Development of an air quality station using

low-cost sensors
2023 XIII Brazilian Symposium on Computing Systems Engineering (SBESC) | 979-8-3503-9410-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/SBESC60926.2023.10324138

Mateus Maruzka Roncaglio∗ , Edson Tavares de Camargo∗† , Leila Droprinchinski Martins‡ , Marcio Seiji Oyamada∗
∗ Graduate Program in Computer Science – State University of Western Paraná (UNIOESTE) – Cascavel, Brazil
† Federal Technology University of Paraná (UTFPR) – Toledo, Brazil
‡ Federal Technology University of Paraná (UTFPR) – Londrina, Brazil

Abstract—Air pollution is one of the most important health each year and is the greatest environmental threat to human
problems causing various diseases. According to the World health [6]. Nearly every person on Earth (99% of the world’s
Health Organization (WHO), it is estimated that more than population) breathes air that exceeds World Health Organi-
7 million deaths are due to air pollution. For this reason,
air quality monitoring is an important indicator for guiding zation (WHO) guidelines. In terms of the Sustainable Devel-
public policy. However, government stations are widely scattered opment Goals (SDGs) [7], air pollution contributes massively
and their high cost makes it unprofitable to invest in higher to climate change (SDG 13), threatens our health and well-
resolution. Low-cost air quality monitoring sensors can overcome being (SDG 3), jeopardizes food security (SDG 2), and makes
this problem, but also bring new challenges. This paper presents urban dwellers unsafe (SDG 11). In this regard, air quality
the development of a low-cost air quality monitoring device.
The low-cost station collects the following measurements: carbon monitoring is a critical factor to track concentrations of air
monoxide (CO), nitrogen dioxide (NO2 ), sulfur dioxide (SO2 ), pollutants (e.g., NO2 , O3 , CO2 , N2 O, PM10, PM2.5, etc.) [3].
ozone (O3 ), and particulate matter. The collected data is corrected In Brazil, in particular, air quality monitoring is carried
using a machine learning model and sent to the Internet in real out almost exclusively by public agencies through air quality
time via the LoRaWAN protocol. Mean square error (MSE), monitoring stations distributed throughout the cities. The main
root mean square error (RMSE), mean absolute error (MAE),
linearity coefficient (R2 ), and Pearson r were used to compare problem is that 20 of Brazil’s 27 states do not measure air
model performance. Our results show that linear models such quality or do so inadequately [8], [9]. The cost of acquiring and
as the Alphasense equation and linear model regression cannot maintaining conventional monitoring stations makes networks
accurately describe the sensor response to the reference gas with adequate spatial distribution uneconomical, especially in
sensor, whereas the RF model performs better in each metric. developing or less developed countries. Recent advances in
The performance of the RF model demonstrates the potential to
improve air quality monitoring and the decision-making process. low-cost air quality sensor technology combined with remote
Index Terms—Machine learning, Air-quality, Low-cost sensor sensing and traditional monitoring methods provide a new way
to understand and inform the air quality [10], [11]. Optical
and electrochemical sensors integrated into circuit boards are
I. I NTRODUCTION capable of quantifying pollutants that make up the air quality
Cities are the centers of economic activity and human needs. index. With the ability to communicate in real time with the
As these places continue to grow, especially in terms of popu- Internet, there is an opportunity to develop low-cost stations
lation, the concept of smart cities has emerged as an alternative that monitor the air quality of a region in real time.
to solve the pressing problems of urban life. Although a smart However, the challenge of building low-cost sensor monitor-
city can be defined in many ways, one of its main goals is ing stations is enormous. Low-cost sensors typically result in
to improve the quality of life of its population [1], [2]. For a trade-off between the price, size, and power consumption of
some authors, building a smart city depends heavily on the the device and the accuracy and precision of the measurements
use of information and communication technologies, which made. Because of their design and composition, low-cost sen-
include the Internet of Things (IoT) paradigm [3], [4]. In the sors are susceptible to variations in humidity and temperature,
IoT, devices with sensors and embedded computers monitor as well as interference from other gasses [12]. In addition,
a variety of activities in an urban area and communicate sensor selection, programming of components and interfaces,
with the Internet through a mostly wireless network. The vast calibration, and verification of the accuracy of measurements
amount of data collected is stored in a database, analyzed, made with low-cost sensors must all be considered when
and ultimately serves as the basis for public information and designing a station. Calibration involves, among other things,
government decision making. comparing measurements made with the low-cost sensor to
One of the biggest problems facing cities is pollution, measurements made at official air quality stations to correct
especially air pollution. As Martina Otto of the United Nations the data.
Environment Program (UNEP) describes, air pollution is an This work aims to develop an air quality monitoring solution
emergency, both for public health and for climate [5]. Air using low-cost sensors (electrochemical and optical), and to
pollution is responsible for about 7 million deaths worldwide correct the sensor measurement using a machining learning

