Evolving trends in application of low-cost air quality sensor networks: challenges and future directions

Abstract

Low-cost air quality sensors (LCS) are increasingly used to complement traditional air quality monitoring yet concerns about their accuracy and fitness-for-purpose persist. This scoping review investigates topics, methods, and technologies in the application of LCS networks in recent years that are gaining momentum, focusing on LCS networks (LCSN) operation, drone-based and mobile monitoring, data fusion/assimilation, and community engagement. We identify several key challenges remaining. A major limitation is the absence of unified performance metrics and cross-validation methods to compare different LCSN calibration and imputation techniques and meta-analyses. LCSN still face challenges in effectively sharing and interpreting data due to a lack of common protocols and standardized definitions, which can hinder collaboration and data integration across different systems. In mobile monitoring, LCS siting, orientation, and platform speed are challenges to data consistency of different LCS types and limit the transferability of static calibration models to mobile settings. For drone-based monitoring, rotor downwash, LCS placement, flight pattern, and environmental variability complicate accurate measurements. In integrating LCS data with air quality models or data assimilation, realistic uncertainty quantification, ideally at the individual measurement level, remains a major obstacle. Finally, citizen science initiatives often encounter motivational, technological, economic, societal, and regulatory barriers that hinder their scalability and long-term impact.

Introduction

Air pollution is a global environmental challenge with negative impacts on human health, ecosystems, and quality of life^1,2,3. The conventional approach for air quality monitoring relies on networks of expensive, high-quality, fixed-reference monitoring stations. Reference monitoring networks are sparse due to the high cost, complexity of the equipment, and need for trained operators⁴. However, for more than a decade, researchers have investigated the potential of low-cost sensors (LCS) to fill in the gaps and expand participation in air quality monitoring⁵. The widespread deployment of LCS, characterized by their affordability (<USD 2,500 as defined by the US EPA Sensor Toolbox⁶), portability, and ease of use, has resulted in a new era in air quality monitoring and management.

The adoption of LCS, has empowered a range of actors, including citizens, researchers, and policy-makers, to actively engage in data collection and analysis, even in low- and middle-income countries^7,8. The cost of LCS can vary depending on the pollutant species and the manufacturing process, as well as the features included (i.e., inlet drier, solar panels, software, etc.). Existing LCS deployments have made it possible to gather real-time-high-resolution data⁹, improving our understanding of air pollution patterns at local scales. At street, city, regional, and even global scales, LCS measurements can have benefits for validating atmospheric composition models¹⁰.

Individuals and communities have deployed LCS in various environments, both indoors and outdoors, to collect data on air pollutants such as particulate matter (PM)¹¹, volatile organic compounds, and gases such as O₃ (Ozone) and NO₂ (Nitrogen dioxide)^12,13. LCS have also been integrated into smart city initiatives^14,15, environmental monitoring networks^16,17, and public health systems^5,18,19. The data obtained from these LCS can inform evidence-based decision-making, policy formulation, and the development of targeted interventions to improve air quality¹⁴.

However, despite the many advancements, uncertainties and limitations of LCS remain¹⁵. These include concerns relating to data accuracy, cross-sensitivity to ambient pollutants, sensor drift, data standardization, and interoperability between different LCS providers²⁰. The most important LCS limitation that is extensively highlighted in literature is their lower accuracy compared to reference instruments. LCS must be calibrated for their data to be used robustly. Field calibration, or the correction of LCS, occurs by co-locating one or more sensors with a reference monitor for an extended time period which can range from ~ weeks to months, and then developing a model to correct the LCS so that the measurements closely represent those from the reference monitor. The logistics of deploying and calibrating large LCS networks (LCSN) can be challenging. Finally, the short operating lifetimes of these instruments (typically between 2 months to 2 years) require them to be replaced often¹⁸.

An additional, equally important but often overlooked aspect is the absence (at least until recently in the European Union) of a standardized framework for evaluating LCS performance. In Europe, CEN (European Committee for Standardization) has published standards for gases, CEN/TC 264/WG 42, and for particles, CEN/TS 17660-2:2024, which aim at simplifying the use of LCS as part of the regulatory framework resulting from the Clean Air for Europe Directive 2008/50/EC²¹, and its sub-directives. However, a new updated directive was recently put in place (2024/2881) that specifies that LCS measurements can act as indicative. Indicative measurements can supply sufficiently reliable data for assessing compliance with limit values, target values, critical levels, and alert or information thresholds. They can also provide the public with relevant information, potentially helping to reduce the number of reference mointoring stations needed.

Due to the rapid advancements and increasing interest in the application of LCS, numerous review papers have been published on different aspects of LCS technology. These papers cover field and laboratory sensor applications and development^{4,5,7,19,22,23,24,25,26,27,28,29}, field calibration^30,31,32,33 techniques, data quality, and validation^8,15,34,20, protocols, and enabling technologies³⁵, air quality data crowd-sourcing^14,29,36, and drone-based measurements^37,38.

In this paper, we focus on reviewing some of the most emergent LCS applications found in the literature and use our domain knowledge and expertise to identify and explore the “evolving trends” in the application and operation of LCS. We identify key challenges that need to be overcome to guide future LCS research. By “evolving trends” in the application of LCS, we refer to new developments, methodologies, or shifts in how these sensors are used or perceived in environmental monitoring (Fig. 1). We selected the following trends that are currently evolving: (1) LCS operational data quality evaluations via advanced computing technologies (Industry 4.0), (2) LCS in mobile air quality monitoring (ground and aerial), (3) LCS data integration with air quality models and remote sensing datasets, and (4) LCS in citizen science.

Fig. 1: Overview of “Evolving Trends” in the application of air quality low-cost sensors (LCS).

On the left: available methods and technologies for monitoring and estimating air quality. On the right: established and evolving trends in the application of LCS, in combination with the methods and technologies presented on the left side.

