Networked Low-Cost Sensor Systems for Urban Air Quality Monitoring: A Long-Term Use-Case in Bari (Italy)

Abstract

A sensor network based on 10 stationary nodes distributed in Bari (Southern Italy) has been deployed for urban air quality (AQ) monitoring. The low-cost sensor systems have been installed in specific sites (e.g., buildings, offices, schools, streets, ports, and airports) to enhance environmental awareness of the citizens and to supplement the expensive official air-monitoring stations with cost-effective sensor nodes at high spatial and temporal resolution. Continuous measurements were performed by low-cost electrochemical gas sensors (CO, NO₂, O₃), optical particle counter (PM₁₀), and NDIR infrared sensor (CO₂), including micro-sensors for temperature and relative humidity. The sensors are operated to assess the performance during a campaign (July 2015–December 2017) of several months for citizen science in sustainable smart cities. Typical values of CO₂, measured by distributed nodes, varied from 312 to 494 ppm (2016), and from 371 to 527 ppm (2017), depending on seasonal micro-climate change and site-specific conditions. The results of the AQ-monitoring long-term campaign for selected sensor nodes are presented with a relative error of 26.2% (PM₁₀), 21.7% (O₃), 25.5% (NO₂), and 79.4% (CO). These interesting results suggest a partial compliance, excluding CO, with Data Quality Objectives (DQO) by the European Air Quality Directive (2008/50/EC) for Indicative (Informative) Measurements.

1. Introduction

1.1. Background and Motivation

Urban air quality is a major issue for public health, especially in megacities. The UN Environment Program Report 2016 [1] reported that many worldwide cities do not meet the requirements for cleaner air. Air pollution is still the principal environmental cause of premature deaths in the European Union, as reported by the European Environment Agency (EEA) Report 2017 [2] and Report 2024 [3]. In fact, air pollution is causing more than 400,000 premature deaths each year. Furthermore, millions of Europeans suffer from respiratory and cardiovascular diseases caused by air pollution: NO₂, O₃, SO₂, and particulate matter (PM) are identified as key species affecting quality of life and mortality rates, as reported by the World Health Organization (WHO) Report 2006 [4]. Also, CO, NO₂, and fine particulate matter are known to affect respiratory and cardiovascular diseases. The costs of air pollution in the European Union are well over €20 billion a year. To mitigate these negative effects of air pollution on human health, environment, and eco-systems, the European Commission adopted the Ambient Air Quality Directive (2008/50/EC) [5] and the successive Directive EU 2024/2881 of the European Parliament and of the Council of 23 October 2024 on Ambient Air Quality and Cleaner Air for Europe [6] that fixed stringent thresholds for each pollutant.

Currently, low-cost sensors (LCS) are gaining importance in monitoring air quality at high spatial and temporal resolutions [7]. In fact, the World Meteorological Organization (WMO) supported LCS for the measurement of atmospheric composition as a future application for environmental monitoring [8]. Also, the US Environmental Protection Agency (EPA) deliberated standards and performance targets for air-quality sensors [9]. A unified terminology for LCS has been proposed to label different levels of sensor systems and sensor networks [10]. Some limitations in air sensors’ performance still exist, such as sensitivity, selectivity, stability, limit of detection, drift, repeatability, and so on. Current trends to overcome these low operations explore functional materials, nanomaterials, and nanotechnologies [11], and advanced and hybrid transducers [12].

Furthermore, climate change monitoring is crucial to implement abatement strategies of greenhouse gases and CO₂ emissions for a low-carbon footprint in sustainable cities. In recent years, low-cost sensor systems [13,14,15] and sensor networks [16,17,18,19] for air-quality monitoring have been demonstrated in the real world for performance assessment, environmental awareness, personal exposure, compliance, legislative purposes, and decision support. In this study, we have performed a long-term use case using air sensors deployed in Bari city, Italy.

1.2. Case-Studies of City Sensor Networks for Air-Quality Monitoring

In recent years, an increasing number of case studies related to the use of portable sensors and sensor networks for urban air-quality monitoring have been demonstrated. The MESSAGE (Mobile Environment Sensor System Across GRID Environments) project, funded by the UK during 2006–2009, included small cost sensors (CO, NO₂, and NO) to increase the amount of data available for modelling pollution emission and dispersion [20,21]. The University of Cambridge, Newcastle University, and Imperial College London were key research leaders. Furthermore, the SNAQ-Heathrow (Sensor Network Air Quality-Heathrow) project has deployed a high-density air quality sensor network in and around London Heathrow airport. This used state-of-the-art low-cost sensors for selected gases and for particulates, providing an unprecedented dataset. The project was funded by NERC from January 2011 to December 2014 [22]. The objective was the creation of a calibrated, high spatial and temporal resolution dataset including NO, NO₂, CO, CO₂, SO₂, O₃, VOCs, and size-speciated PM for further scientific and policy studies. Furthermore, the OPEN-SENSE project analyzed the largest spatially resolved UFP (Ultra Fine Particle) dataset containing over 50 million measurements. The project [23] collected the measurements throughout more than two years using mobile sensor nodes installed on top of public transport vehicles in the cities of Zurich and Basel, Switzerland. Based on these data, land-use regression models have been developed to create pollution maps with a high spatial resolution of 100 mt × 100 mt. Fishbain et al. [24] reviewed the mathematical tools available for assessment of the performance of air quality micro-sensing units using data collected in eight European cities of the FP7 project CITI-SENSE. Recently, the Copenhagen Wheel developed an electric bike technology living inside a red hub connected to a smartphone. There are no buttons, wires, or external batteries. The Wheel’s custom sensors, computers, battery, and motor do the rest for measuring air quality in Copenhagen [25]. L. Sun et al. [26] developed a sensor network used to monitor air quality during a marathon in 2015 in Hong Kong. In Beijing, China, a city network based on 2500 sensor nodes for PM_2.5 detection is operated by the local environmental authority for decision support. J. Zhang et al. [27] developed a portable farmland information collection system with multiple sensors for air-quality monitoring in China. S. Choi et al. [28] developed a micro-sensor node for air-pollutant monitoring in Korea. T.F. Villa et al. [29] developed an Unmanned Aerial Vehicle (UAV)-based system for air pollution measurements in Australia. J. Kim et al. [30] developed a sensor network based on 50 nodes distributed at Berkeley (California) for air-quality monitoring (CO, NO, NO₂, O₃, CO₂, and PM₁₀). The CLAIRITY project [31] developed an air quality network near the Massachusetts Institute of Technology (MIT) in Boston, USA.

Further case studies of low-cost sensor networks deployed for urban air-quality monitoring have been extensively investigated worldwide. In Europe, R.L. Jones’ team [32] studied a wireless sensor network (WSN) for source attribution of air pollution and operated in Cambridge city, UK; V. van Zoest et al. [33] investigated a field-calibrated NO₂-based WSN in Eindhoven city, Holland. Kaminska et al. [34] reported a study conducted in Wrocław (Poland) in October 2021, combining data on eBC (equivalent of black carbon) concentration (measured by microaethalometer), air quality (from the national environmental monitoring system), and traffic (from the municipal traffic management system). The aim of this research was to estimate, by an indirect approach, the relation between eBC concentration and other air pollutants in order to create more sustainable cities. Also, in Israel, Yuval et al. [35] demonstrated an optical particle counter (OPC) network for assessment of emission impact as operated in Elad city.