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on December 14,2023 at 15:13:35 UTC from IEEE Xplore. Restrictions apply.
model. The station was designed to measure atmospheric variables were the concentrations of CO, NO2 and O3 , using
particulate matter with aerodynamic diameters of 10 and 2.5 the MOX sensors TGS2442, MiCS-2714, and MiCS-2614,
micrometers (µ) (PM10 and PM2.5), ozone (O3 ), carbon respectively. The area with the highest traffic volume was
monoxide (CO), nitrogen oxides (NOx , NO and NO2 ), sulfur found to be the most polluted. Furthermore, data from zone
dioxide (SO2 ), temperature and relative humidity, and wind 3 were compared with an official reference station, as it is
speed. Data are transmitted using an LoRaWAN network [13]. located near uSense node. The authors claim that although the
The project of an air quality monitoring device included sensor device does not have the same accuracy as the professional
selection, circuit board design, component programming, com- station, it can still provide useful information on air quality. It
parison of measurements from low-cost sensors with official is also emphasized that the success of low-cost air monitoring
stations, and data analysis. This work is part of a large project devices depends heavily on the accuracy of the sensors. In the
called EnvCity (FAPESP), whose goal is to establish an air and paper, a two-point calibration was performed to obtain a linear
water quality monitoring network that uses low-cost sensors equation to estimate the concentration from the MOX sensors.
and real-time data transmission to make the data available to In [18], different calibration models for Real-time Afford-
society [14]. The machine learning model was trained using able Multi-Pollutant (RAMP) sensors were investigated using
data collected for 4 months with the low-cost station installed commercial sensors from Alphasense. For measurements, the
next to a reference station in Sao Paulo city, Brazil. CO-B41, Ox-B431, NO2-B43F models were used, measuring
This paper is organized as follows. Section II introduces CO, O3 and NO2 , respectively. For the measurement of CO2 ,
the low-cost air quality sensors and presents related work. the SST-CO2-A sensor was used, which has integrated sensors
Section III presents the developed prototype. Section IV for temperature and relative humidity. Calibration techniques
describes the reference station and data set. Section V presents were evaluated in the study using linear regression and random
the results and finally Section VI concludes the work and forest (RF). Measurements were taken over six months at
points out future directions. Carnegie Mellon College. The authors conclude that it is pos-
sible to significantly improve the quality of the samples using
II. BACKGROUND AND R ELATED W ORK machine learning, including the calibration model using RF
This section briefly introduces low-cost sensors for air to meet the requirements of the United States Environmental
quality and related work. Protection Agency.
In [19], a LoRaWAN device was developed for air moni-
A. Low-cost sensors for air-quality monitoring toring in combination with sensor calibration using machine
The main low-cost commercial technologies for measuring learning. Two sensor calibration methods were considered for
the concentration of gases usually related to air quality are the Alphasense CO-B4 carbon monoxide sensor. The first
electrochemical sensors and MOX (Metal Oxide) sensors [15]. uses linear regression, while the second uses a Multi-Layer
Electrochemical sensors work on the principle of oxidation and Perceptron (MLP) neural network with a hidden layer. The
reduction. In these processes, there is an exchange of electrons sensors were placed near to a reference station managed by the
between substrates. Electrochemical sensors generally come in City of Auckland composed of the Teledyne API Model 300A
2, 3, and 4-electrode configurations. In its simplest form, with CO analyzer. In both cases, the temperature, humidity and the
two electrodes, the sensor has one electrode labeled Working measured CO value were used to correct the data. The neural
Electrode (WE) and another labeled Counter Electrode (CE). network approach was superior to the linear regression method
A redox reaction occurs on the WE and CE electrodes and and proved to be a promising method for sensor calibration.
electric charges are generated. When a reduction occurs at In this work, the sensors are read at 5-second intervals
one electrode, oxidation occurs at the other electrode. The and an exponential moving average (EMA) filter (α = 0.3)
current produced is usually proportional to the concentration is applied to remove noise from the readings. The collected
of the gas involved. In sensors with 3 electrodes, there is data is sent over a LoRaWAN provider and stored in a cloud
another electrode called Reference Electrode (RE). It is used to platform at a rate of 8 minutes, so that even small variations
maintain a constant potential at the WE electrode, controlling in gas concentration caused by local traffic can be captured.
the operating range. The circuit used to maintain a constant This transmission rate is due to the limit of messages that can
voltage between these electrodes is called a potentiostatic be sent over the private LoRaWAN service. Similar to [18],
circuit. In addition, another transimpedance amplifier [16] in this work we also use a RF model to correct the value
circuit is used to convert the current into an electrical voltage of the low-cost sensor. However, in this work, we use a grid
that can be read by the microcontroller. search step to tune the RF hyperparameters to find the best
configuration of the RF, including the preprocessing applied
B. Related work to the data, which was not mentioned in the referenced work.
In [17], a low-cost system was developed to monitor air
quality in urban areas, called uSense. The devices used Wi- III. P ROTOTYPE
Fi communication to send data to the cloud. Three nodes To integrate the different sensors a signal acquisition board
were installed in three areas with different characteristics of was developed. For the sensing, the prototype uses the elec-
automobile traffic in the city of Lucca, Italy. The observed trochemical sensors from the Alphasense B family and its

