The Design and Deployment of a Self-Powered, LoRaWAN-Based IoT Environment Sensor Ensemble for Integrated Air Quality Sensing and Simulation

Abstract

The goal of this study is to describe a design architecture for a self-powered IoT (Internet of Things) sensor network that is currently being deployed at various locations throughout the Dallas-Fort Worth metroplex to measure and report on Particulate Matter (PM) concentrations. This system leverages diverse low-cost PM sensors, enhanced by machine learning for sensor calibration, with LoRaWAN connectivity for long-range data transmission. Sensors are GPS-enabled, allowing precise geospatial mapping of collected data, which can be integrated with urban air quality forecasting models and operational forecasting systems. To achieve energy self-sufficiency, the system uses a small-scale solar-powered solution, allowing it to operate independently from the grid, making it both cost-effective and suitable for remote locations. This novel approach leverages multiple operational modes based on power availability to optimize energy efficiency and prevent downtime. By dynamically adjusting system behavior according to power conditions, it ensures continuous operation while conserving energy during periods of reduced supply. This innovative strategy significantly enhances performance and resource management, improving system reliability and sustainability. This IoT network provides localized real-time air quality data, which has significant public health benefits, especially for vulnerable populations in densely populated urban environments. The project demonstrates the synergy between IoT sensor data, machine learning-enhanced calibration, and forecasting methods, contributing to scientific understanding of microenvironments, human exposure, and public health impacts of urban air quality. In addition, this study emphasizes open source design principles, promoting transparency, data quality, and reproducibility by exploring cost-effective sensor calibration techniques and adhering to open data standards. The next iteration of the sensors will include edge processing for short-term air quality forecasts. This work underscores the transformative role of low-cost sensor networks in urban air quality monitoring, advancing equitable policy development and empowering communities to address pollution challenges.

1. Introduction

Aerosols are airborne particulates composed of both solid and liquid particles [1]. These aerosols, commonly known as Particulate Matter (PM), are linked to a multitude of issues pertaining to global climate effects and human health [2,3,4,5]. The concentration of particulate matter in the air is generally measured using two key parameters: PM_2.5 and PM₁₀. PM_2.5 refers to the total mass of all airborne particles with an aerodynamic diameter of up to 2.5 microns, known as fine particulate matter, while PM₁₀ encompasses all airborne particles with an aerodynamic diameter of up to 10 microns, known as coarse particulate matter.

The physiological effects of exposure to PM can be profound and varied, depending on the size and composition of the particles. Fine particulate matter (PM_2.5) can penetrate deep into the lungs [6] and enter the bloodstream [7], causing short-term effects such as respiratory irritation [8], exacerbation of asthma [9], and acute cardiovascular events [10]. Long-term exposure to PM has been associated with chronic health issues, including chronic obstructive pulmonary disease (COPD) [11], lung cancer [12], and cardiovascular diseases [13], leading to significant morbidity and mortality [14,15]. Moreover, vulnerable populations—such as children, the elderly, and individuals with pre-existing health conditions—face heightened risks [16,17]. Children are particularly susceptible due to their developing respiratory systems, while the elderly often have compromised health that makes them more vulnerable to the deleterious effects of air pollution [18,19,20].

In addition to direct health impacts, PM exposure carries substantial socio-economic effects [21]. The economic burden associated with PM-related health issues can be significant, encompassing healthcare costs for treatment and management of chronic diseases, lost productivity due to illness, and diminished quality of life for affected individuals and families [22]. This burden disproportionately affects low-income communities [23], where access to healthcare may be limited. Economic disparities can hinder these populations’ ability to mitigate the impacts of PM exposure, making it essential to develop targeted public health interventions that address these inequalities.

Indoor PM, especially fine particles (PM_2.5), has been linked to various adverse health effects, including respiratory and cardiovascular diseases [24]. Long-term exposure can worsen asthma, reduce lung function, and increase the risk of heart disease and lung cancer [25]. Major sources of indoor PM include cooking, heating, and tobacco smoke [26].

Geographic disparities in PM exposure further complicate the health landscape. Urban areas, particularly those with high traffic and industrial activities, often experience higher concentrations of PM, resulting in worse health outcomes compared to rural areas [27,28]. Studies have shown that low-income neighborhoods in urban centers frequently encounter greater PM exposure due to environmental injustices, where socio-economic factors influence both the location of pollution sources and access to resources for health mitigation [29]. For instance, research has highlighted that communities situated near highways or industrial zones are more likely to suffer from higher rates of respiratory diseases and other health complications linked to PM exposure. Understanding these geographic disparities highlights the importance of implementing effective, low-cost monitoring systems—such as LoRaWAN IoT, Wi-Fi-based, or LTE/GSM-based solutions—to identify pollution hot-spots and enable targeted interventions aimed at protecting vulnerable populations. The current article provides a comprehensive description of a LoRaWAN-based sensor illustrated in Figure 1.

Measurements that sufficiently capture the spatial and temporal variability of atmospheric components, like particulate matter, are essential for accurately characterizing these components at the neighborhood scale. In most cases, rather than the actual physical scales of variability, the available fiscal and computational resources determine the resolution of atmospheric observations and simulation. In medium and large cities, for example, only a small number of monitoring stations are typically present. The Dallas-Fort Worth metroplex, which has a population of seven million people, has only three regulatory grade airborne particulate sensors. The ease of sensor placement is influenced by a given location’s access to power and network resources.