In the Americas, D.H. Hagan et al. [36] studied the calibration and assessment of the electrochemical air quality sensors by co-location with regulatory-grade instruments in Hawaii, USA. S.N. Feinberg et al. [37] examined the spatiotemporal variability of urban particulate matter using data from a network of low-cost air pollution sensors operated in the Memphis area, USA. Kimbrough et al. [38] proposed a Local-Scale Air Quality Study integrating low-cost sensors and reference-grade monitoring in a complex metropolitan area in Kansas City, USA. G. Miskell et al. [39] demonstrated reliable data from low-cost ozone sensors in a hierarchical network as deployed in the Los Angeles area, USA. A. Collier-Oxandale et al. [40] evaluated the field and laboratory performance of 28 gas-phase air quality sensors by the AQ-SPEC (Air Quality Sensor Performance Evaluation Centre) programme as deployed in the Los Angeles area, USA. Also, in Brazil, D. da Silva Sousa et al. [41] developed an automated, self-powered, and integrated analytical platform for online and in situ air-quality monitoring as tested in two Brazilian cities from the San-Paulo region.

In Asia, Wei et al. [42] studied the impact of temperature and humidity on electrochemical sensors’ (CO, NO, NO₂, and O₃) response in ambient air-quality monitoring in Hong Kong, China. Wang et al. [43] demonstrated the application of machine learning for the in-field correction of a PM_2.5 low-cost sensor network in Taipei, Taiwan. Cao et al. [44] deployed a distributed air-sensor network to investigate the spatiotemporal patterns of PM_2.5 concentrations in Dezhou city, China.

1.3. Ambient Air Quality EU Directive

The Ambient Air Quality EU Directive (2008/50/EC) defines two different regimes of measurements: (a) Fixed Measurements and (b) Indicative Measurements (Informative Measurements). The former mean measurements were taken at fixed sites to determine the air pollution levels in accordance with the relevant Data Quality Objectives (DQO), provided by reference instrumentation only. The fixed measurements are mandatory in zones and agglomerations where the upper assessment thresholds are exceeded. The latter mean measurements meet DQOs that are less strict than those required for fixed measurements. The DQOs of the Indicative Measurements can be addressed by air quality low-cost sensors. The uncertainty for fixed measurements is as follows: 15% (NO_x/NO₂/NO, CO, SO₂); 25% (Benzene); 15% (O₃); and 30% (PM₁₀/PM_2.5). On the contrary, the uncertainty for Indicative Measurements is as follows: 25% (NO_x/NO₂/NO, CO, SO₂); 30% (Benzene); 30% (O₃); and 50% (PM₁₀/PM_2.5). The EU Directive does not include CO₂ as a compound of interest, but CO₂ is an indicator for traffic and greenhouse emissions. A rough estimation of the uncertainty should range from 5 to 10%, as provided by the actual limits of the CO₂ sensor performance for real applications.

Further details on uncertainties that are useful for Indicative Measurements to be achieved by low-cost sensors for air-quality monitoring are reported in Appendix B, as recently updated by Directive of the European Parliament and of the Council on Ambient Air Quality and Cleaner Air for Europe [6], dated 14 October 2024 (PE-CONS 88/24).

1.4. Aim of Work

Bari is a metropolitan city in Southern Italy of about five hundred thousand people (ninth position in Italy and third position in Southern Italy) with an economy based on commerce, manufacturing, and a knowledge economy. The transportation system is globally inter-operative, with a local surface network, port, and international airport. Bari is the headquarters of two Public Universities (Sciences, Engineering, Literature, Social, Law, Economy, and Medicine) with about sixty thousand students. The city is a member of the European Covenant of Mayors for Climate and Energy. This urban eco-system is a perfect real scenario for a long-term campaign for the air quality sensor network.

The present study on air quality sensor networks for urban monitoring is strongly linked to previous works from the author team related to the application of low-cost sensors in urban areas such as city centres, industrial zones, and airports [45,46,47].

In this study, the performance of the sensor network for urban air-quality monitoring has been assessed in the city of Bari by 30-month long-term measurements. The sensor data is compared to the reference environmental public data to evaluate the accuracy and data quality objectives of the European Air Quality Directive (2008/50/EC) for Indicative Measurements.

2. Materials and Methods

2.1. Sensor Nodes

The sensor nodes, AIRBOX, shown in Figure 1, were designed to be small in size (23 cm width × 30 cm height × 10 cm depth), thermoplastic weatherproof (IP 65/66), able to operate on their own, and connected to power. The compact size of the sensor node was useful to facilitate deployment and reduce the installation footprint. Each sensor node includes a suite of 9 low-cost sensors: 4 electrochemical gas sensors: NO₂, O₃, CO, and SO₂ (not considered in this study); 1 Near Dispersive Infra-Red (NDIR) sensor: CO₂; 1 Photo-Ionization Detector (PID): total Volatile Organic Compounds (tVOCs); 1 Particulate Matter sensor: PM₁₀; 1 temperature (T) sensor; and 1 relative humidity (RH) sensor. The components are listed in

Table 1. List of gas/particulate matter and environmental sensors installed in each AIRBOX sensor node for urban air-quality monitoring.

Pollutant	Sensor Type	Manufacturer	Characteristics	Principle of Operation
CO	CO-B4	Alphasense, UK	0–20 ppm	Electrochemical
NO2	NO2-B4	Alphasense, UK	0–2 ppm	Electrochemical
O3	O3-B4	Alphasense, UK	0–2 ppm	Electrochemical
SO2	SO2-B4	Alphasense, UK	0–2 ppm	Electrochemical
CO2	CO2-IRC-A1	Alphasense, UK	0–5000 ppm	NDIR
VOCs	VOCs-PID-A1	Alphasense, UK	0–100 ppm	Photo-ionization
PM10	PPD20V	Shinyei, Japan	0–100 µg/m3	Light scattering
Temperature	TC1047A	Microchip	−40 ± 125 °C	Thermo-converter
Relative Humidity	HIH5031	Honeywell	0–90%	Capacitive

. These sensors were selected due to their low-power consumption (<10 mW for the electrochemical sensor, and <60 mW for the NDIR sensor) and their low-voltage operation (3.3–5.0 Vcc), including good sensitivity and low limit of detection at the ppb level of environmental interest.

The sensor node incorporates a Raspberry Pi-based computer (Model B+) with 4 USB ports, 1 Ethernet port, and 512 Mb RAM, using the open-source operating system Linux for data acquisition management by an SD card (2 Gb) for on-board data logging.

The remote node uses a GPRS modem (Key HSPA Huawei E303) with a SIM card, operated by a public provider, to transmit data to a central server (Dell, Model PowerEdge R320, 6 core-Intel Xeon, 16 Gb RAM, 6 discs by 500 Gb in raid5 configured) with a pack sampling rate to be configured properly.

Additionally, a Wi-Fi key (Edimax USB 2.0, 802.11, 2.4 GHz, 150 Mbit/s) is installed for city operation covered by a digital network. Finally, a Global Positioning System (GPS) receiver (Garmin, Model GPS 18× USB) was used for convenient deployment, mainly for mobile sensing purposes.

2.2. Sensor Network

A sensor network based on 10 AIRBOX stationary nodes has been deployed for urban air-quality monitoring in the city of Bari (Southern Italy).

The sensor nodes have been installed in the external areas (roofs, windows, terraces, and balconies) of the specific sites (buildings, offices, schools, streets, ports, and airports), selecting the available operating site as close as possible to a reference air-quality monitoring station managed by ARPA-Puglia to validate sensor data.

All nodes, shown in Figure 2, are deployed in an urban area of 24 km²: 6 nodes are located in the city area with a mean distance of less than 1 km from the Adriatic sea; and 4 nodes are located in the tangential road area at high vehicular traffic, including 3 nodes operated in the industrial zone.

Table 2. List of the sensor nodes deployed in Bari city and managed by ENEA in the period of the campaign from 1 July 2015 to 31 December 2017. The air-quality monitoring stations of the local environmental agency (ARPA-Puglia) closest to the deployed sensor nodes are indicated with the related Universal Transverse Mercator (UTM) coordinate system.

