sensors

Article
Estimating CO2 Emissions from IoT Traffic Flow Sensors and
Reconstruction
Stefano Bilotta and Paolo Nesi *

DISIT Lab, Department of Information Engineering, University of Florence, 50139 Firenze, Italy;
[email protected]
* Correspondence: [email protected]

Abstract: CO2 emissions from burning fossil fuels make a relevant contribution to atmospheric
changes and climate disruptions. In cities, the contribution by traffic of CO2 is very relevant, and the
general CO2 estimation can be computed (i) on the basis of the fuel transformation in energy using
several factors and efficiency aspects of engines and (ii) by taking into account the weight moved,
distance, time, and emissions factor of each specific vehicle. Those approaches are unsuitable for
understanding the impact of vehicles on CO2 in cities since vehicles produce CO2 depending on
their specific efficiency, producer, fuel, weight, driver style, road conditions, seasons, etc. Thanks to
today’s technologies, it is possible to collect real-time traffic data to obtain useful information that
can be used to monitor changes in carbon emissions. The research presented in this paper studied
the cause of CO2 emissions in the air with respect to different traffic conditions. In particular, we
propose a model and approach to assess CO2 emissions on the basis of traffic flow data taking into
account uncongested and congested conditions. These traffic situations contribute differently to the
amount of CO2 in the atmosphere, providing a different emissions factor. The solution was validated
in urban conditions of Florence city, where the amount of CO2 is measured by sensors at a few points
where more than 100 traffic flow sensors are present (data accessible on the Snap4City platform). The
solution allowed for the estimation of CO2 from traffic flow, estimating the changes in the emissions
factor on the basis of the seasons and in terms of precision. The identified model and solution allowed
Citation: Bilotta, S.; Nesi, P.
the city’s distribution of CO2 to be computed.
Estimating CO2 Emissions from IoT
Traffic Flow Sensors and Keywords: smart city; vehicle CO2 emissions factor; traffic flow; reconstruction algorithm; traffic
Reconstruction. Sensors 2022, 22, 3382. congestion; regression CO2 model; seasonal changing
https://doi.org/10.3390/s22093382

Academic Editor: Reza Malekian

Received: 1 April 2022 1. Introduction

Accepted: 26 April 2022 Traffic assessment and control is a central topic for intelligent transportation systems
Published: 28 April 2022
(ITS). Thanks to today’s technologies, real-time data can be collected and used to monitor
Publisher’s Note: MDPI stays neutral and control vehicular traffic. The knowledge of real-time traffic information enables the
with regard to jurisdictional claims in development of a relevant number of services and improvements in many areas: congestion
published maps and institutional affil- detection and reduction, dynamic network traffic control, improved information services
iations. (e.g., traffic information, dynamic route guidance, road digital signage), planning for future
investments on mobility solutions, and reducing fuel consumption and pollution emissions.
Recently, regarding the latter, there has been a deeper understanding of the environmental
parameters (for example, PM10 (particulate matter), PM2.5, CO (carbon monoxide), CO2
Copyright: © 2022 by the authors.
(carbon dioxide), SO2 (sulphur dioxide), O3 (ozone), H2 S (hydrogen sulphide), NO (nitric
Licensee MDPI, Basel, Switzerland.
oxide), NO2 (nitric dioxide), and NOx (nitric monoxide and dioxide)) and how much they
This article is an open access article
are influenced by a city’s structures, what the causes of high values of pollutants are, the
distributed under the terms and
reasons for the registered high values from the IOT Data Network, and the dynamic of their
conditions of the Creative Commons
diffusion and propagation [1,2]. For more on this, see the normative of the European Com-
Attribution (CC BY) license (https://
mission regarding the conformant of the environmental value with respect to the reference
creativecommons.org/licenses/by/
4.0/).
values (2008/50/EC Directive on Ambient Air Quality and Cleaner Air for Europe and

Sensors 2022, 22, 3382. https://doi.org/10.3390/s22093382 https://www.mdpi.com/journal/sensors

Sensors 2022, 22, 3382 2 of 20

2004/107/EC Directive on Heavy Metals and Polycyclic Aromatic Hydrocarbons in Ambi-

ent Air). Greenhouse gas (GHG) emissions make a relevant contribution to atmospheric
changes and climate disruptions. Many gases contribute to global warming, and the CO2
emissions from burning fossil fuel represents the primary contribution in the transportation
field. Traffic emissions from fuel combusted in vehicles are typically estimated by multi-
plying activity data by emission factors (EFs) [3]. For example, the distance travelled by a
vehicle has a large influence on emissions, since in general more activity leads to greater
emissions. Vehicle speed is another important influence on emissions, because road traffic
EFs are strongly dependent on speed [4,5]. Emissions also depend on vehicle category.
Different vehicle categories have different EFs due to factors such as vehicle mass, fuel
specification, engine size, aerodynamics, and emissions control technology. One of the
most commonly used methodologies for fleet-wide emissions assessment is the emissions
inventorying methodology included in the EMEP/EEA guidebook [6]. A specific report
on car emissions of CO2 can be found in [7], where traffic emissions of CO2 is analysed,
identifying the typical emissions factors for different kinds of vehicles and their trends
over time. In that context, different emissions factors have been identified according to
fuel type, efficiency, producer, power, weight, etc. Therefore, almost all CO2 emissions
in vehicular traffic depend on a variety of vehicle- and traffic-related parameters, such as
vehicle characteristics and motorisation, driving behaviour, and traffic conditions. Thus,
most of the approaches for measuring CO2 are based on (i) energy models that take into
account the conversion of total consumed energy by vehicles, taking into account efficiency,
carbon content per unit of heat, carbon oxidation factor, and rate [8], or on (ii) transport
models that take into account the average distance, the emissions factor of the vehicles, and
the weight moved [7].

1.1. Literature Review

The present research studies the cause of CO2 emissions in the air, which are primarily
related to the impact of vehicular traffic emissions in uncongested and congested situations,
assuming that the population of the vehicle type depends on the city area but is stationary
over time. Roughly speaking, traffic congestion is the deterioration of smooth free-flowing
traffic conditions due to increased travel demand and/or reduced traffic movement capac-
ity [9]. It is commonly accepted that under the so-called “stop-and-go” traffic conditions
associated with congestion there is an increase in the number of accelerations and decelera-
tions performed by vehicles, which result in increased emissions [10,11]. The study in [12]
measured instantaneous speed, acceleration, travel time, fuel consumption, and exhaust
emissions in a field test to find that fuel consumption depends largely on traffic flow pat-
terns and traffic conditions. Fuel consumption under congested flow conditions is higher
than under free-flow conditions due to the frequency of operations such as acceleration,
deceleration, and “stop-and-go” driving. Various tools exist for calculating CO2 emissions
and simulating vehicle operations on a vehicle-specific level (micro-level). For example,
the European Union recently introduced a vehicle simulation model, CO2 MPAS, for CO2
certification purposes [13]. Other works combined VISSIM traffic simulation software
with different fuel consumption and emissions models to study vehicle fuel consumption
and emissions at a given simulated signalized intersection [14,15]. In [16], a model for
estimating the amount CO2 in grams (briefly gCO2 from traffic has been proposed. The
model assessed a mean emissions factor of 203 gCO2 /km per vehicle for an “urban slow”
pattern, for “stop-and-go A” pattern a mean emissions factor of 460 gCO2 /km per vehicle,
and for “stop-and-go B” an emissions factor of 738 gCO2 /km per vehicle. The driving
modalities assumed that when the number of cars is at a minimum there is no traffic
congestion and the urban slow pattern can be considered. When the traffic corresponds
to the mean observed value, the “stop-and-go A” model is applied, and when the traffic
reaches its maximum highest emissions factor, the “stop-and-go B” pattern can be used.
In [17], an approach to annual road traffic CO2 emissions estimation for an urban area with
street-level resolution was conducted. In the mentioned study, the CO2 emissions factors
Sensors 2022, 22, 3382 3 of 20

were computed by applying a TRL (Transport Research Lab) dataset [18] to the fleet vehicle
composition in an urban area, and the related modelling uncertainties, running from 6.5%
to 12%, focused on peak traffic periods, when significant emissions of CO2 occur.
In order to control and reduce emissions, national and international organisations have
defined guidelines and targeted limits to be respected currently, and to be progressively
reduced over the years/months. In this regard, the European Union has set limits for the
yearly mean value of the concentration of pollution. In [19], a model and tool to compute
long-term predictions was proposed, up to 180 days in advance, for the progressive mean
value of NO2 with a precision needed to enable decision-makers to perform corrections. In
this context, the TRAFAIR project (Understanding Traffic Flows to Improve Air Quality)
has been focused on the short- and mid-term prediction of NOx on the basis of traffic flow
emissions [20].

