Outline

  1. Using meteorological normalisation to detect interventions in air quality time series
  2. 1. Introduction
  3. 1.1. Objectives
  4. 2. Methods
  5. 2.1. Data
  6. 2.1.1. Port of Dover SO2
  7. 2.1.2. London Marylebone Road NO2 and NOx
  8. 2.2. Modelling and the hyperparameters
  9. 3. Results and discussion
  10. 3.1. Port of Dover SO2
  11. 3.1.1. Models
  12. 3.1.2. Influence of sulfur fuel limits on SO2 concentrations
  13. 3.2. London Marylebone Road NOx
  14. 3.2.1. Models
  15. 3.2.2. Changes in primary NO2
  16. 4. Conclusions
  17. Competing interests
  18. Acknowledgments
  19. References
Previous PDF in this issueNext PDF in this issue
Science
of
the
Total
Environment
653
(2019)
578–588
Contents
lists
a
v
ailable
at
ScienceDirect
Science
of
the
Total
Environment
journal
homepage:
www.elsevier.com/locate/scitotenv
Using
meteorological
normalisation
to
detect
interventions
in
air
quality
time
series
Stuart
K.
Grange
a
,
*
,
David
C.
Carslaw
a
,
b
a
Wolfson
Atmospheric
Chemistry
Laboratories,
University
of
York,
York
YO10
5DD,
United
Kingdom
b
Ricardo
Energy
&
Environment,
Harwell,
Oxfordshire
OX11
0QR,
United
Kingdom
HIGHLIGHTS
•
Detecting
the
influence
of
air
quality
interventions
is
important.
•
Changes
in
meteorology
over
time
complicate
air
quality
intervention
analysis.
•
Meteorological
normalisation
was
applied in
two locations
to explore
interventions.
•
The
changes
detected
in
the
nor-
malised
time series
were
associated
to
interventions.
•
The
non-black-box
nature
of
the
pro-
cedure
allows
for
interpretation
of
results.
GRAPHICAL
ABSTRACT
ARTICLE
INFO
Article
history:
Received
14
August
2018
Received
in
revised
form
25
October
2018
Accepted
25
October
2018
Available
online
28
October
2018
Editor:
Pavlos
Kassomenos
Keywords:
Air
pollution
Data
analysis
Management
Machine
learning
Random
forest
ABSTRACT
Interventions
used
to
improve
air
quality
are
often
difficult
to
detect
in
air
quality
time
series
due
to
the
complex
nature
of
the
atmosphere.
Meteorological
normalisation
is
a
technique
which
controls
for
meteo-
rology/weather
over
time
in
an
air
quality
time
series
so
intervention
exploration
(and
trend
analysis)
can
be
assessed
in
a
robust
way.
A
meteorological
normalisation
technique,
based
on
the
random
forest
machine
learning
algorithm
was
applied
to
routinely
collected
observations
from
two
locations
where
known
inter-
ventions
were
imposed
on transportation
activities which
were expected
to
change ambient
pollutant
concentrations.
The
application
of
progressively
stringent
limits
on
the
content
of
sulfur
in
marine
fuels
was
very
clearly
represented
in
ambient
sulfur
dioxide
(SO
2
)
monitoring
data
in
Dover,
a
port
city
in
the
South
East
of
England.
When
the
technique
was
applied
to
the
oxides
of
nitrogen
(NO
x
and
NO
2
)
time
series
at
London
Marylebone
Road
(a
Central
London
monitoring
site
located
in
a
complex
urban
environment),
the
normalised
time
series
highlighted
clear
changes
in
NO
2
and
NO
x
which
were
linked
to
changes
in
primary
(directly
emitted)
NO
2
emissions
at
the
location.
The
clear
features
in
the
time
series
were
illuminated
by
the
meteorological
normalisation
procedure
and
were
not
observable
in
the
raw
concentration
data
alone.
The
lack
of
a
need
for
specialised
inputs,
and
the
efficient
handling
of
collinearity
and
interaction
effects
makes
the
technique
flexible
and
suitable
for
a
range
of
potential
applications
for
air
quality
intervention
exploration.
©
2018
The
Authors.
Published
by
Elsevier
B.V.
This
is
an
open
access
article
under
the
CC
BY
license
(http://creativecommons.org/licenses/by/4.0/).
* Corresponding
author.
E-mail
address:
stuart.grange@york.ac.uk
(S.K.
Grange).
1.
Introduction
Across
all
spatial
and
temporal
scales,
weather
influences
con-
centrations
of
atmospheric
pollutants
and
in
turn
ambient
air
quality
(Stull,
1988;
Monks
et
al.,
2009).
The
effects
of
weather
(or
https://doi.org/10.1016/j.scitotenv.2018.10.344
0048-9697/ ©
2018
The
Authors.
Published
by
Elsevier
B.V.
This
is
an
open
access
article
under
the
CC
BY
license
(http://creativecommons.org/licenses/by/4.0/).
Previous PDF in this issueNext PDF in this issue
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
579
meteorology)
on
air
quality
are
often
much
greater
than
intervention
or
management
efforts
to
control
air
pollution
and
therefore
inter-
vention
events
can
be
very
difficult
to
detect
and
quantify
within
an
observational
record
(Anh
et
al.,
1997).
Similarly,
when
considering
trends
in
ambient
air
pollution,
it
can
be
difficult
to
know
whether
a
change
in
concentration
is
due
to
meteorology
or
a
change
in
emis-
sion
source
strength.
Meteorological
variation
can
therefore
frustrate
the
analysis
of
trends
in
different
pollutant
species.
If
meteorology
is
not
controlled
or
accounted
for,
the
changes
in
pollutant
concentra-
tions
observed
may
be
contaminated
with
meteorological
variation
rather
than
emission
or
chemically
induced
perturbations
which
can
lead
to
erroneous
conclusions
concerning
the
efficacy
of
air
qual-
ity
management
strategies
(Libiseller
et
al.,
2005;
Wise
and
Comrie,
2005).
This
issue
is
often
acknowledged,
but
infrequently
addressed.
Meteorological
normalisation
is
one
technique
which
can
be
used
to
control
for
meteorology
over
time
in
air
quality
time
series.
The
central
philosophy
of
meteorological
normalisation
is
to
reduce
vari-
ability
in
an
air
quality
time
series
with
statistical
modelling.
The
reduction
of
variability
is
achieved
by
training
a
model
which
can
explain
some
of
the
variation
of
pollutant
concentrations
through
a
number
of
independent
variables.
The
independent
variables
used
are typically surface-based meteorological
observations and time
variables
which
act
as
proxies
for
regular
emission
patterns
such
as
hour
of
day
and
season
(Derwent
et
al.,
1995).
However,
in
prac-
tice, any independent variable which could
explain variations in
pollutant
concentrations
could
be
used.
Once
the
model
has
been
trained
and
it
is
found
that
it
can
explain
an
adequate
amount
of
the
dependent
variable’s
variation,
the
model
can
be
used
to
remove
the
influence
the
independent
variables
have
on
the
dependent
vari-
able
by
sampling
and
predicting.
The
time
series
which
results
can
then
be
exposed
to
further exploratory
data analysis
(EDA) tech-
niques
such
as
formal
trend
analysis
and/or
intervention
exploration
(Grange
et
al.,
2018).
The
normalised
time
series
is
in
the
pollutant’s
original
units
and
can
be
thought
of
as
concentrations
in
“average”
or
invariant
weather
conditions.
There
has
been
some
air
quality
research
conducted
which
uses
the
idea
of
change-point
analysis
to
investigate
changes
in
atmo-
spheric
pollutant
concentrations
(for
example
Carslaw
and
Carslaw,
2007;
Carslaw
et
al.,
2006).