Node ID	Node Location	Sensor Node Coordinates		Closest ARPA-Puglia Air Quality Monitoring Station	AQM Station UTM33 Coordinates		Distance Sensor Node and AQM Station (km)
		Long. East (UTM33)	Lat. North (UTM33)		Long. East	Lat. North
Node 1	ENELVia Capruzzi	16.87370 (657305)	41.11720 (4553459)	ARPAViale Kennedy	656105	4551478	2.1
Node 2	ENEAViale Japigia	16.89963 (659494)	41.11259 (4552995)	ARPAVia Caldarola	658520	4553079	1.4
Node 3	PERONIVia Bitritto	16.84361 (654830)	41.09558 (4551005)	ARPAModugno EN4	650120	4553064	6.1
Node 4	AMTABVia Jacobini	16.80469 (651516)	41.11529 (4553125)	ARPAModugno EN4	650120	4553064	6.0
Node 6	AIRPORT WojtylaViale Ferrari	16.77646 (649090)	41.13979 (4555796)	ARPAModugno EN2	648305	4555516	8.8
Node 7	MOBILEon a public bus	-	-	-	-	-
Node 8	POLITECNICOVia Amendola	16.88335 (658130)	41.11126 (4552817)	ARPAVia Caldarola	658520	4553079	0.8
Node 9	COFELYStrada Vassallo	16.87113 (657148)	41.09260 (4550724)	ARPACarbonara	654377	4598816	2.2
Node 10	UFFICI COMUNALIPiazza Ferrarese	16.87206 (657145)	41.12676 (4554518)	ARPAVia Cavour	657197	4554020	0.8
Node 11	FAIRLungomare Starita	16.839167 (654361)	41.136389 (4555528)	ARPACUS Bari	654877	4555353	0.7
Node 12	PORTMolo San Vito	16.5100 (626852)	41.0748 (4548159)	ARPACUS Bari	654877	4555353	0.7

summarizes the location of sensor nodes to be referenced with the closest air quality station.

2.3. Virtual Private Network and Server

A Virtual Private Network (VPN) has been implemented to communicate by secure digital certificate with each remote node (nasuspi-x) using TCP/IP protocol (IPv4). The VPN scheme is shown in Figure 3. The software used for data acquisition management is open-source Python (version 3.11). The sampling rate for each node is 10 s. The data are locally stored in each node and transmitted every 60 min towards the central server (Dell, Model PowerEdge R320, 6 core-Intel Xeon, 16 Gb RAM, 6 discs by 500 Gb in raid5 configured). Data from each remote node has been uploaded and stored on the central server, hosted by ENEA, and displayed on a private web page accessed with signed credentials. The web page displays data in a tabular format to support the direct download. The data acquired by each virtual node (nasuspi-x) is useful for automatic post-processing in the dedicated webGIS page.

2.4. Global Information System

Figure 4 shows the screenshot of the web Global Information System to manage the visualization of the sensor network in the framework of the national project RES-NOVAE, including measurement setup for individual nodes, selection of node to be visualized, mean at hourly, daily, monthly, and yearly steps for individual AQI, and related absolute value of gas concentration of any measured pollutant in restricted areas. Moreover, it is possible to select data, both in tabular and graphic form, related to the period of the measurements of interest for analysis and post-processing.

Finally, an APP RES-NOVAE in the Android environment has been developed to visualize the AQI and gas concentration of a given specific node upon daily format.

2.5. Sensor Calibration

The air quality sensors used in the Bari campaign consisted of commercial low-cost electrochemical sensors for gas (NO₂, O₃, and CO) detection and a commercial cost-effective optical particle counter (OPC) for particulate matter (PM₁₀) detection, including miniaturized sensors for meteorology parameters (T and RH). Usually, these chemical sensors (gases and PM) arrive uncalibrated from the manufacturer, and thus, it is necessary to calibrate using a dedicated reference analyzer. Laboratory evaluation and in-field calibration are important to convert the output sensor signal (V) into gas concentration (ppb) and into PM concentration (µg/m³). The lab calibration is a preliminary test before a more accurate in-field calibration to minimize the cross-contamination effects and environmental interferent (T and RH) compensation. Therefore, before the Bari campaign (October 2014), ENEA calibrated the prototypes by co-located measurements with reference analyzers in an official air-quality monitoring station at JRC, located in Ispra, Italy (45°48.881′ N, 8°38.165′ E). The in-field calibration by the co-location of sensors and reference equipment took place in Ispra from 28 January to 8 June 2014 for NO₂, O₃, CO, SO₂, and PM₁₀ in order to assess the mean error of the sensors compared to the reference instrument. During this measurement period, the NO₂ concentrations varied from 5 to 40 ppb with a mean error ranging from 3 to 15 ppb, depending on environmental conditions and measured concentration range. The O₃ concentrations varied from 10 to 50 ppb with a mean error ranging from 5 to 20 ppb, depending on environmental conditions and measured concentration range. The CO concentrations varied from 200 to 600 ppb, with a mean error ranging from 90 to 170 ppb, depending on the environmental conditions and measured concentration range. The SO₂ concentrations were very low in the rural area at Ispra, and no estimation of the concentration mean error was applicable. The PM₁₀ concentrations varied from 5 to 68 µg/m³ with a mean error ranging from 3 to 10 µg/m³, depending on environmental conditions and the measured concentration range.

The in-field calibration of the sensors, already calibrated in other similar sites at Ispra, has been performed in Bari by reference instrumentation of the closest air-quality monitoring station. The calibration curves have been adapted for the on-site campaign at Bari. The air sensors are sensitive to temperature and humidity changes, as expected. A weak, not a strong, reaction of the deployed air sensors with ambient conditions has been measured in the Bari campaign; therefore, we do not exclude sensor data that are affected under very humid conditions (>75%), as reported by Mueller et al. [48].

Finally, the field calibration of the used air sensors has been adapted at the Bari sites by an initial period of training and successive optimization depending on the specific location where the air-sensor box was deployed for experimentation.

For electrochemical gas sensors, the concentration C (ppb) was calculated as follows:

$$C=\mathrm{S}\left[\right(V_w-{\mathrm{V}}_{w0})-(V_a-{\mathrm{V}}_{\mathrm{a}0}\left)\right]$$

(1)

where S is sensor sensitivity (ppb/V or ppm/V); V_w is the working electrode voltage (V); V_a is the auxiliary electrode voltage (V); V_w0 is the zero-air working electrode voltage (V); and V_a0 is the zero-air auxiliary electrode voltage (V). The auxiliary electrode voltages affect the sensitivity of the electrochemical gas sensors ranging from 10 to 100 nA/ppm, depending on gas concentration under test and environmental ambient conditions.

For particulate matter sensors, the concentration c (µg/m³) was calculated as follows:

$$c={\mathrm{A}}_0+\mathrm{S}\times \mathrm{V}$$

(2)

S is the sensor sensitivity [(µg/m³)/V]; A₀ is a bias constant (µg/m³); V is the sensor output voltage (V).

2.6. Methodology

The sensors have been deployed in Bari city at selected sites for a long-term campaign of urban air-quality monitoring. The criteria for selection of the demonstration sites are as follows: (i) pragmatic availability to host a sensor node on the roof of the offices, buildings, schools, port, airport, and city centre by project partners; (ii) power availability; (iii) coverage of the urban area; and (iv) closeness of the sensor-site to an air quality monitoring station (AQMS) by ARPA-Puglia in the Bari city. The map of the distributed sensor nodes and the related network of the reference AQMS is reported in Figure 5. The campaign started on 1 July 2015 and stopped on 31 December 2017. The huge database is composed of about 5 Gb of big data. The data of the sensor network are managed by means of ENEA VPN, WebGIS, and a server, while the public official environmental data were periodically downloaded from the official site of ARPA-Puglia [49]. The final aim of the campaign was the validation of the sensor data compared to the reference data for performance assessment.

