Introduction
Ozone is a hazard to human health and
plants and animals and
a greenhouse gas . Prediction of ozone air quality on local
and regional scales is key for providing prior warning of impending ozone
exceedances . Knowledge of the processes
that control the variability in ozone precursors is vital for understanding
and predicting ozone air quality.
Currently, a wide variety of techniques are used to predict ozone
concentrations ranging from statistically based models
and neural networks to prognostic models of atmospheric
processes that include data
assimilation .
For prognostic models, uncertainties result from meteorology, the limitations
of the photochemical mechanisms, wet and dry deposition, uncertainties in the
emissions of ozone precursors, and, for data assimilation, observation
uncertainty . Most current statistical
and data assimilation air quality forecasting techniques rely primarily on
surface observing networks, but satellite observations are increasingly
coming to the fore .
Ozone pollution can develop under different polluted photochemical regimes.
Under low to moderate levels of NOx (NO and NO2) pollution, such as
can be found in rural and suburban environments, increases in NOx lead to
proportional increases in ozone, which is why this regime is classed as
NOx limited . Under much higher
levels of NOx pollution, i.e. those present in densely populated regions,
increases in NOx bring about decreases in ozone. Under these conditions,
the only means by which ground-level ozone can increase are via increases in
volatile organic compound (VOC) emissions , and
consequently this regime is considered to be VOC limited. Further, studies
show that the sensitivity of ozone to either NOx or VOCs can vary with
time, e.g. during different days of the
week . The priorities to monitor and
observe ozone and its different precursors therefore vary according to
location and time.
Observations and models, and their combination through data assimilation,
comprise essential tools for air quality
prediction . Observations are an
essential part of such systems, so it follows that their characteristics
could directly affect their performance. We seek to address this connection
in our study. Given this, we will now attempt to review the relevant
characteristics of the current and planned (in the near term) state of the
air quality monitoring network in order to explain the motivation for our
work and, later, to place some of our findings in context.
The US national surface air quality observing network typically observes a
wide range of chemical species. For instance, surface monitoring sites within
California (http://www.arb.ca.gov/adam/) have instruments that can
measure in situ ozone, CO, NO2, nitrogen oxide, particulate matter with
diameters of 2.5 and 10 µm, sulfur dioxide (SO2), methane, total
hydrocarbons, and hydrogen sulfide. The surface network is also usually able
to make observations at least at an hourly temporal resolution. However, due
to the spatial limitations of the surface air quality monitoring network,
space-borne remote-sensing observations, which typically have greater spatial
sampling, are also able to support air quality research and operational air
quality forecasting .
Surface station in situ data are made at a high spatial resolution (a few
metres up to a tens of kilometres), which is typically much higher than most
air quality models. As a result, this introduces the problem of having
representativity errors between the model, which is unable to represent
fine-scale variability, and the observations that can measure this
variability. This problem therefore limits the efficacy of data assimilation
and systems need to be carefully designed to take this type of error into
account.
For this study, the spatial characteristics of observations from different
platforms are not considered, but the advantages satellite data offer in
terms of increased spatial coverage have been recognised. Consequently,
various studies have been conducted that highlight the benefits of
satellite-borne instruments for air quality research
.
Further, satellite observations of air pollutants have been used within data
assimilation models to advance air quality research
.
Excluding the issue of spatial sampling, there are considerable differences
between remote-sensing observations and the existing surface observing
network. Each individual ground station is able to observe a wider range of
species at the surface (see above) but only at a single point. On the other
hand, space-based remote-sensing techniques can only observe a limited number
of species that have relevance to air quality (such as ozone, CO, NO2,
SO2, CH4, glyoxal, and HCHO), have coarser horizontal spatial
resolution observing with a footprint ranging from several to up to tens of
kilometres, and have (with current capabilities) only limited vertical
resolution and sensitivity to the surface or boundary layer. Also, all of the
studies cited above used instruments onboard satellites in low earth orbit
(LEO). Due to the orbital configuration, LEO-borne instruments are only able
to observe the same location on a far more infrequent basis compared to the
temporal sampling of the ground-based network.
Instruments onboard geostationary (GEO) satellites can also offer good
spatial coverage (on the continental and regional scale) without sacrificing
temporal sampling. This makes them potentially ideal to support future air
quality research and forecasting. However, in order to achieve this goal,
developments must be made to improve satellite instrument sensitivity to the
boundary layer and surface gas phase composition . Various
strategies have been proposed to achieve this aim (primarily for CO and
ozone). They typically consist of either combining wavelength bands that have
been previously exploited, i.e. ultraviolet (UV), visible (VIS), and IR
(infrared) , or
by focusing on new wavelength bands, i.e. the Chappuis bands for ozone in the
visible range that offer potential novel benefits. The UV
and the Chappuis band in the visible range were combined theoretically to
determine the benefit of such an approach during the development of the TEMPO
instrument and as part of a European
initiative .
As a result of the perceived benefits, several GEO missions are currently in
the various stages of planning. These include the Geostationary Coastal and
Air Pollution Events (GEO-CAPE) planned by NASA to cover the North American
continent (http://science.nasa.gov/earth-science/decadal-surveys/).
Sentinel 4 (http://www.esa.int/esaLP/SEM3ZT4KXMF_LPgmes_0.html) is
planned by ESA to cover Europe, and the Geostationary Environment
Spectrometer (GEMS) is aimed at providing coverage of East
Asia. Further, NASA's decadal survey and
state that GEO-CAPE and GEMS will observe the
following trace gases: ozone, CO (not with GEMS), NO2, HCHO, and
SO2.
GEO-based observations of trace gases are therefore becoming more relevant
for the study of air quality and for operational air quality forecasting. For
the planned GEO missions, various choices exist regarding which wavelength
bands to observe in, and these will influence the already limited range of
observable species in the troposphere. In addition, instrument design choices
affect how often observations can be made, at what time of day, and how well.
For instance, thermal infrared (TIR)-based instruments cannot measure
NO2, and UV–VIS instruments cannot observe during the night-time. Thus,
instrument design choices will affect the future capabilities of these
missions.