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on December 14,2023 at 15:13:35 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
S ENSORS USED IN THE PROPOSED STATION

Sensors Measuring Range

Anemometer 0 32,4 m/s

ISB – CO-B4 1000 ppm (max)
ISB – NO2-B43F 20 ppm (max)
ISB – OX-B431 20 ppm (max)
ISB – SO2-B4 100 ppm (max)
Particulate matter PM1.0, PM2.5, PM10 µg/m3
DHT22 -40 to 80 ◦ C / 0 to 100 %RH

Fig. 2. EnvCity Air Quality Prototype

ISB (Individual Sensor Board), the DHT22 temperature and
humidity sensor, a particulate matter sensor and an anemome- TABLE II
ter. For ADC readings, a 16 bit ADS1115 analog converter DATA FROM THE CHARACTERIZATION OF SENSORS USED IN STATION 1
is used. However, this ADC has only 4 analog inputs and
NO2
each electrochemical sensor requires two of them. Therefore, Gain W0 A0 Sensitivity We Ae
Sensor Sensitivity
an analog CD74HC4067 multiplexer is used to expand the (mV /nA) (mV ) (mV ) (mV /ppb) (mV ) (mV )
(nA/ppm)
number of analog inputs of the system [20], [21]. The LoRa CO-B4 0.8 330 316 0.408 336 321 -
communication is performed by a RFM95W radio. SO2-B4 0.8 361 350 0.29 335 342 -
Table I presents the sensors used in the prototype. The NO2-B43F -0.73 222 212 0.31 230 220 -
OX-B431 -0.73 229 234 0.369 237 242 -587
sensors were selected to obtain measurements of gasses and
particulate matter, which are used as a reference for measuring
air quality. Sensors are selected based on the measurement
A. Calculating the Gas Concentration
range and precision required to calculate the air quality.
In addition to the data transmission and sensor connection To calculate the concentration of a gas of interest at the four
requirements, the board is designed to support both wall power electrode sensors used in this work, the Equation 1 is used,
supply or li-ion battery operation. A second version of the where [GAS]P P B is the concentration of the gas of interest in
board is under development that will also allow the use of ppb, Wu and Au are the voltage measured at the WE pin and
solar cells to charge the batteries. AE pin, respectively, We and Ae are the electronic offset of
Figure 1 shows the signal acquisition board developed in the sensor, which is usually between 200 mV and 300 mV, so
this work. The connectors allow quick and easy integration of that the output voltage is positive even at low concentrations.
the Alphasense B family of sensors, to which up to six sensors S is the sensor sensitivity expressed in mV per ppb. W0 and
can be connected. This board is connected to the processing A0 are the output voltage of the sensor in zero air. [22].
board via a flat cable. In this project we have developed the
firmware for the boards Raspberry PI 4 and ESP32-TTGO. (Wu − We − W0 ) − (Au − Ae − A0 )
[GAS]PPB = (1)
S
For the sensors used in this work, the sensor parameters are
presented in Table II. Note that the O3 sensor also responds
to NO2 , so is given a NO2 sensitivity for this type of sensor.
IV. T EST SETUP
In this work, we evaluated the electrochemical sensors for
CO, O3 , NO2 and SO2 gasses. Particulate matter is evaluated in
[23]. To evaluate the data obtained, the prototype was installed
near to a reference station located at the Faculty of Medicine of
Fig. 1. Signal acquisition board the University of Sao Paulo (FM-USP). The reference station
is composed of four different modules: Thermo-Fisher Model
The board and sensors were mounted in a plastic enclosure 42i NOx Analyzer, Model 48i CO Analyzer, Model 49i Ozone
with dimensions 42 cm per 30 cm and 15 cm with IP6 certifi- Analyzer and Model 43i SO2 Analyzer.
cation (see Figure 2). In our tests, we found that without air
circulation, the humidity gets stuck at high values, e.g. 90% A. Collected data
RH. So we installed two 12V coolers, one of which removes The low-cost station developed in this work was placed near
air from the box and the other draws air into the box. A buck a reference station in Sao Paulo, Brazil, from March 2023 to
boost module was used to control the voltage of these coolers June 2023. In this type of station each analyzer costs more
and thus the airflow. than 10,000 USD. In contrast, each Alphasense sensor and its