Fixed LCSN data integration with regulatory environmental monitoring networks

Direct field calibration

Typically, the deployment of an LCSN begins with co-locating some sensors (one or more), or all if the fleet is small, alongside a high-precision reference instrument for a specific period, often 1–4 weeks³⁹, depending on the measurement frequency, to develop field calibration functions. The co-location can involve either circulating a mobile reference station between sites or positioning sensors near a fixed reference station before relocating them to their final monitoring locations, where ongoing access to reference data is generally unavailable. Such direct, site-specific calibration forms the foundation of improving the accuracy of LCS, as demonstrated by Barcelo-Ordinas et al.⁴⁰, De Vito et al.⁴¹, Ferrer-Cid et al.⁴², and Bisignano et al.⁴³. However, reliance on initial baseline calibration alone has inherent limitations, including vulnerability to sensor drift or changing local conditions that could render the initial calibration insufficient. As networks expand, and/or become mobile⁴⁴, maintaining performance often necessitates integrating direct co-location with proxy-based, transfer, or network-wide calibration strategies.

Proxy-based field calibration

When access to fixed reference stations is limited or infeasible, proxy-based calibration provides an adaptable alternative for extending calibration coverage. This semi-blind approach³¹ relies on using one or more mobile LCS as temporary proxies: first, co-locating the proxy sensor with a reference station to obtain a valid calibration function, then relocating it sequentially to other uncalibrated units to transfer the calibration indirectly. This chain or node-to-node method can be repeated across multiple sites, forming a practical solution for sparse reference networks or challenging terrain. Studies by Kizel et al.⁴⁵, Maag et al.⁴⁶, Sá et al.⁴⁷, and Vajs et al.⁴⁸ demonstrate the feasibility and operational benefits of proxy-based strategies under real-world conditions. However, these methods can be sensitive to cumulative error propagation if the chain is extended too far without periodic re-anchoring to a high-quality reference. Consequently, proxy-based calibration is best applied in combination with direct or periodic co-location to maintain accuracy across the network.

Transfer-based field calibration

Transfer-based calibration extends the use of an initial site-specific calibration function to other sensors of the same type deployed under comparable conditions. For the successful deployment of this strategy, similar climatological and source contexts^49,50,51, and low intra-sensor variability⁵² are important in enabling a model trained at one location to be applied elsewhere with minimal additional reference data. Studies by Cheng et al.⁵³, Laref et al.⁵⁴, Liu et al.⁵⁵, Cui et al.⁵⁶, Villanueva et al.⁵⁷, and DeSouza et al.⁵⁸ illustrate this approach and show its efficiency when full co-location is impractical or prohibitively costly. Transfer calibration can reduce operational effort and costs; however, its performance depends strongly on the similarity between training and deployment environments⁵². Mismatches in pollutant mix, weather patterns, or local sources can degrade model accuracy, requiring periodic adaptation or supplemental local correction^49,51.

A more generalized variant, global or multi-site transfer calibration (Fig. 2), aims to develop calibration models that remain robust across diverse geographic contexts and temporal conditions⁵⁹. Instead of relying on a single reference site, this strategy combines data from multiple co-location pairs to build an aggregated model capable of capturing wider spatio-temporal variability. For example, Miquel-Ibarz et al.⁶⁰, Chu et al.⁶¹, Vikram et al.⁶², De Vito et al.⁶³, Solórzano et al.⁶⁴, Bagkis et al.⁶⁵, and Villanueva et al.⁵⁷ demonstrate how global calibration frameworks can extend the validity of local LCSN, and at the same time, reduce the frequency of site-specific recalibration. Despite their scalability, these approaches require sufficiently diverse and representative training datasets to minimize location-specific biases. In practice, global models are most effective when paired with adaptive updates to account for emerging concept drift or unexpected local deviations.

Fig. 2: Schematic visualization of the global calibration approach.

Given N co-located pairs and six distributed air quality low-cost sensors (LCS), the data are retrieved into a central server. There, a global calibration function is learned with machine learning, trained on the combined dataset from all the co-located pairs, and the function is then applied to the six LCS. We give the credit to²²³ and the following: https://www.wunderground.com/sensors, from those, some images are used and modified to generate this figure.

Connectivity-based field calibration

Connectivity-based approaches integrate the physical and functional relationships between sensor nodes into the calibration process, often using graph theory to represent how pollutant concentrations disperse spatially and temporally across a network. By modeling each sensor as a node and its relationship with neighboring sensors as edges, graph-based methods can propagate local corrections, detect anomalies, and help stabilize measurements against drift or unexpected local variation. Studies by Zhang et al.⁶⁶, Wu et al.⁶⁷, Jin et al.⁶⁸, Iyer et al.⁶⁹, Li et al.⁷⁰, and Ferrer-Cid et al.⁷¹ illustrate how connectivity-based models can enhance calibration performance in dense urban deployments and complex environments. Although these methods are promising for using the spatial structure of large LCSN, they require careful design of graph topology, sufficient data for training, and computational resources to operate in near real time.

Deep Learning as a cross-cutting layer

Deep learning methods increasingly complement traditional calibration by capturing non-linear relationships, managing domain adaptation, and improving performance accuracy (Supplementary Table 1). Architectures such as multi-layer perceptrons, recurrent neural networks, and graph neural networks have been deployed for both calibration and gap-filling tasks. They offer improved accuracy under changing ambient conditions. Recent studies, including Dey et al.⁷², Schlund et al.⁷³, Zhang et al.⁶⁶, Chen et al.⁷⁴, and Okafor and Delaney⁷⁵, demonstrate how deep learning can improve the stability and transferability of LCSN calibration and imputation pipelines. Combined with connectivity-based frameworks, deep learning can further strengthen adaptive calibration by exploiting both local network structure and spatiotemporal dynamics.

Graph-based modeling of LCSN

Graph-based modeling is increasingly recognized as a framework for capturing the spatial and temporal dependencies inherent in LCSN⁷⁶. By representing each sensor as a node and linking them with edges that encode physical distances, wind flows, or other dispersion relationships, graph-based approaches align naturally with how pollutants propagate in urban and regional settings. Graph Neural Networks have emerged as tools for anomaly detection, data imputation, and spatiotemporal prediction within these networks. Recent studies show this momentum: Wu et al.⁶⁷ applied Graph WaveNet and spatiotemporal graph convolutional networks to identify faulty sensors in Taiwan’s extensive LCSN, outperforming rule-based and conventional ML baselines. Xu et al.⁷⁷ used dynamic, wind-informed graphs for short-term PM_2.5 forecasting, and Jin et al.⁶⁸ advanced this further using a spatio-temporal multi-attention multi-graph framework to improve long-term predictions. Iyer et al.⁶⁹ combined message-passing recurrent networks with geostatistical techniques to generate fine-grained air quality maps for Delhi, robustly handling sparse or intermittent measurements. Li et al.⁷⁰ proposed the GCN-Informer based on the informer⁷⁸ architecture, integrating graph convolutional networks with transformer-based architectures for air quality prediction. Ferrer-Cid et al.⁷¹ focused on outlier detection, introducing the Volterra Graph-based Outlier Detection (VGOD) mechanism to identify and correct erroneous sensor data. Together, these studies highlight how graph-based methods can strengthen calibration, anomaly detection, and prediction pipelines by embedding domain knowledge of pollutant transport and inter-sensor relationships within LCSN.