We have demonstrated that a range of possible capabilities and
characteristics exist for both the current and planned air quality observing
systems (ground and satellite based). Within the scope of this paper, we
study how the frequency and specific timing during the day of observation,
the species that are measured, and how well they are measured affect the
ability to conduct air quality research and to aid air quality forecasting
using a data assimilation system. This interaction between observation
characteristics and data assimilation system performance is interesting and
needs to be studied. Therefore, addressing this question will be of interest
to the current air quality observing network and to the planned or future GEO
air quality focused missions. In order to do this, we carry out a series of
sensitivity analyses using different sets of pseudo-observations to test the
influence that various observation characteristics have upon the ability to
predict ozone within an idealised model. This model consists of a
photochemical box model, its adjoint, and a 4D-variational data assimilation
system set-up to constrain ozone precursor emission uncertainties (NOx,
CO, and VOCs). This framework thereby mimics a state of the art air quality
forecasting system. We conduct an uncertainty analysis using a linear
estimation technique for each of our sensitivity tests. We are able to
perform the uncertainty analysis owing to the fact that we use a box model
because it limits the size of the matrices we solve for. Within the context
of a summertime ozone pollution episode that emerges during stagnant
anticyclonic conditions, we attempt to address the following specific
questions:
How does the ability to predict ozone vary across three separate observing
scenarios? The first uses only CO and NO2 observations (CN), the second
uses Ozone, CO, and NO2 (OCN), and the third uses HCHO, CO, and NO2
(HCN).
What are the effects of both observing frequency and the choice of when
to observe on the prediction of ozone within our framework?
How does observation noise, when applied evenly onto each observation,
affect ozone prediction in our system?
How are the results of these sensitivity tests affected by photochemical
regime (i.e., either NOx- or VOC-limited regimes)?
Ignoring ozone prediction, which combination of observed species allows
the best constraint on ozone precursor emissions?
In order to support our conclusions regarding the aims above we carry out a
variety of complementary analyses
to demonstrate that the 4D-variational data assimilation scheme can
solve the full non-linear retrieval of the emission parameters;
to test the robustness of our methodology to choices regarding our
assumed diurnal emission profile;
to test whether the assumed VOC emission uncertainties can be represented
using different VOCs.
Section 2 describes all aspects of the methodology, Sect. 3 describes the
results from each of the analyses, Sect. 4 discusses our results, Sect. 5
details our conclusions.
Methodology
Overview
We use a photochemical box model run over 3 days to represent a worsening
period of ozone air quality during a stagnation event. Meteorological
stagnation events under hot, sunlit conditions over urban areas typically
lead to poor ozone air quality . We assume
that the idealised mixing and transport represented in the box model are
sufficient to represent the meteorology during anticyclonic conditions. For
each of the different sensitivity tests that we perform we use different sets
of pseudo-observations of ozone, HCHO, CO and NO2 (see Sect. 2.3 and
examine Fig. to see an example of the
pseudo-observations relative to the true ozone state) in order to separately
constrain the ozone precursor emissions with the 4D-variational data
assimilation system. The ozone precursor emissions have known a priori
errors. We then make a prediction of ozone using the a posteriori emissions.
Within the model framework, days 1–2 represent the period over which
observations are made and the assimilation is carried out and the final day
represents the prediction and monitoring period. Within this final phase, we
compare the ozone prediction, based upon the a posteriori emissions, to the
ozone true state in order to assess the assimilation performance. We support
this assessment using a range of statistics and diagnostics that shall be
discussed shortly.
The use of 4D-variational data assimilation to solve the ozone precursor
emission inversion problem is consistent with the current state of the art in
prognostic air quality forecast modelling development. For example, the
Community Multi Scale Air Quality modelling system (),
the Sulfur Transport Eulerian Model (), and
are all developing such assimilation capabilities. Thus,
our model framework is relevant to and is reflective of the current and
future direction of air quality forecasting.
In order to establish the utility of more complex air quality forecasting
systems that might use 4D-variational data assimilation, our prototype
forecasting system is demonstrated theoretically. Since the emission
inversion problem that we explore only becomes more complex as the model
state space increases and additional sources of uncertainty are introduced, a
failure to show sufficiently reduced prediction error in this simplified
setting would indicate that more complex systems are unlikely to fare better.
Sufficient prediction model error within our framework is therefore a
necessary but not sufficient condition for more complex 4D-variational data
assimilation forecasting systems using air quality observations to be
successful.
One other advantage of selecting a photochemical box model is that we are
able to generate a Jacobian describing the model response to emission
parameter perturbations, which can be used within an analytical modelling
framework to conduct uncertainty analysis. It would be very difficult to
produce a Jacobian within regional or global chemical transport models in a
timely fashion given the size of the model state space. Therefore, we use an
analytic model (derived from the photochemical box model) that is simplified
relative to the full assimilation framework. This is a linear estimation
technique based upon . To support our analyses we
calculate the following diagnostics using this method: a posteriori ozone
prediction error covariance, a posteriori emission parameter error
covariance, the emission averaging kernel, and the associated degrees of
freedom of signal.
The 4D-variational (4D-var) data assimilation and uncertainty analysis using
the linear estimation are therefore complementary methods, and we use both
techniques to achieve our aim of exploring the effect of observing
characteristics on ozone prediction. In addition, we conduct a series of
supporting analyses to test some of our assumptions.
Photochemical box model
A pseudo 1-D photochemical box model was built using the Kinetic
Pre-Processor (KPP) . The model is
not truly 1-D in the vertical because we use a parameterisation to describe
variability in the boundary layer height and mixing volume. The Rosenbrock
solver is used to integrate the KPP-generated ordinary differential equations
required to calculate trace gas concentrations . The
photochemical mechanism consists of 171 gas phase species and 524 chemical
reactions simulating the degradation of hydrocarbons from C1 to C5
including isoprene and is based upon the Master Chemical Mechanism v3.1
(http://mcm.leeds.ac.uk/MCM/). In addition, the
model includes dry deposition for all relevant chemical species, it contains
a two-parameter photolysis scheme, and it simulates the emission of ozone
precursors including NOx, CO, and VOCs.
The various different profiles of the temporal variability emission
factor, k(t), used in the analysis of the emission solution sensitivity to
diurnal emission variability. The red dashed and the solid black lines
indicate the alternative and standard emissions variabilities, respectively.
The different profiles of variability are indicated at the top of each panel
in bold text.
Coastal urbanised southern California (SC) has historically been, and
continues to be, an interesting area of study for air quality owing to the
large-scale urbanisation and population, the resulting anthropogenic
emissions, and the meteorological conditions during summertime that are
favourable for the development of photochemical smog conditions. We therefore
set up the box model to study conditions that are analogous to this region
and environment. Consequently, we situate the box model at 33∘ N,
run it from 30 June to 2 July, and use an atmospheric humidity equivalent to
a volume mixing ratio of 0.0162. In addition, we use anthropogenic (NOx,
CO, and VOCs) and biogenic (isoprene) emissions that result in a range of
atmospheric mixing ratios typical for urbanised SC.