The spatial scales are dependent on variables such as local weather, distribution of sources, and terrain, as evidenced by street level surveys conducted with data at less than a meter resolution [30]. The smart city paradigm is synonymous with having an increased number of devices that can provide adequate coverage within a given geographic area [31,32].

As low-cost sensing systems gain traction, environmental monitoring initiatives are being expanded to engage the public more effectively. Traditional air quality monitoring systems typically utilize a wider array of sensors, eschew low-power networks, and tend to be more expensive. Low-power wireless technologies become increasingly valuable in scenarios where energy sources are unavailable or present challenges. Recently, research has shifted towards low-power wide area network (LPWAN) technologies, including Long Range Wide Area Network (LoRaWAN), Narrowband IoT (NB-IoT), LTE-M, and Sigfox [33,34,35,36]. Although all these technologies are energy-efficient, we have chosen LoRaWAN for its scalability [37], robustness [38], and cost-effectiveness [39]. In fact, LoRaWAN has been adopted for many IoT-related applications since its inception in 2013.

2. Design

The main elements of the sensing configuration are depicted in Figure 2.

2.1. Main Controller Unit (MCU)

The MCU unit is where the main computing of the system is enacted. The MCU is comprised of three main components:

• ATSAMD21G18: 32-Bit ARM Cortex M0+: This microcontroller unit serves as the end device controller, managing communication between the LoRaWAN radio module (RHF76-052DM) and the application, handling sensor data acquisition, and processing the data for transmission to the LoRaWAN module.
• RHF76-052DM: LoRaWAN Module acts as a communication interface between the microcontroller unit (ATSAMD21G18) and the LoRaWAN network.
• L70: GPS Module delivers precise location data.

The integrated unit is marketed as the Seeeduino LoRaWAN with GPS by the vendor, Seeed Technology Co., Ltd. (Seeed Studio, Shenzhen, Guangdong, China), and is considered an Arduino development board that uses the LoRaWAN protocol. ATMSAMD21G18 operates at 48 MHz maximum clock frequency and 3.3 V [40].

The incorporated LoRaWAN Module (RHF76-052DM) has dual band technology with 20 dBm max power being provided at 434–474 MHz and 14 dBm max power being provided at 868–915 MHz. LoRaWAN protocols Class A and Class C are supported by the module. All of these characteristics make it suitable for ultra-long range communications (up to 7 km in rural areas) and low power consumption [41]. However, the range of LoRaWAN is typically limited in urban settings due to high-density environments and signal attenuation caused by obstacles such as buildings and walls. This limitation becomes even more significant when there is no direct line of sight (NLOS), as signal propagation is further degraded by diffraction, reflection, and scattering. Studies have shown that while LoRaWAN can achieve ranges of several kilometers in rural or line of sight (LOS) conditions, urban NLOS scenarios often reduce this range significantly [42,43]. The design presented in this paper uses the frequency band range of 902.3–914.9 MHz (USA LoRaWAN frequency spectrum) at 14 dBm and LoRaWAN protocol Class A for purposes of communication.

Additionally, the unit incorporates a low powered, small sized, GPS module. The L70 can communicate with a GPS receiver by NMEA 0183 (a combined electrical and data specification) [44]. This communication provides location information for the current application.

2.2. Power Control Unit (PCU)

The PCU is marketed as the Sunny Buddy—Maximum Power Point Tracking (MPPT) Solar Charger by the vendor, SparkFun Electronics (SparkFun, Niwot, CO 80503, USA). It features the LT3652 circuit, a complete monolithic step-down battery charger that operates within an input voltage range of 4.95 V to 32 V. The PCU is connected both to the solar panel and to the battery, and it regulates the power to the MCU intelligently while keeping the battery charged throughout the day. A voltage regulation loop in the LT3652 reduces the charge current when the input voltage falls below a programmed level, which is determined by a resistor divider that we have set. If sufficient power is supplied by a solar panel, an input loop is used to maintain the LT3652’s maximum output. Charge current is terminated when the LT3652 falls below a tenth of the maximum charge current. Once charging is terminated, the LT3652 enters a low-current (85 μA) standby mode. If the battery voltage falls 2.5% below the programmed float voltage, an auto-recharge feature initiates a new charging cycle [45].

2.3. INA219s—Power Sensors

The INA219 is an integrated I²C Bus and a shunt resistor. It monitors shunt current, shunt voltage (0 to 26 V), bus voltage and power by measuring the voltage drop across the shunt resistor. Operational voltage is 3–5.5 V. INA219 can be operated from −40–125 °C. I²C or SMBUS-compatible interface is used to communicate data with the MCU. Two separate INA219s are connected to the battery and the solar panel, and the flowing current from both sources is fed through to the PCU. For the I²C interface, two separate slave addresses are used to maintain uniqueness between the two INA219s.