Methods
such
as
these
rely
on
regime
changes where a
time series abruptly shifts from
one regime to
another
(Lyubchich
et
al.,
2013).
In
the
air
quality
domain,
this
rarely
happens,
since
changes
are
usually
nuanced
and
occur
progressively
with
much
variability
which
makes
the
generality
of
this
approach
for
investigating
intervention
efforts
poor.
Meteorological
normali-
sation
is
potentially
a
more
general
approach
which
enables
its
use
in
a
greater
range
of
applications.
Atmospheric processes
are complex,
non-linear, and
observa-
tions
commonly
record
collinearity
with
other
observations.
These
attributes
make
the
process
of
statistical
modelling
very
challenging,
especially
so
with
parametric
methods
(Barmpadimos
et
al.,
2011).
With
the
rise
of
machine
learning
algorithms,
these
attributes
can
be
much
more
easily
accommodated
due
to
the
non-parametric
and
robust
nature
of
these
techniques
(Friedman
et
al.,
2001).
The
mete-
orological
normalisation
technique
used
here
uses
random
forest,
an
ensemble
decision
tree
machine
learning
method
as
the
modelling
algorithm.
Random
forest
has
been
described
very
well
and
in
depth
else-
where
(see
Breiman,
2001;
Friedman
et
al.,
2001;
Tong
et
al.,
2003;
Ziegler
and
König,
2013;
Jones
and
Linder,
2015;
Grange
et
al.,
2018).
However
in
brief,
a
single
decision
tree
is
formed
from
a
series
of
binary splits
which results in homologous
or “pure” groups. The
splitting process
is
recursive
which
means splitting
occurs
until
purity
is
achieved
if
the
tree
is
allowed
to
grow
to
its
maximum
depth.
Decision
trees
make
no
assumptions
on
the
input
data
struc-
ture
(they
are
non-parametric),
allow
for
interaction
and
collinearity
among
variables,
and
will
ignore
variables
which
are
irrelevant
to
the
dependant
variable
(Ziegler
and
König,
2013).
Decision
trees
are
fast
to
train,
fast
to
make
predictions,
and
are
conceptually
simple
to
understand.
However,
they
suffer
heavily
from
overfitting,
an
issue
where
the
model
represents
the
training
set
well,
but
does
not
gen-
eralise
to
sets
which
were
not
used
for
training
(Jones
and
Linder,
2015).
Using
a
model
which
predicts
pollutant
concentrations
and
suffers
from
overfitting
would
result
in
the
model
being
contam-
inated
with
noise
from
the
training
set
and
unreliable
predictions
would
impede
analyses.
Random
forest
is
an
algorithm
which
controls
for
the
tendency
of decision
trees to overfit. The
algorithm achieves this
by sam-
pling
(with
replacement)
the
training
set
with
a
process
called
bagging
(bootstrap
aggregation)
(Breiman,
1996).
In
modern
usage,
sampling
of
the
independent
variables
is
usually
done
during
bag-
ging
too.
Bagging
results
in
a
new,
sampled
set
called
out-of-bag
(OOB)
data. A
decision tree
is
then grown
on the
OOB data.
The
bagging-then-tree
growth
is
repeated,
generally
a
few
hundred
times.
Because
OOB
data
is
sampled,
all
the
decision
trees
are
grown
on differing observations and
independent variables which leads
to a
“forest” of
decorrelated trees.
After training,
all the
individ-
ual
trees
within
the
forest
are
used
to
predict,
but
their
predictions
are
aggregated
as
a
mean
(or
the
mode
for
categorical
dependent
variables) and that
forms the single ensemble
prediction for the
model.
The
meteorological
normalisation
technique
is
pragmatic
in
respect to
the input
variables required for
many common appli-
cations.
Generally,
routinely
accessible
surface
meteorological
vari-
ables
are
very
effective
for
the
process
and
specialised
or
obscure
variables
are
generally
not
necessary
for
the
technique
to
be
applied.
Although
traffic
counts,
upper
air
data,
and
outputs
from
weather
models
will
usually
strengthen
a
model’s
explanatory
power,
the
existence
or
access
to
such
variables
is
not
a
pre-
requisite,
an
attribute
which
is
very
useful
for
most
situations
where
such
inputs
are
not
available.
For
pollutants
which
are
pri-
marily controlled by regional
scale processes, most
notably par-
ticulate
matter
(PM)
and
ozone
(O
3
),
additional
variables
such
as
boundary layer
height, air
mass cluster, or
back trajectory infor-
mation
would
however
be
beneficial
to
include
if
possible
and
examples
can
be
found
elsewhere,
for
example
Grange
et
al.
(2018).
The
temporal
variables
used
as
independent
variables
in
the
meteorological
normalisation
models:
Julian
day,
weekday,
and
hour
of
year
are
included
not
for
their
direct
influence
on
atmospheric
concentrations,
but
because
they
act
as
proxies
for
cyclical
emis-
sion patterns. Hour of day for example offers a term to explain
emissions
with
a
diurnal
cycle
such
as
traffic-related
rush
hour
emis-
sions or domestic heating
phases, while Julian day is
a seasonal
term
which
represents
emissions
or
atmospheric
chemistry
which
varies
seasonally. These
processes are
generally strong
drivers
of
concentrations
of most
atmospheric pollutants
(Henneman et
al.,
2015).
Random
forest’s
ability
to
handle
collinearity
and
interaction
between
these
and
the
other
independent
variables
used
and
the
lack
of
need
of
specialised
or
exotic
inputs
results
in
a
flexible
tool
kit
for
probing
the
influences
of
interventions
on
air
quality
time
series.
1.1.
Objectives
The
primary
objective
of
this
paper
is
to
apply
a
meteorological
normalisation
technique
based
on
random
forest,
a
machine
learn-
ing
algorithm
to
detect
interventions
in
air
quality
monitoring
data.
This
is
done
to
gain
understanding
of
what
physical
and
chemical
processes
are
driving
ambient
pollutant
concentrations
and
highlight
the
suitability
and
potential
of
the
technique
to
other
applications.
Two
case
studies
are
presented
using
routine
data
sets
in
Dover,
South
East
England
where
sulfur
fuel
limits
of
ships
were
imposed
580
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
and
changes
in
ambient
sulfur
dioxide
(SO
2
)
concentrations
are
expected
and
in
Central
London
where
congestion
charging
and
local
bus
fleet
management
has
perturbed
oxides
of
nitrogen
(NO
x
)e
m
i
s
-
sion
sources.
The
changes
in
concentrations
and
emissions
are
then
explained
in
respect
to
implementation
of
policy
which
would
be
dif-
ficult
to
detect
with
other
EDA
techniques
where
no
meteorological
normalisation
is
performed.
2.
Methods
2.1.
Data
2.1.1.
Port
of
Dover
SO
2
Hourly
SO
2
concentrations
were
analysed
from
the
Port
of
Dover,
a
major
port
located
in
Kent
in
the
South
East
of
England.
Fig.
1.
Maps
of
the
study
sites
with
a
United
Kingdom
insert
for
country-scale
context.
The
Port
of
Dover
complex
is
displayed
in
(a)
and
the
internal
lines
indicate
roads
and
Greater
London
is
shown
in
(b),
with
the
London
Boroughs
and
City
of
London
indicated
with
internal
polygons.
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
581
Table
1
Details
of
the
air
quality
monitoring
sites
in
Dover
and
London
used
in
this
analysis.
Sites
without
end
dates
are
still
operational.