1.2. Paper Aim and Structure

In this paper, we present a model to estimate the CO2 emissions from traffic flow data,
characterising the specific city traffic flow in terms of emissions factor, which is regarded as
the amount of CO2 produced per vehicle for unit distance (gCO2 /km per car). As reported
in the above analysis, the literature on the state of the art estimates the CO2 in (i) controlled
experiments for each single vehicle type, as in [12–15], and (ii) a general approach on
the basis of the energy consumed (fuel) by taking into account of vehicle distribution, as
in [16,17]. Please note that a global measurement of CO2 could be performed by satellites,
but those measurements are affected by multiple problems, since they (1) cannot be precisely
attributed to a specific city area since they perform the measurement on the basis of the
column of air from the ground to satellite quote, and (2) are imprecise in the presence of
clouds, wind, air humidity, etc. The proposed method, unlike the approaches presented
in [12–17], allows the CO2 emissions to be estimated on city roads. In particular, the
proposed method:
• Allows the CO2 produced by vehicles in a city to be estimated, which is a relevant
contribution of CO2 produced in cities, directly measuring the impact of vehicle
population on the production of CO2 , thus increasing the precision in measuring CO2 .
The advantage of identifying the function from traffic flow to CO2 is that it can be
used to estimate the CO2 in city areas based on knowing the traffic flow with a certain
precision, and thus the total CO2 production of the city, which currently can only be
coarsely guessed by using a very limited number of CO2 sensors, whereas most cities
have hundreds of traffic flow sensors.
• Is based on (i) the identification of the relationships from the measured traffic flow
and determining the emissions factors, taking into account different traffic behaviours,
from fluid traffic to “stop-and-go” conditions (congested and uncongested traffic
situations); (ii) the changes in the emissions factor at different periods of the year;
(iii) an approach to statistical validation by means of the CO2 measurements taken
from specific sensors and traffic flow data, allowing for an assessment of the precision
of the indirect estimation of CO2 on the basis of traffic flow.
• Provides a solution for computing CO2 emissions locally in the city and not only for
specific vehicles or globally at city level, as mentioned above in the literature review.
Moreover, in the validation approach, a number of problems have been solved, such
as (a) the positions of traffic flow sensors typically not being aligned with those of air
quality, (b) the measurements of CO2 not being temporally aligned with those of traffic flow
sensors, and (c) different areas possibly impacting CO2 in a different manner depending
on the city. The model presented in this paper and the related results were produced and
validated by exploiting the Snap4City framework for smart cities, mobility and transport,
and data analytics, also using Km4City/Sii-Mobility models and tools. The research
was funded by the National Ministry of Research [21]. The algorithms were executed by
exploiting the Snap4City data aggregator and its semantic model [22,23] and the Snap4city
platform [24,25].
 research was funded by the National Ministry of Research [21]. The algorithms were exe-
cuted by exploiting the Snap4City data aggregator and its semantic model [22,23] and the
Sensors 2022, 22, 3382 4 of 20
Snap4city platform [24,25].
The paper is structured as follows and is depicted in Figure 1, which describes the
data flows and algorithms put in place and the corresponding sections of the paper. In
The paper is structured as follows and is depicted in Figure 1, which describes the data
Section
flows and 2, algorithms
we presentput theindata
placedescription and the related
and the corresponding sections problems. In Section
of the paper. In Section3, 2,
the anal-
ysis of traffic flow data is reported to pose the basis to identify
we present the data description and the related problems. In Section 3, the analysis of the traffic flow modalities
and needed
traffic normalisations.
flow data is reported toIn Section
pose 4, thetomodel
the basis identify forthe
thetraffic
identification of congested
flow modalities and and
uncongested traffic flows
needed normalisations. is presented.
In Section In Section
4, the model for the5, identification
the problemsofrelated to the
congested anddifferent
uncongested
sensor traffic
locations forflows is presented.
measuring CO2 InandSection
traffic5, flow
the problems
data area related to the different
addressed, presenting the
sensor flow
traffic locations for measuring
reconstruction CO2 andto
algorithm traffic flow traffic
estimate data area addressed,
flow far frompresenting
the sensorthe locations.
traffic flow reconstruction algorithm to estimate traffic flow far from
Section 6 presents the model approach to compute the amount of CO2 in the air on the the sensor locations.
Section 6 presents the model approach to compute the amount of CO2 in the air on the basis
basis of the measured traffic flow in uncongested and congested situations in terms of
of the measured traffic flow in uncongested and congested situations in terms of emission
emission
factors. Thefactors. The experimental
experimental results6.4)
results (see Section (see Section 6.4)
demonstrate (i) demonstrate (i) factors
that the emission that the emis-
sion factors change in different periods of the year, and (ii) the precision
change in different periods of the year, and (ii) the precision of the indirect estimation of theofindirect
estimation of COof2 on
CO2 on the basis theflow.
traffic basisConclusions
of traffic flow. Conclusions
are drawn in Section are7. drawn in Section 7.

Figure 1.
Figure Data flows
1. Data flows and
andsolutions
solutionstotopass
passfrom traffic
from flow
traffic to CO
flow estimation,
to 2CO withwith
2 estimation, paper sec- section
paper
tion mapping.
mapping.
2. Data Description and Related Problems
2. Data Description and Related Problems
As mentioned in the introduction, the main goal of the present work is to find a
As mentioned
relationship betweenin the introduction,
vehicular traffic flow the main
sensors andgoal
COof the present
2 , which can alsowork is to find a rela-
be measured.
tionship
Thus, the between
validationvehicular
of the model traffic
couldflow sensors
be viable and CO
in specific 2, whichor
conditions, can alsogeneral
a more be measured.
modelthe
Thus, hasvalidation
to be defined. Both
of the air quality
model couldsensors
be viable andin vehicular
specific traffic sensorsorare
conditions, taken general
a more
model has to be defined. Both air quality sensors and vehicular traffic sensorsare
into account for the present dissertation, and a description of the measurements that are taken
typically performed is needed before discussing the model. Typically, CO2 measurements
into account for the present dissertation, and a description of the measurements that are
are performed either by the count of the particles in the air (commonly, part per million,
typically performed is needed before discussing the model. Typically, CO2 measurements
ppm) or by means of the CO2 weight in a given air volume (mg/m3 or g/m3 ) for certain
are performed
temporal windows either by the
in a given count In
location. ofany
thecity,
particles in the
the number of air
CO2(commonly, part per
sensors is limited. For million,
ppm) or by means of the CO 2 weight in a given air volume (mg/m
example, in Florence, which is a metro city area of 1.5 million inhabitants, about 10 sensors or g/m ) for certain
temporal
are present.windows
However,ininathegiven location.
downtown In only
area, any city, thepresent,
four are numberdenotedof CO2by sensors is limited.
their IDs:
For example,
SMART09, in Florence,
SMART27, whichand
SMART28, is aSMART29
metro city area 2a
(Figure of for
1.5 the
million inhabitants,
respective locationsabout 10
in the municipality
sensors are present. of However,
Florence, from [25]).
in the Those sensors
downtown area,are onlynotfour
all inare
critical locations
present, denoted by
their IDs: SMART09, SMART27, SMART28, and SMART29 (Figure 2a forwhereas
for traffic or for pollutants. SMART28 and SMART29 are in dense traffic roads, the respective
SMAR27 and
locations SMART09
in the are in mid-range
municipality of Florence,trafficfrom
areas.[25]).
For each
Those air-quality
sensorssensor,
are not theall
datain critical
are registered every 2 min. In most European cities the number of CO sensors is not much
locations for traffic or for pollutants. SMART28 and SMART292 are in dense traffic roads,
higher. Those sensors were chosen since they are in critical points of the city and they are
whereas SMAR27 and SMART09 are in mid-range traffic areas. For each air-quality
 sensor, the data are registered every 2 min. In most European cities the number of CO2
sensors is not much higher. Those sensors were chosen since they are in critical points of
the city and they are located close to major roads. Therefore, it is reasonable to assume
that the CO2 measured by a sensor represents the measurement taken in a given area
around the sensor itself. They are calibrated in this sense. The presence of almost collo-
Sensors 2022, 22, 3382 5 of 20
cated CO2 and of traffic flow sensors allowed us to compute the emission factors as de-
scribed in the following section. On the other hand, the absence or significant reduction
of them could be overcome with the usage of rented CO2 sensors or by using the typical
located
values ofclose
the to major roads.
emission factorsTherefore,
in the area,it diminishing
is reasonabletheto precision
assume that themeasurement.
of the CO2 measured
by a sensor represents the measurement taken in a given area around
Typically, traffic sensor data are simultaneously registered every 10 min, the sensor itself.
and their
They are calibrated in this sense. The presence of almost collocated CO and
number is much higher than those of CO2 and in general for air quality assessments. The
2 of traffic flow
sensors allowed us to compute the emission factors as described in the following
data exploited refer to (about) 100 traffic flow devices located in the municipality of Flor- section.
On theasother
ence, hand,onthe
depicted theabsence or significant
right side of Figure 2b.reduction
They areofspire
themand
could be overcome
virtual with
spire sensors
the usage of rented CO sensors or by using the typical values of the emission factors
that produce values of2 counting, traffic density, and thus, local traffic flows. All of them in the
area, diminishing
are well calibratedthe
and precision
produceofcoherent
the measurement.
results.

Figure 2. The left side depicts the locations of the air-quality sensors considered in the present work
Figure 2. The left side depicts the locations of the air-quality sensors considered in the present work
(4
(4of
ofthe
the10
10 in
in the
the area
area of
of Florence). The right
Florence). The right side
side depicts
depictsthe
themap
mapof
ofthe
thetraffic
trafficsensors
sensorsininFlorence
Florence
Municipality. (a) CO
Municipality. (a) CO22 sensors; (b) traffic flow sensors.
sensors.

Typically,
Analysingtraffic sensor
their geo data are simultaneously
distribution (see Figure 2a,b),registered
it can be every 10 min,
observed that and their
the CO 2
number is much higher than those of
sensor locations have one or more traffic sensors CO 2 and in general for air quality assessments.
in their proximity, but they are not pre-
The data
cisely exploitedThus,
co-located. referthe
to (about) 100 traffic
actual traffic in theflow
CO2devices
sensor located
positioninhas
thetomunicipality
be estimated.of
Florence,
To this end, the so-called technique for traffic flow reconstruction [26,27] mayspire
as depicted on the right side of Figure 2b. They are spire and virtual help sensors
in this
that produce
sense, values of
as described in counting,
the followingtraffic density,
section. and
The thus,flow
traffic localmeasurements
traffic flows. All areofstrongly
them are
well calibrated
dependent on aand produce
number coherent
of road results.
features: road relevance (primary, secondary, etc.), num-
ber of lanes, speed limits, presence of(see
Analysing their geo distribution speedFigure
meters,2a,b), it can be
distance fromobserved that the etc.
road crossings, CO2
sensor locations have one or more traffic sensors in their proximity, but
Moreover, a certain class of roads (e.g., the so-called primary/main roads) may provide they are not precisely
co-located. Thus, the
higher capability withactual
respecttraffic in the
to local and CO 2 sensor position
single-lane has toflow
cases. Traffic be estimated. To this
sensors provide
end, the so-called technique for traffic flow reconstruction [26,27]
at each time slot different measurements regarding vehicular traffic flow, such as:may help in this sense, as
described in the following section. The traffic flow measurements are strongly dependent
• Vehicular traffic flow: number of vehicles crossing the supervised location during a
on a number of road features: road relevance (primary, secondary, etc.), number of lanes,
given period of time (which is usually referred to in terms of hours, that is, #cars/h);
speed limits, presence of speed meters, distance from road crossings, etc. Moreover, a
• Vehicular average speed: average speed of the vehicles crossing the supervised lo-
certain class of roads (e.g., the so-called primary/main roads) may provide higher capability
cation (measured in km/h);
with respect to local and single-lane cases. Traffic flow sensors provide at each time slot
• Vehicular density: number of vehicles in terms of road occupancy (measured in
different measurements regarding vehicular traffic flow, such as:
#cars/km).
• Vehicular traffic flow: number of vehicles crossing the supervised location during a
given period of time (which is usually referred to in terms of hours, that is, #cars/h);
• Vehicular average speed: average speed of the vehicles crossing the supervised location
(measured in km/h);
• Vehicular density: number of vehicles in terms of road occupancy (measured in
#cars/km).
• Travel time: average time that vehicles take to transit the supervised area (reported
in s).
 Sensors 2022, 22, x FOR PEER REVIEW 6 of 20