2.4. TPL5110—Low-Power Timer

Designed for battery operated or duty cycled applications, the TPL5110 nano timer is a low power, low current timer with an integrated MOSFET driver. Using only 35 nA of power, the TPL5110 is able to enable the power supply line so as to drastically reduce the overall system standby current during sleep time. As a result of its reduced power consumption, it can also be used with smaller batteries, which makes it well suited for energy harvesting and wireless sensor applications, such as the one being implemented [46]. TPL5110 functions to shut down an MCU after a predetermined period of time and reactivate it after another predetermined period. This is particularly helpful if the system enters a sleep cycle when manual intervention is not possible in the field. However, the main purpose of the TPL5110 is to preserve energy efficiently. TPL5110 serves as a bridge between the PCU and MCU. In the design architecture of the system, embedded firmware is run inside of a MCU during its life cycle. TPL5110 allows us to choose the period in which the system will reset and rerun the firmware. A potentiometer is used to alter the period. In the current configuration, we used an external resistor to achieve this. The external resistor is set to limit the life cycle to a maximum of 15 min. However, in most cases, the timer will cut off power to the MCU much sooner, depending on the battery’s stored energy and the solar panel’s real-time output. This is accomplished through a signal sent through to the TPL5110 from the MCU after the power metrics are measured via the INA219s.

2.5. External Sensing Unit (ESU)

The ESU is housed within a solar radiation shield, which safeguards the sensing ensemble from climatic factors while ensuring accurate readings by promoting airflow around the sensing elements. The ESU incorporates two sensors: the IPS7100 for particulate matter measurements and the BME280 for climate-related measurements.

• IPS7100: This sensor is an Optical Particle Counter (OPC), also known as an Intelligent Particle Sensor (IPS). It provides particle counts for sizes ranging from 0.1 to 10 $\mathsf{\mu }$m. Although it cannot directly count particles ≤0.1 $\mathsf{\mu }$m, it estimates particle counts for 0.1 $\mathsf{\mu }$m diameters through extrapolation. Operating at a rate of two cycles per second, the sensor classifies particles into seven size bins within this range. Additionally, the IPS7100 provides conventional particle mass fractions used to quantify air pollution, including PM₁, PM_2.5, and PM₁₀, as well as measurements for PM_0.1, PM_0.3, PM_0.5, and PM₅. The IPS7100 can be used with both UART and I²C hardware communication protocols. In the present application, the hardware communication protocol I²C is used. With a low power consumption of 0.325 W during measurements and only 0.001365 W in sleep mode, this device is well-suited for solar and battery applications, such as the one described here. The sensor draws air using a fan and is calibrated to the U.S. EPA-approved FEM (Federal Equivalent Method) reference instrument, the GRIMM EDM180, through the Piera Automated Calibration (PASCAL) system under a constant temperature (25 °C) and relative humidity (50%) [47].
• BME280: Optical particle counters usually consider temperature, pressure, and humidity as bias factors. We added the Bosch BME280 sensor to account for such biases as well as report climate data. It is designed specifically for mobile and wearable applications. Compact and power-efficient, it provides us with crucial climate data. BME280 is a digital humidity, pressure, and temperature sensor that supports both SPI and I²C interfaces. It operates within a voltage range of 1.71–3.6 V [48]. In our configuration the device is set to run at 3.3 V using the I²C interface.

3. Operation

The sensing system is intended to obtain sensor measurements and transmit the data in a sustainable manner. The duration of a single firmware life cycle is 15 min, and its operation is depicted in Figure 3. During a single life cycle, the first step is to read current battery and solar power voltages using the two INA219s and select a power mode. Figure 4 illustrates the different power modes. In the event that the sensor does not receive sufficient power, power mode 0 is set and the TPL5110 is used to set the system to sleep for the remaining duration of the current life-cycle.

In the event that the sensor has sufficient power, a LoRaWAN connection is established between the sensor node and the central gateway. Such a life cycle involves setting one of the following power modes: 1, 2, 3, 4 or 5. In order to ensure maximum security, each node has a unique ID and an application key that is specific to that ID. If the specifics match, the connection is made. By default, the data rate is set to 2 (3125 bits per second). In some cases, the network may adjust the data rate automatically in response to transmission characteristics. Once a successful connection has been established, power mode information is transmitted as the initial transmission of the current life cycle. Sensors are then initialized and the system is set in motion to collect data. For a given life cycle, the power modes also govern the number of sampling cycles and frequency of sampling for each sensor—for instance, the primary sensor (IPS7100) will vary its sampling frequency between 1 and 2 min based on the power mode selected for a given life cycle.

After the 15 min of the current life cycle is exhausted, the TPL5110 resets the system so that it initiates a new life cycle for the system.

4. Power Management

Careful steps have been taken to have sufficient power to run the system through out its field deployment. This is predominantly controlled via the power mode. Figure 4 gives the means in which power modes are set for each life cycle. In power mode 5, which is when the battery is fully charged as well as sufficient amount of sunlight is present, the firmware runs continuously without halting the system. On all the other power modes, after a predefined number of sensing cycles, the sensing system is halted to preserve continuous measurement of data throughout a given day, in all weather conditions. For example on power mode 1 and 2 the system will shut itself down after 1 sensing cycle which would only take around 7 min. And hence the system will sleep after for 8 (15–7) min of its ‘single life cycle’. And likewise on power mode 3, and 4 the system will shut itself down after 2 sensing cycles which would only take around 10 min. Although, power modes 1 and 2 as well as 3 and 4 are operationally similar for the current application, they are given specific power modes to cater future developments of the system, which is intended to have more sensing units. A summary of the sensing cycles for each power mode is provided in

Table 1. Sensing Cycles for each Sensor.