LCS in mobile monitoring

Ground measurements

Mobile monitoring has been used to complement stationary measurements of air pollution to capture hyperlocal (street-by-street) variations in pollutants. Although some pollutants like PM_2.5 display more variation at the regional scale, concentrations of other traffic-related pollutants like NO₂ and CO can vary over a few meters^79,80,81. LCS are increasingly being adopted for mobile monitoring purposes. Although there have been several reviews of mobile air pollution monitoring using higher-quality instruments^82,83,84, none of these reviews have focused on the application of LCS.

Mobile monitoring using high-quality instruments often requires specialized vehicles operated by trained staff. A key advantage of using LCS for mobile applications is the possibility of using non-specialized vehicles, such as bicycles^85,86, trains⁸⁷, trash trucks⁸⁸, people⁸⁹, and taxi cabs⁹⁰ as monitoring platforms. LCS can be deployed for pedestrian-based air quality monitoring as well^9,91. The use of these routine vehicles for sampling purposes can increase the spatial areas and time of sampling cost-effectively. O’Keeffe et al.⁹² and Anjomshoaa et al.⁹³, for example, demonstrated a methodology to evaluate the number of routine vehicles needed to be outfitted with sensors to cover a given space-time sampling cross-section of a city. The portability of LCS has also enabled the development of personal exposure assessments for epidemiologic and awareness-raising applications^94,95,96. Kappelt et al.⁹⁷ and Russell et al.⁹⁸ employed LCS to determine the PM levels inside the metro system (i.e., onboard trains) of Copenhagen and showed that the PM concentration levels increased when compared with the ambient measurements out of the metro system. Mobile measurements from LCS have also been successfully integrated into models to develop high-resolution exposure assessments for different cities^99,100,101 as well as with measurements from other sources (stationary monitors, satellite instruments, and models)¹⁰².

Aerial-based measurements

An increasingly important development in atmospheric science is monitoring the vertical distribution of air pollutants¹⁰³, from ground level to several hundred meters¹⁰⁴. The use of LCS on-board drones (or Unmanned Aerial Vehicle) has emerged for three-dimensional measuring of key pollutants such as PM₁, PM_2.5, PM₁₀, NO₂, CO, and O₃ at altitudes (from ground level to several hundred meters) and is currently evolving as a promising technology to collect data at different altitudes^{105,106,107,108} (see Supplementary Note 1).

Conventional ground-based monitoring networks cannot resolve the vertical gradients in pollutant concentrations that shape local exposure, urban heat islands, and near-source pollution dynamics. Recent research shows that copter drone-based LCS monitoring can deliver high-resolution datasets when combined with data analytics^{106,108,109,110,111}. Hemamalini et al.¹¹², for example, integrated drone-mounted LCS data with deep learning to forecast air pollution levels near open dumpsites in India. Or Villa et al.³⁸ evaluated the utility of using small drones for ambient mobile monitoring.

Recent studies have also demonstrated the use of fixed-wing drones for vertical profiling of air pollutants, particularly in peri-urban or special event contexts^113,114,115. Although fixed-wing drones provide greater flight endurance and spatial reach, regulatory constraints and safety concerns make them less suitable for urban environments. They have been successfully used to monitor pollutants near airports, assess Saharan dust events, and measure PM₁, PM_2.5, and PM₁₀ profiles in rural India. However, their application in dense urban environments remains limited compared to copter-based drones. In addition to air quality monitoring, the recent deployment of drone-based LCS systems has shown promise in validating numerical models such as Weather research forecasting coupled with chemistry (WRF-Chem), Community Multiscale Air Quality (CMAQ), or EURopean Air pollution Dispersion – Inverse Model (EURAD-IM) models^{116,117,118,119}.

LCS data integration with air quality models and remote sensing datasets

Integrating LCS systems with models via data assimilation¹²⁰ and data fusion¹²¹ techniques has the potential to enhance the accuracy and resolution of air quality¹²² mapping at local and regional scales and is thus a promising application of LCSN. Although individual LCS accuracy is typically limited, their fusion with model data (e.g., local-scale dispersion or chemical transport models) using these techniques can offer valuable complementary insights, provided the sensors are carefully calibrated, quality-controlled, and uncertainties are properly assessed. This integration adds value to the LCSN by robustly and objectively interpolating point observations in space and simultaneously enhances models by constraining simulations with real observations^102,123.

We distinguish here two main approaches for combining LCS observations with models or auxiliary data: (1) Data fusion, which relies on geostatistics¹²⁴ or similar traditional statistical methods, typically applied offline as post-processing and (2) data assimilation, stemming from the field of numerical weather prediction, which is typically conducted online and actively updates the model state during runtime¹²⁵.

Data fusion techniques

Data fusion techniques traditionally refer to combining multiple observations, but here we use the term to also include integrating observations with models to clearly distinguish from data assimilation methods (e.g., Shetty et al.¹²⁶). Data fusion, which in our context typically refers to approaches such as geostatistics, land-use regression, or machine learning (ML), allows for merging data from various air quality sources to create an improved overall mapping result. These techniques typically use statistical approaches applied as post-processing rather than directly assimilating data into running models. (1) Spatial Interpolation: Includes simpler methods like nearest neighbor resampling or Inverse Distance Weighting (IDW), as shown, for example, by Chu et al.⁶¹ in mapping PM_2.5 with LCS. (2) Geostatistics: These methods use spatial correlations among measurements to estimate pollutant levels at unsampled locations. Kriging techniques commonly interpolate pollutant concentrations, potentially incorporating auxiliary data from dispersion or chemical transport models. Such approaches generate high-resolution air quality maps accounting for spatial variability. Examples include studies by Li et al.¹²⁷, Castell et al.¹²⁸, Schneider et al.¹⁰², Schneider et al.¹²⁹, and Gressent et al.¹³⁰. (3) Land-Use Regression (LUR): LUR models correlate air quality data with land-use and emission-related variables (e.g., traffic density, population density). Using statistical regression incorporating spatial and temporal predictors, LUR estimates pollutant levels at unmonitored sites. Studies using this method include Adams et al.¹³¹, Coker et al.¹³², Jain et al.¹³³, Lim et al.¹³⁴, and Weissert et al.¹³⁵. (4) Machine Learning (ML): ML is widely employed for calibrating LCS, yet its use in combining sensor data with models remains less explored. Demonstrations of ML methods integrating simulated LCS observations with models are provided by Coker et al.¹³², Guo et al.¹³⁶, Jain et al.¹³³, Liang et al.¹³⁷, Lim et al.¹³⁴, and Lopez-Ferber et al.¹³⁸.