Location
Site
name
Site
type
Latitude
Longitude
Elevation
Date
start
Date
end
Dover
Langdon
Bay
Meteorological
51.133
1.350
117
1973-03-08
Dover
Dover
Langdon
Cliff
Urban
background
51.132
1.339
98
2001-03-17
2010-03-05
Dover
Dover
Docks
Urban
industrial
51.127
1.336
6
2006-11-17
2013-01-03
London
London
Heathrow
Meteorological
51.478
−0.461
25
1948-12-01
London
London
Marylebone
Road
Traffic
51.523
−0.155
35
1997-01-01
Two
air
quality
monitoring
sites,
Dover
Docks
and
Dover
Langdon
Cliff’s
SO
2
data
were queried
from the
Kent
Air Quality
database
(Ricardo
Energy
and
Environment,
2018).
A
nearby
meteorological
site,
Langdon
Bay
located
to
the
west
of
the
port
was
used
to
provide
surface
meteorological
observations
and
were
accessed
from
NOAA’s
Integrated Surface Database (ISD)
(NOAA, 2016)(
Fig. 1 (a)). The
monitoring
sites
had
different
commissioning
and
decommissioning
dates
and
neither
site
is
still
operating
(Table
1).
SO
2
observations
are
available
between
March
2001
and
December
2012.
The
data
capture
rates
for
SO
2
at
Dover
Langdon
Cliff
and
Dover
Docks
for
their
online
period
were
92
and
82%
respectively.
These
monitor-
ing
sites
are
of
interest
because
marine
fuels
in
British
and
European
waters
have
been
subject
to
a
series
of
sulfur
content
fuel
limits.
The
introduction
and
continued
enforcement
of
these
sulfur
fuel
limits
were
expected
to
influence
ambient
SO
2
concentrations.
The
details
of
these
interventions
are
discussed
further
in
Section
3.1.2.
2.1.2.
London
Marylebone
Road
NO
2
and
NO
x
Hourly
NO
2
and
NO
x
data
from
London’s
Marylebone
Road
air
quality monitoring
site
were
accessed
from
smonitor
Europe,
a
European
database
containing
the
observations
and
metadata
from
the
AirBase
and
Air
Quality
e-Reporting
(AQER)
repositories
(Grange,
2016,
2017). NO
x
concentrations
have been
monitored since
July
1997
and
the
final
year
of
reporting
sourced
from
the
European
data
repositories
used was
2016.
Data capture
rates
for NO
x
and
NO
2
for
the
analysis
period
were
97%.
London
Heathrow,
a
large
airport
located
at
the
far
west
of
Greater
London
was
used
for
surface
mete-
orological
observations
sourced
from
NOAA’s
ISD
(Fig.
1
(b)).
London
Marylebone
Road
is
situated
in
a
complicated
central
urban
envi-
ronment.
The site
is
located
one metre
south
of the
kerb
on
the
A501
trunk
road
and
sits
within
an
irregularly
shaped
street
canyon.
London
Marylebone
Road
is
a
prominent
and
often
analysed
site
due
to
its
long
observational
record
and
diverse
suite
of
pollutants
which
are
monitored
at
the
site
(Jeanjean
et
al.,
2017).
NO
x
and
NO
2
concentrations
across
European
cities
are
a
signif-
icant
issue
and
many
member
states
are
non-compliant
to
the
legal
European
ambient
air
quality
limits
(Weiss
et
al.,
2012;
Grange
et
al.,
2017).
Almost
all
locations
which
are
non-compliant
are
classified
as
roadside
(or
‘traffic-influenced’)
(European
Environment
Agency,
2016).
London
has
some
of
the
highest
roadside
concentrations
of
NO
x
and
NO
2
in
Europe
and
London
Marylebone
Road
(Fig.
1
(b))
is
an
often
referenced
monitoring
site
for
its
high
concentrations.
To
combat
the
issue
of
traffic
congestion,
Greater
London
author-
ities
imposed the
Congestion
Charge
Zone (CCZ),
which
was
first
enforced
in
February
2003
(Atkinson
et
al.,
2009).
Since
that
time,
the
London
Low
Emission
Zone
(LEZ),
and
the
Emissions
Surcharge
(bet-
ter
known
as
the
T-Charge)
have
also
been
implemented
to
combat
air
pollution
(Transport
for
London,
2018).
The
details
and
start
dates
of
these
various
measures
are
displayed
in
Table
2.
All
these
inter-
ventions
are
significant
investments
with
large
amounts
of
planning
and
resources
to
execute
and
maintain.
2.2.
Modelling
and
the
hyperparameters
For
both
examples,
the
meteorological
normalisation
procedure
was
conducted
in
the
same
way
and
the
rmweather
Rp
a
c
k
a
g
e(
v
e
r
-
sion
0.1.2)
was
used
for
this
process
(R
Core
Team,
2018;
Grange,
2018).
The
number
of
trees
for
the
random
forest
models
was
fixed
at
300,
the
minimal
node
size
was
five,
and
the
number
of
variables
split
at
each
node
was
the
default
for
regression
mode:
the
rounded
down
square
root
of
the
number
of
independent
variables
which
in
these
examples was three (rmweather’s function arguments n_trees,
min_node_size,a
n
dmtry
respectively).
The
independent
variables
used
were:
Unix
date
(number
of
seconds
since
1970-01-01)
as
the
trend
term,
Julian
day
as
the
seasonal
term,
weekday,
hour
of
day,
air
temperature,
relative
humidity,
wind
direction,
wind
speed,
and
atmospheric
pressure.
Training
was
only
conducted
on
observations
which
had
non-missing
wind
speed
and
the
pollutant
being
mod-
elled.
Three
hundred
predictions
were
used
to
calculate
the
meteoro-
logically
normalised
trend.
The
normalised
trends
were
aggregated
to
monthly
resolution
for
presentation
in
Section
3.
A
conceptual
representation
of
the
meteorological
normalisation
processes
is
dis-
played
in
Fig.
A1.
For
the
Dover
SO
2
examples,
models
were
calculated
using
the
full
observational
set,
but
after
investigating
the
models
(discussed
in
Section
3.1.1),
the
observations
were
filtered
to
wind
directions
which
were
sourced
from
the
port
and
these
models
are
the
ones
which were
used for
the time
series analysis
(Section 3.1.2).
For
observations
at
London
Marylebone
Road,
no
filtering
was
under-
taken.
In the
case of
London
Marylebone Road,
there
are a
large
number
of potential
events
which could
influence pollutant
con-
centrations
and
emissions.
To
objectively
identify
events,
the
mete-
orologically
normalised
time
series
were
tested
for
breakpoints
or
changes
in
structure.
The
structural
change
algorithm
is
described
in
Zeileis
et
al.
(2003),
Zeileis
et
al.
(2002)
and
was
implemented
with
the
strucchange
Rp
a
c
k
a
g
e
.
The
random
forest
algorithm
does
not
directly
offer
the
ability
to
determine
error
or
uncertainty
of
estimates.
However,
uncertainty
is
important
to
consider
in
many
situations.
To
enable
uncertainty
to
be
evaluated
for
the
case
studies,
50
random
forest
models
were
grown
for
each
example
with
the
hyperparameters
described
above,
but
with randomly
sampled (bootstrapped)
input sets.
The boot-
strapping
of
the
observational
data
ensured
the
models
were
grown
on
different
training
sets.
The
importance
values
(a
measure
of
the
Table
2
Details
of
interventions
within
Greater
London
to
counter
traffic
congestion.