Sensors 2022, 22, 3382 6 of 20

• Travel time: average time that vehicles take to transit the supervised area (reported
in s).
Eachtraffic
Each trafficsensor
sensordatum datummay maypresent
presentaaspecific
specificbehaviour
behaviourdepending
dependingon onthethecontext.
context.
At each traffic sensor location, the measured vehicular traffic
At each traffic sensor location, the measured vehicular traffic flow (density) depends on the flow (density) depends on
the number
number of lanes
of lanes of theof road
the road
where where
the the traffic
traffic sensor
sensor is placed.
is placed. In In general,
general, themeasured
the measured
vehiculartraffic
vehicular trafficflow
flow(density)
(density)comingcomingfrom froma atraffic
trafficsensor
sensorlocated
locatedinina amultiple-lane
multiple-laneroad road
isisgreater
greaterwithwithrespect
respecttotothe thetraffic
trafficflow
flow(density)
(density)value valuecoming
comingfrom fromaasingle-lane
single-laneroad.road.
Moreover,the
Moreover, thetraffic
trafficsensors
sensorsoften oftenconsist
consistofofoptical
opticalcameras,
cameras,and andthe therelated
relatedmeasured
measured
data depend on specific parameters that are set up on the traffic sensor itself, that is,is,the
data depend on specific parameters that are set up on the traffic sensor itself, that the
selectedarea
selected areatotobebemonitored
monitoredand andthethelength
lengthofofthe theobserved
observedroad roadsegment.
segment.For Forexample,
example,
when
whena atraffic
trafficsensor
sensorisislocated
locatedon ona along,
long,straight
straightroad,road,ititisissetsetup
upininorder
ordertotomonitor
monitora a
long
longroad
roadsegment.
segment.However,
However,when whenthe thetraffic
trafficsensor
sensorisisplaced
placedininthe theproximity
proximityofofa acurvecurve
ininthe
theroad
roadorora aroad
roadjunction,
junction,the thesupervised
supervisedarea area is is
smaller.
smaller.
Each
Eachtraffic
trafficsensor
sensor monitors
monitors a fixed fixed supervised
supervisedarea areathatthatisis constituted
constituted byby means
means of a
ofgiven
a given road segment length. Then, each traffic sensor
road segment length. Then, each traffic sensor admits specific travel time measure- admits specific travel time
measurements.
ments. In the absence In the absence
of trafficofcongestion
traffic congestion
at the trafficat the trafficlocations,
sensor sensor locations,
the measured the
measured
travel time travel(totime
cross(to the
crosssegment)
the segment) is higher
is higher when when thetherelated
related monitored
monitored area areaisis
greater/longer,
greater/longer,since sincethethevehicles
vehiclesneed needmore moretime timetototravel.
travel.MoreMoreprecisely,
precisely,ininthetheabsence
absence
ofoftraffic
traffic congestion, each traffic sensor admits a minimum (monitored) travel time,which
congestion, each traffic sensor admits a minimum (monitored) travel time, which
can
canbebedefined
definedasasthe thevehicular
vehiculartime timeneeded
neededtotoacross acrossthe thesupervised
supervisedarea areawithin
within(at (atmost)
most)
the
thespeed
speedlimit
limitoccurring
occurringatatthe therelated
relatedtraffic
trafficsensor
sensorlocation.
location.Thus,
Thus,different
differentminimum
minimum
cross/travel
cross/travel time timecan canbe beregistered
registeredfor fordifferent
differenttraffic
trafficsensors.
sensors.InInorderordertotocompare
comparethe the
amount
amount of traffic data coming from the whole urban network, different trafficbehaviours
of traffic data coming from the whole urban network, different traffic behaviours
were
wereconsidered.
considered. For example,
For example, cumulative
cumulative data were
data collected
were at several
collected timesteps
at several timestepsand the
and
dataset was reordered according to increasing travel time.
the dataset was reordered according to increasing travel time. Figure 3 shows the moni- Figure 3 shows the monitored
data
toredtrend
dataoftrend
aboutof50 traffic
about 50sensors during during
traffic sensors seven days seven ofdays
observation, from Monday
of observation, from Mon-to
Sunday (24 h for day), taking into account the relationship between
day to Sunday (24 h for day), taking into account the relationship between the measured the measured vehicular
traffic flow traffic
vehicular with respect
flow with to the measured
respect to thetravelmeasuredtime. travel time.

Figure3.3.The
Figure Thecollection
collectionofofmonitored
monitoredtraffic
trafficdata
datacoming
comingfrom
fromabout
about5050traffic
trafficsensors
sensorsatatsparse
sparse
locations in the network over 7 days of observation. Each traffic sensor location presents a specific
locations in the network over 7 days of observation. Each traffic sensor location presents a specific
traffic behaviour depending on the travel time.
traffic behaviour depending on the travel time.

Whena large
When a large number
number of traffic
of traffic sensors
sensors are taken
are taken into into account
account in an in an urban
urban road net-
road network,
work,
the the complexity
complexity of the problem
of the problem increases,
increases, and a and a uniform
uniform approach
approach is needed
is needed in order
in order to
properly handle and compare the amount of traffic data at different traffic locations. On
this basis, it can be supposed that the increment in travel time at a given point may be due
Sensors 2022, 22, 3382 7 of 20

to the inception of crowding conditions, and that not all sensor locations may be prone
to crowding conditions. In the next section we introduce a uniform approach to formally
understand when uncongested and congested situations occur in the road network by
analysing a large amount of traffic data.

3. Traffic Flow Data Analysis

As reported in the introduction, the focus is to propose a general approach to deter-
mine a model to pass from traffic flow data to CO2 emissions, taking into account (i) a
uniform approach that allows the measurements to be compared according to the physical
context of sensor placement, (ii) the different traffic conditions (congested vs. uncongested),
and (iii) seasonal changes. To achieve the proposed scope, a traffic data alignment is
needed in order to handle the mentioned measurements by means of appropriate data
normalisation. Since all the measured traffic data are affected by the physical context of the
sensor placements, the traffic data must be compared in a specific way. In order to do that,
the following normalisation approach was conducted for each i-th traffic sensor:
• Measured vehicular traffic flow, denoted by F (t) at a given timestamp t, is normalized
with respect to the number of lanes, denoted by C, of the road of the location. Thus, the
normalized vehicular traffic flow at a given timestamp t, denoted by Fn(t), is given by
Fn(t) = F (t)/C;
• Measured travel time, denoted by T (t) at a given timestamp t, is normalized with
respect to the minimum travel time, denoted by Tm, occurring in the absence of conges-
tion in the sensor location. Tm can be defined as the travel time needed to cross the
area at the speed limit of the observed segment. Thus, the normalized travel time at a
given timestamp t, denoted by m(t), is given by m(t) = T (t)/Tm.
According to the above normalisation, the traffic flow data in Figure 3 can be compared
as depicted in Figure 4. In this way, all the curves in Figure 4 started from the origin and
the data were aligned. For example, consider two distinct road segments monitored by
means of a given traffic sensor. Suppose that the first monitored road segment is a two-lane
street 0.35 km long with a speed limit of 50 km/h and a minimum travel time of 25.2 s;
the second monitored road segment is a single-lane street 0.25 km long with a speed limit
of 30 km/h and a minimum travel time of 30 s. At a given time, the sensor located in
the first monitored road segment registers 800 vehicles/h and the related travel time is
37.8 s. The sensor in the second monitored road segment registers 400 vehicles/h and the
related travel time is 45 s. After the above normalisation, both the traffic situations in the
example admit the same behaviour: 400 vehicles/h cross the single-lane street at 1.5 times
their minimum travel times. The measurements are now aligned and comparable. The
normalised measurements are represented by means of the same point/contribution in the
traffic data alignment shown in Figure 4.
The normalized travel time is a dimensionless value denoted by m, and it can be
considered a multiplier factor of a given minimum travel time. So, we set m = 1 in the
absence of traffic congestion for each traffic sensor location and the related traffic data were
comparable.
In order to describe the traffic data in the whole network, the average traffic flow value
was estimated with respect to the corresponding average travel time in the network. Since
each travel time at a given location can be expressed in terms of multiplier factor of its
minimum travel time, then an average value can be estimated by considering all the traffic
sensor locations in the network. Figure 5 shows such average behaviour by considering the
data in Figure 4. This approach allows the traffic flow to be characterized in the whole city
in a given period of time (or time slot) in which the traffic flow data are collected.
Sensors 2022, 22, 3382 8 of 20
Sensors 2022, 22, x FOR PEER REVIEW 8 of 20

Figure 4. Traffic data alignment referring to the traffic trends depicted in Figure 3. It is performed
by means of a normalisation approach that involves the number of road lanes, minimum travel time,
and speed limit for each traffic sensor location.

In order to describe the traffic data in the whole network, the average traffic flow
value was estimated with respect to the corresponding average travel time in the network.
Since each travel time at a given location can be expressed in terms of multiplier factor of
its minimum travel time, then an average value can be estimated by considering all the
traffic sensor locations in the network. Figure 5 shows such average behaviour by consid-
Figure4.
ering
Figure 4.Traffic
the Traffic data
datadata
in alignment
Figure 4. This
alignment referring
approach
referring to the
to the traffic
allows
traffic trends
the
trends depicted
traffic flow
depicted inbe
to Figure
in Figure 3. It 3.
is It
characterizedis performed
performed in by
the
by means of a normalisation approach that involves the number of road lanes, minimum travel time,
wholeofcity
means in a given period
a normalisation approachof that
timeinvolves
(or timetheslot) in which
number of roadthe traffic
lanes, flow data
minimum are
travel col-
time,
and speed limit for each traffic sensor location.
lected.
and speed limit for each traffic sensor location.
In order to describe the traffic data in the whole network, the average traffic flow
value was estimated with respect to the corresponding average travel time in the network.
Since each travel time at a given location can be expressed in terms of multiplier factor of
its minimum travel time, then an average value can be estimated by considering all the
traffic sensor locations in the network. Figure 5 shows such average behaviour by consid-
ering the data in Figure 4. This approach allows the traffic flow to be characterized in the
whole city in a given period of time (or time slot) in which the traffic flow data are col-
lected.

Figure5.5.The
Figure Theaverage
averagetraffic
traffictrend
trendthat
thatconstitutes
constitutesa amean
meanbehaviour
behaviourby
bytaking
takinginto
intoaccount
accountthe
the
different trends depicted in Figure 4.
different trends depicted in Figure 4.