Power Mode	Sensing Cycles	Sensing Period (min)
		IPS7100	BME280	INA219s	L70
0	0	-	-	-	-
1	1	-	-	-	-
2	1	-	-	-	-
3	2	2	2	4	4
4	2	2	2	4	4
5	∞	1	1	3	3

For each life cycle of the sensing system, each sensor makes a predetermined number of measurements based on the power mode. When more than one measurement is made for a given sensor, the interval between each measurement is also controlled. This table reflects the number of sensing cycles along with the sensing periods.

.

Power mode 3 is the most commonly allocated mode for the sensor. In this mode, a single 15-min life-cycle consists of several sensor packets. Specifically, the sensor transmits two packets of IPS7100 data (56 bytes each), two packets of BME280 data (24 bytes each), one packet containing the pair of INA219 data (32 bytes), and one packet of L70 data (55 bytes). This totals 247 bytes transmitted every 15 min. The data rate of 16.47 bytes per minute refers only to the sensor data payload, excluding additional information such as headers and metadata within each packet. Package sizes of each sensor payload is further discussed on

Table 2. LoRaWAN Data Packets.

Sensor ID	FPORT	Packet Size (Bytes)	Parameters Measured
IPS7100	15	56	Standard particulate matter mass fractions and particulate counts for particulate diameters 0.1, 0.3, 0.5, 1.0, 2.5, 5.0 and 10 $\mathsf{\mu }$m.
BME280	21	12	Climate readings of atmospheric temperature, pressure and humidity.
INA219s	3	32	Voltages, currents, and power readings for both the battery pack and solar panels.
LP70	5	55	Longitude, latitude, altitude, as well as the UTC time stamps.

In LoRaWAN Sensing systems, the Port field (FPort) is commonly used to differentiate between different types of messages. To distinguish between the sensors used, we use the FPort field. Although the current infrastructure only uses four specific sensors, the sensing system is designed to accommodate multiple sensors. In the table are the fields identifying each sensor, the packet sizes associated with the LoRaWAN payload, and the specific measurements made for each sensor.

.

The power modes are designed to prevent the battery from being completely depleted. In conditions of very low power, when the battery drops below 10%, and while in power mode 0 (see Figure 4), the sensor predominantly stays in sleep mode for the majority of the 15-min cycle, only briefly waking to check the battery status. In the unlikely event that the sensor operates for an extended period without sunlight, causing the battery to reach 0%, the sensor will remain inactive until the solar panels recharge the battery sufficiently. Once the battery reaches a sufficient power level, the firmware will reactivate the system, enabling the sensor to resume operation.

The firmware used for the LoRaWAN nodes in this study has been made publicly available in the Zenodo data repository and can be accessed at https://doi.org/10.5281/zenodo.6727816 (last accessed on 23 June 2022) (Supplementary Materials).

5. Communication

This system is considered to be an end node within the LoRaWAN network. A central communication gateway, a component of a more comprehensive sensing suite referred to as a ‘Central Node’, receives the data collected by the system. Integrated into the Central Node is a Dragino LIG16, manufactured by Dragino Technology Co., Limited (Dragino, Shenzhen 518116, China), featuring an SX1302 LoRa concentrator. In this system, the nodes and the gateway are interconnected through a LoRaWAN network operating in Sub-Band 8 (913.5–914.9 MHz) for uplinks.

In this architecture, the gateway is placed at the center, while the end nodes are arranged around it following a star topology. To activate end devices, this network uses Over-the-Air (OTAA) activation, since this is the most secure and reliable method. The system data is transmitted via packets with packet size and contents set by the sensor type. Details regarding the specific data packets are provided in Table 2. LoRaWAN gateways gather data packets and send them up to a LoRaWAN cloud where they are sorted out and distributed using an MQTT pipeline. The MQTT pipeline is used to cater both the SharedAirDFW public portal and an extensive analytical toolbox. Figure 5 provides a comprehensive overview of the complete network architecture.

6. Analytical Toolbox

Over the past three years, our sensor network has accumulated approximately 6 TB of data, with storage demands expected to grow as new sensors are deployed each month. Of this total, only a small fraction (10%) originates from the LoRaWAN nodes examined in this study.

Most other sensing systems in our network, unconstrained by LoRaWAN’s bandwidth limitations, generate substantially larger data volumes. This is primarily due to their integration of a greater number of sensors and higher-frequency data collection. Our infrastructure includes extensive outdoor monitoring systems that operate continuously, measuring key environmental parameters such as particulate matter, meteorological variables (temperature, humidity, pressure, and precipitation), and gas concentrations (e.g., carbon dioxide and ozone). Additionally, our network includes light sensing suites and acoustic monitoring systems, with the latter utilizing edge computing for the analysis of bird calls. Beyond stationary monitoring, similar sensing technologies are deployed on mobile platforms, including an electric vehicle, a fully autonomous drone, and a fully autonomous boat, to enable dynamic data collection. We also develop wearable devices designed to capture biometric data, including heart rate, heart rate variability, and skin temperature, for physiological monitoring applications.