The diurnal emission variability in anthropogenic compounds is prescribed
according to the National Atmospheric Emissions Inventory (NAEI)
(http://www.naei.org.uk/emissions/) for an urbanised area (see
Fig. ), and the isoprene emission variability is
parameterised to correlate to solar zenith angle offset by 2 h to consider
both temperature and photon flux effects . The
isoprene emissions have an average daily emission of
1.7 × 1010 molecules m2 s-1 and an afternoon peak
of 4.6 × 1010 molecules m2 s-1, which yields
modelled isoprene mixing ratios less than 10 pptv (parts per trillion by
volume) typical for this region. The diurnal variability in the isoprene
emissions is separate and distinct to the anthropogenic VOCs. From now on,
when we discuss VOCs we are referring to anthropogenic VOCs unless otherwise
stated. The VOC speciation is defined according to NAEI and the total peak
emission of carbon via VOCs (excluding isoprene) is
2.3 × 1012 carbon atoms m-2 s-1 and the average
emission is 1.2 × 1012 carbon atoms m-2 s-1.
These anthropogenic VOC emissions are typical for urbanised regions. Boundary
layer dynamics are described with a prescribed variability in mixing height
ranging from 500 to 1500 m and mixing between the boundary layer and free
troposphere equivalent to a constant 10 % mass exchange per hour. In our
model, the vertical extent represents the full depth of the boundary layer.
Background free tropospheric concentrations of long-lived species are assumed
to remain constant and are defined in Table .
Background free-tropospheric concentrations of trace gases mixed
into the boundary layer in the photochemical model. NMHCs indicate
non-methane hydrocarbons.
Chemical species
Background mixing ratio
Ozone
30 ppbv
NO
100 pptv
NO2
50 pptv
CO
80 ppbv
CH4
1.76 ppm
NMHCs
100–200 pptv each
The model is run under a range of photochemical conditions typical for
urbanised SC. This is achieved by varying the NO emissions across nine
different scenarios that span the full range of modelled ozone responses with
respect to changing NOx concentration (i.e. from NOx- to
VOC-limited conditions). We use the same emissions for the other species
across all of these different NO emission scenarios. For the purposes of the
emission inversion, we define our ozone precursor emissions in a simplified
form (excluding emitted species not considered in the inversion) as
ϕi(t)=xiEi(t),i=NO,CO,VOC,
where xi represents the time-independent emission scaling factors for
the emitted species, i, and Ei(t) represents the emissions with a
prescribed and repeating diurnal cycle for each emitted species. The emission
inversion solves for xi, the time-independent emission scaling factors,
which can be represented as a vector, x, for the emitted species,
i, as shown by
xi=xi,i=NO,CO,VOC.
Further, we define the true state of the emission scaling factors as
xt. The variability in ENO(t) is shown in
Fig. , and this variability is represented by
Ei(t)=eik(t),
where k(t) is the temporal variability emission factor for all of the
emitted species and ei is the time-independent emission for each
species. Note then that all of the anthropogenic emissions (NO, CO, and VOCs
– Ei(t)) share the same temporal variability. The variability in k(t)
is shown in Fig. as the “standard emission
variability”. Table shows the values of
eNO, eCO, and eVOC used in our model
simulations.
Values of the different parameters and emissions used in the
photochemical box model. The emissions are shown with the corresponding units
of molecules m-2 s-1. Since k(t)‾ is 1.89, the
average emissions, E(t)‾, are a factor of 1.89 larger than
ei. For E(t)‾NO, the value shown outside the
brackets is equivalent to xNO=1, and the values in the
brackets (same units) denote the range in the emissions that arise from using
the full range of xNO (0.5–2.5).
Model
variable
Parameter or emission value
k(t)‾
1.89
xNO
0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0, 2.25, 2.5
eNO
4.8 × 1010 molecules m-2 s-1
eCO
2.6 × 1012 molecules m-2 s-1
eVOC
4.3 × 1010 molecules m-2 s-1
E(t)‾NO
9 × 1010 molecules m-2 s-1 (4.5 × 1010–2.3 × 1011)
E(t)‾CO
5 × 1012 molecules m-2 s-1
E(t)‾VOC
8.2 × 1010 molecules m-2 s-1
Simulated range in peak NOx mixing ratios that result from the
different photochemical scenarios using different xNO
(0.5–2.5). Also shown are the ranges of peak CO and HCHO that result from
emissions of CO and VOCs, respectively.
Chemical species
Modelled peak mixing ratio range
NOx
4.0–24.0 ppbv
NO
1–11.3 ppbv
NO2
3–16.9 ppbv
CO
590–820 ppbv
HCHO
6.5–8.1 ppbv
In the emission inversion calculations, we represent VOC emissions via ethene
emissions. We selected ethene because it is a sufficiently reactive gas that
is emitted in abundance through the course of anthropogenic activity. Thus,
the adjoint sensitivities to ethene emissions are sufficiently high to allow
the 4D-var system to find adequate solutions for the VOC emission parameter.
Table describes the set-up of the photochemical
model for the range of different NO emission scenarios that we investigate
and shows the values of k(t)‾, and, for each species, e and
E(t)‾. Note that for E(t)‾ the overbar indicates the
mean value of a variable.
The NO emission scalings shown in Table are chosen
to represent a wide range of photochemical conditions and given the VOC
burden in the model, xNO emission scalings 0.5, 0.75 and 1.0
represent NOx-limited conditions, 1.25, 1.5 and 1.75 represent
transitional conditions, and 2.0, 2.25, and 2.5 represent VOC-limited
conditions. The mixing ratios of NOx that result from these different NO
emission factors, and the mixing ratios of CO and HCHO that result from the
CO and VOC emissions are all summarised in
Table .
A schematic showing how both the a priori and a posteriori emissions
relate to the true emissions of NO and the modelled peak afternoon ozone that
results from these emission variabilities. Note that the same emission
variability is used for all of the anthropogenic chemical species emitted in
the model. The a priori and a posteriori emissions are scaled relative to the
true emissions, and these differences can be characterised as being due to
different emission scaling factors (i.e. xNO) for the a priori,
a posteriori and true emissions. The solid black, green dashed and red dashed
lines show the truth, a posteriori, and a priori emissions,
respectively.