Name
Abbreviation
Start
date
Area
covered
Operation
Congestion
Charge
Zone
CCZ
2003-02-17
Central
London
07:00–18:00
Mo–Fr
London
Low
Emission
Zone
(first
phase)
LEZ
2008-02-04
Greater
London
24/7
London
Low
Emission
Zone
(second
phase)
LEZ
2012-01-03
Greater
London
24/7
Emissions
Surcharge
T-Charge
2017-10-23
Central
London
07:00–18:00
Mo–Fr
Ultra
Low
Emission
Zone
(planned)
ULEZ
2019-04-08
Central
London
24/7
582
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
Table
3
Mean
random
forest
model
performance
statistics
four
the
four
sets
of
models
grown
for
the
analysis.
Location
Model
nR
2
Dover
Dover
Docks
SO
2
34,224
0.67
Dover
Dover
Langdon
Cliff
SO
2
53,535
0.63
London
London
Marylebone
Road
NO
2
131,677
0.82
London
London
Marylebone
Road
NO
x
131,677
0.83
variables’
strength
or
influence
on
prediction),
partial
dependencies,
and
predictions
for
each
of
the
50
models
were
then
summarised.
The
summaries
used
from
the
“ensemble
of
the
ensembles”
were
the
mean,
and
the
2.5%
and
97.5%
quantiles
of
the
50
estimates
i.e.
a
range
that
spans
the
95%
confidence
interval
in
the
mean.
The
model
performance
statistics
for
the
four
sets
of
models
are
displayed
in
3.
Results
and
discussion
3.1.
Port
of
Dover
SO
2
3.1.1.
Models
The
random
forest
models
grown
for
SO
2
at
the
two
Dover
sites
had
R
2
values
of
63
and
67%
(Table
3),
therefore,
the
models
had
mod-
erate
explanatory
ability
for
Dover’s
SO
2
concentrations.
However,
it
should
be
noted
that
predicting
concentrations
over
such
short
time
periods
with
intermittent
source
strength
is
challenging
and
data
capture
was
less
than
ideal
for
these
monitoring
sites.
The
moder-
ate
performance
can
be
explained
by
SO
2
at
this
location
containing
large
amounts
of
variation
due
to
ship
movements
and
if
winds
were
in
a
favourable
direction
to
transport
emissions
from
the
port
com-
plex
to the
monitoring
sites (southerlies).
Indeed,
wind direction
was
the
most
important
variable
for
SO
2
explanation
for
the
random
forest
models
(Fig.
2).
Partial
dependence
plots
of
decision
tree
models
allow
the
learn-
ing
process
to
be
interpreted
and
a
data
user
to
examine
how
variables
are
being
handled
in
the
predictive
model.
Fig.
3
demon-
strates
a
two-way
partial
dependence
plot
for
SO
2
concentrations
at
Dover
Landon
Cliff
using
wind
direction
and
date
(the
trend
term)
as
the
independent
variables.
The
feature
which
is
most
clear
is
the
band
of
increased
SO
2
dependence
between
150
and
210
◦
.O
u
t
s
i
d
e
of
this
band
of
southerly
winds,
there
were
low
levels
of
dependence
on
SO
2
concentrations.
The
Dover
Landon
Cliff
monitoring
site
was
located
north
of
the
Port
of
Dover
docks
and
very
slightly
to
the
east
Fig.
2.
Variable
importance
plot
for
SO
2
at
Dover
Langdon
Cliff
between
2001
and
2010
calculated
by
50
random
forest
models.
(Fig.
1
(a)).
The
partial
dependence
on
wind
direction
is
consistent
with
this
location
and
indicates
that
wind
direction
was
handled
sen-
sibly
in
the
random
forest
model.
This
observation
can
be
confirmed
further
with
a
bivariate
polar
plot
of
mean
SO
2
concentrations
by
wind
direction
and
speed
at
the
monitoring
site
(Fig.
4).
The
first
sul-
fur
content
fuel
change
in
mid-August
2006
can
also
be
seen
in
the
two-way
partial
dependence
plot
as
a
clear
reduction
in
SO
2
depen-
dence
when
winds
were
sourced
from
the
port
(the
south;
discussed
further
in
Section
3.1.2;
Fig.
3).
Another
clear
feature
isolated
by
the
partial
dependence
plots
was
that
SO
2
concentrations
increased
with
increasing
air
tempera-
ture
at
the
Dover
monitoring
sites
(Fig.
5).
This
relationship
was
an
unexpected
outcome
because
generally,
pollutant
concentrations
are
inversely
related
to
air
temperature
because
emissions
are
more
effi-
ciently
diluted
during
warmer
periods
owing
to
increased
thermal
turbulence.
For
some
sources
such
as
heating,
emissions
are
greater
at
lower
temperatures,
but
when
considering
shipping
emissions,
this
would
be negligible.
At
Dover,
the
SO
2
relationship
between
concentrations
and
air
temperatures was
indicative of
convective
thermal
mixing
being
an
important
physical
process
which
resulted
in
SO
2
emitted
by
ships
to
be
mixed
towards
the
measurement
site
at
the
cliff
top.
This
turbulent
mixing
at
high
temperatures
resulted
in
high
SO
2
concentrations
at
the
surface
and
this
feature
cannot
be
easily
observed
in
the
hourly
observational
data.
The
illumination
of
such
physical
processes
is
a
major
advantage
of
the
random
for-
est
algorithm
compared
to
other
machine
learning
methods
such
as
support
vector
machines
(SVM)
or
artificial
neural
networks
(ANNs)
because
they
do
not
offer
the
same
amount
of
model
legibility.
3.1.2.
Influence
of
sulfur
fuel
limits
on
SO
2
concentrations
Since
the
early
2000s,
there
has
been
a
number
of
increasingly
stringent sulfur based fuel limits
imposed on ships operating in
British
and
European
Union
(EU)
waters
due
to
their
status
as
Sul-
fur
Emission
Control
Areas
(SECAs)
or
Emission
Control
Areas
(ECAs).
The
most
important
events
for
sulfur
control
were
implemented
on
August
11, 2006
and
January 1,
2010.
In August
2006,
the MAR-
POL
Annex
IV
regulations
were
applied
which
introduced
a
1.5
%
Fig.
3.
Partial
dependence
of
wind
direction
and
date
on
SO
2
concentrations
at
Dover
Landon
Cliff
between
2001
and
2010.
The
Dover
Landon
Cliff
monitoring
site
was
located
north
of
the
Port
of
Dover
(Fig.
1
(a)).
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
583
Fig.
4.
Bivariate
polar
plot
of
mean
hourly
SO
2
concentrations
at
Dover
Landon
Cliff
between
2001
and
2010.
The
Dover
Landon
Cliff
monitoring
site
was
located
north
of
the
Port
of
Dover
(for
a
location
map,
see
Fig.
1
(a)).
sulfur
limit
on
fuel
oils
used
by
vessels
moving
between
EU
ports
(International
Maritime
Organization,
2005).
The
pre-August
2006
sulfur
content
for
British
vessels
has
been
estimated
at
2.7
%
which
represents
a
reduction
in
sulfur
content
of
44
%
(Entec,
2010).
At
the
start
of
2010
an
additional
limit
was
imposed
for
all
vessels
at
berth
where
such
vessels
were
required
to
be
operated
with
maximum
fuel
sulfur
content
of
1
%.
These
changes
should
be
evident
in
the
SO
2
time
series
of
the
nearby
ambient
monitoring
sites.
However,
if
a
time
series
is
plotted,
the
influence
of
these
changes
are
subtle
and
not
clear
due
to
the
high
amounts
of
variation
within
SO
2
concentrations
(Fig.
6).
The
meteorologically
normalised
SO
2
time
series
for
the
Dover
sites
are
displayed
in
Fig.
7,
after
the
observations
were
filtered
to
wind
directions
which
came
for
the
port,
hence
the
tight
95
%
confi-
dence
intervals.
The
dates
when
changes
in
sulfur
fuel
content
were
Fig.
5.