The high degree of spatial variability in environmental conditions necessitates a robust computational infrastructure capable of managing thousands of individual time series. This infrastructure must support real-time analysis, pollution alerts, and, ultimately, particle tracking and forecasting capabilities. To that end, we have developed a containerized solution comprising three open-source docker images: Node-RED, InfluxDB, and Grafana [49,50,51]. Node-RED allows us to easily ingest and process incoming sensor data in a simple graphical pipeline. This data is then injected by Node-RED into InfluxDB, a time-series database optimized for handling high frequency data from thousands of sources. Finally, Grafana is then connected to the database to enable dynamic, highly configurable, real-time visualization dashboards. These dashboards provide an interactable web interface to the data beyond what the current website enables.

6.1. Node-RED

Node-RED is a graphical programming tool that provides a collection of interfaces to hardware, data APIs, and more. Using a selection of processing nodes, the user generates a flow diagram representing the data ingestion and processing pipeline, which can be easily deployed with one click. A key advantage worth highlighting is that Node-RED’s low/no-code approach allows for an easily maintainable and scalable solution that can be adapted to new sensors and architectures. In Figure 6, the menu of nodes is visible on the left, while the deployment flow is shown in the center.

As discussed in the previous section, sensor data is published to a collection of MQTT topics in JSON format. The first step of the processing pipeline is to define a set of MQTT listeners that identify packets unique to each sensor type. When a message is identified, the packet is then processed into a JSON dictionary and subsequently parsed using a JavaScript node that allows for custom processing code. In this step, the relevant entries are parsed from the JSON dictionary and organized to be injected into the database. For instance, when a data packet from the BME280 sensor is received, it is first structured as a JSON dictionary that includes the payload data, FPort information, timestamp, and other metadata. The process begins by extracting and isolating the relevant objects from the packet, focusing on the payload data and associated details. A null check is performed to ensure the availability of data. Next, the flow uses the FPort field to identify the type of sensor data contained in the packet. For the BME280, the FPort value is 21 as indicated in Table 2. Once identified, the payload is decoded using a JavaScript function to convert the raw byte data into meaningful measurements, such as temperature, pressure, or humidity. The decoded data is then assigned a unique node ID derived from the sensor’s MAC address. Finally, the processed data is stored in the InfluxDB database for further analysis and visualization. The full pipeline is illustrated on Figure 6.

Credentials for connecting to the MQTT broker as well as database access tokens are automatically loaded from user-supplied configuration files. The entirety of the flows themselves are saved as JSON files, enabling easy version-control via git. Once a pipeline is developed in the web interface, the relevant files are stored and loaded automatically by a docker upon deployment.

6.2. InfluxDB

InfluxDB was chosen as the sensor network database for its ease of use, time-series optimizations, and open-source status. In particular, a machine with a 4 core CPU and 4 GB of RAM can support up to 5000 writes per second. Like Node-RED, InfluxDB has a web interface to enable data exploration and query generation. During the build process, our docker implementation generates a default bucket for the time series as well as default tokens to enable writing by Node-RED and reading by Grafana. A sample of the data explorer web interface is shown in Figure 7.

Complicated queries can be constructed using InfluxDB’s Flux query language, as it applies mathematical functions to compute statistical values and parameters. Queried data can be downloaded directly into CSV files for further analysis, or users can connect directly to the database by access token in their programming language of choice. Official client libraries exist for most popular languages, including Python, C++, and JavaScript.

6.3. Grafana

The final piece of the analysis toolbox is Grafana. When connected to InfluxDB, this tool enables the creation of a suite of real-time dashboards that visualize sensor data with a wide selection of graph types.

The dashboard is shown in Figure 8, in which a variety of visualizations have been selected to show real-time data for 7 consecutive days. Users can select a desired time range to query, or the dashboard can be set to automatically update at set rate (in the case of Figure 8, every 5 s). The size and location of each panel can be arranged by the user to suit their individual liking, and each panel can also expand for a full-screen view. Currently, dashboards exist for both Central and LoRaWAN nodes. Dashboards depicting data from multiple nodes are also available.

These three tools combine together to form a comprehensive analysis and visualization platform. Future work will incorporate automatic alerts for pollution exceedances as well as additional external data sources providing users with unparalleled access to local air quality data.

7. Deployment

Each LoRaWAN device is designed for deployment in real-world field environments. For the upcoming deployment, 10 LoRaWAN End Nodes will be set up with a single central gateway. The system is intended for use across the Dallas-Fort Worth metroplex, with plans to deploy 10 such systems, totaling 100 Nodes. The system has been publicly utilized in the city of Joppa, TX, as part of the Shared Air DFW initiative. Figure 9 shows the first two sensors deployed for public use in Joppa, TX, visible on a digital map available to the public at sharedairdfw.com (accessed on 18 May 2022).

Additionally, Figure 10 depicts a LoRaWAN Node deployment at the University of Texas at Dallas.

8. Sensor Calibration

In this study, we focus on the use of low-cost air quality sensing systems. Governmental agencies primarily rely on Federal Reference Method (FRM) or Federal Equivalent Method (FEM) instruments to enforce air quality regulations due to their high accuracy. However, these systems typically cost between $25,000 and $100,000 USD, making widespread deployment financially challenging [52]. Additionally, FRM and FEM instruments often have lower sampling rates. This proposal introduces a low-cost sensing solution calibrated using co-located data collection with an FEM system employing an identical sensing package. This approach aims to provide a cost-effective alternative for air quality monitoring while maintaining data reliability and frequency.