Forecasting framework and 4D-variational data assimilation
Several NOx emissions scenarios are simulated to cover a wide range of
photochemical conditions (xNO = 0.5–2.5). Each emission
scenario is represented mathematically as a forward model,
F(x,t), which represents the concentrations as a function of
time-evaluated emissions, x. Depending on the scenario, either
pseudo-observations of CO, NO2, O3, or HCHO are used in various
combinations (see Fig. for a representation of the
ozone pseudo-observations relative to the true state for ozone). In order to
derive the pseudo-observations the model true state is sampled at 3-hourly
intervals in the standard scenarios (used as default unless specified) and at
intervals between 1 and 24 h in scenarios characterising the impact of
observing frequency on prediction error. The sampled species concentrations
are then combined with an additive-noise model to generate the
pseudo-observations, y, represented by
y=F(x,t)+n,
where n is the noise
n=F(x)‾×β×ϵ,
where F(x)‾ is the average species concentration
(values shown in Table ), β is the noise scaling
factor, and ϵ is a random number with a normalised Gaussian
distribution with a standard deviation of 1 and a mean of 0. The modelled
concentrations for all species and times resulting from F(x) can
be represented as a vector, q,
q=F(x,t)
or for specific species, z, at time t as qz(x,t),
qz(x,t)=F(x,t)z,
where z can be O3, NO2, CO or HCHO. We define a priori emission
scaling factors, xa, with specified errors relative to
xt (Table provides a summary of the values of x used
for both xt and xa), which are combined with the
model to yield the a priori model state, F(xa). Note
that within our framework the a priori is also the initial guess.
Values of F(x)‾ used to calculate
y. The overbar indicates that this represents the mean value.
F(x^)
Mixing ratio
Ozone
44.4 ppbv
CO
620 ppbv
NO2
6.5 ppbv
HCHO
3.9 ppbv
Values of x and xa (in terms of unitless
emission scaling factor) used in the 4D-variational data assimilation model.
x
xa
NO
CO
VOC
NO
CO
VOC
0.5
1.0
6.5
0.475
0.95
0.1
0.75
–
–
0.7125
–
–
1.0
–
–
0.95
–
–
1.25
–
–
1.1875
–
–
1.5
–
–
1.425
–
–
1.75
–
–
1.8375
–
–
2.0
–
–
2.1
–
–
2.25
–
–
2.3625
–
–
2.5
–
–
2.625
–
–
The assimilation is started at the first iteration with the forward model
using the initial guess and is thus described as
F(xa) after one iteration. A cost function,
which is a scalar, J(x), is then evaluated:
J(x)=12((y-F(x))TSn-1(y-F(x))+12(x-xa)TSa-1(x-xa)),
where Sa is the a priori constraint matrix and
Sn is the observation error covariance (where the
superscript T indicates the transpose). The 4D-variational data assimilation
method seeks the solution for x, x^, that minimises
J(x),
x^=minxJ(x),
such that the gradient of the cost function with respect to x is 0 if
the solution x^ is equal to the true state, xt, (though
this is never fully achieved):
∇xJ=KTSn-1(y-F(x^))-Sa-1(x^-xa)=0,
where K is the Jacobian matrix (see Eq. ) describing the
forward-model response to perturbations to the emission parameters and
∇xJ is the adjoint
sensitivity , which was calculated by the
Rosenbrock solver and which indicates the sensitivity of
the cost function to the emission parameters. The cost function and its
adjoint sensitivities are passed to the quasi-Newton L-BFGS
algorithm . The L-BFGS algorithm iteratively determines the
optimal state of x, x^, that minimises the difference
between the model and observations subject to the a priori constraints.
A representation of the ozone prototype forecasting framework and
the 4D-variational data assimilation results for scenario OCN with
β=0.1. The observation period covers the first 48 h period of the
assimilation, during which time pseudo-observations are made (at a frequency
of every 3 h in this case) and are used within the assimilation. The
observations are used to constrain the emissions of ozone precursors, which
in turn allows the forecasting model to produce the a posteriori ozone
prediction. During the prediction and monitoring period the model true state
now plays the monitoring role allowing comparisons to be made to the ozone
forecast. The a posteriori ozone prediction represents the forecast for ozone
concentrations 1 day in the future. D represents the a posteriori
prediction model error and G represents the a priori and initial-guess
prediction error. The black solid line, red solid line, green dashed line,
and blue diamonds represent the truth, a priori, a posteriori, and
pseudo-observations, respectively.
Using the estimated emissions, x^, the forward model,
F(x^), provides the air quality prediction of the ozone
concentration, qO3(x,t), on the afternoon of the third
day of the simulation during the prediction and monitoring period. The
relevance of qO3(x,t) to the prediction and monitoring
period is shown in Fig. .
Figure shows how the a priori emissions,
xa, relate to the true emissions xt and the a
posteriori emissions, x^, after the 4D-variational data
assimilation, as well as the a priori, the true and the a posteriori ozone
levels (i.e. qO3(xa,t),
qO3(xt,t), and qO3(x^,t),
respectively). The left panel of Fig. shows the
a priori emission error for NO emissions and the right panel shows the a
posteriori NO emission error. The a posteriori emission parameter error can
be defined more generally as a vector x̃.
x̃=x^-xt
Figure provides an example representation of the
pseudo-observation ozone prediction, qO3(x^,t),
relative to the true state, qO3(xt,t), during the
prediction and monitoring period on the third day. In
Fig. D represents the a posteriori ozone
prediction error at time tμ (tμ is 15:00 LT on day 3 during the prediction and monitoring period),
defined by
D=qO3(x^,tμ)-qO3(xt,tμ).
In Fig. G represents the a priori ozone
prediction error defined by
G=qO3(xa,tμ)-qO3(xt,tμ).
The air quality prediction error over the entire prediction and monitoring
period for each of the species, z, can be defined as a vector,
q̃:
q̃zj=qz(x^,tj)-qz(xt,tj),j=3,6…21,24,
where j is the hour of day on the third day during the prediction and
monitoring period.
Uncertainty analysis
Overview
The uncertainty analysis has two foci: the evaluation of the performance of
the emissions estimates and an estimation of the a posteriori ozone
prediction error. Note that there is a direct synergy between these two
analyses since uncertainties in the emissions estimate directly impact upon
ozone prediction uncertainty. The diagnostics that we calculate in the
analysis of the emissions uncertainties include the a posteriori emission
parameter error, the emission averaging kernel matrix, and the emission
inversion degrees of freedom of signal.