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on December 14,2023 at 15:13:35 UTC from IEEE Xplore. Restrictions apply.
connection board costs about 100 USD. The cost of reference TABLE IV
sensors is high because they use different technologies such as D IVISION OF THE DATASET
chemiluminescence, photomertry, or infrared absorption, and Model Train Test Validation Total
also a controlled temperature and humidity room and a robust NO2 20515 6839 6839 34193
airflow capture, resulting in more precise and stable readings. CO 20916 6972 6972 34860
The data from the low-cost station was transmitted using SO2 20487 6829 6829 34145
OX 20863 6955 6955 34773
the LoRaWan protocol in a private network. The device was
installed at a distance of about 900 meters from the nearby
gateway. The only spreading factor over which messages were V. R ESULTS
received was the SF12, which limits the message size to 51
bytes. For the gas sensors, only the voltage measurements of Table V presents the results for the CO concentration,
the electrodes were sent and the concentration was calculated between the reference value and the raw value of the sensor,
in the cloud platform. and the predicted values by the linear regression and RF
We used the Scikit-learn machine learning library to fit a RF model. For the predictions of CO, the RF model had a
regressor model and a linear regression model for comparison. significantly lower root mean square error (RMSE) compared
We also evaluated the raw conversion equation provided by to the others estimation models. Compared to the model
Alphasense. By comparing the performance of these three provided by Alphasense, both the RF model and the linear
models, we aimed to show that the RF model can improve regression have a better RMSE value. The same trend is seen
the data provided by low-cost sensors and that linear models when comparing the mean squared error (MSE) and the mean
perform poorly with these sensors. absolute error (MAE), as shown in Table V. Finally, both
The performance evaluation of the models developed in this the RF model and linear regression had similar performance
study was compared using the root mean square error (RMSE), in the training, testing, and validation data set, suggesting
mean squared error (MSE), mean absolute error (MAE), R- that they are not overfitting. This is also observed in the
squared (R2 ), and Pearson correlation coefficient (Pearson r). R2 and Pearson r metrics that the RF model perform better
For the RF, we also performed a grid search to select the than the other evaluated models. For the linear regression,
best configuration, including the number of regressors, that the Pearson r indicates some correlation between the sensor
is, the number of trees, and which scaling method performed and the reference, but the R2 does not. This suggests that
better, e.g., min-max scalers and standard scalers. The whole something is affecting these values. It may be a bias, noisy
parameters grid used on the optimization is presented in the data, or a non-linear relationship. In the Figure 3a we can
Table III. The number of trees was chosen around the value see that the data has some distortion or offset and contains
used in [18] based on the trade-off between performance and much more noise. The IAG station (Institute of Astronomy,
model complexity. The input of the RF model are: WE and Geophysics) represents the reference value. Figure 4 presents
AE voltages, the concentration determined by the Alphasense a 15 day span comparing the concentration of CO at the
equation, temperature, and humidity, and the output is the reference station and the estimated by the RF model.
corrected concentration. The WE and AE voltages were used The same trend can be observed in the OX-B431 sensor
as inputs because the auxiliary electrode is not in contact with (O3 ). In this case, the RF model also had better performance
the gas and, therefore, its response is not affected by external in each metric. Table VI shows the RF model has strong cor-
factors. The only exception is the ozone (O3 ) model. The OX relation with the reference, considering both R2 and Pearson
-B431 is also sensitive to NO2 , so the WE voltage of the NO2- r and a small error, close to 3 ppb for the MAE metric. Figure
B43F is also used as input. For training all models, a 5-fold 5 presents a 15 day span comparing the concentration of O3
cross-validation scheme was used and the data set was split, in the reference station and the estimated by the RF model.
with 80% of the total data set used for training, 10% for the Figures 3a and 3b show that the Alphasense raw equation is
test data set, and the last 10% for validation. not precise to calculate the concentration of the CO and O3 .