3. Results

3.1. Stationary Sensor Network Versus Reference Methods

Figure 6, Figure 7, Figure 8, Figure 9, Figure 10 and Figure 11 show the time series of the daily mean of a given sensor node (Node 1: ENEL) compared to the closest air quality monitoring station (AQMS) in Bari city for the selected months of August 2015, December 2015, March 2016, October 2016, April 2017, and December 2017. The previously mentioned six months cover all four seasons during the experimental campaign (1 July 2015–31 December 2017) in order to study the different climate conditions affecting urban air pollution. The monitored air pollutants, including PM₁₀, NO₂, O₃, and CO, have been selected due to official public reference data available from the Environmental Protection Agency. The comparison of the performance between the sensor node and reference analyzer is expressed by a simple figure of merit, the Mean Absolute Error (MAE). It varies for PM₁₀ from 4.88 μg/m³ (August 2015) to 14.67 μg/m³ (December 2017); for NO₂ from 9.22 ppb (October 2016) to 15.18 ppb (April 2017); for O₃ from 4.81 ppb (December 2016) to 16.62 ppb (August 2015); and, finally, for CO from 0.12 ppm (April 2017) to 0.40 ppm (October 2016). These results indicate a good performance of the PM₁₀, O₃, and CO sensors of Node 1, while a worse performance has been recorded for the NO₂ sensor. Additionally, the official rate of the released air-pollutant concentration is the daily mean, so no real variability during the whole day can be appreciated by this time resolution. Moreover, the CO and NO₂ concentration during the month under investigation, as measured by the Node 1 (ENEL) in Figure 6, Figure 7, Figure 8, Figure 9, Figure 10 and Figure 11, is almost constant due to the location of the sensor-node (Node 1) near an urban street at high traffic, as referenced by the closest air-quality monitoring station (ARPA-Viale Kennedy). Furthermore, the PM₁₀ peak, recorded on 23 March 2016 (see Figure 8) as 78 μg/m³ measured by the sensor and 81 μg/m³ measured by the analyzer, is attributed to Saharan dust in Bari city, as certified by ARPA-Puglia.

3.2. Sensor Network Performance

Figure 12 shows the 2016 and 2017 yearly means of the PM₁₀, NO₂, O₃, and CO of the sensor network (eight nodes) compared to the selected air quality monitoring station (AQMS) in Bari city. The ARPA-Viale Kennedy AQMS is of the urban background in the city centre and is equipped with PM₁₀, NO₂, and O₃ analyzers, but without a CO analyzer, referenced alternately by the ARPA-Via Cavour AQMS located in the city centre and categorized as an urban traffic station. The mean of the selected reference analyzers has been chosen in June 2016 and June 2017 as annual mid-term.

For PM₁₀ assessment, all nodes, excluding Node 2 in 2017, show a yearly mean comparable to the reference mean, exhibiting excellent accuracy in long-term operations. For NO₂ assessment, generally, the sensor nodes show values higher than reference analyzers with lower accuracy in real operations. For O₃ assessment, the sensor nodes, excluding Node 3 in 2017, show values comparable to reference analyzers with good accuracy in real operations. For CO assessment, some sensor nodes (Nodes 1, 2, 3, and 4) show values higher than reference analyzers with lower accuracy in real operations, and the remaining sensor nodes (Nodes 6, 8, 9, and 10) show values comparable to reference analyzers with better accuracy in real operations.

The performance discrepancy of the sensor nodes and reference analyzers could be attributed to local conditions of air pollution. In the next sections, a detailed analysis of each monitored air pollutant will be presented.

3.2.1. Particulate Matter (PM₁₀)

Figure 13 shows the 2016 and 2017 trend of the monthly mean of the PM₁₀ measured by eight sensor nodes (Shinyei, PPD20V) distributed in Bari city and compared to the monthly mean of the public reference data by air-quality monitoring stations operated by the local environmental agency. The monthly mean values do not exceed the limit of 50 μg/m³, as regulated by EU Directive (2008/50/EC) in 2016 and 2017.

However, in such cases, the daily mean values can exceed the regulated threshold, as shown in Figure 14. In fact, as an example, during February 2016 the comparison between sensor Node 2 (ENEA) and closest air quality station (Via Caldarola) reported exceedance over a threshold of 50 μg/m³ in five specific days (15, 16, 17, 28, and 29 February) due to Saharan dust in Bari city, while the exceedance of 52 μg/m³ that occurred on 24 February was due to an anthropogenic source: vehicular traffic. The high regression coefficient R of 0.84 and low Mean Absolute Error (MAE) of 8.1 μg/m³ made the sensor node able to monitor the particulate matter in real time with good accuracy at the given location (ENEA office) for long-term operation.

Additionally, during May 2017, the comparison between sensor Node 6 (Airport) and the closest air quality station (Modugno EN2) reported exceedance over a threshold of 50 μg/m³ in two specific days (12 and 13 May) due to Saharan dust in Bari city, while no exceedance occurred due to anthropogenic sources. The high regression coefficient R of 0.88 and low Mean Absolute Error (MAE) of 4.6 μg/m³ made the sensor node able to monitor the particulate matter in real time with good accuracy at the airport for long-term operations.

Table 3. Performance of the air quality sensor network (8 nodes) assessed by means of specific indicators related to PM10 in the years 2016 and 2017.

Node Info		2016					2017
Node ID	Node Location	Sensor Node PM10 Yearly Mean (µg/m3)	Closest AQMS PM10 Yearly Mean (µg/m3)	(*) Sensor Node PM10 Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)	Sensor Node PM10 Yearly Mean (µg/m3)	Closest AQMS PM10 Yearly Mean (µg/m3)	(*) Sensor Node PM10 Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)
Node 1	ENELVia Capruzzi	25.85	22.56	4.49	19.90	0.58	26.22	21.36	7.67	35.89	0.36
Node 2	ENEAViale Japigia	22.61	25.91	3.41	13.17	0.62	38.23	27.49	14.12	51.35	0.25
Node 3	PERONIVia Bitritto	22.32	26.67	7.78	29.16	0.30	20.31	29.85	10.30	34.52	0.76
Node 4	AMTABVia Jacobini	21.40	26.67	8.16	30.60	0.29	18.77	29.85	11.09	37.14	0.28
Node 6	AIRPORT WojtylaViale Ferrari	22.37	27.31	5.61	20.54	0.14	20.69	27.18	6.49	23.86	0.63
Node 8	POLITECNICOVia Amendola	23.14	25.91	3.29	12.68	0.67	22.42	27.49	5.10	18.55	0.48
Node 9	COFELYStrada Vassallo	23.39	24.98	2.49	9.97	0.71	n/a	27.16	n/a	n/a	n/a
Node 10	UFFICI COMUNALIPiazza Ferrarese	25.94	24.69	4.85	19.63	0.65	33.47	26.34	7.78	29.52	0.62
	MEAN	23.38	25.59	5.01	19.45	0.49	25.73	27.09	8.93	32.97	0.48

(*) MAE = Mean Absolute Error. (**) Sensor Node Relative Error (SNRE) is defined as sensor MAE divided by AQMS Yearly Mean. (°) Sensor performance using the Min value instead of the Avg value for Node 10 only. In 2017, the node operated for only 3 months. (§) Node 9 was removed in December 2016. No data for 2017. n/a = not available.

reports on the performance of the sensor network, deployed in Bari city, for PM₁₀ assessment during the years 2016 and 2017 using some specific indicators, such as yearly mean, Mean Absolute Error (MAE), relative error (RE), and regression coefficient (R) by comparison with the public official data of the closest air-quality monitoring station (AQMS), managed by the local environmental agency. The yearly means of each sensor node and closest AQMS are comparable with a MAE low value ranging from the lowest 2.49 μg/m³ (COFELY) to the highest 8.16 μg/m³ (AMTAB) in 2016, and from the lowest 5.10 μg/m³ (POLITECNICO) to the highest 14.12 μg/m³ (ENEA) in 2017. Usually, for each node, the MAE and RE tend to increase from 2016 to 2017 after continuous operations, indicating that the sensor performance becomes worse over time due to poisoning after 30 months of measurements. Finally, in 2016, the coefficient regression (R) shows high values (0.58–0.71) for five nodes (ENEL, ENEA, POLITECNICO, COFELY, and UFFICI COMUNALI) and low values (0.14–0.30) for three nodes (PERONI, AMTAB, and AIRPORT); while in 2017 R shows high values (0.48–0.76) for four nodes (PERONI, AIRPORT, POLITECNICO, and UFFICI COMUNALI) and low values (0.25–0.36) for three nodes (ENEL, ENEA, and AMTAB), with one node (COFELY) stopped in 2017 for an unavailable location. As a general rule, if R is lower than 0.5, then recalibration should be needed to enhance accuracy.