Partial
dependence
of
SO
2
on
air
temperature
at
Dover
Landon
Cliff
between
2001
and
2010
calculated
by
50
random
forest
models.
implemented
are
displayed
as
vertical
lines
in
Fig.
7
and
the
influence
of
sulfur
fuel
changes
are
clear
(compared
with
Fig.
6).
At
Dover
Langdon
Cliff,
the
monitoring
site
which
was
online
dur-
ing
the
MARPOL
1.5
%
fuel
sulfur
limit
transition
during
August
2001
shows
the
shift
in
ambient
SO
2
very
clearly
(Fig.
7).
The
mean
meteo-
rologically
normalised
SO
2
concentrations
for
the
pre-
and
post-fuel
change periods were 48
and 26
l
gm
−3
respectively. This differ-
ence
represented
in
percentage
change
is
45%
and
the
corresponding
estimated
change
in
sulfur
fuel
content
was
44
%.
This
extremely
good
agreement
between
sulfur
content
fuel
changes
and
normalised
ambient
SO
2
concentrations
suggests
that
the
Port
of
Dover
activities
and
ship
movements
remained
constant
during
the
transition
phase
and
the
source
of
SO
2
at
this
location
was
almost
exclusively
from
the
port.
The
second
sulfur
fuel
content
change
was
implemented
on
Jan-
uary
1,
2010
and
this
intervention
is
also
clearly
displayed
in
the
meteorologically
normalised
SO
2
concentrations
of
the
Dover
Docks
monitoring
site
(Fig.
7).
The
percentage
change
in
fuel
sulfur
content
was 33
% and
the percentage
change in
ambient SO
2
concentra-
tions
was
32
%.
Like
the
previous
intervention,
these
two
percent-
age
changes
match
almost
exactly,
which
is
somewhat
surprising
because
the
intervention
was
applied
only
to
berthed
vessels
which
would
only
make
up
a
component
of
the
Port
of
Dover
activities.
3.2.
London
Marylebone
Road
NO
x
3.2.1.
Models
The
random
forest
models
of
NO
x
and
NO
2
at
London
Marylebone
Road
performed
well
and
had
R
2
values
of
82
and
83
%
respectively
(Table
3).
This
good
performance
can
be
explained
by
hour
of
day
being
a
very
good
predictor
for
traffic
flows
and
therefore
emissions
at
this
location
for
these
(mostly)
traffic-sourced
pollutants
(Fig.
8).
The
performance
of
the
random
forest
models
would
be
rather
dif-
ficult
to achieve
with
dispersion
or deterministic
models in
such
a
complicated
location.
For
example,
the
dispersion
models
evalu-
ated
in
Carslaw
et
al.
(2013)
struggled
to
represent
the
street
canyon
environment,
even
when
traffic
information
was
taken
into
account.
The
importance
plots
for
the
London
Marylebone
Road
models
also
show
that
wind
direction
is
the
most
important
variable
to
predict
NO
2
and
NO
x
concentrations.
London
Marylebone
Road
is
located
in
a
street
canyon
and
is
subjected
to
complex
flows,
including
venti-
lation,
vortices,
and
leeward
accumulation
of
pollutants,
(primarily)
dependent
on
wind
direction
(Carslaw
and
Carslaw,
2007;
Catalano
et
al.,
2016).
This
complexity
is
demonstrated
in
the
importance
of
wind
direction
in
explaining
NO
x
and
NO
2
concentrations
(Fig.
8)a
n
d
this
has
been
noted
before
at
this
location
(Charron
and
Harrison,
2005;
Westmoreland
et
al.,
2007).
3.2.2.
Changes
in
primary
NO
2
Using the
predictive models
for meteorological
normalisation
results
in
very
clear
and
almost
noiseless
meteorologically
nor-
malised
trends
shown
in
Fig.
9.
It
is
immediately
clear
that
NO
x
and
NO
2
are
not behaving
the same
way
at this
monitoring
location.
This
is
because
of
changes
in
vehicular
primary
(directly
emitted)
NO
2
during
the
analysis
period
(1997–2016)
(Carslaw,
2005;
Carslaw
et
al.,
2016;
Grange
et
al.,
2017).
The
vertical
lines
on
Fig.
9
show
the
breakpoints identified
by
structural change
analysis
after the
meteorological
normalisation
procedure.
NO
x
concentrations decreased after the introduction of a bus
lane adjacent to
the monitoring site
in 2001 but have
remained
near
constant
since
the
introduction
of
the
CCZ
in
February
2003
(Fig.
9
and
Table
2).
Despite
the
progressively
stringent
vehicular
emission
controls
being
applied
across
Europe
between
2003
and
2016
(the
last
year
of
data
in
analysis),
they
have
had
little
effect
to
NO
x
at
London
Marylebone
Road.
This
observation
could
be,
at
584
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
Fig.
6.
Daily
SO
2
concentrations
at
two
monitoring
sites
in
Dover
between
2001
and
2012.
least
partly,
explained
by
the
disconnect
between
laboratory
test-
ing
and
real-world
emissions
of
NO
x
which
become
a
public
issue
after
the
diesel
emission
scandal
in
September
2015
(Brand,
2016;
Schmidt,
2016).
However,
heavy
duty
vehicles
are
also
very
impor-
tant
to
consider
alongside
passenger
vehicles
at
this
Central
London
location
(Laybourn-Langton
et
al.,
2016;
Greater
London
Authority,
2017).
NO
2
concentrations
at
London
Marylebone
Road
have
increased
since
1997
and
were
at their
maximum between
2002 and
2008
(Fig.
9).
The
changes
observed
can
be
explained
by
changes
to
the
vehicle
fleet
using
the
adjacent
A501
road
resulting
from
the
intro-
duction
of
congestion
charging,
London’s
Low
Emission
Zone,
and
evolution
of
the
local
bus
fleet.
The
rapid
increase
of
NO
2
concentra-
tions
was
observed
in
the
meteorologically
normalised
time
series
between
July
2002
and
July
2003
(Fig.
9).
The
CCZ
was
introduced
in
mid-February
2002;
right
in
the
middle
of
the
period
of
increas-
ing
NO
2
and
within
six
months
of
the
suggested
breakpoint
(October
2012).
The
increase
in
NO
2
concentrations
was
due
to
increased
pri-
mary
NO
2
because
no
change
in
the meteorologically
normalised
NO
x
was
observed
at
the
same
time.
The
implementation
of
the
CCZ
was
accompanied
with
a
retrofitting programme of Euro III local buses with continuously
regenerating
diesel
particulate
filters
(CRDPF,
also
known
by
their
commercial
name:
CRT
filters).
CRDPF
are
passive
devices
and
have
two
components:
an
upstream
oxidation
catalyst
and
a
particulate
matter
(PM) filter.
The oxidation
catalyst oxidises
NO
within the
exhaust
stream
to
NO
2
and
this
NO
2
is
then
used
as
a
PM
oxidant
in
the
filter-proper.
The
observations
show
that
these
retrofitted
pas-
sive
devices
were
not
optimised
because
much
of
the
generated
NO
2
was
not
reduced
within
the
PM
filter
and
was
therefore
emitted
into
Fig.
7.
Meteorologically
normalised
SO
2
concentrations
at
two
monitoring
sites
in
Dover
between
2001
and
2012
as
calculated
by
50
random
forest
models.
The
vertical
lines
show
the
start
dates
of
when
changes
in
marine
sulfur
fuel
content
were
implemented.
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
585
Fig.
8.
Variable
importance
plot
for
50
NO
2
random
forest
models
for
London
Marylebone
Road.
The
uncertainty
among
the
importances
of
the
50
models
was
very
small
and
therefore
the
quantiles
are
not
shown.