Data assimilation techniques

Data assimilation techniques for atmospheric composition integrate observations from multiple sources, such as LCS, traditional monitoring stations, and satellites, with mathematical models to improve air pollutant concentration estimates¹³⁹. These methods assimilate observed data directly into the model’s internal state estimation process, typically represented by a vector of model variables. (1) Optimal Interpolation (OI): OI minimizes differences between model predictions and observations using statistical interpolation to weight observations by proximity. It can function offline without directly updating the model’s internal state vector. Examples using OI with LCS data include Mijling¹⁴⁰, Schneider et al.¹²², and Hassani et al.¹⁴¹ (see Fig. 3 as an example). (2) Ensemble Kalman Filter (EnKF): EnKF iteratively updates an ensemble of model states by integrating observations with model predictions. The observations adjust ensemble members, enhancing the predictions. Lopez-Restrepo et al.¹⁴² demonstrated EnKF use with LCS. (3) 3D-Var, 4D-Var, and 4D-EnVar: These variational methods minimize a cost function^143,144 to assimilate observations into models, often employing adjoint models to efficiently calculate gradients. 3D-Var updates at single time points, while 4D-Var assimilates data across a temporal window. Although primarily designed for surface observations or satellites, these systems can be adapted for LCS data, despite limited applications to date. Lopez-Ferber et al.¹³⁸ used 3D-Var with simulated sensor observations.

Fig. 3: Integration of quality-controlled low-cost sensor observations with model information using data assimilation.

This is the example of average PM_2.5 concentrations from December 1, 2020, to February 28, 2021, for the city of Kristiansand, Norway. Top left panel: Model a priori dataset (background) together with sensor observations (symbols); top right panel: depiction of innovation, showing disparities between model predictions and sensor data at deployment sites; bottom left panel: the resultant concentration field from data assimilation (the “Analysis”); bottom right panel: visualization of differences between analysis and the model, highlighting spatial corrections made during assimilation. Base map copyright OpenStreetMap contributors, map tiles by Stamen Design under CC BY 3.0. Modified from a similar figure in Hassani et al.¹⁴¹.

Distinctions between these techniques can overlap; for example, OI shares similarities with geostatistical methods, and many studies combine techniques like land-use regression with ML. Each approach has strengths and limitations, but all are valuable for urban air quality mapping using LCSN, depending on specific requirements and data availability.

Low-cost sensors data integration with satellite observations

Due to the growing interest in integrating LCS data with satellite observations, we have focused more on this topic. Earth observation satellites have become key tools in air quality monitoring¹⁴⁵. They overcome the spatial limitations and uneven distribution of ground monitors by providing daily, global data on air pollutants¹⁴⁵. Instruments like TROPOMI on Sentinel 5-Precursor (S5-P), Moderate Resolution Imaging Spectroradiometer (MODIS) on Terra and Aqua, the Ozone Monitoring Instrument on Aura, and VIIRS on the Suomi NPP/NOAA satellites retrieve key atmospheric components such as NO₂, O₃, CO, Methane (CH₄), SO₂, formaldehyde, as well as Aerosol Optical Depth/Thickness (AOD/AOT) and other aerosol properties through spectral radiance measurements¹⁴⁶.

Satellites provide columnar measurements of atmospheric constituents, representing total atmospheric content¹⁴⁷. However, these measurements cannot be directly used for urban-scale health exposure assessments, which require information specifically about surface-level concentrations¹⁴⁸. Many studies convert satellite-derived columnar measurements into spatially continuous surface concentration estimates by integrating them with ground-based air quality observations (such as reference stations, LCS, etc.) and/or outputs from chemical transport models through data fusion or assimilation techniques^149,150. Supplementary Table 2 summarizes a selection of recent studies that combine LCS and satellite observations to estimate or analyze surface air pollutant levels. These studies can broadly be categorized based on the methodology used for LCS integration: (1) Spatial Interpolation-Based Approaches. LCS data are used in conjunction with satellite-derived variables (e.g., AOD or tropospheric NO₂ columns) in spatial interpolation models to estimate surface concentrations across a region. (2) Regression and ML-Based Approaches. LCS data are integrated as either independent or dependent variables, or serve in both roles, alongside satellite inputs and auxiliary parameters (e.g., meteorology, land use), to enhance spatial mapping and predictive accuracy of surface pollution levels.

Fewer studies have adopted interpolation-based approaches for integrating LCS and satellite data (e.g., Li et al.¹²⁷, Chao et al.¹⁵¹, and Gupta et al.¹⁵²). Most LCS networks are concentrated in urban areas, where spatial interpolation alone may not be optimal, as it tends to smooth out localized pollution peaks, leading to underestimation. This highlights the need for data integrations and alternative approaches, with ML being the most adopted approach due to its robustness and ability to model complex, non-linear relationships. Tree-based ML algorithms, particularly the Random Forest, are among the most widely used. For example, Huang et al.¹⁵³ found that Random Forest models outperformed or matched other ML methods such as Support Vector Machines, Gradient Boosting, and neural networks, citing superior predictive accuracy and greater computational efficiency, along with ease of training and interpretability.

Neural networks, particularly Convolutional Neural Networks (CNNs), can capture spatial and temporal dependencies through their layered structure but are often computationally intensive and require large, well-curated datasets for training. As a result, their application in this context has been limited. Nevertheless, Liang et al.¹³⁷ demonstrated the potential of CNNs by integrating LCS data, AOD, and traffic indicators to estimate PM_2.5 at a high spatial resolution of 30 m over Denton County, Texas, ideal for capturing intra-urban pollution variability.