Let us start to study the traffic conditions by considering the described mean traffic
behaviour in order to observe uncongested and congested traffic situations. The vehicular
traffic flow and travel time data are largely influenced by congested traffic conditions. More
precisely, congested traffic situations have a higher vehicular traffic flow in the monitored
road section. Moreover, the travel time depends on the vehicular average speed, which
comes close to 0 when traffic congestion occurs. Higher travel time reduces the vehicular
Figure 5. The average traffic trend that constitutes a mean behaviour by taking into account the
flow, and it is inversely proportional to the number of vehicles passing. The uncongested
different trends depicted in Figure 4.
and congested traffic situations are implicitly determined by means of the volume of
vehicular flow that passes the supervised area in the unit of time. To consider mean/typical
 behaviour in order to observe uncongested and congested traffic situations. The vehicular
traffic flow and travel time data are largely influenced by congested traffic conditions.
More precisely, congested traffic situations have a higher vehicular traffic flow in the mon-
itored road section. Moreover, the travel time depends on the vehicular average speed,
which comes close to 0 when traffic congestion occurs. Higher travel time reduces the
Sensors 2022, 22, 3382 9 of 20
vehicular flow, and it is inversely proportional to the number of vehicles passing. The
uncongested and congested traffic situations are implicitly determined by means of the
volume of vehicular flow that passes the supervised area in the unit of time. To consider
behaviour
mean/typicalof the whole network
behaviour for a network
of the whole such value,
for awe canvalue,
such take into account
we can take the
intoaverage
account
values of the values
the average normalised
of thevehicular flow,
normalised by Fn,
denotedflow,
vehicular and theby
denoted 𝐹𝑛, and the
normalised travel time m.
normalised
The
travel time 𝑚.
vehicular flow rate
The vehicular t canrate
at time flow be defined
at time tas follows:
can be defined as follows:

Fn𝐹𝑛(𝑡)
(t)
FR𝐹𝑅(𝑡)
(t) = = 𝑚(𝑡)
m(t)
which is, in terms of fluid dynamics, the volumetric flow rate [28] (also known as volume
which is, in terms of fluid dynamics, the volumetric flow rate [28] (also known as volume
flow rate, rate of fluid flow), that is, the volume of a fluid passing in the time unit. Coming
flow rate, rate of fluid flow), that is, the volume of a fluid passing in the time unit. Coming
fromthis
from this general
general definition,
definition, wewe introduced
introduced aasimilarsimilarconcept
conceptininthe thetraffic
trafficflow
flowtheory
theory to
estimate and observe the variation of traffic flow in terms of vehicular
to estimate and observe the variation of traffic flow in terms of vehicular flow rate. For flow rate. For in-
stance, the
instance, thetraffic
trafficsituations
situations described
described in in the mentioned example
the mentioned example after
after the
thenormalisation
normalisation
approach admit the same contribution in terms of vehicular flow rate, which isisequal
approach admit the same contribution in terms of vehicular flow rate, which equalto to
266.6 vehicles/h. Such a normalisation is necessary to compare the
266.6 vehicles/h. Such a normalisation is necessary to compare the flow rate in each data flow rate in each data
observationand
observation andobserve
observethe therelated
relatedvariation.
variation.
In Figure 6, the general behaviour
In Figure 6, the general behaviour of ofthe
themean
meanflow
flowrate
rate(MFR)
(MFR)isispresented
presentedby bytaking
taking
intoaccount
into accountthe thedata
datain inFigure
Figure5.5.ItItconsiders
considersthe the(average)
(average)values
valuesof ofthe
thenormalised
normalisedtraffic
traffic
flow and the corresponding normalised travel time in a given period
flow and the corresponding normalised travel time in a given period of time, where the of time, where the
observations are sorted according to increasing travel time measurements
observations are sorted according to increasing travel time measurements in the whole in the whole
network.The
network. Theflow
flowrate
ratecan
canalso
alsobe
becomputed
computedthroughthroughthe thedata
datacoming
comingfromfromaasingle
singletraffic
traffic
sensor in order to understand the variation in traffic modality at
sensor in order to understand the variation in traffic modality at a specific location a specific location of the
of
road
the network.
road network. Nevertheless,
Nevertheless, the the
described
described normalisation
normalisationapproach
approach has has
beenbeen
always con-
always
ducted to compare
conducted to compare different local
different traffic
local variations.
traffic variations.

Figure6.6. The
Figure The mean
mean flow
flow rate
rate (MFR)
(MFR)considers
considerstraffic
trafficobservations
observationsthat
thatare sorted
are according
sorted to to
according in-
creasing travel time measurements in the traffic sensors depicted in Figure 5.
increasing travel time measurements in the traffic sensors depicted in Figure 5.

MFRscan
MFRs canimmediately
immediatelyprovide
provideevidence
evidenceof
ofthe
thewhole
wholenetwork
networkchanging,
changing,and
andthey
they
allowdifferent
allow differenttraffic
trafficbehaviours
behaviourstotobe
bedetermined
determinedaccording
accordingtotothe
theseasonal
seasonalchanges.
changes.In
In
this work, the data taken into account were those from March, May, July, and October 2021
in order to analyse different situations in different seasons. More precisely, each season
observation is represented by means of a period of 7 days, from Monday to Sunday, 24 h a
day. Figure 7 shows the seasonal changes according to the mean flow rate by considering all
of the traffic sensors in the network of Florence, where the related behaviours in March,
May, July, and October are described by means of the curves coloured blue, green, yellow,
and orange, respectively.
 this work, the data taken into account were those from March, May, July, and October
2021 in order to analyse different situations in different seasons. More precisely, each sea-
son observation is represented by means of a period of 7 days, from Monday to Sunday,
24 h a day. Figure 7 shows the seasonal changes according to the mean flow rate by consid-
ering all of the traffic sensors in the network of Florence, where the related behaviours in
Sensors 2022, 22, 3382 10 of 20
March, May, July, and October are described by means of the curves coloured blue, green,
yellow, and orange, respectively.

Figure7.7.The
Figure Theseasonal
seasonalchanging
changingcurves
curvesare
areshown
shownininterms
termsofofMFR,
MFR,where
wherethe
thebehaviours
behavioursofofthe
the
seasons are depicted in blue (March), green (May), yellow (July), and orange (October).
seasons are depicted in blue (March), green (May), yellow (July), and orange (October).

Asexpected,
As expected,the thegeneral
general behaviour
behaviour of the
of the MFRMFR in March
in March is similar
is similar toone
to the thein
one in Oc-
October,
tober,
and theand
MFRthe MFRisinsimilar
in July July istosimilar
the onetointhe
May.oneOfincourse,
May. the
Of course, the meteorological
meteorological conditions
conditions
influence influence
vehicle usagevehicle
in termsusage in terms
of driving of driving
behaviour andbehaviour and volume.
volume. During During
the cold the
seasons
cold
the seasons in
variation thetraffic
variation in traffic
modality seemsmodality seems
to be more to be morewith
emphasised emphasised
respect towith
the respect
warm
to the warm
seasons, duringseasons,
whichduring which
a relevant a relevant
number number
of city of city
users get userswith
around get around withand
motorbikes mo-
torbikes
bikes, andinfluence
which bikes, whichless influence
or none thelessflow
or none the flow
counting. counting.
Different Different
seasonal seasonalcan
behaviours be-
haviours
also cansalso
be observed be observed
numerically by numerically
means of the by meansdeviation
standard of the standard
of each deviation
mentionedofMFR,
each
N
2 (𝑀𝐹𝑅(𝑡) − 𝑀𝐹𝑅) , where N is the number of ob-
mentioned byMFR, 1 computed by
computed N ∑ MFR ( t ) − MFR , where N is the number of observations and MFR
t =1
servations and 𝑀𝐹𝑅 is the mean, which is equal to 134.52, 122.85, 114.55, and 135.92 for
is the mean, which is equal to 134.52, 122.85, 114.55, and 135.92 for March, May, July, and
March, May, July, and October, respectively, according to the values in Figure 7. The
October, respectively, according to the values in Figure 7. The greatest difference can be
greatest difference can be observed between the curves representing the behaviour in Oc-
observed between the curves representing the behaviour in October and July, in orange
tober and July, in orange and yellow, respectively, in Figure 7. Table 1 shows such a devi-
and yellow, respectively, in Figure 7. Table 1 shows such a deviation in terms of absolute
ation in terms
differences of absolute
at some differences
corresponding at some corresponding points.
points.
Table 1. Numerical values coming from the seasonal changing curves in terms of mean flow rate
Table 1. Numerical values coming from the seasonal changing curves in terms of mean flow rate
depicted in Figure 7.
depicted in Figure 7.
Mean Flow Rate 100 250 400 550 700 850 950
Mean Flow Rate 100 October 250 7.3 40045.1 550
178.3 700
289.0 320.0 850 340.4 950
353.7
October 7.3 July 45.1 38.0 178.3
110.9 289.0
197.3 320.0
271.4 312.2 340.4335.6 353.7
349.4
July 38.0 110.9 30.7
ABS. DEVIATION 197.365.8 271.4
19.0 312.2
17.5 7.7 335.6 4.8 349.4
4.2
ABS . DEVIATION 30.7 65.8 19.0 17.5 7.7 4.8 4.2

4. Traffic Flow Modalities: Congested and Uncongested

The present section is devoted to formally identifying traffic flow modalities: un-
congested and congested traffic situations that occur, and when we can detect being in
one of the two cases. The traffic flow data analysis reported in the previous section may
allow us to understand when a congested situation occurs at a given traffic sensor location.
In particular, by considering the MFR behaviour, two distinct situations in the diagrams
were identified. From the behaviour of FR (which is absolutely similar to the trend of
MFR), a change in concavity in the related diagram was identified. In particular, for each
traffic sensor a specific FR trend was present, and there was a point in which the FR curve
changed concavity. That point is the point in which the travel time started to decrease and
Sensors 2022, 22, 3382 11 of 20

the traffic started to congest. Such a concavity changing in each diagram can be determined
by means of the unique inflection point, and the corresponding value in the flow rate can
be identified when a reduction in the vehicles passage is taking place. More precisely, when
the concavity of the flow rate is upwards, the vehicles passage assumes a quick increase and
the traffic flow proceeds unimpeded. Then, the situation of uncongested traffic is assumed.
Otherwise, when the concavity of the flow rate is downwards, the vehicle passage reduces
to a quasi-constant condition. Such a traffic modality occurs when the flow is slowed
down, and stop-and-go situations arise since the road capability is limited. Therefore, a
congested traffic situation can be assumed for such behaviour. Such a concavity change
may be determined by identifying the inflection point in the FR trend, as detailed in the
following section. In order to formally estimate the inflection point, we analyse the function
f, which is defined by means of a given FR behaviour. More precisely, each monitored traffic
location provides an FR trend, which can be compared with other trends by performing
the traffic data alignment described in the previous sections. Therefore, it is possible to
model the FR trend with a function f—for example, for a time period of one week (7 days,
from Monday to Sunday, 24 h a day, with traffic measurements every 10 min by the traffic
sensors). In order to identify function f, we proceed to perform a polynomial approximation
to minimise the worst-case error. Thus, the polynomial approximation P of function f min-
imises |P(x) − f(x)|, where x varies over the chosen interval. Therefore, the approximation
P of function f is obtained by using a third-order polynomial form: P(x)=ax3 +bx2 +cx+d.
Different values of the coefficients (a, b, c, d) of P determine a different traffic behaviour
in terms of FR. For each FR(T,i) observed in a given period T at the i-th traffic location, a
unique polynomial approximation is represented by (ai (T), bi (T), ci (T), di (T)). For exam-
ple, by assuming the period T, running from 2021-05-03T00:00:00 to 2021-05-09T23:59:59,
we obtained the characterisation of the polynomial approximation, as seen in Table 2.