The Jacobian matrix
The Jacobian matrix can be used to help characterise the variance in
x̃ and q̃. Therefore, it is advantageous to
determine K. Within our framework, each element of K
represents the forward-model response, ∂qz(x,t)/∂xi, at time t and for observed species, z, to perturbations in
emissions of species, i, in the case of the OCN scenario (using
pseudo-observations of ozone, CO, and NO2). It is defined by
K=∂qO3(x,t1)/∂xNO∂qO3(x,t1)/∂xCO∂qO3(x,t1)/∂xVOC∂qO3(x,t2)/∂xNO∂qO3(x,t2)/∂xCO∂qO3(x,t2)/∂xVOC.........∂qO3(x,tNt)/∂xNO∂qO3(x,tNt)/∂xCO∂qO3(x,tNt)/∂xVOC∂qCO(x,t1)/∂xNO∂qCO(x,t1)/∂xCO∂qCO(x,t1)/∂xVOC∂qCO(x,t2)/∂xNO∂qCO(x,t2)/∂xCO∂qCO(x,t2)/∂xVOC.........∂qCO(x,tNt)/∂xNO∂qCO(x,tNt)/∂xCO∂qCO(x,tNt)/∂xVOC∂qNO2(x,t1)/∂xNO∂qNO2(x,t1)/∂xCO∂qNO2(x,t1)/∂xVOC∂qNO2(x,t2)/∂xNO∂qNO2(x,t2)/∂xCO∂qNO2(x,t2)/∂xVOC.........∂qNO2(x,tNt)/∂xNO∂qNO2(x,tNt)/∂xCO∂qNO2(x,tNt)/∂xVOC=∂F(x,t)∂x,
where K has dimensions Ni×N. Ni is the number of
species in the emission factor state vector, x, and is thus always 3.
We define N as the total number of observations for all species:
N=Nt×Ny,
where Nt is the number of points in time at which the model
perturbations are sampled and Ny is the number of species whose
perturbations are used in the Jacobian. In the case of Eq. ()
y = O3, CO and NO2; therefore, Ny = 3. y
includes HCHO in the HCN scenario.
These plots show the columns of the Jacobian matrix, K,
that correspond to the perturbations of the three observed species in
scenario OCN. Ozone is shown on the left, CO in the middle, and NO2 on
the right. This Jacobian is for the xNO=1.25 emission
scenario. The shaded area represents observations made during the night.
NO2 observations made using visible remote-sensing instruments can only
function during the daytime, so there is no need to include a row in the
Jacobian corresponding to night-time NO2 observations. The blue, red,
and green solid lines represent qZ(x,t)/∂xNO,
qZ(x,t)/∂xCO, and qZ(x,t)/∂xVOC, respectively. The y axes on the left and right represent
the different perturbations to x.
Figure plots columns of the Jacobian, and it shows that
ozone is more sensitive to changes in emissions during the afternoon and that
CO and NO2 respond to changes in emissions during the rush hour periods.
The key assumption in using the Jacobian is that changes in the emissions can
be described (see ) approximately by
F(x)-F(x+δx)≈Kδx.
This assumption has been validated using finite differencing (results not
shown) to compare to solutions derived from the right side of
Eq. ().
Emission error characterisation
We calculate various statistics to determine the emission estimation
performance. First, we determine the a posteriori emission parameter error
covariance, which is defined (see ) by
Ex̃x̃T=(Sa-1+KTSn-1K)-1.
Next, we calculate the emission averaging kernel defined by
A=(Sa-1+KTSn-1K)-1KTSn-1K
and the degrees of freedom of signal that is calculated via
d.o.f.=Tr(A),
where both of these diagnostics provide information on the resolution of the
emission retrieval, i.e. the ability of the estimate to uniquely distinguish
between the emissions of individual species. The notation
Tr(A) indicates the trace of a matrix. While the diagonals
of A represent the sensitivity of x^i to xi, the
d.o.f. represents the number of separate emission parameters that can be
uniquely retrieved.
Ozone a posteriori prediction errors across the complete range of
parameter space for xNO (0.5–2.5) on the x axis and β
(0.1–5) along the y axis with each panel presenting the results from the
three observing scenarios CN, OCN and HCN. The coloured contours represent
the a posteriori prediction error in units of ppbv. The green and red colours
indicate low and high levels of a posteriori ozone prediction error,
respectively.
Ozone prediction error characterisation
Using the a posteriori emission error, we can determine the a posteriori
ozone prediction error during the prediction period. In order to do this we
need to define a new Jacobian matrix, K′, that defines the forward
photochemical response during the prediction and monitoring period (day 3) to
perturbations in the emissions. Thus, K and K′ simply
differ because K describes the model response during the
observation period as opposed to the prediction and monitoring period. Each
element of K′ is ∂qz(x,tj)/∂xi, where j is the index of time denoting when the
model is sampled on the third day. The a posteriori ozone prediction error
covariance for the third day can be determined by
Eq̃q̃T=K′Ex̃x̃TK′T.
List and details of all of the experiments carried out as part of
the uncertainty analysis. The experiment details include the observed
species, xNO emission factors (see
Table for the full list), the observation noise,
β, and the observing frequency. The eight different values of β
are 0.01, 0.05 , 0.1, 0.25, 0.5, 1.0, 2.5, and 5.0. These fractional errors
are relative to the average species mixing ratios over all of the
photochemical scenarios (see Table ). The observing noises
are identical for each compound within a particular scenario unless otherwise
stated. All of the results from these experiments are described in Sect. 3.1.
We also include short notes describing other aspects of the experiments. The
table includes a list of the precise sections where the different experiments
are discussed.
