air quality is relatively
unclear due to the lack of long-term observations. Meanwhile, the long years
of routine radiosondes remain underutilized as a tool for urban pollution
studies. In this study, the SOM was applied to 5-year (2013–2017)
radiosonde-based θv profiles to classify the state of the ABL over
Beijing. The classified ABL types were then evaluated in relation to
near-surface air quality, with an attempt to understand the roles of the
changing ABL structure in air quality variation in Beijing. The main
findings are as follows.
The SOM provides a continuum of nine ABL types (i.e., SOM nodes), and each
is characterized with distinct dynamic and thermodynamic conditions within
the ABL. Node 1 represents a near neutral layer, with the lowest θv gradient and the highest wind speed. Node 3 features a strong static
stability in the upper ABL. In contrast, Node 9 and Node 7 respectively
correspond to the moderate and strong static stability in the lower ABL.
The self-organized ABL types are capable of characterizing the influence of
nocturnal mixing on near-surface pollutant loadings. From the near neutral
(i.e., Node 1) to strong stable conditions (i.e., Node 9), the average
concentrations of PM2.5, NO2 and CO increased dramatically during
all seasons except summer. Meanwhile, the diurnal cycles of these pollutant
species are strongly modulated by ABL dynamics. Although the modulation
effect varies from season to season, the higher peak concentrations commonly
occur under the more stable conditions. However, increasing stability has
an opposite effect on O3, resulting in the lowest O3 level in Node 9.
For SO2, the highest average concentrations are typically observed in
Node 3. The pre-noon SO2 peaks are more significant after a strong
stable night.
Analysis of three typical wintertime pollution months (i.e., January 2013,
December 2015 and December 2016) suggests that the ABL types are one of the
primary drivers of day-to-day PM2.5 variations in Beijing. During the
three pollution months, the frequency of the stable ABL types (i.e., Nodes 9
and 3) increases significantly compared to the 5-year winter mean. Using a
meteorology–environment method, the relative (absolute) contributions of
the ABL anomaly to elevated PM2.5 concentrations are estimated to be
58.3 % (44.4 µg m-3) in January 2013, 46.4 %
(22.2 µg m-3) in December 2015 and 73.3 % (34.6 µg m-3) in
December 2016.
This work revealed the common pattern of the ABL influences on Beijing's air
quality. The established linkages between ABL type and air quality could be
useful for developing an operational forecast and warning system. In
addition, this work demonstrated that the SOM-based ABL classification
scheme is a helpful tool for understanding urban air pollution. Since the
SOM technique is good at feature extraction, the coarse-resolution
radiosondes can be taken as input to classify the state of the ABL.
Therefore, the SOM-based ABL classification scheme can take advantage of the
long-term available radiosondes, making it a simple and economical
alternative to other approaches to stability classification. We believe that
the pollution-related ABL research and the formulation of pollution control
measures could benefit from application of the SOM analytical tool.