Table 2. An example of the coefficients related to the polynomial approximations of the MFR(Spring)
and FR(Spring,i) for 3 traffic flow sensors during the (same) selected period of time in spring.

Polynomial Approx. a b c d
MFR(Spring) −0.0000011 0.0014465 0.0686048 −17.38807
FR(Spring,1) −0.0000019 0.0027314 −0.140624 −9.454277
FR(Spring,2) −0.000001 0.0016093 −0.247455 2.96993
FR(Spring,3) −0.0000019 0.002776 −0.471167 11.87328

As is well known, since the polynomial approximation P is differentiability class C2 for

each set of finite coefficients that differs from zero (P, its first derivative P0 , and its second
derivative P00 , exist and are continuous), the condition P00 = 0 can be used to find the
desired inflection point in the considered interval of vehicular traffic behaviour. Thus, in a
fixed period of observation, both at a given traffic sensor location and in the whole network
(by assuming average traffic behaviour), it is possible to compute the inflection point that
determines a change in the traffic modality from congested to uncongested. This allows the
different contributions of traffic flow to be taken into account in terms of CO2 . Then, it is
possible to consider uncongested and congested situations, both at a given traffic location
and in the whole network (see Table 3 for the identified inflection point according to the
traffic sensors with the polynomial approximations in terms of flow rate listed in Table 2).
Sensors 2022, 22, 3382 12 of 20

Table 3. An example of the coefficients related to the polynomial approximations of the MFR(Spring)
and FR(Spring,i) for 3 traffic flow sensors during the (same) selected period of time in spring.

Traffic
Traffic Flow Inflection Point Flow Rate Traffic Flow
Density
Sensors (T-TH OBS) on P (#Cars/h) (#Cars/h)
(#Cars/Km)
All city sensors 438 215.8 255 7.24
S1 (near SMART27) 479 366 366.88 7.12
S2 (near SMART28) 536 186.14 216 6.35
S3 (near SMART29) 487 234.24 240 4.21

5. Traffic Flow Reconstruction into CO2 Sensor Locations

In order to identify a model that can allow us to compute CO2 values on the basis
of traffic flow data, the most effective approach could be to validate the model in specific
points for long time windows. In [16], a general model for the whole city, estimating the
amount CO2 from traffic in grams (briefly gCO2 ), was proposed. The model assessed
a mean emissions factor of 203 gCO2 /km per vehicle for the “urban slow” pattern, for
the “stop-and-go A” pattern a mean emission factor of 460 gCO2 /km per vehicle, and for
the “stop-and-go B” an emission factor of 738 gCO2 /km per vehicle. The amount of CO2
measurements by an air quality sensor is related to the space/volume surrounding the
location of the sensor itself. Therefore, the measured CO2 would depend on the amount of
traffic flow passing in the road segment volume, which is of pertinence to the air-quality
sensor. In the previous section, we demonstrated that the FR function presents two main
modalities, also considering the MFR for the whole city. As highlighted in Section 2,
locations in which CO2 sensors are located typically do not present corresponding traffic
sensors in their precise proximity; they may be hundreds of meters away. Moreover, CO2
sensor SMART09 also does not present traffic sensors in the area. On the other hand, the
traffic of the road segments closest to the CO2 sensor impacts the measured CO2 directly. A
traffic reconstruction technique [26,27] can be used to estimate the traffic flow density in the
road segments closer to the CO2 sensor over time. To this end, traffic flow reconstruction
can be used. The exploitation of traffic flow data for CO2 estimation is addressed in
Section 6.
Traffic flow reconstruction [26,27] is the process of producing a value of traffic density
(flow)—e.g., vehicle per meter (vehicles per minute)—for each road (or road segment, or
a large number of road segments) by starting from a limited number of traffic sensors
measuring traffic density (flow) on the road. The measures of traffic density are typically
obtained by stationary sensors in strategic positions. The problem of traffic flow reconstruc-
tion is regarded as the solution of the LWR (Lighthill–Whitham–Richards) model [29,30],
which models the traffic density in terms of the partial differential equation (PDE). The
solution of the LWR model is not a trivial matter for large networks due to its computational
complexity and constraints [31–33]. The estimation of the traffic distribution on junctions
plays a crucial role in the effectiveness of the LWR model application in a real context in
order to have a complete description of the road segments composing the urban network
in terms of traffic density. Traffic flow reconstruction is performed by solving a nonlinear
model based on the conservation of vehicles, described by the following scalar hyperbolic
conservation law. On a single road, we have:

∂ρ(t, x) ∂f(ρ(t, x))

+ = 0,
∂t ∂x
where ρ(t, x) is the traffic density of vehicles, which admits values from 0 to ρmax , where
ρmax > 0 is the maximum traffic density; the f(ρ(t, x)) function is the vehicular flow, which
is defined by means of the product ρ(t, x) v(t, x), where v(t, x) is the vehicle speed, and
the boundary conditions are ρ(t, h) = ρh (t), ρ(t, k ) = ρk (t) and the initial values are
ρ(0, x) = ρ0 (x), where x ∈(h, k). In the case of first-order approximation, we assume
Sensors 2022, 22, 3382 13 of 20

that v(t, x) is a decreasing function, depending on the density, and then the correspond-
ing flux is a concave function.
 Thus, we consider
 the local speed of the vehicles to be
ρ ρ
v(ρ) = vmax 1 − ρmax and then f(ρ) = vmax 1 − ρmax ρ, where vmax is the speed limit
on a given road segment (these assumptions are referred to in the literature as Greenshield’s
Model). The solution is obtained by an iterative process at finite differences on the basis of
the traffic flow data in the sensor points. For each timestep, the traffic flow reconstruction
is performed by producing a value of traffic density in each city road segment of the graph,
which are typically 20 mt.
The accuracy of the described solution primarily depends on the computation of
the so-called traffic distribution matrices (TDM), that is, the traffic flow distribution at
junctions. In order to model the traffic distribution at junctions, a distribution matrix can
be used to describe the percentage of vehicles leaving each outcoming road with respect
to those entering each incoming road. Thus, the traffic distribution matrix is defined as
n+m
TDM = wji j=n+1,...,n+m,i=1,...,n , so 0 < wji < 1 and ∑ wji = 1, for i = 1, . . . , n and

j=n+1
j = n + 1, . . . , n + m, where wji is the percentage of vehicles arriving from the i-th incoming
road and taking the j-th outcoming road (assuming that, at each junction, the incoming
flux coincides with the outcoming flux). The real values of wji may depend on the time
of the day, the road size, the crossing light settings, etc., and thus, are unknown a priori.
The values of wji are estimated by giving the lower mean error by means of the stochastic
relaxation technique as described in [26]. The traffic flow reconstruction algorithm has to
be computed progressively and in a parallel architecture, since the estimation of traffic
flow density for the city (in Florence there are about 30,000 segments) at time instant t
would depend on traffic flow at time t − 1 and on the new measurements coming from
the sensors. For each traffic sensor update, we have a complete description of the road
segments composing the urban network in terms of traffic density.

6. Computing CO2 Emissions Factors from Traffic Flow Data and Modalities
In this section, we are going to identify how uncongested and congested traffic situa-
tions each contribute to the amount of CO2 measured. The amount of CO2 measured by
a sensor is related to the area surrounding the sensor. The idea is to consider the amount
of traffic contained in the road segments in which the CO2 sensors are located. The traffic
flow reconstruction algorithm allows the amount of traffic to be computed in terms of
vehicular density in each road segment with a length of 20 m every 10 min. By choosing,
from the total number of reconstructed road segments, the closest one to each air-quality
sensor, we can analyse the impact of traffic on the CO2 measurements in different zones
of the city. In substance, we identified a model (see Section 6, Equation (1)) with which
the computation of CO2 from traffic would be possible. The computation of CO2 depends
on the traffic flow and of the emissions factors, which are different for congested and
uncongested traffic flow cases. Therefore, we found a way to detect the conditions to
compute the number of vehicles passing close to the CO2 control point in the different
conditions (Section 4). This approach allows us to demonstrate (validate) the model by
computing the emissions factors.
The measurements of CO2 and traffic flow refer to different systems of measurement,
different time intervals, and different acquisition methods.

6.1. Time Alignment of Traffic Flow and Measured CO2 Data

Traffic flow data are available every 10 min, and CO2 data every 2 min. In order to
align and compare the data coming from the air-quality sensors with the traffic data sensors,
the CO2 measurements were carried over to the timesteps of the traffic sensors. Thus, the
average of the CO2 data over the interval considered by the traffic sensors was taken into
account (see the left side of Figure 8). Moreover, the standard traffic flow measurements
are expressed in terms of #cars/h. Thus, we need to divide by 6 to get #cars/10 min; see an
example on the right side of Figure 8.
 Traffic flow data are available every 10 min, and CO2 data every 2 min. In order to
align and compare the data coming from the air-quality sensors with the traffic data sen-
sors, the CO2 measurements were carried over to the timesteps of the traffic sensors. Thus,
the average of the CO2 data over the interval considered by the traffic sensors was taken
into account (see the left side of Figure 8). Moreover, the standard traffic flow measure-
Sensors 2022, 22, 3382 14 of 20
ments are expressed in terms of #cars/h. Thus, we need to divide by 6 to get #cars/10 min;
see an example on the right side of Figure 8.