Experiment
Section
Observed
xNO
Observation
Observing
Notes
species
scenarios
noise (β)
frequency
CN
First and third subsection of Sect. 3.1.1
CO and NO2
Nine xNO scenarios (0.5–2.5)
Eight β values (0.01–5.0)
3 h
OCN
First and third subsection of Sect. 3.1.1
Ozone, CO and NO2
Nine xNO scenarios(0.5–2.5)
Eight β values (0.01–5.0)
3 h
HCN
First and third subsection of Sect. 3.1.1
HCHO, CO and NO2
Nine xNO scenarios (0.5–2.5)
Eight β values(0.01–5.0)
3 h
HOCN
First subsection ofSect. 3.1.1
HCHO, ozone, CO and NO2
Nine xNO scenarios(0.5–2.5)
Eight β values(0.01–5.0)
3 h
Results not shown in any figure
Comparison betweenHCN and OCN(EHCN–EOCN)
Second subsection of Sect. 3.1.1
HCHO, ozone, CO and NO2
Nine xNO scenarios (0.5–2.5)
Eight β values(0.01–5.0)
3 h
Three different scenarios tested each using different HCHO observation noise
Observing frequencyexperiment
Sect. 3.1.2
Ozone, CO and NO2
Nine xNO scenarios (0.5–2.5)
β=0.25
Six frequencies tested: 1, 3, 6, 12, 18, and 24 h
Observing timeexperiment
Sect. 3.1.2
Ozone, CO and NO2
Nine xNO scenarios(0.5–2.5)
β=0.25
3 h
16 different scenarios tested; observations are removed at different times in each case
Summary of experiments
We describe all of the experiments that we perform for the uncertainty
analysis (Sect. 3.1) in Table . In each experiment we
test a range of different observation characteristics using different
parameters. To give an example, for the CN observing scenario we test the
model forecast uncertainties across the nine values of xNO
(i.e. 0.5–2.5 with increments of 0.25) and for eight different levels of
observing error (β=0.01–5; equivalent to 1, 5, 10, 25, 50, 100, 250,
and 500 % relative error). Thus, we perform 72 separate tests for this
experiment and for the OCN and HCN scenarios as well. However, for the
experiment comparing HCN and OCN we carry out three separate tests where we
scale HCHO observation noise relative to the other species. We test three
different scalings: 50 % lower, the same, and 50 % higher noise.
Section 3.2 is dedicated to sensitivity studies using the full 4D-var data
assimilation forecast system. In Sect. 3.2.1 we demonstrate the ability of
the 4D-var data assimilation forecast system to forecast ozone when using the
three observation scenarios CN, OCN, and HCN. For these experiments we use
observations made at 3 h intervals and using β=0.1.
Next, in Sect. 3.2.2, we define a range of different k(t) scenarios in
order to probe the emission solution and ozone forecast sensitivity to the
assumed diurnal emission variability. These alternative k(t) scenarios and
the standard emission variability are shown in
Fig. . In each test we perform the 4D-var data
assimilation forecast using the alternative k(t) scenario while still
assuming that the standard emission variability is representative of the true
state. We perform this test using the OCN scenario, observing at 3 h
intervals and with β=0.1.
When conducting the VOC emission inversion, we represent VOC emission
uncertainties as ethene emission uncertainties (rather than a more diverse
range of VOCs). In Sect. 3.2.3 we test that assumption using a sensitivity
analysis by assuming VOC emission errors for ethane instead of ethene. Again,
we perform this test for the OCN scenario, observing at a 3 h frequency and
with β=0.1.
Discussion and conclusions
We addressed a set of key questions to determine how characteristics of
observations of ozone and its precursors affect one's ability to constrain
ozone precursor emissions and consequently to predict ozone when using an
idealised prognostic air quality model coupled to a data assimilation
framework. These questions consisted of which species to observe, how well to
observe them, how often to make observations, when to make them during the
diurnal cycle, and how long to observe before making a prediction. Further to
this, we were interested in how the answers to these questions changed
according to varying photochemical regimes (from NOx- to
VOC-limited conditions for ozone formation). These questions are relevant
to determining, in a very coarse way, how the various observing platforms
(e.g. LEO and GEO satellites)
and ground monitoring networks are able to support air quality research and forecasting.
We used a framework consisting of a photochemical box model
using idealised meteorology, its adjoint, and a 4D-variational data assimilation system set-up
to constrain ozone precursor emission uncertainties (NOx, CO, and VOCs). The photochemical
box model used idealised meteorology that represented stagnant summer weather conditions.
Using linear analysis to assess the framework's prediction uncertainties, we carried out a series
of sensitivity analyses to test the performance of the forecasting framework under a range of different
observing scenarios. This consisted of using various sets of pseudo-observations. We examined the effect
of changing which four species were observed (CO, NO2
and HCHO, CO, and NO2), of varying the observation noise, of changing the observing frequency,
and of changing the time during the day when observations are made.
We were able to demonstrate that the 4D-var framework was able to constrain
ozone precursor emissions and consequently that it was able to reduce ozone
prediction uncertainties and provide an adequate ozone forecast under the
idealised conditions that we used. This therefore demonstrated our
framework's relevance to future air quality forecasting systems that might
utilise state of the art assimilation and observations made using either the
ground station network or from orbiting satellites. Clearly, more
difficulties and challenges remain before such a framework can be used in a
real-world setting, such as how to incorporate averaging kernels of satellite
retrievals into the assimilation system or accounting for representativity
errors. Also, using the linear analysis to estimate the prediction
uncertainties, we were able to derive a series of general conclusions that
are discussed below.
The effect of changing the observed species
Our results show that the variability in ozone prediction error with both
photochemical regime and observing species scenario (CN, OCN and HCN) is
complex and no single observed species is ideal for all photochemical
conditions.
Under NOx-limited conditions ozone prediction error is strongly
controlled by the a posteriori NO emission errors, and therefore observations
of NO2 and ozone would be highly advantageous. Ozone provides a
particularly good constraint upon NO emissions under very NOx-limited
and VOC-limited conditions. The value of NO2 observations in
constraining NO emissions improves as the NOx lifetime increases under
the somewhat less NOx-limited conditions (xNO=1.0–1.25). Much of the troposphere is in fact highly NOx-limited
outside of the most polluted areas .
Under VOC-limited conditions ozone prediction error is sensitive to both a
posteriori xNO (due to the anticorrelation of ozone to
NOx) and xVOC errors, and thus observations of ozone, HCHO
and NO2 allow significant improvements in ozone prediction error.
Assimilating ozone, therefore, allows constraints to be placed upon VOC and
NO emission uncertainties. HCHO provides an excellent constraint upon
reactive VOC emissions, which due to their reactivity are more relevant to
air quality compared to less reactive VOCs. NO2 provides an excellent
constraint upon NO emissions under VOC-limited conditions; more than under
NOx-limited conditions due to the longer NOx lifetime. Despite the
fact that large geographical portions of the US are NOx-limited, a
disproportionately large percentage of the population lives within or is
exposed to ozone arising from VOC-limited conditions due to the significant
extent of urbanisation within the US. Large urbanised areas of the south-west
of the US that lack significant native vegetative biomass typically have a
larger VOC-limited regime that extends over the urban as well as suburban
areas. In contrast, US cities in the east are located in regions with often
dense vegetative biomass, e.g. Atlanta, and thus the VOC-limited region is
far more geographically limited to the urban centre itself. Therefore,
improving ozone predictive skill within VOC-limited conditions will not yield
forecasting improvements over a wide geographical area but will yield
improvements within certain regions with large populations.