The
importances
for
the
NO
x
models
were
very
similar.
the
roadside
atmosphere
and
thus
significantly
increased
ambient
NO
2
concentrations
(Fig.
9).
NO
2
concentrations
remained
approximately
stable
until
February
2008
when
London’s
Low
Emission
Zone
(LEZ)
was
introduced
and
NO
2
concentrations
began
to
decrease
(Fig.
9).
The
second
NO
2
break-
point
was detected
for
February 2008
giving
some evidence
that
the
LEZ
reduced
NO
2
concentrations
at
London
Marylebone Road
(although
no
corresponding
change
in
NO
x
was
observed).
However,
during
this
period
the
local
bus
fleets
were
also
being
progressively
replaced
with
newer
buses
compliant
to
the
later
Euro
IV,
V,
and
VI
heavy
vehicle
emission
standards
(Finn
Coyle,
Tom
Cunnington,
and
Gabrielle
Bowden
(Transport
for
London),
personal
communica-
tion,
March
2018)
as
well
of
natural
vehicle
turnover
removing
older
and
more
polluting
vehicles
from
the
in-service
fleet.
The
third
NO
2
breakpoint
identified
coincided
with
route
18,
the
bus
route
with
the
highest
peak
vehicle
requirements
(PVR),
shifting
from
Euro
III
to
Euro
V
vehicles
in
late
2010
(Fig.
9).
After
2011,
NO
2
concentrations
con-
tinued
to
decline
with
the
introduction
of
Euro
VI
and
hybrid
buses
servicing
the
453,
27,
and
205
routes.
By
the
end
of
2016,
NO
2
had
declined
to
almost
pre-CCZ
concentrations.
The
features
displayed
in
the
normalised
time
series
were
not
clear
in
the
raw
concentration
data
(displayed
in
Fig.
A2)
and
the
breakpoints
identified
were
unable
to
be
resolved
without
the
meteorological
normalisation
technique.
The
tandem
use
of
the
meteorological
normalisation
procedure
and
breakpoint
analysis
is
powerful
and
can
revel
many
changes,
but
in
many
cases
there
may
not
be
sufficient
information
or
metadata
to
help
explain
the
changes
observed.
In
this
Central
London
example,
many
of
the
factors
driving
pollutant
concentrations
are
known
due
to
the
site’s
prominence.
London
Marylebone
Road
also
monitors
ozone
(O
3
),
something
which
is
rare
for
roadside
monitoring
locations
in
Europe.
The
NO
2
,
NO
x
,a
n
dO
3
complement
allows
for
the
estimation
of
primary
NO
2
with
an
independent
method
by
determining
the
total
oxidant
(OX;
NO
2
+O
3
)
within
NO
x
(Jenkin,
2004;
Carslaw
and
Beevers,
2005).
Fig.
10
shows
monthly
estimates
of
the
primary
NO
2
fraction
at
London
Marylebone
Road
with
robust
linear
regression.
Fig.
10
is
consistent
with
Fig.
9
with
a
rapid
increase
in
primary
NO
2
during
2002
and
a
reduction,
but
at
a
slower
rate
after
2008
thus
further
confirming
and
validating
that
the
trends
observed
in
Fig.
9
are
driven
by
changes
in
primary
NO
2
emissions.
The
reason
why
the
trend
is
similar
in
Figs.
10
and
9
is
that
at
this
particular
site
increased
emissions
of
primary
NO
2
were
sufficient
to
have
a
measurable
effect
on
ambient
concentrations.
Fig.
9.
Meteorologically
normalised
NO
x
and
NO
2
at
London
Marylebone
Road
between
1997
and
2016
as
calculated
by
50
random
forest
models
(for
each
pollutant).
The
vertical
lines
on
show
the
breakpoints
identified
by
structural
change
analysis.
586
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
Fig.
10.
Monthly
total
oxidant
(OX;
NO
2
+O
3
)
at
London
Marylebone
Road
between
1997
and
2016.
Slope
and
errors
were
calculated
with
robust
linear
regression.
4.
Conclusions
Controlling
for
changes
of
meteorology
is
an
important
compo-
nent
to
consider
when
conducting
air
quality
data
analysis
over
time.
A
meteorological
normalisation
technique
using
random
forest
was
used
to
investigate
interventions
in
routine
air
quality
monitoring
data
from
two
areas.
The
interventions
applied
to
marine
fuel
con-
tent
changes
were
explored
in
Dover,
a
port
city
in
the
South
East
of
England
and
the
interventions
were
represented
in
the
meteoro-
logically
normalised
time
series
almost
exactly.
The
non-black
box
nature of
the random forest
models was used to
investigate the
dependence
of
pollutant
concentrations
on
meteorological
variables
such
as
air
temperature
and
wind
direction
which
highlighted
the
benefit
of
the
technique
where
physical
and
chemical
atmospheric
processes
can
be
illuminated,
understood,
and
explained.
In
the
example
of
the
implementation
of
congestion
charging
in
Central
London,
very
clear
changes
in
primary
NO
2
emissions
were
displayed
in
the
meteorologically
normalised
time
series.
The
perfor-
mance
of
these
roadside
models
was
high
due
to
the
models’
ability
to
use
wind
direction
and
hour
of
day
very
effectively,
something
which
dispersion
or
deterministic
models
struggle
with
when
used
for modelling
street canyon environments. The
case studies pre-
sented
are
both
examples
where
there
is
significant
ability
to
cross
check
the
observed
features
with
available
information
on
changes
in
the
sites’
local
environments
to
validate
the
outputs.
The
meteorological
normalisation
technique
is
very
relevant
for
exploring
the
influence
of
interventions
or
management
activities
on
local
air
quality.
The
combination
of
a
non-parametric
method,
the
lack
of
need
for
specialised
measurements,
and
the
effective
use
of
proxy
variables
lends
the
technique
to
a
wide
range
of
air
quality
data
analysis
applications.
Competing
interests
The
authors
declare
no
competing
interest.
Acknowledgments
S.K.G. was supported by Anthony Wild
with the provision of
the Wild
Fund Scholarship.
This work
was also
partially
funded
by
Natural
Environment
Research
Council (NERC)
[grant number:
NE/N007115/1].
Appendix
A
Fig.
A1.
The
framework
for
the
meteorological
normalisation
technique.
The
training
and
validation
phase
is
iterative
to
ensure
the
model
does
not
overfit
and
adequate
performance
is
achieved.
After
the
technique
has
been
completed,
other
analyses
are
conducted
on
the
normalised
time
series.
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
587
Fig.
A2.
Daily
NO
2
and
NO
x
concentrations
at
London
Marylebone
Road
between
1997
and
2016.
References
Anh,
V.,
Duc,
H.,
Azzi,
M.,
1997.
Modeling
anthropogenic
trends
in
air
quality
data.
J.
Air
Waste
Manag.
Assoc.
47
(1),
66–71.
https://doi.org/10.1080/10473289.1997.
10464406.
Atkinson,
R.,
Barratt,
B.,
Armstrong,
B.,
Anderson,
H.,
Beevers,
S.,
Mudway,
I.,
Green,
D.,
Derwent,
R.,
Wilkinson,
P.,
Tonne,
C.,
Kelly,
F.,
2009.
The
impact
of
the
congestion
charging
scheme
on
ambient
air
pollution
concentrations
in
London.
Nov.
Atmos.
Environ.
43
(34),
5493–5500.
http://www.sciencedirect.com/science/article/pii/
S1352231009006268.
Barmpadimos,
I.,
Hueglin,
C.,
Keller,
J.,
Henne,
S.,
Prévôt,
A.S.H.,
2011.
Influence
of
meteorology
on
PM
10
trends
and
variability
in
Switzerland
from
1991
to
2008.