K.K. Johnson et al. [50] used the proposed PPD20V sensor to measure the fine particulate matter PM_2.5 at roadside sites in the cities of Atlanta (USA) and Hyderabad (India) to estimate the on-road emission factors by pairing PM_2.5, black carbon (BC), and carbon dioxide (CO₂) measurements.

Finally, in the present study, all PM₁₀ sensor nodes, excluding Node 2 (ENEA) in 2017, exhibit a relative error of less than 50% (uncertainty of the EU Ambient Air Directive for Indicative Measurements) and, therefore, the Data Quality Objective (DQO) has been achieved by the sensor network during the 30-month operations (July 2015–December 2017).

3.2.2. Ozone (O₃)

Figure 15 shows the 2016 and 2017 trend of the monthly mean of the O₃ measured by 8 sensor nodes (Alphasense, O₃-B4) distributed in the Bari city and compared to the monthly mean of the public reference data by air-quality monitoring station (ARPA-Viale Kennedy) operated by the local environmental agency. The monthly mean values do not exceed the daily limit of 90 ppb (180 μg/m³), as regulated by the EU Directive (2008/50/EC) in 2016 and 2017. As an example, during February 2017, the comparison between sensor Node 1 (ENEL) and the closest air quality station (Viale Kennedy) reported no exceedance over the threshold of 90 ppb (180 μg/m³), as shown in Figure 16. Despite the low regression coefficient R of 0.23 (not shown), the combination of the low Mean Absolute Error (MAE) of 8.9 ppb and the relative error (RE) of 25.3% made the sensor node able to monitor the ozone in real time with an accuracy fitting the Indicative Measurements requirements at the given location (ENEL) for long-term operation.

Table 4. Performance of the air-quality sensor network (8 nodes) assessed by means of specific indicators related to O₃ in the years 2016 and 2017. The closest AQMS selected as a reference station is the ARPA-Viale Kennedy (urban background).

Node Info		2016					2017
Node ID	Node Location	Sensor Node O3 Yearly Mean (µg/m3)	Closest AQMS O3 Yearly Mean (µg/m3)	(*) Sensor Node O3 Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)	Sensor Node O3 Yearly Mean (µg/m3)	Closest AQMS O3 Yearly Mean (µg/m3)	(*) Sensor Node O3 Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)
Node 1	ENELVia Capruzzi	32.07	44.40	12.33	27.78	0.90	33.38	45.47	12.10	26.60	0.89
Node 2	ENEAViale Japigia	38.89	44.40	6.71	15.10	0.63	34.66	45.47	11.03	24.26	0.82
Node 3	PERONIVia Bitritto	44.16	44.40	16.40	36.93	0.17	55.10	45.47	17.64	38.80	0.10
Node 4	AMTABVia Jacobini	40.28	44.40	4.88	11.00	0.96	42.23	45.47	4.22	9.28	0.90
Node 6	AIRPORT WojtylaViale Ferrari	38.19	44.40	6.86	15.45	0.70	37.64	45.47	3.80	8.36	0.91
Node 8	POLITECNICOVia Amendola	n/a	44.40	n/a	n/a	n/a	n/a	45.47	n/a	n/a	n/a
Node 9	COFELYStrada Vassallo	28.66	44.40	11.58	26.08	0.93	n/a	45.47	n/a	n/a	n/a
Node 10	UFFICI COMUNALIPiazza Ferrarese	n/a	44.40	n/a	n/a	n/a	n/a	45.47	n/a	n/a	n/a
	MEAN	37.04	44.40	9.79	22.05	0.71	40.60	45.47	9.75	21.46	0.72

(*) MAE = Mean Absolute Error. (**) Sensor Node Relative Error (SNRE) is defined as sensor MAE divided by AQMS Yearly Mean. (°) Sensor performance using the Min value instead of the Avg value for Node 10 only. In 2017 nodes operated for 3 months only. (§) Node 9 was removed in December 2016. No data for 2017. n/a = not available.

reports on the performance of the sensor network, deployed in Bari city, for O₃ assessment during the years 2016 and 2017 using some specific indicators, such as yearly mean, Mean Absolute Error (MAE), relative error (RE), and regression coefficient (R) by comparison with the public official data of the closest air-quality monitoring station (ARPA-Viale Kennedy), classified as an urban background station and managed by local environmental agency. The ozone (O₃) is monitored by three air-quality monitoring stations only (Viale Kennedy Bari, CUS Bari, and EN2 Modugno) in Bari. The selection of the reference analyzer operated in Viale Kennedy has been made due to its having the best location in the centre of the city and its being symmetrically distant from all sensor nodes distributed in Bari. In fact, the other two air-quality monitoring stations (CUS Bari and EN2 Modugno) are asymmetrically located in the East and West urban peripheries. The yearly means of each sensor node and the selected AQMS are comparable with a MAE low value ranging from the lowest 4.88 ppb (AMTAB) to the highest 16.40 ppb (PERONI) in 2016, and from the lowest 3.80 ppb (AIRPORT) and 4.22 ppb (AMTAB) to the highest 17.64 ppb (PERONI) in 2017. Usually, for each node, the MAE and relative error (RE) tend to increase from 2016 to 2017 after continuous operations, indicating that the sensor performance became worse over time due to poisoning after 30 months of measurements. Finally, in 2016 the coefficient regression (R) shows high values (0.63–0.96) for five nodes (ENEL, ENEA, AMTAB, AIRPORT, and COFELY), and a low value (0.17) for one node (PERONI), and there were two nodes (POLITECNICO and UFFICI COMUNALI) with data not available; while in 2017 R shows high values (0.82–0.91) for four nodes (ENEL, ENEA, AMTAB, and AIRPORT), and a low value (0.10) for one node (PERONI), and there were three nodes (POLITECNICO, COFELY, and UFFICI COMUNALI) with data not available. As a general rule, if R is lower than 0.5, then recalibration should be needed to enhance accuracy. These R values, measured for ozone by this study, are comparable (R = 0.84) to the 18-day short-campaign (7–25 August 2015), using the same electrochemical O₃ sensor operated on the campus of the University of York (UK) [51].

Finally, in the present study, all O₃ sensor nodes, excluding Node 3 (PERONI) in 2016 and 2017, exhibit a relative error (RE) of less than 30% (uncertainty of the EU Ambient Air Directive for Indicative Measurements) and, therefore, the Data Quality Objective (DQO) has been achieved by the sensor network during the 30-month operations (July 2015–December 2017).