Our findings with respect to the utility of NO2 and HCHO observations
for constraining NOx and VOC emissions, respectively, and in turn for
improving ozone estimation are broadly consistent with the findings of
, who used satellite observations of NO2 and HCHO in
conjunction with 4D-variational data assimilation to solve for NO2 and
HCHO emissions and to improve the model's ozone estimation. One should note,
however, that our work goes further by demonstrating how the efficacy of
NO2 and HCHO observations varies according to photochemical regime.
Similar to , we demonstrate the use of ozone in
this regard. Our work offers an extension to by
considering the photochemical regime and by considering other observations
simultaneously.
Note that the statements above regarding the need to constrain NO and VOC
emissions under NOx- and VOC-limited conditions, respectively, are
consistent with expectations since ozone is more sensitive to both sets of
emission uncertainties under the respective conditions. Further, the use of
ozone to constrain either NOx or VOC emissions in either of the
respective photochemical regimes is fully consistent with existing theory
relating to ozone control strategies and our
understanding of factors controlling ozone on regional and continental
scales . This was one motivation for us to explore this
problem.
There is one further advantage to observations of ozone and HCHO made under
VOC-limited conditions. Often, plumes of NOx-polluted and VOC-limited
air can be exported from regions that are VOC limited into areas that are
NOx limited, and this can lead to significant temporal variability in
the photochemical regime in the regions surrounding an urban centre.
Therefore, observations of HCHO and ozone in addition to NO2
observations could help to understand such events and in turn reduce ozone
prediction errors.
We have indirectly performed a sensitivity test to see if CO observations
affect ozone a posteriori prediction errors. We can address their potential
impact within the OCN scenario by examining the Jacobian matrix (see
Fig. ). This shows that ozone is relatively insensitive to
perturbations in CO emissions and, therefore, also to a posteriori CO
emission uncertainties. In fact, it appears that only the β=5.0 noise
scenario has sufficiently large a posteriori CO emission error to cause
significant a posteriori ozone prediction error (about 5 ppbv). The Jacobian
predicts perturbations in CO associated with such emission error to be over
700 ppbv. Such large changes in CO mixing ratios can occur in reality in
urban areas from the influence of wildfires. For instance, CO mixing ratios
were as high as 10 ppmv during the summer of 2010 as a direct result of the
rare and extreme fire events occurring in Russia that
summer . Episodic perturbations of
only ∼ 700 ppbv are therefore
more likely to result from the more frequent and less severe wildfire events
that occur within Europe on an annual basis.
Observation error
We now make some broad conclusions regarding the observation uncertainties.
Both the OCN and standard HCN scenarios achieve a posteriori ozone prediction
errors of 2.4–6.1 and 1.9–6.3 ppbv, respectively, when absolute errors
equivalent to 33 % of the average over polluted regions were used. Even
though the OCN and HCN scenarios compared favourably to one another in terms
of their a posteriori ozone prediction errors, when we considered more
realistic observational noise on the HCHO observations, the performance of
the HCN scenario was degraded to 2.2–6.9 ppbv (33 % noise level). In
comparison, for the same noise level, the CN scenario achieved ozone
prediction errors of 2.5–8.4 ppbv. Only when the noise level was reduced to
25 % were the OCN and HCN scenarios able to achieve ozone prediction errors
of 5 ppbv or less. At 10 % noise, ozone prediction errors of less than
2.5 ppbv were consistently attained for both OCN and HCN. This strongly
points towards there being a good payoff in forecast accuracy with reducing
observation error. Further work in a 3-D framework would be required in order
to determine how these ozone forecast errors translate into the context of
real air quality forecasting. For instance, it might be possible to calculate
the probability of detection or false-alarm rate statistics in a similar way
to the work carried out by .
Connecting this to real instrument profiles and real observations, and how
these might perform in a real assimilation system, is beyond the scope of
this study. The furthest we can take this point is to note that the resulting
prediction uncertainties for a particular observation noise scenario are
optimistic and represent the lowest error that could be expected. This is
because of reduced complexity in our model's representation of its spatial
domain and its meteorology and because of the way we represented the errors
in our observations, which in reality would be more complex.
Temporal considerations
Concerning the temporal sampling of observations, there is a strong
sensitivity of ozone prediction error to observation removal in the daytime,
particularly in the afternoon, and therefore observations made during the day
present greater returns in terms of improved forecasting ability. The
NOx-limited regimes favour observations made throughout the day with
increased observing density close to 15:00 LT The VOC-limited regimes favour
a greater concentration of observations within the afternoon even up to
18:00 LT in the most VOC-limited cases. These differing results for the two
different photochemical regimes are consistent with existing knowledge about
photochemistry and NOx lifetime. The main underlying factors controlling
this are the changing time at which ozone peaks and the time of day that
emissions occur that contribute to that peak. Under VOC-limited conditions
ozone peaks later in the day due to the reduced ozone lifetime and the slower
recovery of HOx radicals (suppressed by NOx) that occurs after the
night-time period. The NOx-limited scenarios also show a smaller peak in
the morning. This smaller peak is present due to the observations of ozone
and NO2 during the morning rush hour that better allow NOx
emissions to be constrained. The presence of the smaller peak also indicates
that peak afternoon ozone concentrations are sensitive to the morning rush
hour emissions of NOx; this is possible due to the longer ozone lifetime
present under NOx-limited conditions.
We demonstrate that the ozone prediction error is sensitive to the frequency
of observation. We show that ozone prediction errors vary between negligible
values and up to 12.5 ppbv as the observing frequency varies between once
per hour to once per day, respectively. The ozone prediction error is
maximised within either the NOx-limited or VOC-limited regimes.
We find very similar levels of ozone prediction error for the scenarios that
observe once every hour and every 3 h (1.8–3.2 ppbv compared to
2.2–4.8 ppbv, respectively), and we also find that ozone prediction errors
greater than 5 ppbv only emerge for observing scenarios using a frequency of
6 h or more. The fact that our forecasting system performs best using
observations made at a frequency of 3 h or less highlights the temporal
sampling advantage posed by the ground observation network relative to
observing systems with lower observing frequency, i.e. a satellite in LEO
configuration.
It is likely that there is an effect on ozone prediction error due to the
interaction between observing frequency and observing time.
Figure implies that observing scenarios measuring at
the same frequency could yield different prediction errors due to when they
actually sampled during the diurnal cycle. However, in each test we made at a
particular observing frequency, the observations were made at a fixed
specific set of times, and so our work does not address this issue. We do
think that this is relevant to evaluating different types of observing
scenario, and we would therefore like to explore this problem in a future
paper.