Feb..
Atmos.
Chem.
Phys.
11
(4),
1813–1835.
http://www.atmos-
chem-
phys.net/
11/1813/2011/.
Brand,
C.,
2016.
Beyond
‘Dieselgate’:
implications
of
unaccounted
and
future
air
pollu-
tant
emissions
and
energy
use
for
cars
in
the
United
Kingdom.
Energy
Policy
97,
1–12.
http://www.sciencedirect.com/science/article/pii/S030142151630341X.
Breiman,
L.,
1996.
Bagging
predictors.
Mach.
Learn.
24
(2),
123–140.
https://doi.org/
10.1007/BF00058655.
Breiman,
L.,
2001.
Random
forests.
Mach.
Learn.
45
(1),
5–32.
https://doi.org/10.1023/
A:1010933404324.
Carslaw,
D.,
Apsimon,
H.,
Beevers,
S.,
Brookes,
D.,
Carruthers,
D.,
Cooke,
S.,
Kitwiroon,
N.,
Oxley,
T.,
Stedman,
J.,
Stocker,
J.,
2013.
Defra
Phase
2
Urban
Model
Evaluation.
UK AIR:
Air Information Resource.
https://uk-
air.defra.gov.uk/library/reports?
report_id=777.
Carslaw,
D.C.,
2005.
Evidence
of
an
increasing
NO
2
/NO
x
emissions
ratio
from
road
traffic
emissions.
Aug.
Atmos.
Environ.
39
(26),
4793–4802.
http://www.
sciencedirect.com/science/article/pii/S1352231005005443.
https://doi.org/10.
1016/j.atmosenv.2005.06.023.
Carslaw,
D.C.,
Beevers,
S.D.,
2005.
Estimations
of
road
vehicle
primary
NO
2
exhaust
emission
fractions
using
monitoring
data
in
London.
Jan.
Atmos.
Environ.
39
(1),
167–177.
http://www.sciencedirect.com/science/article/pii/
S1352231004008775.
Carslaw,
D.C.,
Carslaw,
N.,
2007.
Detecting
and
characterising
small
changes
in
urban nitrogen dioxide
concentrations. Atmos. Environ. 41
(22), 4723–4733.
http://www.sciencedirect.com/science/article/pii/S1352231007002919.
https://
doi.org/10.1016/j.atmosenv.2007.03.034.
Carslaw,
D.C.,
Murrells,
T.P.,
Andersson,
J.,
Keenan,
M.,
2016.
Have
vehicle
emissions
of
primary
NO
2
peaked?
Faraday
Discuss.
189
(0),
439–454.
https://doi.org/10.1039/
C5FD00162E.
Carslaw,
D.C.,
Ropkins,
K.,
Bell,
M.C.,
2006.
Change-point
detection
of
gaseous
and
particulate traffic-related
pollutants at
a roadside
location. Nov.
Environ. Sci.
Technol.
40
(22),
6912–6918.
https://doi.org/10.1021/es060543u.
Catalano,
M.,
Galatioto,
F.,
Bell,
M.,
Namdeo,
A.,
Bergantino,
A.S.,
2016.
Improving
the
prediction
of
air
pollution
peak
episodes
generated
by
urban
transport
networks.
Jun.
Environ.
Sci.
Pol.
60,
69–83.
http://www.sciencedirect.com/science/article/
pii/S1462901116300594.
Charron,
A.,
Harrison,
R.M.,
2005.
Fine
(PM
2.5
)a
n
dc
o
a
r
s
e(
P
M
2.5−10
)
particulate
matter
on
a
heavily
trafficked
London
Highway:
sources
and
processes.
Oct.
Environ.
Sci.
Technol.
39
(20),
7768–7776.
https://doi.org/10.1021/es050462i.
Derwent,
R.,
Middleton,
D.,
Field,
R.,
G
oldstone,
M.,
Lester,
J.,
Perry,
R.,
1995.
Analysis
and
interpretation
of
air
quality
data
from
an
urban
roadside
location
in
Central
London
over
the
period
from
July
1991
to
July
1992.
Atmos.
Environ.
29
(8),
923–946.
http://www.sciencedirect.com/science/article/pii/135223109400219B.
Entec,
2010.
Defra
UK
Ship
Emissions
Inventory–Final
Report.
Doc
Reg
No.
21897-
01.
https://uk-
air.defra.gov.uk/assets/documents/reports/cat15/1012131459_
21897_Final_Report_291110.pdf.
European
Environment
Agency,
2016.
Air
Quality
in
Europe
–
2016
Report.
EEA
Report.
No
28/2016.
http://www.eea.europa.eu/publications/air-
quality-
in-
europe-
2016.
Friedman, J., Hastie, T., Tibshirani, R., 2001.
The Elements of Statistical Learning.
2nd
ed.,
Data
Mining,
Inference,
and
Prediction,
1.
Springer
series
in
statistics
Springer,
Berlin.
Grange,
S.K.,
2016.
smonitor:
A
Framework
and
a
Collection
of
Functions
to
Allow
for
Maintenance
of Air
Quality
Monitoring
Data.
https://github.com/skgrange/
smonitor.
Grange, S.K., 2017.
Technical Note: smonitor
Europe. Tech. rep.,
Wolfson Atmo-
spheric
Chemistry
Laboratories,
University
of
York.
https://doi.org/10.13140/RG.
2.2.20555.49448/1.
Grange,
S.K.,
2018.
rmweather:
Tools
to
Conduct
Meteorological
Normalisation
on
Air
Quality
Data.
R
package
version
0.1.2.
https://CRAN.R-
project.org/package=
rmweather.
Grange,
S.K.,
Carslaw,
D.C.,
Lewis,
A.C.,
Boleti,
E.,
Hueglin,
C.,
2018.
Random
forest
mete-
orological
normalisation
models
for
Swiss
PM
10
trend
analysis.
May.
Atmos.
Chem.
Phys.
18
(9),
6223–6239.
https://www.atmos-
chem-
phys.net/18/6223/2018/.
Grange,
S.K.,
Lewis,
A.C.,
Moller,
S.J.,
Carslaw,
D.C.,
2017. Lower
vehicular primary
emissions
of
NO
2
in
Europe
than
assumed
in
policy
projections.
Dec.
Nat.
Geosci.
10
(12),
914–918.
https://doi.org/10.1038/s41561-
017-
0009-
0.
Authority,
Greater
London,
2017.
London
Atmospheric
Emissions
Inventory
(laei)
2013.
https://data.london.gov.uk/dataset/london-
atmospheric-
emissions-
inventory-
2013.
Henneman,
L.R.,
Holmes,
H.A.,
Mulholland,
J.A.,
Russell,
A.G.,
2015.
Meteorolog-
ical
detrending of
primary
and
secondary pollutant
concentrations:
method
application
and
evaluation
using long-term
(2000–2012)
data
in
Atlanta.
Oct.
Atmos.
Environ.
119,
201–210.
http://www.sciencedirect.com/science/article/pii/
S1352231015302521.
588
S.
Grange,
D.
Carslaw
/
Science
of
the
Total
Environment
653
(2019)
578–588
International
Maritime
Organization,
2005.
Revised
MARPOL
Annex
VI.
Annex
VI
of MARPOL
addresses
air
pollution
from
ocean-going
ships.
http://www.
imo.org/en/OurWork/Environment/PollutionPrevention/AirPollution/Pages/Air-
Pollution.aspx.
Jeanjean,
A.P.R., Buccolieri,
R.,
Eddy,
J., Monks,
P.S.,
Leigh,
R.J.,
2017.
Air quality
affected
by
trees
in
real
street
canyons:
the
case
of
Marylebone
neighbourhood
in
central
London.