TABLE III TABLE V

PARAMETERS USED ON THE GRID SEARCH CO M ODELS COMPARISION

Model MSE RMSE MAE R2 Pearson r

Parameter Values
train 0.000090 0.009469 0.000862 0.999238 0.999619
Number of trees 32, 64, 128 Random
test 0.003678 0.060643 0.025682 0.969293 0.984530
Max features log, sqrt Forest
valid 0.003398 0.058290 0.023933 0.968397 0.984073
Normalization MinMaxScaler, StandardScaler
train 0.033342 0.182598 0.126312 0.716582 0.846512
Linear
test 0.032578 0.180494 0.127936 0.727976 0.853655
Regression
The number of points obtained after cleaning the data, valid 0.031462 0.177375 0.125703 0.707367 0.841058
removing outliers, and dealing with NaNs is displayed on Raw valid 3.244431 1.801230 1.771182 -29.176909 0.604245
Table IV. Due to LoRa communication and outliers, some
messages are partially valid, resulting in a different number In the SO2 sensor, linear regression does not improve sensor
of samples for each analyzed gas. performance because it yields little correlation when consid-

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on December 14,2023 at 15:13:35 UTC from IEEE Xplore. Restrictions apply.
 300 500
Alphasense Equation Alphasense Equation
Linear Regression 250 160 Linear Regression
3 250
Random Forest Random Forest
1:1 Line 0 140 1:1 Line
200
2 250 120
150
Predicted Values

Predicted Values

Predicted Values

Predicted Values
Alphasense Equation 100
500 Linear Regression
1 100 Random Forest
750 80
1:1 Line
50
1000 60
0
Alphasense Equation 0 1250 40
Linear Regression
1 Random Forest 50 1500 20
1:1 Line
100 1750 0
1.0 1.5 2.0 2.5 3.0 0 20 40 60 80 0 50 100 150 200 250 300 350 0 20 40 60 80 100 120 140 160
IAG - CO IAG - OX IAG - SO2 IAG - NO2

(a) CO models comparison (b) O3 models comparison (a) SO2 models comparison (b) NO2 models comparison

Fig. 3. Models applied to the validation data set Fig. 6. Models applied to the validation data set

CO Concentration from 2023/06/15 to 2023/06/30

3000
CO EnvCity
3000
TABLE VII
IAG
2500 SO2 M ODELS COMPARISION
Concentration (ppb)

2500
2000
2000 IAG 1500 Model dataset MSE RMSE MAE R2 Pearson r
1500 1000
500 train 46.746783 6.837162 2.949419 0.965109 0.984278
1000 Random
0 test 303.326829 17.416281 7.562499 0.776708 0.881582
0 0 0 0
Forest
6-1
5
6-1
6
6-1
7
6-1
8
6-1
9
6-2
0
6-2
1
6-2
2
6-2
3 100 200 300 valid 282.557657 16.809451 7.616631 0.781761 0.884312
3-0 3-0 3-0 3-0 3-0 3-0 3-0 3-0 3-0 CO EnvCity
202 202 202 202 202 202 202 202 202
Date train 1216.333799 34.875977 24.272176 0.092145 0.303553
Linear
test 1235.706595 35.152619 24.481475 0.090342 0.300677
Regression
Fig. 4. CO Concentration reference vs RF model over 15 days valid 1173.783486 34.260524 24.112676 0.093407 0.305650
Alphasense Raw 2976753.814686 1725.327162 1714.670597 -2298.150543 0.045593