In ML applications, LCS data have been used as targets, predictors, or both. When used as targets, LCS data offer higher spatial density, improving model performance, especially in residential areas, compared to using only reference-grade data^153,154,155. However, differences in data accuracy can introduce bias¹⁵⁴. Weighted approaches, like Bi et al.¹⁵⁵’s Random Forest model, reduce this bias by giving more weight to reference data. As predictors, LCS help enhance spatial detail. For example, Fu et al.¹⁵⁶ integrated LCS and satellite NO₂ data to improve high-resolution pollution mapping. Some studies use LCS both as inputs and targets by incorporating interpolated LCS layers (e.g., via kriging) to capture fine-scale variability^137,157. In addition, LCS data have also been used to derive correction factors for satellite-based PM_2.5 estimates¹⁵⁸.

Satellite data functions as an independent covariate in all reviewed studies (Supplementary Table 2). MODIS MAIAC (Multi-Angle Implementation of Atmospheric Correction) AOD¹⁵⁹ retrieved from the Terra and Aqua satellites is mostly used in PM_2.5 studies, mainly due to its fine 1 km spatial resolution, while TROPOMI is preferred for NO₂ monitoring due to its high spatial resolution of trace gas retrievals. The importance and predictive value of satellite variables can vary across studies and pollutants. For example, Huang et al.¹⁵³ and Yu et al.¹⁵⁴ identified satellite AOD as highly important for PM_2.5 estimation, while Liang et al.¹³⁷ and Bi et al.¹⁵⁷ observed lower importance, likely due to the complex spatio-temporal AOD-PM_2.5 relationship. The effectiveness of satellite AOD as a predictor can depend on atmospheric conditions, aerosol type, vertical distribution, and the presence of meteorological covariates in the model¹⁶⁰. In contrast, for pollutants like NO₂, where columnar measurements are more directly related to surface concentrations, satellite products such as TROPOMI NO₂ may provide better predictive power.

Across studies, RMSE values range from 1.15 – 21 µg m^-3 and R² values from 0.7 and 0.86, mostly based on K-fold cross-validation. Particularly, the RMSEs for ML-based studies were lower at 1.15−9.3 µg m^-3. Most studies focus on urban-scale PM_2.5 mapping, especially in the United States and China, with limited work at regional scales. This emphasis reflects the dense availability of PM_2.5 LCSN. The most common spatial-temporal resolution across studies is daily averages at a 1 km spatial scale, though some studies reach finer spatial resolutions of 30−100 m.

Low-cost sensors application in citizen science and community engagement projects

Citizen science projects using LCS technology have proliferated globally, raising greater public engagement and awareness of environmental issues. This part of our review aims to provide an overview of the current state of LCS air quality monitoring within citizen science projects, including both the successes and obstacles encountered in these initiatives.

Kruger and Shannon¹⁶¹ define citizen science as the involvement of citizens in science as active researchers. In the literature, “citizen science” is also referred to by several equivalent terms, including participatory science, community science, crowdsourced science, public science, and amateur science. In some cases, depending on the context and the specific focus of the research, open science, volunteer science, distributed science, and Do It Yourself (DIY) science can also be relevant. The emergence of LCS has opened new possibilities for citizen science and community engagement in the air quality field, allowing individuals to actively participate in monitoring and addressing air pollution-related issues^162,163. The evolving application of LCS in citizen science projects in air quality monitoring has brought numerous benefits and opportunities. LCS have improved air quality data collection by spreading independent community-based networks for air quality monitoring²⁵. Through citizen science projects, individuals can install sensors in their homes, workplaces, or public spaces, providing a vast amount of real-time data^5,164. This crowd-sourced approach enhances spatial coverage^4,165, enables the identification of pollution hotspots^141,166, and helps prioritize the required interventions. Additionally, engaging citizens in air quality monitoring raises awareness about the effects of air pollution on human health and the environment^167,168. This awareness motivates behavioral changes regarding indoor and outdoor air pollution, such as reducing vehicle emissions, adopting sustainable transportation modes, or modifying personal habits (e.g., Park et al.¹⁶⁹, leading to a collective effort to mitigate pollution). Engaging citizens in air quality monitoring projects empowers them to actively participate in environmental decision-making processes^14,170. Through collecting and analyzing personalized air quality data, citizens become stakeholders and contributors to scientific research, policy-making, and environmental management²⁰. This empowerment raises a sense of ownership and responsibility for environmental issues.

LCS bridge the gap between citizen scientists and the scientific community. Involving citizens in data collection facilitates knowledge exchange, collaboration, and co-creation of solutions¹⁷¹. The data produced by citizen science projects is made decision-ready through user-friendly formats such as real-time data streams, interactive maps, and reports (e.g., Longo et al.¹⁷² or Barros et al.¹⁷³). Scientists can benefit from datasets provided by citizen scientists, while citizens and communities gain access to scientific expertise, validation of their findings, and a deeper understanding of the scientific process (see e.g., Hassani et al.⁸⁶, Hassani et al.¹⁴¹, Hassani et al.¹⁶⁶, and Connolly et al.¹⁷⁴). Citizen science projects using LCS can establish connections between communities and various institutions, including governmental bodies, academic institutions, and environmental organizations.

Tracking and reporting all citizen science projects, initiative stories, and community-based networks on air quality (Fig. 4) can be a challenging task due to the overwhelming number of projects in existence (uRADMonitor¹⁶⁴, AirQo¹⁷⁵, PRAISE-HK¹⁷⁶, CAIRSENSE¹⁷⁷, CanAirIO [https://canair.io/, accessed 07.04.2025] are just a few). However, some inventories and databases provide information and resources related to citizen science activities and initiatives (see Supplementary Note 2).

Fig. 4: Spatial distribution of fixed outdoor low-cost sensors used in three international citizen engagement-related initiatives: PurpleAir, Sensor.Community, and SmartCitizen.

Only sensors with at least 30 days of data availability (not necessarily continuous) are displayed on the map. PurpleAir was initially founded with a strong emphasis on citizen science and community engagement, allowing individuals to contribute air quality data through crowdsourcing. Over time, the network has expanded, and not all sensors are necessarily owned by individual citizens, some are operated by institutions, research projects, or local authorities. However, it remains one of the largest networks with significant community participation, and its public data availability continues to support local engagement in air quality monitoring.