A significant challenge for low-cost particulate matter sensors is the lack of integrated heaters to prevent hygroscopic growth, which often leads to the overestimation of particle sizes and consequently, PM mass concentrations. This issue is particularly pronounced when ambient temperature and dew point are closely aligned. Thus, the initial phase of our calibration process involves applying a humidity correction to the particle counts, followed by adjustments to the PM concentrations to ensure accurate measurement under varying humidity conditions. The humidity correction is based on the methodology developed in [53]. The entire process of the sensor calibration is illustrated on Figure 11.

Previous research indicates that low-cost particulate matter sensors are affected by a range of environmental factors, such as humidity, temperature, and atmospheric pressure [54,55,56]. To improve the accuracy of these sensors, advanced machine learning algorithms are employed to calibrate humidity-adjusted readings, resulting in a more accurate representation of actual PM concentrations. The calibration process involved co-locating a low-cost sensing system, referred to as a MINTS Node (developed at the University of Texas at Dallas as part of the Multi-Scale Integrated Intelligent Interactive Sensing (MINTS) research program), with a sensing package similar to that of the LoRaWAN nodes near the Hawks Athletic Center in Fort Worth, Texas. This setup was placed alongside a Federal Equivalent Method (FEM) device, specifically the BAM 1022—Beta Attenuation Mass Monitor, which continuously collected particulate data for comparison.

The dataset, consisting of input features gathered from the BME280 and IPS7100 sensors of the MINTS Node, along with the corresponding target variable (PM_2.5 values from the BAM 1022), is first consolidated into a unified structure to optimize processing efficiency. To address the potential bias due to the unequal distribution of values, particularly the over-representation of smaller PM_2.5 values, the data is divided into bins based on the range of each feature and the target variable. Within each bin, random sampling is employed to ensure a roughly equal number of data points across bins, guaranteeing a representative sample of the overall dataset, especially for underrepresented values. The dataset is then shuffled and split into training and independent validation subsets, with 80% of the data used for training and 20% reserved for validation. This approach ensures that the validation set remains separate for assessing model performance on unseen data, thereby providing an unbiased evaluation of the model’s generalization capabilities.

A random data split, rather than a sequential one, was performed to prevent potential bias, as the dataset does not yet cover the full range of atmospheric and particulate matter concentrations. Given the significant seasonal climate variations in Texas, a sequential split might have skewed the dataset and compromised the model’s performance. Once sufficient data is collected, a sequential split approach will be considered.

The machine learning model, trained on a sensing package similar to that of LoRaWAN nodes, can be applied to data from these nodes, enabling it to map to the reference-grade monitor as well. The machine learning pipeline is deployed on the University of Texas at Dallas (UTD) cloud infrastructure via an MQTT pipeline endpoint. The corrected data is subsequently fed back into the analytical toolbox as an additional data point. The dataset, collected from December 2023 to July 2024, will continue to grow, allowing for ongoing model refinement. As new data from Fort Worth’s co-located sensors becomes available, we will retrain the model to improve its coverage of climate and particulate matter variability.

A comparative analysis of several non-linear, non-parametric machine learning methods—including neural networks, support vector regression, ensemble decision trees, and a baseline linear regression model—was conducted. The Degrees of Freedom (DoF) for each model were considered to provide insight into their complexity and generalization capacity. Hyper-parameter optimization was applied to select models, and the results, as shown in

Table 3. Performance Comparison of Machine Learning Models on Coefficient of Determination (R²) and Root Mean Square Error (RMSE) Metrics.

Machine Learning Model	DoF	Training		Independent Validation		Combined		Rank
		R2	RMSE	R2	RMSE	R2	RMSE
HPO Random Forest	≈4,999,334	0.97	1.72	0.82	11.88	0.92	5.58	1
Random Forest	≈1,246,877	0.97	1.84	0.82	12.25	0.91	5.81	2
HPO Neural Network	9857	0.86	9.53	0.74	17.36	0.83	11.36	3
Neural Network	2881	0.61	26.41	0.54	30.36	0.59	27.26	4
HPO Support Vector Machine	≈25,145	0.53	31.97	0.48	34.75	0.51	32.61	5
Support Vector Machine	≈29,697	0.48	34.88	0.44	37.05	0.47	35.35	6
Linear Regression	12	0.28	48.68	0.23	51.11	0.27	49.03	7

Multiple Machine Learning (ML) models were deployed to identify the most effective approach for predicting PM_2.5 levels from the BAM 1022 using the sensor ensemble integrated into the LoRaWAN Nodes. Root Mean Square Error (RMSE) values and Coefficient of Determination (R²) scores were recorded for the training, independent validation, and combined (training + independent validation) datasets across all considered models. The models were ranked based on R² values derived from an independent validation dataset. It was found that the Hyper-parameter Optimized (HPO) Random Forest Regressor, an ensemble of decision trees, provided the best performance for this task. Furthermore, the Degrees of Freedom (DoF) were incorporated into the analysis to quantify the number of independent parameters contributing to model flexibility. For Random Forest models, DoF was estimated as the product of the number of trees and the effective number of splits per tree. For Neural Networks, DoF was computed as the total number of trainable parameters, including weights and biases. For Support Vector Machines (SVMs), DoF was approximated based on the number of support vectors. Finally, for Linear Regression, DoF corresponded to the number of predictor variables plus the intercept term. The inclusion of DoF provides additional insight into model complexity and generalization capability.