ering Pearson r, but shows no relationship when considering

R2 , as shown in Table VII. In contrast, the model RF shows a considering the difference between the error metrics and the
strong correlation when considering these two metrics. A small correlation metrics in the training data set and the test data set,
mean error (MAE) is also obtained with the RF model, which as shown in Table VIII and Figure 6b. Despite this overfitting,
is close to 17 ppb. Figure 6a presents the estimated values the RF model outperformed the other models in all metrics.
of the all models evaluated in this work compared to SO2 The general concentration at which the sensor was used was
concentration reported by the reference station, and Figure 7 low and close to the lower detection limit, which may explain
shows the response of the model RF compared to the reference the poor overall performance, as presented in Figure 8, where
value over a period of 15 days. the response of the RF model is compared to the reference
Finally, for NO2-B43F sensor, the RF model showed good station over a period of 15 days.
improvement compared with other models. The RF model The RF model consistently outperformed both the equation-
showed a lower error between sensor data and the NO2 based Alphasense model and linear regression for all pollu-
reference sensor, resulting in a better response. However, it tants. This result demonstrates the superior predictive capabil-
can be observed that the RF model exhibited overfitting when ities of the RF model in capturing the complex relationships
between measured characteristics and pollutant concentrations.

TABLE VI SO2 Concentration from 2023/06/15 to 2023/06/30

O 3 MODELS COMPARISON 350 SO2 EnvCity 350
300 IAG
300
Concentration (ppb)

250 250
Model MSE RMSE MAE R2 Pearson r 200 200
IAG

train 1.192306 1.091928 0.606860 0.991973 0.996070 150 150

Random 100 100
test 6.479777 2.545541 1.536031 0.956311 0.977918
Forest 50 50
valid 7.099862 2.664557 1.549073 0.953464 0.976477 0
0
train 73.386601 8.566598 6.520073 0.505942 0.711296 5 6 7 8 9 0 1 2 3 0 50 100 150 200 250
6-1 6-1 6-1 6-1 6-1 6-2 6-2 6-2 6-2
Linear 2 3-0 3-0 3-0 3-0 3-0 3-0 3-0 3-0 3-0 SO2 EnvCity
test 72.361439 8.506553 6.525381 0.512113 0.715640 2 0 202 202 202 202 202 202 202 202
Date
Regression
valid 74.061396 8.605893 6.564901 0.514562 0.717615
Alphasense Raw 16413.507293 128.115211 46.904596 -106.582942 0.146659
Fig. 7. SO2 Concentration reference vs RF model over 15 days

Ox Concentration from 2023/06/15 to 2023/06/30 NO2 Concentration from 2023/06/15 to 2023/06/30

OX EnvCity NO2 EnvCity
40 IAG 500 IAG 500
40
400
Concentration (ppb)

Concentration (ppb)

400
30 30
300 300
IAG

IAG

20 20
200 200
10 10 100 100

0 0 0 0
-15 -16 -17 -18 9 0 1 2 3 0 10 20 30 40 -15 -16 -17 -18 -19 -20 -21 -22 -23 0 100 200 300 400 500
-06 -06 -06 -06 6-1 6-2 6-2 6-2 6-2 6 6 6 6 6 6 6 6 6
3 3 3 3 3-0 3-0 3-0 3-0 3-0 OX EnvCity 3-0 3-0 3-0 3-0 3-0 3-0 3-0 3-0 3-0 NO2 EnvCity
202 202 202 202 202 202 202 202 202 202 202 202 202 202 202 202 202 202
Date Date

Fig. 5. O3 Concentration reference vs RF model over 15 days Fig. 8. NO2 Concentration reference vs RF model over 15 days