Figure 8.
Figure 8. On
Onthe
theleft
leftside: thethe
side: time alignment
time of CO
alignment data data
of 2CO in a specific location,
in a specific passing
location, from ppm/2
passing from
2
min to ppm/10 min, at a certain time of day. On the right side: the time alignment of traffic flow
ppm/2 min to ppm/10 min, at a certain time of day. On the right side: the time alignment of traffic
data in a specific location, passing from #cars/h to #cars/10 min, at the same time of day.
flow data in a specific location, passing from #cars/h to #cars/10 min, at the same time of day.
6.2. Pollutant Data Type
6.2. Type
CO22 measurements are typically estimated either by the percentage
CO percentage of of its
its particles
particles in
in
the total
the total amount of the gasses in the air (commonly, part per million) or by means
(commonly, part per million) or by means of the of the
CO22 weight in a given
CO given volume
volume ofof air
air(commonly, 𝐦𝐠/𝐦𝟑3 or
(commonly, mg/m 𝐠/𝐦3𝟑).). These
or g/m These standard
standard units
units
of measurement
of measurement admitadmit aa direct
direct conversion
conversion ofof one
one into
into the
the other
other and vice-versa,
vice-versa, taking
taking into
into
account certain
account certain parameters
parameters such
such as the mole of the particulate
particulate matter under
under consideration.
consideration.

6.3.
6.3. From
From Traffic
TrafficFlow
FlowData
Datato toCO
CO22
In
In order
orderto
torelated
relatedthe theamount
amountofofCOCO2 in
2 ina agiven
givenroad
roadsegment
segment with
withthethe
vehicle density
vehicle den-
in itsin
sity proximity, it is necessary
its proximity, to associate
it is necessary the COthe
to associate 2 measurement
CO2 measurementwith a volumetric section
with a volumetric
of road segment.
section The measured
of road segment. emissions
The measured of CO2 of
emissions in CO
the 2volume should should
in the volume be equivalent to
be equiv-
the CO produced by vehicles according to the traffic behaviour, which
alent to the CO2 produced by vehicles according to the traffic behaviour, which can be
2 can be identified
according to what was
identified according presented
to what in Section
was presented in 4. Therefore,
Section we canwe
4. Therefore, assume to have
can assume two
to have
different contributions of emissions according to a classification of traffic behaviours.
two different contributions of emissions according to a classification of traffic behaviours. Thus,
the
Thus,following equation
the following holds: holds:
equation
S(𝑆(𝑧)𝐺(𝑡,
z) G (t, z)𝑧)
==K1𝐾(z(𝑧)𝐹
) F1 (t,(𝑡,
z)𝑧)𝐿(𝑧)
L(z) ++K2𝐾(z(𝑧)
) F2𝐹(t,(𝑡,z)𝑧)𝐿(𝑧)
L(z) (1)
(1)
where t 𝑡isisthe
where thetime
timeinterval
intervaland 𝑧 the
andz is is the
COCO 2 sensor ID.
2 sensor ID.
•• S𝑆(𝑧)𝐺(𝑡,
(z) G (t, z𝑧)) isis the
the amount
amount of 𝑔𝐶𝑂 ininaavolumetric
of gCO
2
volumetricsection
section of
of the
the road
road segment
segment at
at aa
given time t,
given time t, where: where:
o
# 𝐺(𝑡,
G (𝑧)
t, z)isisthe
themeasurement
measurement of ofCO 2 from the sensor in 𝑔𝐶𝑂 /𝑚
CO in3 the time in-
2 from the sensor in gCO2 /m in the time
terval (these values are measured by the CO 2 sensors);
interval (these values are measured by the CO2 sensors);
o
# 𝑆(𝑧)
S(z)isisthe
thearea
areaininwhich
whichthe
thesensor
sensorcollects
collectsthe valuesinin 𝑚
thevalues and it
m3 and it is
is estimated
estimated
onon
thetheroad
roadsegments
segmentsclose
closetotothe
theCO
CO2 2sensor
sensorlocation.
location.More
Moreprecisely,
precisely, we
we have:
have:

S(z) = L(z) C (z) W (z) H (z) (2)

where C (z) is the number of lanes, W (z) is the width of the road lane, and H (z)
is the height of the volume, which depends on the position of the CO2 sensor
(typically at 3 m).
# L(z) is the road length corresponding to the amount of m or Km performed by
the vehicles in that specific area of the CO2 sensor, supposing that the vehicles
change neither road nor behaviour in the segment.
Sensors 2022, 22, 3382 15 of 20

• Contribution coming from vehicles/cars moving in uncongested conditions:

# F1 (t, z) is the traffic count in uncongested conditions in terms of #cars in the
time interval. This can be measured on the basis of traffic sensors and/or
estimated as solutions of the above-presented LWR PDE via the traffic recon-
struction model (Section 5);
# K1 (z) is an emissions factor to be determined, which is the amount of gCO2 /km
per car in uncongested conditions.
• Contribution coming from vehicles/cars moving in congested conditions:
# F2 (t, z) is the traffic count in congested conditions in terms of #cars in the time
interval. This can be measured on the basis of traffic sensors and/or estimated
as solutions of the above-presented LWR PDE via the traffic reconstruction
model);
# K2 (z) is an emissions factor to be determined, which is the amount of gCO2 /km
per car in congested conditions.
In Equation (1), for a certain time interval and sensors (CO2 and traffic), K1 (z) and
K2 (z) are unknowns. They can be estimated by means of a multilinear regression providing
a large number of measurements for both CO2 and traffic in the same contextual conditions:
location, traffic, season, etc. Once K1 (z) and K2 (z) are determined, it is possible to apply
Equation (1) to estimate CO2 emissions (i.e., G (t, z)) in other parts of the city provided that
the other factors are known. Among them are the typical traffic behaviour, the distribution
of vehicle types, etc., which can be estimated to be very similar in the same city and in
closed areas/roads.
Moreover, in Equation (1), F1 (t, z) and F2 (t, z) are obtained by classifying the vehicles
passing in the road segment closest to the z-th CO2 sensor location according to the general
condition of congested or uncongested. Thus, they are computed as the product between
the traffic density (#cars/km) and the vehicular speed (km per time interval):

F1 (t, z) = D1 (t, z) V1 (z) (3a)

F2 (t, z) = D2 (t, z) V2 (z) (3b)

According to the following cases of congested and uncongested:
ρ(t,z)
• If C (z)
≤ q(z), then D1 (t, z) = ρ(t, z) and D2 (t, z) = 0;
ρ(t,z)
• If C (z)
> q(z), then D1 (t, z) = 0 and D2 (t, z) = ρ(t, z).
where:
• ρ(t, z) is the traffic density (#cars/km) as observed on the traffic sensors and/or
estimated as solutions of the above-presented LWR PDE via traffic flow reconstruction;
• q(z) is the value on the inflection point of traffic density in the proximity of the z-th
air-quality sensor location to detect congested and uncongested cases, as described in
Section 4, Table 3;
• C (z) is the number of road lanes;
• V1 (z) and V2 (z) are the average vehicular speeds in z-th location in the cases of
uncongested and congested situations, respectively, if the specific velocity cannot be
measured or reconstructed. For example, V2 (z) is the vehicular speed in the case of
a traffic congestion situation, and we assume that the value of V2 (z) varies from 0 to
5 km/h in accordance with the road characteristics at the z-th location.
Therefore, according to Equation (1), for each time instant, it is possible to classify the
traffic as congested or uncongested and thus to compute the contributions of CO2 . This
approach may lead to computing the whole city’s traffic flow CO2 production over the day,
as well its distribution on the map.
 Sensors 2022, 22, x FOR PEER REVIEW 16 of 20

Therefore, according to Equation (1), for each time instant, it is possible to classify the
Sensors 2022, 22, 3382 traffic as congested or uncongested and thus to compute the contributions of CO16 2. This
of 20
approach may lead to computing the whole city’s traffic flow CO2 production over the
day, as well its distribution on the map.
6.4. Assessing Seasonal Changes of Estimations: Experimental Results
6.4. Assessing Seasonal Changes of Estimations: Experimental Results
Equation (1) should be satisfied for each observation of traffic flow and CO2 in each
Equation (1) should be satisfied for each observation of traffic flow and CO2 in each
segment. It is possible to estimate the unknowns K1 (z) and K2 (z) for each location in
segment. It is possible to estimate the unknowns 𝐾 (𝑧) and 𝐾 (𝑧) for each location in
which CO2 and traffic flow data are known. The estimation can be performed by means
which CO2 and traffic flow data are known. The estimation can be performed by means
of a multilinear regression in which the dependent variable is the amount of gCO2 in a
of a multilinear regression in which the dependent variable is the amount of 𝑔𝐶𝑂 in a
volumetric section of the road segment at a given time t, and the explanatory variables are
volumetric section of the road segment at a given time t, and the explanatory variables are
the traffic count in each condition (congested and uncongested): Fi (t, z) L(z).
the traffic count in each condition (congested and uncongested): 𝐹 (𝑡, 𝑧)𝐿(𝑧).
To this end, we estimated K1 (z) and K2 (z) for a number of sensors, and in all of
To this end, we estimated 𝐾 (𝑧) and 𝐾 (𝑧) for a number of sensors, and in all of
them the multilinear regression turned out to be significant, producing values for the
them the multilinear regression turned out to be significant, producing values for the co-
−16 and
coefficients with a p-value in the range of p-value = 2 × 10 a t-value > 19 in all cases
efficients with a p-value in the range of p-value = 2 × 10−16 and a t-value > 19 in all cases for
for all coefficients. The R-squared of the models was also statistically significant, typically
all coefficients. The R-squared of the models was also statistically significant, typically
greater than 0.7 in most cases, which means that the models typically explained 70% of
greater than 0.7 in most cases, which means that the models typically explained 70% of
the variability of the response data around their mean at each hour of the day. The mean
the variability of the response data around their mean n at each hour of the day. The mean
1 obs − pred
absolute
absolutepercentage
percentageerror
error(MAPE),
(MAPE),computed
computedas as n ∑∑ | iobs i , is|, commonly
is commonly
i
used
usedas as
a
i =1
loss function
a loss functionforfor
regression
regression problems
problems bybymeans
meansofofitsitsinterpretation
interpretationinintermstermsof ofrelative
relative
error.
error.
The
Theproposed
proposedmethod,
method,unlike unlikethe theapproaches
approachespresented
presentedin in[12–17],
[12–17],allows
allowsthe theCOCO22
emission
emissionto tobe
be estimated
estimated on city city roads.
roads.Following
Followingthethesamesame argument
argument of of [17],
[17], butbut
with with
the
the
aim aim of computing
of computing values
values overover timetimeandand precisely
precisely in space,
in space, the results
the results of theofpresent
the present
work
work
werewere validated
validated in the in
peakthetraffic
peak periods
traffic periods and,[17],
and, unlike unlike [17], the
the related related unknown
unknown CO2 emis-
CO
sions
2 emissions factors (𝑧)
factors 𝐾 (𝑧), 𝐾 1 were
K ( z ) , K ( z ) were computed for each timestamp,
2 computed for each timestamp, estimating the relation- estimating the
relationship between CO sensor measurements and actual traffic
ship between CO2 sensor2 measurements and actual traffic flow data in specific segments flow data in specific
segments
according according to Equation
to Equation (1). Figure (1). 9Figure
shows 9 shows the model
the model resultsresults in terms
in terms of MAPEof MAPEat COat2
CO sensor
sensor
2 locations in traffic periods when uncongested and congested
locations in traffic periods when uncongested and congested traffic situations traffic situations
arose.
arose.TheThemean
meanabsolute
absolutepercentage
percentageerror errorwas
wasclose
closeto to10%10%forforeach
eachsensor
sensorlocation
location and
and
season.
season.Annually,
Annually,thetherelated
relateduncertainty
uncertaintywas was9.1%,
9.1%,admitting
admittingaaminimumminimumofof5%. 5%.