, indicate that neural networks and decision tree ensembles were the top performers. The optimal performance was achieved using the hyper-parameter-optimized ensemble of decision trees, specifically the RandomForestRegressor algorithm from the scikit-learn library, which employs the bagging ensemble method. Key hyperparameters influencing model performance, such as the number of estimators (trees), the minimum samples required to split a node, the minimum samples required at a leaf node, the maximum number of features considered for each split, and the maximum depth of the trees, were tuned using a randomized search approach. The optimal values were determined to be 400 estimators, 2 samples to split a node, 1 sample at a leaf node, 1 feature per split, and a maximum tree depth of 70.

Eleven input variables were used in the multivariate non-linear machine learning regression model. These included humidity-corrected particle counts for particle sizes of 0.1 $\mathsf{\mu }$m, 0.3 $\mathsf{\mu }$m, 0.5 $\mathsf{\mu }$m, 1 $\mathsf{\mu }$m, 2.5 $\mathsf{\mu }$m, 5 $\mathsf{\mu }$m, and 10 $\mathsf{\mu }$m, along with key environmental factors such as atmospheric pressure, temperature, dew point, and relative humidity, collected using the BME280 sensor. The model’s performance, evaluated using the independent validation dataset, is illustrated in Figure 12, with key variables highlighted from the calibration process.

Figure 12 presents the machine learning model’s performance, as assessed on the independent validation dataset. Alongside this, a quantile-quantile plot illustrates model accuracy, and an additional plot highlights the most influential variables identified during the calibration phase.

Results

Figure 12 illustrates the results of a multivariate, non-linear, non-parametric machine learning regression for PM_2.5. The scatter plot (Figure 12a) shows the relationship between the PM_2.5 measurements from the BAM 1022 reference instrument (x-axis) and the PM_2.5 levels predicted by the machine learning calibration of the low-cost instrument (y-axis). For the training dataset, the model achieved a coefficient of determination (R²) value of 0.97 and a root mean square error (RMSE) of 1.72 μg/m³. On the independent validation dataset, the model attained an R² value of 0.82 and an RMSE of 11.88 μg/m³, indicating robust predictive performance and strong generalizability across both training and independent validation datasets. The RMSE of 11.88 μg/m³ represents approximately 12.75% of the dynamic range of PM_2.5, which suggests that the model’s error is a small proportion of the overall variability in PM_2.5 measurements during the relevant time period.

Figure 12b presents quantile–quantile (Q-Q) plots that compare the probability distribution functions (PDFs) of PM_2.5 data from the BAM 1022 reference instrument with those from the machine learning-calibrated low-cost instrument. A straight line in a Q-Q plot indicates identical distributions, as it compares the percentiles of one PDF with the corresponding percentiles of the other. In this instance, the plot shows a straight line from the 25th to the 75th percentiles, confirming the successful calibration of the low-cost instrument.

Figure 12c illustrates the relative importance of the input variables used in the machine learning calibration of the low-cost optical particle counter (IPS7100) as well as the climate sensor (BME280) found within the LoRaWAN System. The importance metric quantifies the increase in error when a specific input variable is omitted. The bar plots are arranged in descending order of importance, with the top bar representing the most influential variable. This analysis highlights the expected significance of the particle count for 2.5-micron particles, along with environmental factors such as temperature, pressure, dew point, and humidity, all of which contribute to the accurate calibration of the low-cost instrument against the reference instrument (BAM 1022).

After developing the machine learning model, we applied it successfully to data collected from sensors across our network. Figure 13 shows a time series from a LoRaWAN sensor deployed at the University of Texas at Dallas in Richardson, Texas, for October 2024. The figure compares the raw PM_2.5 measurements from the IPS7100, the humidity-corrected PM_2.5, and the machine learning-corrected PM_2.5 values.

9. Empowering Policy Control and Urban Health

Air quality management using low-cost PM_2.5 sensor systems, as discussed in this article, presents an innovative approach to enhancing policy control and improving urban health outcomes. The system outlined here provides real-time, high-resolution air quality data, coupled with climate data and machine learning outputs, which significantly enrich the data stacks used by policymakers. Traditionally, policy frameworks have relied on expensive, large-scale monitoring systems. In contrast, the proposed system offers granular spatial and temporal assessments of air quality, capturing episodic pollution spikes, such as those caused by wildfires, temperature inversions, or urban industrial pollution. Similar approaches have been successfully implemented in California and Sarajevo, where properly calibrated low-cost sensors aligned with regulatory networks, demonstrating their utility in localized air quality monitoring and public engagement [57,58].

The LoRaWAN sensors, being small and cost-effective, enable widespread deployment and maximum ground coverage. This extensive network allows for proactive identification and simulation of air pollution hot-spots, aiding in urban planning and pollution control measures. Furthermore, the system’s ability to simulate real-time scenarios, such as the spread of a biological agent through a city, mimics plume dispersion dynamics with high accuracy. This capability enhances emergency response strategies and public safety measures, offering critical insights during potential crises.