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on December 14,2023 at 15:13:35 UTC from IEEE Xplore. Restrictions apply.
 TABLE VIII [2] C. S. Lai, Y. Jia, Z. Dong, D. Wang, Y. Tao, Q. H. Lai, R. T. K. Wong,
NO 2 M ODELS COMPARISION A. F. Zobaa, R. Wu, and L. L. Lai, “A review of technical standards
for smart cities,” Clean Technologies, vol. 2, no. 3, pp. 290–310, 2020.
Model Dataset MSE RMSE MAE R2 Pearson r [Online]. Available: https://www.mdpi.com/2571-8797/2/3/19
train 1.414767 1.189440 0.149140 0.998438 0.999218 [3] P. Bellini, P. Nesi, and G. Pantaleo, “Iot-enabled smart cities: A review
Random
Forest
test 579.930001 24.081736 4.756040 0.519663 0.730457 of concepts, frameworks and key technologies,” Applied Sciences,
valid 444.868817 21.091914 4.122299 0.425098 0.652845 vol. 12, no. 3, 2022. [Online]. Available: https://www.mdpi.com/2076-
train 898.727888 29.978791 10.672834 0.007427 0.086180 3417/12/3/1607
Linear
Regr.
test 1198.786454 34.623496 11.524277 0.007085 0.087355 [4] S. L. Ullo and G. R. Sinha, “Advances in smart environment monitoring
valid 768.856989 27.728271 10.244141 0.006410 0.082067 systems using iot and sensors,” Sensors, vol. 20, no. 11, 2020. [Online].
Alphasense valid 72174.154334 268.652479 39.011775 -92.270279 0.009053 Available: https://www.mdpi.com/1424-8220/20/11/3113
[5] United Nations, “The air that keeps us alive is making us sick,” 2022.
[Online]. Available: https://news.un.org/en/story/2022/09/1126101
[6] Our World in Data, “Data review: How many people die from air
In general, our results highlight the clear advantages of pollution?” 2023. [Online]. Available: https://ourworldindata.org/data-
the RF model for environmental monitoring and pollutant review-air-pollution-deaths
[7] United Nations, “SDG - Sustainable Development Goals,” 2023.
concentration prediction. The performance of the RF model [Online]. Available: https://sdgs.un.org/goals
shows its potential to improve air quality assessment and [8] E. d. M. P. A. d. A. Vormittag, S. S. R. Cirqueira, H. Wicher Neto,
decision-making process, making it a valuable tool for various and P. H. N. Saldiva, “Análise do monitoramento da qualidade do ar
no brasil,” Estudos Avançados, vol. 35, no. 102, p. 7–30, May 2021.
environmental applications. The results obtained provide valu- [Online]. Available: https://doi.org/10.1590/s0103-4014.2021.35102.002
able information for selecting an appropriate predictive model [9] B. Felin, “Veja onde É feito o monitoramento da qualidade do ar no
for accurate and reliable pollutant concentration estimation. brasil,” 2023. [Online]. Available: https://www.wribrasil.org.br/noticias/
veja-onde-e-feito-o-monitoramento-da-qualidade-do-ar-no-brasil
[10] United Nations Environment Programme, “Why low-cost
VI. C ONCLUSION sensors? opportunities and challenges,” 2023. [Online]. Avail-
able: https://www.unep.org/explore-topics/air/what-we-do/monitoring-
This work presented the design and the prototyping of a air-quality/why-low-cost-sensors-opportunities-and
low-cost air quality station based on electrochemical sensors. [11] F. Karagulian, M. Barbiere, A. Kotsev, L. Spinelle, M. Gerboles,
This study has shown that the integration of Alphasense low- F. Lagler, N. Redon, S. Crunaire, and A. Borowiak, “Review of the
performance of low-cost sensors for air quality monitoring,” Atmosphere,
cost sensors for carbon monoxide (CO-B4), nitrogen dioxide vol. 10, no. 9, 2019. [Online]. Available: https://www.mdpi.com/2073-
(NO2-B43F), sulfur dioxide (SO2-B4), and ozone (OX-B431) 4433/10/9/506
with a RF model outperforms both the linear regression model [12] “Low-cost pollution sensors - understanding the uncertainties -
defra, uk.” [Online]. Available: https://uk-air.defra.gov.uk/research/
and the raw conversion provided by the Alphasense model. aqeg/pollution-sensors/understanding-uncertainties.php