Remaining challenges and future research directions

Enhancing operational efficiency of low-cost sensor networks/graphs

In the reviewed studies (Supplementary Table 3 to Supplementary Table 7), typically simple train–test splits are used to evaluate calibration models; however, near-real-time LCSN require careful choice of cross-validation to ensure performance under operational conditions. Random cross-validation (RCV) remains common in ML but is poorly suited for LCSN, which collects sequential, strongly auto-correlated data. To maintain accuracy when models are applied in real time, training and validation must respect temporal order and spatial deployment (Supplementary Fig. 1). Practitioners should therefore combine spatial cross-validation (SCV), such as leaving out stations or areas without co-location, with temporal cross-validation (TCV) to test model transfer over time. Forward cross-validation (FCV), which progressively expands the test set, is especially effective for detecting abrupt performance drops due to drift. Together, using SCV, TCV, and FCV helps ensure that calibration models for LCSN remain reliable when new data streams in, including for LCS never directly co-located with a reference instrument.

Across the core themes of field calibration, imputation, concept drift, and graph-based modeling, nearly every major class of ML algorithm, from simple linear regression to advanced spatio-temporal multi-attention graph networks, has now been tested. However, comparing LCS studies remains challenging due to the lack of standardized performance metrics. Studies often report results using varied indicators, from correlation coefficients and error statistics to custom measures, making it difficult to compare models directly or evaluate compliance with regulatory targets. To address this, the adoption of a unified metric such as relative expanded uncertainty could help standardize performance reporting by expressing total measurement uncertainty as a percentage of the true value, aligning with air quality guidelines. Normalizing uncertainty to pollutant concentrations also makes results easier to interpret across contexts. More broadly, combining several metrics and monitoring performance continuously is essential to detect unexpected drifts and ensure reliable sensor data over time. Ideally, a combination of (1) linearity, (2) error, (3) uncertainty, (4) bias, and (5) distribution-based metrics should be considered for a thorough understanding of the calibration behavior.

Although several field calibration approaches exist (e.g., baseline, global), their application depends on the availability of reference stations. The description of a LCSN or, more generally, of a monitoring network as a graph is gaining attraction due to the versatility of graphs to describe relationships between nodes. Based on the reviewed studies in this section, treatment of LCSN as a graph can include spatiotemporal relationships and increase modeling accuracy (e.g., during imputation, outlier elimination, calibration, forecasting). Through adopting the graph approach, researchers can investigate the influence of several parameters represented as connectivity (e.g., edges derived from wind fields, traffic congestion data, dispersion models, land use, altitude, etc.) and make use of the graph computation tools (PyTorch Geometric, networkX, etc.) that are readily available. Therefore, a more suitable term for LCSN that includes connectivity should be LCS graphs.

At least two major operational challenges, missing data imputation and concept drift, should still be addressed for a successful LCSN field calibration design. LCSN data often have missing values due to (1) abrupt malfunctions, (2) power shortages, (3) maintenance, (4) data transmission issues, and (5) removed outliers (see Supplementary Fig. 2)¹⁷⁸. In parallel, concept drift¹⁷⁹, where sensor behaviors shift over time, is a risk to calibration accuracy and transferability (Fig. 5). These issues can be addressed by incorporating an incremental data treatment into the air quality datasets to increase trust in predictions. Many studies implement gap-filling methods to maintain continuous time series, ranging from simple linear interpolation and moving averages to more sophisticated statistical imputation or even ML-based interpolation using neighboring sensors’ data. Such preprocessing steps, though sometimes under-reported, critically influence model reliability. A well-designed preprocessing pipeline ensures that the calibration model learns from representative, reliable data and that real-world deployment of the model is not skewed by transient anomalies. In large networks, automated quality control and data cleaning become essential, as manually vetting data from hundreds of sensors is impractical.

Fig. 5: Two examples of the occurring concept drift.

On the top panel, the daily time series of the PM_2.5 concentrations of a reference instrument is shown, while on the lower panel, the daily time series of the NO₂ concentrations is presented. The red verticals indicate a change in the concept of the series as identified by the adaptive windowing (ADWIN) algorithm²²⁴. The statistical properties of each segment are presented above the time series to demonstrate the effect of the concept drifts.

Therefore, the minimal requirements to successfully operate a static LCSN is to (1) keep one sensor collocated with every reference instrument continuously, (2) calibrate the LCSN as accurately as possible, (3) monitor the performance of calibration models, (4) adapt the calibration models as needed based on triggers, (5) perform a combination of SCV, TCV, and FCV for robust evaluation, and (6) address missing data and concept drift. Finally, the description of LCSN via graphs is an evolving trend in air quality and in LCSN calibration specifically. Seeking new and better connectivity schemes for the description of irregularly distributed monitoring networks, as well as combining meteorological, traffic, and air quality graphs to inform one another, can be two novel research directions.

Challenges with ground-based monitoring using mobile low-cost sensors

A key research question relates to how sensor siting, orientation, and vehicle velocity can impact mobile LCS measurements. So far, there have only been a handful of studies that have systematically addressed this concern. Mui et al.¹⁸⁰ for example, piloted such an evaluation technique for PM_2.5 and gas LCS in California. They found that the PM_2.5 LCS overestimated PM_2.5 concentrations under high wind-speed conditions, suggesting that a flow-directing device was likely necessary to use such sensors reliably for mobile monitoring. Their work also showed that PM loading on the LCS gas inlet resulted in significant noise. Hassani et al.⁸⁶ conducted a study in an urban environment in Norway, using Sinfferbike sensors equipped with Sensirion SPS30 to measure PM_2.5 levels. They found that for every 1 km h^-1 increase in the speed of the bicycles, the standard deviation of the PM_2.5 measurements increased by ~0.03 to 0.04 μg m⁻³ (Fig. 6). This suggests that higher cycling speeds can introduce variability in the air quality data collected by these mobile sensors.