Our infrastructure has been further enhanced to include real-time forecasting by integrating data from additional environmental monitoring systems built and deployed by our laboratory. These systems provide valuable inputs, such as wind speed and direction. When combined with data from the LoRaWAN sensor networks discussed here, alongside open-source climate and pollution data, this integration creates a dynamic and responsive platform for environmental forecasting. This integration supports proactive environmental health strategies, potentially leading to community-level interventions. For example, correlating high PM_2.5 levels with asthma incidents in schools could trigger early warnings, allowing authorities to adjust traffic or industrial activities to mitigate exposure. This approach not only reduces health risks but could also improve key learning outcomes, such as student absenteeism, by minimizing environmental triggers that exacerbate respiratory conditions [59]. By emphasizing human-environment interactions and promoting citizen science, the method advances sustainable development goals and improves access to air quality monitoring.

Furthermore, the use of LoRaWAN sensor networks addresses equity concerns by filling gaps in regulatory networks, particularly in under-monitored areas. The enhanced data can also refine atmospheric models and guide urban planning efforts to reduce pollution sources. Ultimately, these low-cost sensors bridge technological and economic gaps, enabling communities and governments to make informed, data-driven decisions that promote healthier environments.

Our sensor network is already being utilized by organizations such as Downwinders at Risk, the City of Plano, and the City of Gunter, which actively use the sensors to monitor and assess air quality in their communities. Through the integration of these sensors into local environmental monitoring frameworks, these organizations gain real-time data to better understand pollution patterns and identify key areas for intervention. Downwinders at Risk, in particular, advocates for the use of this data to support environmental justice, ensuring vulnerable communities are shielded from the health impacts of air pollution. The City of Plano and the City of Gunter contribute to more informed decision-making, leveraging insights from these low-cost sensors to implement policies that reduce exposure to harmful particulate matter, particularly among vulnerable populations such as children with asthma.

10. Discussion

This study presents a self-powered IoT network leveraging LoRaWAN technology for air quality monitoring, addressing the critical need for scalable, cost-effective environmental sensing systems that offer open-source access to high-frequency data through public portals, alongside transparent software and firmware availability. Although the existing literature [60,61,62,63,64] provides insights into some components of these technologies, this work integrates all essential elements into a comprehensive framework.

Through LoRaWAN, we enable both scientists and the public to collect data and securely transmit it to high-performance computing infrastructures for advanced analysis. This ‘citizen science’ model has gained momentum with the availability of affordable, high-precision sensors that reliably measure pollutants such as particulate matter. Many of these LoRaWAN nodes are now deployed within community areas, making real-time air quality monitoring accessible to local residents.

While FEM/FRM sensor networks offer reliable but low-frequency monitoring, they are costly and can introduce latency. Our objective was to create a low-cost alternative that caters to resource-conscious settings, maintaining comparable data quality through machine learning to achieve higher frequency and minimal latency. Furthermore, our devices utilize LoRaWAN for data transmission, avoiding the need for complex and expensive networking solutions like Wi-Fi.

In this study, we developed and trained a machine learning model using data collected from a specific location, which was then applied to a different geographic region. This approach introduces a geographic variability constraint, as the training and application locations differ, potentially influencing model performance. Atmospheric conditions vary across regions, which can affect the model’s accuracy when applied to new areas. However, our methodology considers exposure to a wide range of air masses originating from various geographic regions. Continuous, second-by-second air quality measurements recorded at a fixed location for training data collection allow us to capture atmospheric conditions shaped by multiple sources, including long-range trans-boundary pollution. For example, air masses reaching the Dallas-Fort Worth (DFW) area are influenced by sources such as the Sahara Desert, the Gulf of Mexico, wildfires in California, and local emissions within the metroplex. This variability in atmospheric transport introduces significant fluctuations in the dataset, which, rather than being a limitation, is an essential characteristic of the data, reflecting the complexity and dynamic nature of real-world atmospheric conditions.

11. Challenges and Future Directions

One of the key challenges encountered during the deployment of these sensors was the limited communication range in urban environments, where access to mounting locations with a clear line of sight was constrained. Consequently, the communication range between the LoRaWAN node and the gateway was restricted to less than one mile. To overcome this limitation, ongoing research is exploring the use of high-powered LoRaWAN radios with a maximum output of 20 dBm, in contrast to the 14 dBm radios used in the initial deployment. In addition, consultations with industry experts have been conducted to identify optimal antenna configurations aimed at improving signal strength and range.

Looking ahead, efforts are underway to expand the proposed system by integrating larger batteries and more energy-efficient solar panels. Currently, the battery life spans from 6 to 12 months, limiting the deployment and data collection capacity. The planned upgrades will enable the integration of additional sensors, allowing for the collection of more comprehensive measurements and enhancing the overall quality of the data. Furthermore, we are advancing our machine learning training processes, which are being conducted in both controlled laboratory environmental chambers and a mobile laboratory mounted on an electric vehicle. This approach is intended to improve the accuracy and reliability of the model, complementing ongoing training efforts with the sensing system deployed in Fort Worth, TX.