The results show that the RF model provides greater accuracy [13] Lora-Alliance, “A technical overview of LoRa and LoRaWAN,”
on correcting the gasses concentrations. Also, in this work 2015. [Online]. Available: https://lora-alliance.org/resource-hub/what-
lorawantm
some improvements are made when compared to ozone (O3 ) [14] FAPESP, “Development of environmental quality monitor-
RF model developed by [18], using the NO2 WE voltage as ing network in smart cities - EnvCity,” 2023. [On-
input leads to a improvement on the R2 metric. Additionally, line]. Available: https://bv.fapesp.br/en/auxilios/108005/development-
of-environmental-quality-monitoring-network-in-smart-cities-envcity/
the study successfully provided a novel model for the SO2 [15] B. Maag, Z. Zhou, and L. Thiele, “A survey on sensor calibration in
sensor, which was not previously analyzed in the [18]. The air pollution monitoring deployments,” IEEE Internet of Things Journal,
low-cost sensor RF approach proves to be a promising and vol. 5, no. 6, pp. 4857–4870, 2018.
[16] “Designing a potentiostatic circuit,” 2022. [Online]. Avail-
cost-effective solution for air quality monitoring, providing able: https://www.alphasense.com/wp-content/uploads/2022/10/AAN
reliable and accurate measurements compared to conventional 105-03 App-Note V0.pdf
regression methods. Therefore, this innovative combination [17] S. Brienza, A. Galli, G. Anastasi, and P. Bruschi, “A low-cost sensing
system for cooperative air quality monitoring in urban areas,” Sensors
holds great potential to improve environmental monitoring, (Switzerland), vol. 15, no. 6, pp. 12 242–12 259, 2015.
contribute to decision making and spatial-temporal resolution [18] N. Zimmerman, A. A. Presto, S. P. N. Kumar, J. Gu, A. Hauryliuk, E. S.
and ultimately advance our understanding of air pollution and Robinson, A. L. Robinson, and R. Subramanian, “A machine learning
calibration model using random forests to improve sensor performance
its impact on public health and the environment. for lower-cost air quality monitoring,” Atmospheric Measurement
Techniques, vol. 11, no. 1, pp. 291–313, 2018. [Online]. Available:
VII. ACKNOWLEDGEMENT https://amt.copernicus.org/articles/11/291/2018/
[19] S. Ali, T. Glass, B. Parr, J. Potgieter, and F. Alam, “Low Cost
We thank the São Paulo Research Foundation (FAPESP), Sensor With IoT LoRaWAN Connectivity and Machine Learning-Based
process 20/05135-5, and the Coordenação de Aperfeiçoamento Calibration for Air Pollution Monitoring,” IEEE Transactions on Instru-
mentation and Measurement, vol. 70, pp. 1–11, 2021.
de Pessoal de Nı́vel Superior (CAPES) Finance Code 001 and [20] “CD74HC4067: High-Speed CMOS Logic 16-Channel Analog
Fundação Araucaria for the research support. Multiplexer/Demultiplexer,” Texas Instruments, 2003. [Online].
Available: https://www.ti.com/lit/ds/symlink/cd74hc4067.pdf
[21] “ADS111x Ultra-Small, 16-Bit ADCs ,” Texas Instruments, 2018.
R EFERENCES [Online]. Available: https://www.ti.com/product/ADS1115
[1] R. Sánchez-Corcuera, A. Nuñez-Marcos, J. Sesma-Solance, A. Bilbao- [22] Alphasense, “Correcting for background currents in four electrode toxic
Jayo, R. Mulero, U. Zulaika, G. Azkune, and A. Almeida, “Smart cities gas sensors.” 2019.
survey: Technologies, application domains and challenges for the cities [23] A. F. Pastório, F. A. Spanhol, L. D. Martins, and E. T. de Camargo,
of the future,” International Journal of Distributed Sensor Networks, “A machine learning-based approach to calibrate low-cost particulate
vol. 15, no. 6, p. 1550147719853984, 2019. [Online]. Available: matter sensors,” in 2022 XII Brazilian Symposium on Computing Systems
https://doi.org/10.1177/1550147719853984 Engineering (SBESC), 2022, pp. 1–8.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on December 14,2023 at 15:13:35 UTC from IEEE Xplore. Restrictions apply.