Figure9.9.Multilinear
Figure Multilinear regression
regression model
model results
results in terms
in terms of MAPE
of MAPE according
according to each
to each sensor
sensor location
location and
and each period of the year. The results were obtained during the peak traffic periods when uncon-
each period of the year. The results were obtained during the peak traffic periods when uncongested
gested and congested traffic situations arise and significant CO2 emissions occur.
and congested traffic situations arise and significant CO2 emissions occur.

The unknown emission factors K1 (z) and K2 (z) for each sensor location and for each
period turned out to have different values in different seasons. Thus, the regressions results
were computed in different seasons, taking a range of days to collect different behaviours,
as described above. The summary of results in terms of emissions factor coefficients and
vehicular speeds is reported in Table 4 for a number of control points. As a result, the
maximum value of K2 (amount of gCO2 /km per car, in congested conditions) was obtained
Sensors 2022, 22, 3382 17 of 20

during the cold seasons, whereas the minimum value of K1 (amount of gCO2 /km per car
in uncongested conditions) occurred during the warm seasons. This makes sense, since in
winter cars tend to consume more due to the heating needs, whereas in warm conditions
heating is needed less. Generally, the average use of cars increases in the cold season
compared to the warm season, as in the latter citizens can choose other means of transport
such as bicycles, scooters, motorbikes, etc., which reduce the polluting impact. The mean
value reported in Table 4 can be considered a characterisation of traffic emissions in the
network during an uncongested and congested traffic situation for the seasons.

Table 4. Summarising numerical results by means of the applied model to identify the desired
emissions factors in case of uncongested and congested traffic situations in different periods of the
year for different sensor locations.

Winter Autumn
Air Sensor K1 K2 V1 V2 K1 K2 V1 V2
SMART09 230.0 681.3 37.4 3.9 317.7 791.5 36.5 3.9
SMART27 160.8 349.7 46.0 1.0 161.7 321.7 44.0 1.0
SMART28 219.6 386.7 36.0 1.0 253.1 352.3 35.0 1.0
SMART29 296.0 732.0 35.1 1.5 355.5 520.0 53.9 1.5
MEAN 226.6 537.4 38.6 1.8 272.0 496.3 42.3 1.8
Summer Spring
Air Sensor K1 K2 V1 V2 K1 K2 V1 V2
SMART09 184.0 709.4 39.8 3.9 217.8 619.2 35.2 3.9
SMART27 133.6 323.3 53.2 1.0 150.4 317.5 51.5 1.0
SMART28 290.3 383.2 36 1.0 274.0 381.5 34.0 1.0
SMART29 264.5 643.7 46.9 1.5 315.3 589.6 57.0 1.5
MEAN 218.1 514.9 43.9 1.8 239.3 476.9 44.4 1.8

Moreover, the present work takes into account only traffic emissions for the purpose
of the CO2 measured in the air. It is known that, in a real-world context, other emissions
influence the total amount of CO2 in the air, such as, for example, natural gas consumption
for domestic heating and cooking. The relationship between traffic emissions and the
amount of CO2 is considered the net of other influences in terms of emissions. Hence, the
present work should be seen as an upper boundary or the worst case of the influence of
the amount of CO2 in the context of the transport sector in the city of Florence, where the
measured CO2 emissions are only related to the vehicular burning of fossil fuels. From an
initial comparison, it seems that the study in [14], conducted about a decade ago in the same
observation area or in the immediate proximity, defined higher CO2 contributions in terms
of emissions factors under high congested conditions with respect to the ones estimated in
the present paper. This bodes well for policies to reduce the impact environmental CO2
combined with the evolution of vehicle engine technologies and the incremental use of
electric vehicles allowing air quality to be improved gradually.
From Equation (1), the estimated (mean) values of K2 and K1 can be also used to
compute the amount of CO2 at the traffic sensor locations, which are far from the air-quality
sensors. This allows the amount of CO2 emissions to be estimated in locations where
air-quality sensors are not placed. More precisely, the CO2 data at a given timestamp in a
given period of the year can be computed by means of both the (mean) emissions factors
estimated in that period (see Table 4) and the traffic data measured in that timestamp
according to Equation (1). Since Florence Municipality admits many traffic sensors in
scattered positions, the CO2 data can also be estimated in the whole area by applying
interpolation methods. For example, Figure 10 shows the estimated CO2 emissions in
the area of Florence at 8 a.m. on a spring workday by using the related traffic data at the
sensors’ locations depicted in Figure 2b.
 estimated in that period (see Table 4) and the traffic data measured in that timestamp
according to Equation (1). Since Florence Municipality admits many traffic sensors in scat-
tered positions, the CO2 data can also be estimated in the whole area by applying interpo-
Sensors 2022, 22, 3382
lation methods. For example, Figure 10 shows the estimated CO2 emissions in the 18 area of
of 20
Florence at 8 a.m. on a spring workday by using the related traffic data at the sensors’
locations depicted in Figure 2b.

Figure10.
10. An
An example
example of
ofaaCO
CO2 heatmap in Florence Municipality. An interpolation method is ap-
Figure 2 heatmap in Florence Municipality. An interpolation method
plied to the CO2 data, which are estimated in accordance with Equation (1) at the traffic sensor lo-
is applied to the CO2 data, which are estimated in accordance with Equation (1) at the traffic
cations depicted in Figure 2b. The traffic data are simultaneously measured in sparse locations at 8
sensor locations depicted in Figure 2b. The traffic data are simultaneously measured in sparse
a.m. on a spring workday. Accessible on Snap4city.org https://www.snap4city.org/dash-
locations at 8 a.m. on a spring workday. Accessible on Snap4city.org
boardSmartCity/view/index.php?iddasboard=MTUzMg== (accessed https://www.snap4city.org/
on 27 April 2022).
dashboardSmartCity/view/index.php?iddasboard=MTUzMg== (accessed on 27 April 2022).
7. Conclusions
7. Conclusions
The emission of CO2 is mainly due to fossil combustion, and therefore from traffic
The emission of CO2 is mainly due to fossil combustion, and therefore from traffic flow
flow
(vehicle (vehicle population,
population, season, season,
traffictraffic behaviour,
behaviour, etc.),
etc.), but but
the the contribution
contribution from from house-
household
hold heating produces a lower impact at ground level where
heating produces a lower impact at ground level where the CO2 sensors are placed and the CO 2 sensors are placed

and where
where peoplepeoplewalk. walk.
In most In medium
most mediumand largeand cities,
large cities,
traffic’straffic’s contribution
contribution to CO2toisCO very 2 is

very relevant.
relevant. The total
The total CO2 CO 2 production of a city could be computed by taking into account
production of a city could be computed by taking into account
the total fuel transformation ininenergy
the total fuel transformation energyusing
usingseveral
severalfactors
factorsandandefficiency
efficiencyaspects
aspectsofofthe the
population of vehicle engines, heating, industries, etc. Moreover,
population of vehicle engines, heating, industries, etc. Moreover, the amount of CO2 the amount of CO 2 pro-

duced bybyvehicles
produced vehiclescan canbebemore
morespecifically
specifically estimated
estimated by taking into
by taking into account
accountthe theweights
weights
moved,the
moved, thedistances,
distances,the theduration
durationofofthe thetrips
tripsand
andemissions
emissionsfactor factorofofeacheachspecific
specificvehicle
vehicle
ororcategory,
category,or orby
by taking
taking into
into account
account thethedistribution
distributionofofvehiclevehicle types
types in in
thetheregion
regionor city
or
without enabling a detailed computation of the CO in the city
city without enabling a detailed computation of the CO2 in the city areas. These approaches
2 areas. These approaches are
unsuitable for estimating and understanding the impact of vehicles in
are unsuitable for estimating and understanding the impact of vehicles in terms of CO2 in the terms of CO 2 in the city area,
since
city area,the vehicles
since produce
the vehicles CO2 depending
produce CO2 depending on their on specific efficiency,
their specific producer,
efficiency, producer, fuel,
weight,
fuel, driver
weight, style,
driver roadroad
style, conditions, etc., etc.,
conditions, which are typically
which are typically different in different
different areas
in different
of the
areas ofcity, and and
the city, thusthus
also also
in different manners
in different mannersin different seasons
in different of the
seasons ofyear. MostMost
the year. cities
havehave
cities a large number
a large numberof traffic flow sensors,
of traffic whereas
flow sensors, the number
whereas the number of CO2of sensors is limited.
CO2 sensors is
In the proposed
limited. model,model,
In the proposed we identified a method
we identified to determine
a method to determine whether vehicular
whether traffic
vehicular
behaviour
traffic behaviour is congested
is congestedor uncongested.
or uncongested. ThisThis
waswas the the
firstfirst
step, since
step, these
since twotwo
these traffic sit-
traffic
uations contribute
situations contributedifferently
differentlytotothe theamount
amountofofCO CO2 2emitted
emittedinto intothe theatmosphere.
atmosphere.
Therefore,we
Therefore, weidentified
identifiedaamethodmethodfor forestimating
estimatingthe theemissions
emissionsfactor factorinindifferent
different
traffic
trafficconditions.
conditions.This Thisallowed
allowedus ustotoidentify
identifyaanew newmodelmodeland andmethod
methodfor forcomputing
computing
COCO 2 2on
onthe thebasis
basisofoftraffic
trafficflow
flowdata
datawith
withrespect
respecttotothe thestate
stateofofthe theart.
art.Therefore,
Therefore,the the
approach
approachpresented
presentedininthe thepaper
paperproposed,
proposed,asasbasicbasicelements
elementsofofthe thesolution,
solution,(i)(i)a amethod
method
for
fordetermining
determiningthe theemission
emissionfactors,
factors,taking
takinginto
intoaccount
accountdifferent
differenttraffic
trafficbehaviours,
behaviours,from from
fluid traffic to “stop-and-go” (uncongested and congested,
fluid traffic to “stop-and-go” (uncongested and congested, respectively); (ii) respectively); (ii) an analysis
an analysis of
changes in the emissions factors in different periods of the year; and (iii) a validation of
the estimation of CO2 from traffic flow data. The validation was performed in the City of
Florence, where a number of CO2 sensors and traffic flow sensors are located. In some cases,
they were collocated, thus creating good conditions for validation. In order to understand
the actual impact of traffic density in such locations, a traffic flow reconstruction algorithm
was applied to estimate the traffic volume in certain road segments of the network, which
Sensors 2022, 22, 3382 19 of 20

were the closest ones to the identified air-quality sensors for validation. The proposed
approach can be easily replicated in other cities to compute their emissions factors according
to the characterisation of the traffic, vehicle population, and behaviour in different parts of
the city and moment in time. From a long-term perspective, the assessment of emissions
factors can provide useful information to verify the effective reduction of CO2 emissions
and the impact of traffic flow due to the push towards electric vehicles. The approach
should also be tested for the estimation of other pollutants produced by vehicles, first
by identifying their corresponding emissions factor with the method presented in this
paper. Moreover, the study could also evolve by taking into account the weather conditions,
which could influence the propagation of emissions based on the wind and humidity. In
the proposed model, those aspects are included in some measure in the seasonal changes
of the emissions factors.