Mar.
Urban
For.
Urban
Green.
22,
41–53.
http://www.
sciencedirect.com/science/article/pii/S1618866716303740.
Jenkin,
M.E.,
2004.
Analysis of
sources and
partitioning
of
oxidant
in
the
UK–part
2:
contributions
of
nitrogen
dioxide
emissions
and
background
ozone
at
a
kerbside
location
in
London.
Sep.
Atmos.
Environ.
38
(30),
5131–5138.
http://
www.sciencedirect.com/science/article/pii/S1352231004006193.
Jones,
Z.,
Linder,
F.,
2015.
Exploratory
Data
Analysis
Using
Random
Forests.
73rd
annual
MPSA
conference,
April
16-19,
2015,
Chicago,
United
States
of
America.
https://
pdfs.semanticscholar.org/e7b7/3565b07a7f1369a20b1055f222423f0feb34.pdf.
Laybourn-Langton, L., Quilter-Pinner, H.,
Ho, H., 2016. Lethal and
Illegal: Solving
London’s
Air
Pollution
Crisis.
Institute
for
Public
Policy
Research.
http://www.
ippr.org/read/lethal-
and-
illegal-
solving-
londons-
air-
pollution-
crisis.
Libiseller,
C.,
Grimvall,
A.,
Waldén,
J.,
Saari,
H.,
2005.
Meteorological
normalisation
and non-parametric smoothing for quality assessment and
trend analysis of
tropospheric
ozone
data.
Environ.
Monit.
Assess.
100
(1),
33–52.
https://doi.org/
10.1007/s10661-
005-
7059-
2.
Lyubchich,
V.,
Gel,
Y.R.,
El
Shaarawi,
A.,
2013.
On
detecting
non-monotonic
trends
in
envi-
ronmental
time
series:
a
fusion
of
local
regression
and
bootstrap.
Environmetrics
24
(4),
209–226.
https://doi.org/10.1002/env.2212.
Monks,
P.,
Granier,
C.,
Fuzzi,
S.,
Stohl,
A.,
Williams,
M.,
Akimoto,
H.,
Amann,
M.,
Baklanov,
A.,
Baltensperger,
U.,
Bey,
I.,
Blake,
N.,
Blake,
R.,
Carslaw,
K.,
Cooper,
O.,
Dentener,
F.,
Fowler,
D.,
Fragkou,
E.,
Frost,
G.,
Generoso,
S.,
Ginoux,
P.,
Grewe,
V.,
Guenther,
A.,
Hansson,
H.,
Henne,
S.,
Hjorth,
J.,
Hofzumahaus,
A.,
Huntrieser,
H.,
Isaksen,
I.,
Jenkin,
M.,
Kaiser,
J.,
Kanakidou,
M.,
Klimont,
Z.,
Kulmala,
M.,
Laj,
P.,
Lawrence,
M.,
Lee,
J.,
Liousse,
C.,
Maione,
M.,
McFiggans,
G.,
Metzger,
A.,
Mieville,
A.,
Moussiopoulos,
N.,
Orlando,
J.,
O’Dowd,
C.,
Palmer,
P.,
Parrish,
D.,
Petzold,
A.,
Platt,
U.,
Pschl,
U.,
Prvt,
A.,
Reeves,
C.,
Reimann,
S.,
Rudich,
Y.,
Sellegri,
K.,
Steinbrecher,
R.,
Simpson,
D.,
ten
Brink,
H.,
Theloke,
J.,
van
der
Werf,
G.,
Vautard,
R.,
Vestreng,
V.,
Vlachokostas,
Ch.,
von
Glasow,
R.,
2009.
Atmospheric
composition
change
-
global
and
r
egional
air
quality.
Atmos.
Environ.
43
(33),
5268–5350.
http://www.sciencedirect.com/
science/article/B6VH3-
4X3N46N-
1/2/1db0fa3c5afafc9418ab227802a71755.
NOAA,
2016.
Integrated
Surface
Database
(ISD).
https://www.ncdc.noaa.gov/isd.
Core
Team,
R., 2018.
R:
A
Language
and Environment
for
Statistical
Computing.
R
Foundation for
Statistical Computing, Vienna,
Austria.
https://www.R-
project.
org/.
Ricardo
Energy
&
Environment,
2018.
Kent
Air
Quality
Database.
http://www.kentair.
org.uk.
Schmidt,
C.W.,
2016.
Beyond
a
one-time
scandal:
Europe’s onging
diesel
pollution
problem.
Jan.
Environ.
Health
Perspect.
124
(1),
A19–A22.
http://www.ncbi.nlm.
nih.gov/pmc/articles/PMC4710587/.
Stull,
R.B.,
1988.
An
Introduction
to
Boundary
Layer
Meteorology.
Kluwer
Academic
Publishers,
London.
Tong,
W.,
Hong, H.,
Fang,
H.,
Xie, Q.,
Perkins,
R.,
2003.
Decision forest:
combining
the
predictions
of
multiple
independent
decision
tree
models.
Mar.
J.
Chem.
Inf.
Comput.
Sci.
43
(2),
525–531.
https://doi.org/10.1021/ci020058s.
Transport
for
London,
2018.
Driving.
https://tfl.gov.uk/modes/driving/.
Weiss,
M.,
Bonnel,
P.,
Kühlwein,
J.,
Provenza,
A.,
Lambrecht,
U.,
Alessandrini,
S.,
Carriero,
M.,
Colombo,
R.,
Forni,
F.,
Lanappe,
G.,
Le
Lijour,
P.,
Manfredi,
U.,
Montigny,
F.,
Sculati,
M.,
2012.
Will
Euro
6
reduce
the
NO
x
emissions
of
new
diesel
cars?
–
insights
from
on-road
tests
with
portable
emissions
measurement
systems
(PEMS).
Dec.
Atmos.
Environ.
62,
657–665.
http://www.sciencedirect.com/science/article/pii/
S1352231012008412.
Westmoreland, E.J., Carslaw, N.,
Carslaw, D.C., Gillah, A., Bates, E.,
2007. Analysis
of
air
quality
within
a
street
canyon
using
statistical
and
dispersion
modelling
techniques.
Dec.
Atmos.
Environ.
41
(39),
9195–9205.
http://www.sciencedirect.
com/science/article/pii/S1352231007006863.
Wise,
E.K.,
Comrie,
A.C.,
2005.
Extending
the
Kolmogorov-Zurbenko
filter:
application
to
ozone,
particulate
matter,
and
meteorological
trends.
Aug.
J.
Air
Waste
Manage.
Assoc.
55
(8),
1208–1216.
https://doi.org/10.1080/10473289.2005.10464718.
Zeileis,
A.,
Kleiber,
C.,
Krämer,
W.,
Hornik,
K.,
2003.
Testing
and
dating
of
structural
changes
in
practice.
Oct.
Comput.
Stat.
Data
Anal.
44
(1-2),
109–123.
http://www.
sciencedirect.com/science/article/pii/S0167947303000306.
Zeileis,
A.,
Leisch,
F.,
Hornik,
K.,
Christian,
K.,
2002.
strucchange:a
nRp
a
c
k
a
g
ef
o
r
testing
for
structural
change
in
linear
regression
models.
J.
Stat.
Softw.
7
(2),
1–38.
http://www.jstatsoft.org/v07/i02/.
Ziegler,
A.,
König,
I.R.,
2013.
Mining
data
with
random
forests:
current
options
for
real-
world
applications.
Dec.
WIREs
Data
Min.
Knowl.
Discovery
4
(1),
55–63.
https://
doi.org/10.1002/widm.1114.
Previous PDF in this issueNext PDF in this issue