3.2.3. Nitrogen Dioxide (NO₂)

Figure 17 shows the 2016 and 2017 trend of the monthly mean of the NO₂ measured by eight sensor nodes (Alphasense, NO₂-B4) distributed in Bari city and compared to the monthly mean of the public reference data by the air-quality monitoring station (ARPA-Viale Kennedy) operated by the local environmental agency. The monthly mean values do not exceed the daily limit of 100 ppb (200 μg/m³), as regulated by EU Directive (2008/50/EC) in 2016 and 2017. As an example, during February 2017, the comparison between sensor Node 1 (ENEL) and the closest air-quality station (Viale Kennedy) reported no exceedance over the threshold of 100 ppb (200 μg/m³), as shown in Figure 18. The combination of the low regression coefficient R of 0.48 (not shown), the low Mean Absolute Error (MAE) of 11.8 ppb, and the relative error of 32% made the sensor node unable to monitor the nitrogen oxide in real time at the given location (ENEL) for long-term operation.

Table 5. Performance of the air quality sensor network (8 nodes) assessed by means of specific indicators related to NO₂ in the year 2016 and 2017. The closest AQMS selected as reference station is the ARPA-Viale Kennedy (urban background).

Node Info		2016					2017
Node ID	Node Location	Sensor Node NO2 Yearly Mean (µg/m3)	Closest AQMS NO2 Yearly Mean (µg/m3)	(*) Sensor Node NO2 Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)	Sensor Node NO2 Yearly Mean (µg/m3)	Closest AQMS NO2 Yearly Mean (µg/m3)	(*) Sensor Node NO2 Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)
Node 1	ENELVia Capruzzi	36.13	32.74	5.72	17.48	0.33	36.13	35.67	4.10	11.51	0.22
Node 2	ENEAViale Japigia	38.16	32.74	7.21	22.04	0.41	40.78	35.67	5.32	14.93	0.16
Node 3	PERONIVia Bitritto	44.69	32.74	15.96	48.75	0.10	37.74	35.67	9.65	27.05	0.10
Node 4	AMTABVia Jacobini	27.33	32.74	9.08	27.74	0.53	25.41	35.67	10.36	29.05	0.29
Node 6	AIRPORT WojtylaViale Ferrari	n/a	32.74	n/a	n/a	n/a	n/a	35.67	n/a	n/a	n/a
Node 8	POLITECNICOVia Amendola	25.99	32.74	5.45	16.66	0.51	32.01	35.67	18.03	50.53	0.27
Node 9	COFELYStrada Vassallo	33.08	32.74	4.35	13.27	0.12	n/a	35.67	n/a	n/a	n/a
Node 10	UFFICI COMUNALIPiazza Ferrarese	n/a	32.74	n/a	n/a	n/a	n/a	35.67	n/a	n/a	n/a
	MEAN	34.23	32.74	7.96	24.32	0.33	34.41	35.67	9.49	26.61	0.20

(*) MAE = Mean Absolute Error. (**) Sensor Node Relative Error (SNRE) is defined as sensor MAE divided by AQMS Yearly Mean. (°) Sensor performance using the Min value instead of the Avg value for Node 10 only. In 2017 node operated for 3 months only. (§) Node 9 was removed in December 2016. No data for 2017. n/a = not available.

reports on performance of the sensor network, deployed in the Bari city, for NO2 assessment during the years 2016 and 2017 using some specific indicators as yearly mean, Mean Absolute Error (MAE), relative error (RE) and regression coefficient (R) by comparison with the public official data of the closest air-quality monitoring station (ARPA-Viale Kennedy), classified as urban background station and managed by local environmental agency. The yearly means of each sensor node and the selected AQMS are comparable with a MAE low value ranging from the lowest 4.35 ppb (COFELY) to the highest 15.96 ppb (PERONI) in 2016, and from the lowest 4.10 ppb (ENEL) to the highest 18.03 ppb (POLITECNICO) in 2017. Usually, for each node the performance tends to decrease from 2016 to 2017 after continuous operations indicating poisoning after 30 months of measurements. Finally, in 2016 the coefficient regression (R) shows high values (0.41–0.53) for 3 nodes (ENEA, AMTAB, POLITECNICO), and a low value (0.10–0.33) for 3 nodes (ENEL, PERONI, COFELY), and 2 nodes (AIRPORT, UFFICI COMUNALI) with data not available; while in 2017 R shows low values (0.10–0.29) for 5 nodes (ENEL, ENEA, PERONI, AMTAB, POLITECNICO), and 3 nodes (AIRPORT, COFELY, UFFICI COMUNALI) with data not available. A periodical calibration should be carried out to enhance performance.

Finally, all NO₂ sensor nodes, excluding Node 3 (PERONI) in 2016 and 2017, Node 4 (AMTAB) in 2016 and 2017, and Node 8 (POLITECNICO) in 2017, exhibit relative error less than 25% (uncertainty of the EU Ambient Air Directive for Indicative Measurements) and, therefore, the Data Quality Objective (DQO) has been achieved by a few sensors of the network during 30-month operations (July 2015–December 2017).

Recently, a valid strategy to improve accuracy in the air quality sensors deployed in the real world is emerging as the field calibration by installing a reference instrument in the sensor network domain, as shown by the 50 nodes networked in the Berkeley Atmospheric CO₂ Observation Network (BEACO₂N) [30] for CO₂, CO, NO, NO₂, O₃ and PM₁₀ monitoring; alternatively by the field stochastic calibration using dynamic neural network architectures demonstrated for NO₂, NO_x and O₃ [52].

3.2.4. Carbon Monoxide (CO)

Figure 19 shows the 2016 and 2017 trend of the monthly mean of the CO measured by 8 sensor nodes (Alphasense, CO-B4) distributed in the Bari city and compared to the monthly mean of the public reference data by air-quality monitoring station (ARPA-Via Cavour) operated by the local environmental agency. The monthly mean values do not exceed the daily limit of 8 ppm (10 mg/m³), as regulated by EU Directive (2008/50/EC) in 2016 and 2017. As an example, during February 2017, the comparison between sensor node 1 (ENEL) and the closest air quality station (Via Cavour) reported no exceedance over the threshold of 8 ppm (10 mg/m³), as shown in Figure 20. The combination of the medium regression coefficient R of 0.56 (not shown), the low Mean Absolute Error (MAE) of 0.23 ppm and the relative error of 17.3% made the sensor node able to monitor the level of the carbon monoxide in real time at the given location (ENEL) for long-term operation. However, a drawback of the CO sensor performance is the absence of seasonal variation as given by the ARPA instrument. In addition, the sensor ageing should be further investigated.

Table 6. Performance of the air-quality sensor network (8 nodes) was assessed by means of specific indicators related to CO in the years 2016 and 2017. The closest AQMS selected as reference station is the ARPA-Via Cavour (urban traffic).

Node Info		2016					2017
Node ID	Node Location	Sensor Node CO Yearly Mean (µg/m3)	Closest AQMS CO Yearly Mean (µg/m3)	(*) Sensor Node CO Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)	Sensor Node CO Yearly Mean (µg/m3)	Closest AQMS CO Yearly Mean (µg/m3)	(*) Sensor Node CO Mean Absolute Error (µg/m3)	(**) Sensor Node Relative Error (%)	Regression Coefficient (R)
Node 1	ENELVia Capruzzi	0.93	0.92	0.35	37.47	0.10	0.91	0.70	0.27	38.10	0.27
Node 2	ENEAViale Japigia	0.92	0.92	0.34	36.75	0.02	0.97	0.70	0.32	45.81	0.18
Node 3	PERONIVia Bitritto	1.26	0.92	0.45	49.16	0.43	1.56	0.70	0.62	88.55	0.52
Node 4	AMTABVia Jacobini	2.46	0.92	1.53	165.95	0.52	2.46	0.70	1.77	252.86	0.46
Node 6	AIRPORT WojtylaViale Ferrari	0.24	0.92	0.68	74.12	0.25	0.19	0.70	0.51	72.45	0.52
Node 8	POLITECNICOVia Amendola	0.19	0.92	0.73	79.56	0.04	0.69	0.70	0.27	38.75	0.39
Node 9	COFELYStrada Vassallo	0.36	0.92	0.56	60.89	0.09	n/a	0.70	n/a	n/a	n/a
Node 10	UFFICI COMUNALIPiazza Ferrarese	0.45	0.92	0.47	51.40	0.72	n/a	0.70	n/a	n/a	n/a
	MEAN	0.85	0.92	0.63	69.41	0.27	1.13	0.70	0.62	89.42	0.39