Implications for emission inversion
Aside from the relevance of these results to air quality forecasting and
research in general, we believe these results are also relevant for emission
and flux estimation via inversion methodologies. Our prototype framework is
very similar to other work using 4D-variational data assimilation
methodologies
and
chemistry transport models that have focused on emission inversion. Much of
the emission inversion performance shown in this study is driven by the
photochemistry, and it is reasonable to suppose that some of our results are
relevant to future work conducted using 4D-variational data assimilation in
emission inversion studies. Note too that Kalman filter methods can also be
used in this application and we should expect that the performance of this
method will be similarly affected by photochemistry. From this premise, we
recommend that emission inversion studies for NOx utilise both
observations of NO2 and ozone since ozone observations add information
to the xNO estimation under both strongly positively and
negatively NOx-limited conditions and NO2 observations constrain
emission parameter uncertainties the most under the more VOC-limited
conditions. Thus, these two observations are complementary to each other.
Likewise, for emission inversions of VOCs, we recommend observations of HCHO
and ozone since HCHO observations can constrain VOC emission uncertainties
under a wide variety of photochemical conditions and ozone can constrain VOC
emission uncertainties under VOC-limited conditions.
Previous studies have shown that NO2
and
HCHO
observations can constrain NOx and VOC emissions, respectively. Although
one could have inferred that combining ozone observations with either
NO2 or HCHO observations would be beneficial, we have shown that it
could be highly advantageous, which is consistent with .
It should be noted that the conclusions regarding VOC emission inversion are
sensitive to our choice of representing VOC emission uncertainties with
ethene. The success of the VOC emission inversion is significantly limited by
solving for ethane instead of ethene emission uncertainties. This is due to
the lack of impact on secondary
chemical species such as HCHO. This is one reason why previous emission inversion modelling
studies have focused on constraining reactive VOCs like
isoprene .
Concerning CO, all of the observing scenarios (CN, OCN, and HCN) performed
equally well at constraining CO emission uncertainties since all these
scenarios included observations of CO. The Jacobian for CO with respect to CO
emission perturbations shown in Fig. clearly shows a
strong sensitivity of CO to changes in its own emissions. On the other hand,
Fig. shows much lower sensitivity of CO to the emissions
of NO or VOCs. These results are fully consistent with expectations due to
the relatively low reactivity of CO and its potential to produce ozone on
short timescales and of the lack of a strong chemical connection between
NOx levels and resulting CO concentrations. In the latter case, there is
a link due to the way that NOx can perturb OH, but due to the relative
unreactivity of CO this leads to only weak sensitivity in the Jacobian.
Consistent with this, there have already been several studies that use
observations of CO to constrain CO
emissions .
In the supporting sensitivity analysis probing emission solution sensitivity
to diurnal emission variability we demonstrate that emission inversions are
potentially highly sensitive to the assumed variability in the emissions and
that even perfect observations would lead to such errors. In our system such
emission inversion errors would be hard to characterise in the absence of any
information regarding the true state of the emissions variability. We
recommend that such uncertainties should be considered and characterised in
emissions inversion studies. Currently diurnal emission variabilities are
determined in the process of building bottom-up emission inventories.
Although our prototype assimilation system can only currently solve for
time-independent scaling factors, it could be modified to solve for
time-dependent scaling factors and the diurnal emissions variability. Future
assimilation forecasting systems should also possess this ability to solve
for time-dependent emission scaling factors. Observations that adequately
capture the diurnal variability in pollutants will be essential to making
this leap from time-independent solutions to time-dependent solutions.
Implications for GEO and LEO satellites
In the previous sections we have motivated the potential utility of surface
or boundary layer ozone, CO, NO2, and HCHO observations either in the
context of improving ozone forecasting or for emission inversions. Ground
station networks that implicitly sample boundary layer air are already in
place across the American and European continents. However, only one of the
current generation of LEO satellite instruments (MOPITT) possesses a reliable
means of attaining unique instrument sensitivity to the boundary layer for
these gases . If full advantage is to be taken of future
GEO stationary satellite instruments' (GEO-CAPE/TEMPO, GEMS, and Sentinel-4)
simultaneous potential for excellent coverage and temporal sampling, with the
aim of fully contributing to state of the art ozone air quality forecasting,
then attaining sensitivity to the boundary layer is essential and should be a
high-priority aim.
The heightened importance of observations made during the morning and mid- to
late afternoon raises the question of whether making more targeted
observations, for instance during the morning and evening rush hours, would
be able to support ozone forecasting even further. There are various
observing systems that would be able to provide this capability, such as
several combined LEO missions or ground stations or a GEO mission with
increased temporal sampling capability during those periods. Investigating
these questions in the future would be of interest to us and the broader
scientific community.
Our forecasting system is better able to improve the ozone prediction using
observations made during the day as opposed to the night. In the context of
satellites, and remembering that our idealised case ignores the effects of
transport, this indicates that instruments capable of observing during the
night, such as those observing in the TIR, do not offer a significant
advantage over instruments restricted to making measurements during the
daytime. Of course, if the effects of transported pollution were to be
considered, such as the night-time mixing of ozone between the boundary layer
and free troposphere, then making observations during the night could offer
additional utility by improving the estimated contribution to the pollution
made by this process. For instance, this could provide advance warning of the
trajectory of a pollution plume. Therefore, a limitation of this work is that
we are not able to explore such effects using a model with only idealised
meteorology.
Our forecasting system (and the emission inversion) performs best using
observations made at a frequency of 3 h or less. This highlights the
temporal sampling advantage posed by satellites in a GEO configuration as
opposed to those in LEO. Currently, the proposed observing frequencies for
the future GEO missions and the current ground monitoring
network are at least 1 h. LEO satellites, on the other hand, cannot attain
high-frequency sampling without a large number of satellites being
employed . In isolation, a single LEO satellite with a
sampling frequency of between 1 and 16 days is perhaps inadequate for the
purpose of constraining precursor emissions on the regional scale or for
supporting air quality forecasting. Another consideration is that observing
frequencies of 3 h or more might not be adequate for studying the diurnal
cycle of pollutants and for forecasting systems that use 3D-var, for
instance, to update ozone concentrations. Note that the nature of our
framework for performing these tests (i.e. a box model using only idealised
meteorology) places limitations on our conclusions such that the performance
of the higher-frequency observing scenarios (3 h or less) may be too
optimistic. Thus, observing at 3 h may be insufficient to constrain ozone
precursor emissions.