Fig. 6

Change in the Standard Deviation of the factory-calibrated PM_2.5 measurements of 11 Snifferbikes (Sensirion SPS30) compared to the bike speed during the on-bike tests conducted on May 31, 2021, and June 10, 2021, in Kristiansand, Norway (after Hassani et al.⁸⁶).

deSouza et al.¹⁸¹ found that the orientation of PM_2.5 LCS for mobile monitoring impacted device performance (deviation of PM_2.5 measurements from those of a higher-quality instrument) in an experiment in Boston; however, device age and choice of calibration approach also proved to be extremely important factors in determining the performance of LCS in a mobile setting. Russell et al.¹⁸² found that the rate of change of pollutant concentrations in different environments, and not vehicle speed, impacted the performance of low-cost PM_2.5 and NO₂ sensors, relative to reference instrumentation, in a mobile sensor deployment in London. Russell et al.¹⁸² also reported that the time resolution of the electrochemical NO₂ sensor they used in their experiment could not capture gradients of NO₂ concentrations in highly variable environments. We note, however, that some research groups have begun to develop strategies to mitigate the slow dynamics of LCS for gaseous species for mobile sensing¹⁸³, although more research is needed.

To obtain reliable patterns of air pollution in a study area, mobile monitoring routes need to be selected such that specific routes are sampled multiple times. Messier et al.¹⁸⁴ found that in San Francisco, for example, 10-20 mobile monitoring runs/road segment were sufficient to obtain robust long-term spatial patterns of traffic-related pollutants in their study area. Hassani et al.⁸⁶, in their Snifferbike study, showed that at least 27 PM_2.5 measurements per 50 m road segment are required to have reliable averages. Opportunistic sampling with routine vehicles may or may not be able to achieve such a desired repeatability. Studies have used simulations to evaluate how many of these vehicles are needed to achieve this sampling strategy¹⁸⁵.

Pedestrian-based LCS deployments have key limitations. Pedestrians often transition rapidly between diverse environments, e.g., between indoor and outdoor spaces¹¹, where environmental conditions such as temperature and humidity can vary significantly. LCS performance is sensitive to these changes, potentially affecting measurement accuracy unless sensors are deployed for monitoring in a specific and controlled environment (e.g., only within a park or defined area). Other concerns about LCS for mobile monitoring are similar to those of stationary sensors: field calibration, drift, and degradation¹⁸¹. Although a wide variety of field calibration models have been used to calibrate LCS for mobile monitoring, research has shown that considering the transferability of calibration approaches from the stationary to the mobile setting is important, specifically because the micro-environments experienced by the sensors during sampling may not represent stationary calibration conditions⁸⁶. As a rule of thumb, deSouza et al.¹⁸¹ found that for short calibration periods, simpler field calibration models were more transferable to the mobile setting. They found that short calibration periods relative to the length of the mobile campaign do not transfer well to the mobile setting because of seasonal changes in aerosol sources. Moreover, deSouza et al.¹⁸¹ also determined that using minute-level co-located data, instead of hourly-averaged data, led to a better-performing calibration model. Other researchers have also successfully adopted on-the-fly or drive-by calibration models for mobile LCS when the sensors drive within a given distance from a fixed reference monitor (for example, Hasenfratz et al.¹⁸⁶).

Challenges associated with low-cost sensor monitoring using drones

Earlier overviews^187,188 summarized the emergence of drone-based chemical monitoring as a complement to stationary networks and manned aircraft. They highlighted applications ranging from volcanic plume sampling to methane leak detection^189,190. However, early studies mostly demonstrated technical feasibility under controlled or low-complexity conditions and did not address the operational, calibration, or regulatory barriers to large-scale, routine drone-based monitoring with LCS, which we cover below.

Influence of environmental variables on the performance of LCS: PM LCS can overestimate concentrations under high humidity. Studies by Savvakis et al.¹⁸⁹ and Bretschneider et al.¹⁰⁷ showed that silica gel-based diffusion dryers placed upstream of the LCS can reduce relative humidity from ambient levels down to about 40% for durations that match typical drone endurance (20–30 min). Masic et al.¹⁹¹ demonstrated that these simple dryers are effective for short flights but may saturate over time and require replacement.

Rotor downwash effects on measurement stability: Hedworth et al.¹⁰⁸ quantified the impact of rotor-induced flow on PM_2.5 readings using wind tunnel and field experiments, showing fluctuation in particle number concentration when sensors are mounted directly beneath propellers. Wu et al.¹⁰⁴ and Liang and Shen¹⁹⁰ recommend extending sampling inlets 50–60 cm vertically above the drone to reduce downwash turbulence. Moormann et al.¹⁹², proposed an alternative approach that the LCS could be fixed 100 cm above the drone frame using two elongated fixed connecting rods, further stabilizing measurements. However, the optimal inlet length depends on rotor design, flight pattern, and local conditions. Another solution to mitigate this issue could be the miniaturization of drones. This includes the development of nano-drones, which may help mitigate environmental and noise concerns and also enable gas-sensing operations within closed spaces. However, as the size of drones decreases, conventional motors become less efficient, necessitating the development of alternative propulsion methods needed for vehicles weighing just a few grams¹⁸⁷.

Sampling flow stability issues in LCS on-board drone: Another consideration for using the PM LCS onboard drones is to improve the flow stability of the LCS. The LCS typically uses a small fan or blower to achieve nominal flow rates under ideal conditions with no upstream tubing or equipment. However, it is recommended that the initial flow configuration of the LCS be modified (by replacing the fan with a miniature pump) to accommodate pressure drops caused by upstream sampling lines (e.g., to facilitate the use of aerosol dryers) and to increase stability against abrupt changes in airspeed typical of drone operations¹⁹².

Limited endurance time of drones (short operating time): Copter drones may be limited to short flights (e.g., 15–20 min), which restrict vertical profiling depth and horizontal coverage¹⁸⁹. Some applications, such as monitoring large industrial sites, necessitate the use of drones that are capable of both prolonged endurance and agile maneuverability. This could be achieved through the deployment of drones equipped with high-power lithium-polymer batteries, fixed-wing platforms, or the use of multiple drones working together in a swarm configuration^110,113 (see e.g., Supplementary Fig. 3). Tethered drones can provide longer hover durations for repeated vertical profiling, although they have reduced flexibility for horizontal transects¹⁸⁷. These drones are well-suited for point-based observations for longer sampling durations, particularly when the field campaign is carefully planned, considering their limitations.

Optimal sampling altitude and mission planning: Different operational and research needs will dictate the optimal height. As an example, Pochwała et al.¹⁰⁹ used decision tree models to optimize drone flight altitude for different PM fractions, identifying that PM₁₀ is best measured around 30 m, and PM₁ and PM_2.5 are best measured between 40–50 m in their Polish case study. However, typically, drone research requires measurements at elevations much higher than 30 m.