(*) MAE = Mean Absolute Error. (**) Sensor Node Relative Error (SNRE) is defined as sensor MAE divided by AQMS Yearly Mean. (°) Sensor performance using the Min value instead of the Avg value for Node 10 only. In 2017, the node operated for 3 months only. (§) Node 9 was removed in December 2016. No data for 2017. n/a = not available.

reports on performance of the sensor network, deployed in the Bari city, for CO assessment during the years 2016 and 2017 using some specific indicators as yearly mean, Mean Absolute Error (MAE), relative error (RE) and regression coefficient (R) by comparison with the public official data of the closest air-quality monitoring station (ARPA-Via Cavour), classified as an urban traffic station and managed by the local environmental agency. The yearly means of each sensor node and the selected AQMS are generally comparable, with a MAE low value ranging from the lowest 0.34 ppm (ENEA) to the highest 1.53 ppm (AMTAB) in 2016, and from the lowest 0.27 ppm (ENEL and POLITECNICO) to the highest 1.77 ppm (AMTAB) in 2017. Usually, for each node, the performance tends to degrade from 2016 to 2017 after continuous operations, indicating poisoning after 30 months of measurement. Finally, in 2016, the coefficient regression (R) shows values ranging from 0.10 to 0.72 for all nodes, with the highest value for node 10 (UFFICI COMUNALI); while in 2017, R shows values ranging from 0.18 to 0.52 for all nodes, with the highest value for Node 3 (PERONI) and Node 6 (AIRPORT), and two nodes (COFELY and UFFICI COMUNALI) with data not available. A periodical calibration should be carried out to enhance performance.

N. Masson et al. [53] studied a metal-oxide-based CO sensor, operated it for 19 days during April 2013 for air-quality monitoring, and installed it in downtown Denver (USA) adjacent to a busy motorway with significant CO pollution, and it reported a RMS error ranging from 0.22 to 0.48 ppm and a CO level ranging from 0.5 to 2.5 ppm. Temperature compensation techniques were used to reduce the sensor drift using a co-location reference analyzer by improving the correlation coefficient (R) up to 0.88.

Finally, in the present study, all CO sensor nodes exhibit a relative error (RE) higher than 25% (uncertainty of the EU Ambient Air Directive for Indicative Measurements) and, therefore, the Data Quality Objective (DQO) has not been achieved by any sensor of the network during the 30-month operations (July 2015–December 2017).

3.2.5. Carbon Dioxide (CO₂), Relative Humidity (RH), and Temperature (T)

Figure 21 shows the 2016 and 2017 trends of the monthly mean of carbon dioxide (CO₂), expressed in ppm, measured by the selected sensor (Alphasense, CO2-IRC-A1), of the urban relative humidity (RH), expressed in % unit, measured by the selected sensor (Honeywell, HIH5031), and of the urban temperature (T), expressed in Celsius degree (°C), measured by the selected sensor (Microchip, TC1047A), in the given Node 1 (ENEL) operated in the Bari city. The yearly mean values of CO₂ varied from 337 to 445 ppm with a yearly mean of 415 ppm in 2016, while the yearly mean values of CO₂ varied from 404 to 476 ppm with a yearly mean of 446 ppm in 2017. The high increase in the CO₂ concentration of 31 ppm between 2016 and 2017 could be attributed to the specific conditions of all the locations of the sensor nodes under test. However, the drift of the sensors could not be excluded and is worthy of further investigation. These data are affected by local micro-climate conditions but the annual trend is generally replicated during both years: in other terms, in the winter (D/J/F), the CO₂ is of a high value of about 480–490 ppm; in the spring (M/A/M), the CO₂ starts a decrease from about 460 to 380 ppm; in the summer (J/J/A), the CO₂ still continues a decrease from about 350 to 310 ppm; and, finally, in the autumn (S/O/N), the CO₂ starts a increase from about 390 to 460 ppm. The monthly lowest value of CO₂ in Node 1 is reached in July, as 312 ppm (2016) and 370 ppm (2017). The monthly highest value of CO₂ in Node 1 is reached in January 2016, at 494 ppm, and in December 2017, at 527 ppm.

Furthermore, the yearly mean values of RH in Node 1 varied from 58.8% to 67.9% with a yearly mean of 63.2% in 2016, while the yearly mean values of RH varied from 55.0% to 63.9% with a yearly mean of 59.3% in 2017. The monthly lowest value of RH in Node 1 is reached in July 2016 as 51.3%, and in August 2017 as 44.8%.

The monthly highest value of RH in Node 1 is reached in November 2016 as 70.1%, and in January 2017 as 73.0%. Additionally, the yearly mean values of T in Node 1 varied from 17.2 °C to 18.8 °C, with a yearly mean of 18.0 °C in 2016, while the yearly mean values of T varied from 17.3 °C to 18.8 °C with a yearly mean of 18.0 °C in 2017.

The monthly lowest value of T in Node 1 is reached in December 2016 as 10.1 °C, and in January 2017 as 8.3 °C. The monthly highest value of T in Node 1 is reached in July 2016 as 27.3 °C, and in August 2017 as 28.0 °C. These data of the RH and T in Bari city (Southern Italy) are typical of a Mediterranean climate with a regular cycle of the seasons.

Sensor networks at low-cost and low power consumption have been used to monitor the urban CO₂ emissions in order to estimate the atmospheric inversion in the Paris area [54] using 70 sensor nodes; they are also used in California’s Bay Area [55], using 34 sensor nodes of the Berkeley Atmospheric CO₂ Observation Network (BEACO2N), and covering an area of roughly 400 km²; and, finally, they are used for urban CO₂ monitoring using 100 sensor nodes [56], operated in the Wuxi city, China. Since there was no common standard in the data format, a comparison cannot be proposed.

4. Discussion

The presented results of the long-term experimental campaign for urban air-quality monitoring in Bari city by a sensor network based on multiple nodes with multi-parametric sensors, such as particulate matter (PM₁₀), gases (NO₂, O₃, CO), and greenhouse gases (CO₂), and meteorological parameters, such as relative humidity (RH) and temperature (T), demonstrate the feasibility of the networked low-cost sensors in the framework of the Indicative Measurements of the EU Ambient Air Quality Directive. A comparison of the sensor performance has been realized using the official public data of the closest air-quality monitoring station (AQMS) of the urban network, operated and managed by the local environmental agency (ARPA-Puglia).

The assessment of the sensor-versus-analyzer inter-comparison has been carried out for PM₁₀, O₃, NO₂, and CO only, due to the availability of the reference public dataset. The results are reported in Figure 22. The figures of merit used for the assessment have been selected as the normalized error, defined as the sensor node relative error normalized to the Indicative Measurements, limits accuracy for each given pollutant, and the relative MAE, defined as the node yearly MAE normalized to the all-nodes mean MAE. If the normalized error is less than 1, then the data quality objective (DQO) for Indicative Measurements (see Section 1.3) has been achieved, while if the normalized error is higher than 1, then the sensor fails the regulatory issue of the data quality. Additionally, if the relative MAE is less than 1, the node sensor exhibits an error lower than the network mean, while if the relative MAE is higher than 1, the node sensor is at high error with poor precision.

By combining the normalized error and relative MAE, the best case for the sensor performance is with both parameters less than 1, and the worst case is with both parameters higher than 1; complementary cases to the best and worst cases are possible with medium performance. To show the long-term sensor assessment, four quadrants corresponding to the different combinations of normalized error with the relative MAE are shown in Figure 22. The performance assessment of the sensor network is summarized in

Table 7. Summary of the performance of the air quality sensor network deployed in Bari, assessed according to DQO as regulated by EU Directive (2008/50/EC). The figures of merit are defined in Appendix A.