Author Contributions: Conceptualization, S.B. and P.N.; methodology, S.B. and P.N.; software, S.B.;
validation, S.B. and P.N.; formal analysis, S.B.; investigation, S.B. and P.N.; resources, P.N.; data
curation, S.B. and P.N.; writing—original draft preparation, S.B. and P.N.; writing—review and
editing, S.B. and P.N.; visualization, S.B. and P.N.; supervision, S.B. and P.N.; project administration,
P.N.; funding acquisition, P.N. All authors have read and agreed to the published version of the
manuscript.
Funding: This work was supported in part by MIUR, the University of Florence, and companies
involved in the co-founding of the Sii-Mobility national project (SCN_00112) on smart city mobility
and transport. Km4City is an open technology and research of DISIT Lab. Sii-Mobility is grounded
and has contributed to the Km4City open solution. The present solution is also adopted in the
Snap4City platform and by the Trafair CEF project (2017-EU-IA-0167).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data accessible on the Snap4City platform. Available online: https:
//www.snap4city.org (accessed on 27 April 2022).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Badii, C.; Bilotta, S.; Cenni, D.; Difino, A.; Nesi, P.; Paoli, I.; Paolucci, M. High Density Real-Time Air Quality Derived Services
from IoT Networks. Sensors 2020, 20, 5435. [CrossRef] [PubMed]
2. Badii, C.; Bilotta, S.; Cenni, D.; Difino, A.; Nesi, P.; Paoli, I.; Paolucci, M. Real-Time Automatic Air Pollution Services from
IOT Data Network. In Proceedings of the 2020 IEEE Symposium on Computers and Communications (ISCC), Rennes, France,
7–10 July 2020; pp. 1–6. [CrossRef]
3. Smit, R.; Ntziachristos, L.; Boulter, P. Validation of road vehicle and traffic emission models—A review and meta-analysis.
Atmos. Environ. 2010, 44, 2943–2953. [CrossRef]
4. Smit, R.; Poelman, M.; Schrijver, J. Improved road traffic emission inventories by adding mean speed distributions.
Atmos. Environ. 2008, 42, 916–926. [CrossRef]
5. Abou-Senna, H.; Radwan, E. VISSIM/MOVES integration to investigate the effect of major key parameters on CO2 emissions.
Transp. Res. Part D Transp. Environ. 2013, 21, 39–46. [CrossRef]
6. European Environment Agency. EMEP/EEA Air Pollutant Emission Inventory Guidebook 2019. Available online: https:
//www.eea.europa.eu/publications/emep-eea-guidebook-2019 (accessed on 27 April 2022).
7. Todts, W. Transport & Environment, European Federation for Transport and Environment AISBL. 2018. Available online:
https://www.transportenvironment.org/ (accessed on 27 April 2022).
8. Zhu, C.; Dawei, G. A research on the factors influencing Carbon emission of transportation industry in “the Belt and Road
Initiative” countries based on panel data. Energies 2019, 12, 2405. [CrossRef]
9. Smit, R.; Brown, A.L.; Chan, Y.C. Do air pollution emissions and fuel consumption models for roadways include the effects of
congestion in the roadway traffic flow? Environ. Model. Softw. 2008, 23, 1262–1270. [CrossRef]
10. Madireddy, M.; De Coensel, B.; Can, A.; Degraeuwe, B.; Beusen, B.; De Vlieger, I.; Botteldooren, D. Assessment of the impact
of speed limit reduction and traffic signal coordination on vehicle emissions using an integrated approach. Transp. Res. Part D
Transp. Environ. 2011, 16, 504–508. [CrossRef]
11. Grote, M.; Williams, I.; Preston, J.; Kemp, S. Including congestion effects in urban road traffic CO2 emissions modelling: Do Local
Government Authorities have the right options? Transp. Res. Part D Transp. Environ. 2016, 43, 95–106. [CrossRef]
Sensors 2022, 22, 3382 20 of 20

12. Choudhary, A.; Gokhale, S. Urban real-world driving traffic emissions during interruption and congestion. Transp. Res. Part D
Transp. Environ. 2016, 43, 59–70. [CrossRef]
13. Fontaras, G.; Valverde, V.; Arcidiacono, V.; Tsiakmakis, S.; Anagnostopoulos, K.; Komnos, D.; Pavlovic, J.; Ciuffo, B. The
development and validation of a vehicle simulator for the introduction of Worldwide Harmonized test protocol in the European
light duty vehicle CO2 certification process. Appl. Energy 2018, 226, 784–796. [CrossRef]
14. Zhao, H.-X.; He, R.-C.; Yin, N. Modeling of vehicle CO2 emissions and signal timing analysis at a signalized intersection
considering fuel vehicles and electric vehicles. Eur. Transp. Res. Rev. 2021, 13, 5. [CrossRef]
15. Gao, Y.; Hu, H. Estimation of queued vehicle emissions at signalized intersections based on vehicle specific power approach.
China J. Highw. Transp. 2015, 28, 101–108.
16. Vaccari, F.P.; Matese, A.; Gioli, B.; Zaldei, A.; Miglietta, F. Carbon Dioxide Emissions of the City Center of Firenze, Italy:
Measurement, Evaluation, and Source Partitioning. J. Appl. Meteorol. Clim. 2009, 48, 1940–1947. [CrossRef]
17. Nejadkoorki, F.; Nicholson, K.; Lake, I.; Davies, T. An approach for modelling CO2 emissions from road traffic in urban areas.
Sci. Total Environ. 2008, 406, 269–278. [CrossRef] [PubMed]
18. Transport Research Laboratory. UK Road Transport Emission Data; Transport Research Laboratory: Leeds, UK, 2005.
19. Bellini, P.; Bilotta, S.; Cenni, D.; Collini, E.; Nesi, P.; Pantaleo, G.; Paolucci, M. Long Term Predictions of NO2 Average Values via
Deep Learning. In Proceedings of the International Conference on Computational Science and Its Applications—ICCSA 2021,
Cagliari, Italy, 13–16 September 2021; pp. 595–610. [CrossRef]
20. Po, L.; Rollo, F.; Viqueira, J.R.R.; Lado, R.T.; Bigi, A.; Lopez, J.C.; Paolucci, M.; Nesi, P. TRAFAIR: Understanding Traffic Flow
to Improve Air Quality. In Proceedings of the 2019 IEEE International Smart Cities Conference (ISC2), Casablanca, Morocco,
14–17 October 2019; pp. 36–43.
21. Sii-Mobility, Italian National Mobility and Transport Action for Sustainable Mobility, Founded by MIUR. Available online:
http://www.sii-mobility.org/ (accessed on 27 April 2022).
22. Km4City Ontology, Knowledge Base 4 the City. Available online: https://www.km4city.org (accessed on 27 April 2022).
23. Bellini, P.; Bilotta, S.; Cenni, D.; Nesi, P.; Paolucci, M.; Soderi, M. Knowledge Modeling and Management for Mobility and
Transport Applications. In Proceedings of the 2018 IEEE 4th International Conference on Collaboration and Internet Computing
(CIC), Philadelphia, PA, USA, 18–20 October 2018; Volume 1, pp. 364–371.
24. Badii, C.; Belay, E.; Bellini, P.; Cenni, D.; Marazzini, M.; Mesiti, M.; Nesi, P.; Pantaleo, G.; Paolucci, M.; Valtolina, S.; et al.
Snap4City: A Scalable IOT/IOE Platform for Developing Smart City Applications. In Proceedings of the 2018 IEEE SmartWorld,
Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud &
Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI),
Guangzhou, China, 8–12 October 2018; pp. 2109–2116.
25. Snap4City. Available online: https://www.snap4city.org (accessed on 27 April 2022).
26. Bilotta, S.; Nesi, P. Traffic flow reconstruction by solving indeterminacy on traffic distribution at junctions. Future Gener. Comp.
Syst. 2021, 114, 649–660. [CrossRef]
27. Bellini, P.; Bilotta, S.; Nesi, P.; Paolucci, M.; Soderi, M. Real-Time Traffic Estimation of Unmonitored Roads. In Proceed-
ings of the 2018 IEEE 16th Intl Conf on Dependable, Autonomic and Secure Computing, 16th Intl Conf on Pervasive
Intelligence and Computing, 4th Intl Conf on Big Data Intelligence and Computing and Cyber Science and Technology
Congress(DASC/PiCom/DataCom/CyberSciTech), Athens, Greece, 12–15 August 2018; pp. 935–942.
28. Mott, R.L. Applied Fluid Mechanics; Prentice Hall: Hoboken, NJ, USA, 2006.
29. Lighthill, M.J.; Whitham, G.B. On kinematic waves II. A theory of traffic flow on long crowded roads. Proc. R. Soc. Lond. Ser. A
Math. Phys. Sci. 1955, 229, 317–345.
30. Richards, P.I. Shock Waves on the Highway. Oper. Res. 1956, 4, 42–51. [CrossRef]
31. Bretti, G.; Natalini, R.; Piccoli, B. A Fluid-Dynamic Traffic Model on Road Networks. Arch. Comput. Methods Eng. 2007, 14,
139–172. [CrossRef]
32. Godunov, S.K. A finite difference method for the numerical computation of discontinuous solutions of the equations of fluid
dynamics. Math. Sb. 1959, 47, 271–290.
33. Kachroo, P.; Sastry, S. Travel Time Dynamics for Intelligent Transportation Systems: Theory and Applications. IEEE Trans. Intell.
Transp. Syst. 2016, 17, 385–394. [CrossRef]