Parameter	Normalized Error (NR)	Relative MAE (RMAE)	Performance Assessment of Sensor Network
PM10	<1 (excluding Node 2 in 2017)	<1 (70% of Nodes)	Very Good
O3	<1 (excluding Node 3 in 2016 and 2017)	<1 (50% of Nodes)	Good
NO2	<1 (excluding Node 3 and Node 4 in 2016; and Node 3, Node 4, and Node 8 in 2017)	<1 (50% of Nodes)	Medium
CO	>1 (all Nodes)	<1 (70% of Nodes)	Poor

.

In terms of monthly average values, the PM₁₀ sensor network exhibits a very good performance, fitting the Indicative Measurement requirements during 2016 and 2017 for all nodes, excluding Node 2 in 2017, with a normalized error less than 1, and a large part of the nodes with relative MAE less than 1. These results demonstrate that the sensor network has successfully reached the DQO for the Indicative Measurements in both years 2016 and 2017, with a low MAE in the majority of the nodes.

In terms of monthly average values, the O₃ sensor network exhibits good performance, fitting the Indicative Measurement requirements during 2016 and 2017 for all nodes, excluding Node 3 in 2016 and 2017, with a normalized error less than 1 and a 50% amount of the nodes with a relative MAE less than 1 only. These results demonstrate that the sensor network has successfully reached the DQO for the Indicative Measurements in both years 2016 and 2017, with a low MAE in half of the nodes.

In terms of monthly average values, the NO₂ sensor network exhibits a medium performance, fitting Indicative Measurement requirements during 2016 and 2017 for all nodes, excluding Node 3 and Node 4 in 2016, and the Node 3, Node 4, and Node 8 in 2017, with a normalized error of less than 1 and a 50% amount of the nodes with a relative MAE less than 1 only. These results demonstrate that the sensor network has successfully reached the DQO for the Indicative Measurements in both years 2016 and 2017, with a low MAE in the minority of the nodes.

In terms of monthly average values, the CO sensor network exhibits a poor performance, fitting Indicative Measurement requirements during 2016 and 2017 for all nodes, with no node at a normalized error less than 1, but 70% of the nodes had a relative MAE less than 1. These results demonstrate that the sensor network has not reached the DQO for the Indicative Measurements in both years 2016 and 2017. This low accuracy is caused by very low CO values measured in Bari city by the deployed sensors, increasing the error and worsening the performance.

Air sensors are useful to monitor the current status of atmospheric pollution at high spatial and temporal resolution for mapping urban and rural zones. Co-location of sensors and reference monitors is a valid approach to in-field calibration. So, the distance between sensors and reference analyzers should be as close as possible, if possible. However, air-quality monitoring stations are not located at high spatial resolution due to high cost. Thus, low-cost calibrated sensors are a good solution to monitor air quality, but they are affected by site specificity and local conditions. In our study, the PM sensors are very much able to detect high concentrations (>50 µg/m³) of particulate matter, especially during the additional event of Saharan dust in Bari, which was measured by sensor Node 2 (ENEA) and referenced by the closest air-quality monitoring station (Via Caldarola), occurring on some days (15, 16, 17, 28, and 29) in February 2016; and this event was also measured by sensor Node 6 (Airport) and referenced by the closest air-quality monitoring station (Modugno EN2), occurring on some days (12 and 13) in May 2017. These PM sensors are at poor resolution in the range of low concentrations (10–40 µg/m³), as reported in Figure 14; but they are very much able to differentiate high concentrations (> 50 µg/m³), which is useful for the early detection of heavy dust in urban areas.

Generally, the performance of each sensor node depends strongly on the specific urban site that is affected by local ambient conditions in terms of temperature, humidity, and other interfering species. The ageing of the used air sensors is surely present in the long-term campaign in urban areas. Each designed sensor node has been deployed in a given urban site for the whole period of the experimental campaign. The exchange of a sensor node from one site to another is not the purpose of this work; the purpose is to check the best sensor performance in terms of ageing and the influence of location.

A comprehensive status of the urban air quality should include other biological pollutants not included in the European Ambient Air Directive (2008/50/EC), such as pesticides, pathogens, toxins, toxic elements, moulds, pollens, and other bio-chemicals. A review [57] of these topics is illustrated for environmental monitoring purposes.

Comprehensive reviews devoted to calibration, methods, and advanced technologies for wireless sensor networks deployed for air-quality monitoring have been proposed and studied in urban areas of all continents, including Europe [58], Australia [59], the USA [60], India [61], and a global overview [62].

5. Conclusions

In this paper, we demonstrate the feasibility of a sensor network based on 10 stationary nodes, distributed in Bari city (Southern Italy), in a city area of 24 km², for urban air-quality monitoring during a 30-month long-term campaign of continuous measurements. The stationary nodes have been installed outside in specific sites (e.g., the roof of buildings, offices, schools, streets, etc.) managed by partners and stakeholders of the Italian national project RES-NOVAE, under the supervision of the Bari energy manager. The location for each sensor-node AIRBOX, designed and developed by ENEA (Brindisi), has been selected, considering the closeness of a public air-quality monitoring station (AQMS) for reference validation. Continuous measurements have been realized for particulate matter (PM₁₀), gases (NO₂, O₃, CO), greenhouse gases (CO₂), and meteorological parameters such as relative humidity (RH) and temperature (T). A sensor-versus-analyzer inter-comparison has been realized in order to assess the sensor performance for selected air pollutants, such as particulate matter (PM₁₀) and gases (NO₂, O₃, CO). The selection of these four pollutants has been limited by official dataset availability as a reference.

In terms of monthly average values, the results show that PM₁₀ sensor network exhibits a very good performance, fitting Indicative Measurement requirements during 2016 and 2017 for all nodes, excluding Node 2 in 2017, with low normalized error and low MAE in the large part of the nodes, successfully achieving the DQO for the Indicative Measurements in both years 2016 and 2017 in majority of the nodes.

In terms of monthly average values, the O₃ sensor network exhibits good performance, fitting Indicative Measurement requirements during 2016 and 2017 for all nodes, excluding Node 2 in 2017, with low normalized error and low MAE (50% of the nodes), successfully achieving the DQO for the Indicative Measurements in both years 2016 and 2017 in half of the nodes.

In terms of monthly average values, the NO₂ sensor network exhibits a medium performance, fitting Indicative Measurement requirements during 2016 and 2017, with low normalized error and low MAE (50% of the nodes), successfully achieving the DQO for the Indicative Measurements in both years 2016 and 2017 in the minority of the nodes.

In terms of monthly average values, the CO sensor network exhibits a poor performance fitting Indicative Measurement requirements during 2016 and 2017, with high normalized error and low MAE (70% of the nodes), failing the DQO for the Indicative Measurements in both years 2016 and 2017. This poor precision is ascribed to the low limit of the urban CO pollution, increasing the error in the measurement and decreasing the performance.

This study is one of a few cases to deploy a wireless sensor network, referenced by public official data, that operated for a 30-month long-term campaign of continuous monitoring, including a full pool of air pollutants using low-cost and low-power consumption sensors in an urban scenario.

Wireless sensor networks for air-quality monitoring are key to enabling technologies to develop the concept of the sustainable city [63] in the era of the Internet of Things (IoT), supporting decision-making and engaging citizens according to the goals of sustainable development.

Current research trends are exploring how to integrate low-cost air quality sensor networks with satellite monitoring systems to study ground-level pollution of main toxic gases and particulate matter at higher spatial and temporal resolutions [64,65].

Finally, our future work has been planned to monitor the chemical threat of fine particulate matter on human health, such as PM_2.5 and PM_1.0, in urban areas.