Lack of in-flight inter-comparison experiments between co-located instruments and LCS on-board drones: Savvakis et al.¹⁸⁹ and Pochwała et al.¹⁰⁶ emphasize the value of in-flight intercomparison (Supplementary Figure 4): drones can hover near reference stations at 10–15 m above ground before sampling to validate sensor accuracy under operational conditions.

Obstacles and challenges of data fusion and assimilation of low-cost sensors

Beyond ensuring robust, high-quality LCS measurements¹⁹³ and transparent processing^79,194,195, accurate uncertainty estimates for both models and sensors are crucial, ideally assessed for each measurement individually. Realistic uncertainty estimates are essential for effective data fusion or assimilation, enabling optimal integration of sensor data and model predictions^123,125,196. However, detailed uncertainty assessment for large-scale LCSN can be prohibitively expensive and time-consuming¹⁹⁷, especially with diverse sensor types. For models, the main challenge is representing error covariance realistically, capturing spatial and temporal propagation of observational signals, rather than overall performance metrics.

Offline data fusion methods, such as geostatistical approaches^124,143, are widely used but only post-process model results, limiting real-time bias correction and forecasting improvements. Standard geostatistical tools often struggle with complex error covariance structures like those related to urban or road-network features¹⁹⁸. Simpler assimilation methods like Optimal Interpolation offer flexibility, but constructing the necessary background error covariance matrix remains challenging^125,143, especially in complex high-resolution urban contexts with dense LCSN. Ensemble methods such as EnKF effectively build this unsupervised matrix but require substantial computational resources, making their use difficult for real-time operational use¹⁹⁹. Variational data assimilation methods present additional challenges, notably in developing and maintaining accurate adjoint models^144,200 required for calculating gradients. Errors in adjoint models can impact assimilation reliability, emphasizing the need for careful validation and expertise.

Overcoming these challenges and moving forward, continued work on innovative algorithms, advancement in algorithmic efficiency, and computational capabilities will be necessary for improved real-time operational integration of LCS data with air quality models. Work on automated and ideally scalable methods for uncertainty quantification and validation should be prioritized. This will improve the reliability of sensor networks and assimilation techniques. New opportunities could include exploring ML techniques for characterizing covariance structures with a potential improvement in assimilation performance. Research on and implementation of standardized LCS data protocols and open-source LCS platforms could further enhance the reproducibility, accessibility, and wider adoption of data fusion and assimilation techniques for better exploiting LCS observations.

Challenges in integration of low-cost sensors with satellite observation data

A comparison of various LCS-satellite integration strategies within ML frameworks is necessary to identify robust/scalable modeling approaches. Standardized metrics such as relative error or normalized RMSE are essential to help inter-study comparisons. A key research direction involves assessing the spatial generalizability of models. Most current studies develop localized models tailored to specific cities or regions. However, building generalized models applicable over larger geospatial domains would enhance their practicality for air pollution monitoring and policymaking.

Although LCS offer high temporal resolution, this potential often remains underutilized due to data discontinuities caused by sensor drift, aging, calibration, or maintenance issues. Additionally, coarse temporal resolution of low-earth polar-orbiting satellites (1–2 observations per day) limits their synergy with LCS. The advent of geostationary satellite (e.g., from Geostationary Environment Monitoring Spectrometer (GEMS), Tropospheric Emissions: Monitoring of Pollution (TEMPO) or the recently launched Sentinel-4), which offer hourly observations of NO₂, O_3, AOD, formaldehyde, SO₂, presents a promising solution to enhance temporal continuity in air quality monitoring, though their integration remains largely unexplored and calls for further research.

A persistent challenge is the spatial mismatch between satellite pixel-based data (several km resolution) and point-based LCS readings. Multiple LCS units can fall within a single satellite pixel, creating a spatial resolution mismatch, as it requires reconciling detailed local measurements with broader, average satellite observations. One solution is to integrate or assimilate the LCS and satellite observations into a numerical model where their unique properties and representativity can be accounted for mathematically through detailed specification of their uncertainty and error propagation.

Another potential research direction is the use of satellite remote sensing products for LCS quality control. For example, products such as NO₂ and O₃ from TROPOMI, AOD from MODIS, and even land surface temperature or vegetation indices (e.g., NDVI) can provide contextual information to flag outliers or validate trends, especially in areas with limited ground-truth reference stations. For instance, if multiple LCS units in a region report unusually high NO₂ values, satellite data can be used to verify whether a regional pollution event is occurring or if sensor malfunction is likely. However, the utility of such comparisons depends on aligning spatial and temporal resolutions; as higher-resolution satellites continue to be launched (e.g., TEMPO), the integration of satellite data into LCS quality assurance workflows will become feasible. Future work could focus on developing frameworks or automated tools for real-time or retrospective LCS validation using satellite inputs.

To make impact, citizen science projects focused on air quality often encounter various barriers, also called obstacles, and bottlenecks²⁰¹. These obstacles include motivational, technological, economic, competition, individual/societal, legal, and regulatory aspects. These obstacles are not peculiar to air-quality citizen science projects, and other community engagements for environmental monitoring also suffer from them.

The “local needs and motivations to monitor” barrier in citizen science projects highlights the importance of clearly defining the research question or the purpose behind gathering data and engaging citizens^177,201. Stakeholders ranging from citizens to scientists and data collectors, along with decision-makers and policymakers at diverse levels, need a thorough comprehension of what motivates, incentivizes, and hinders their engagement²⁰². In some cases, data collection may occur without a specific focus or a clear objective, leading to limited alignment with local needs and engagement from policymakers.

Technological barriers can manifest in the form of limited access to or unfamiliarity with required technologies¹⁶³. For instance, citizens may lack the necessary equipment or software to participate effectively in data collection and analysis²⁰³. This can be particularly challenging for individuals with limited technical expertise or resources. Additionally, establishing a common language between citizens and the project team, which includes developers and data scientists, can be challenging^23,202. This arises due to the differing backgrounds, expertise, and technical knowledge of the involved parties²⁰⁴.

Lack of trust in sensor performance, limited metadata, and absence of quality standards raise concerns about accuracy, reliability, and bias, hindering policy acceptance and limiting integration of citizen science data with official air quality networks and other sources^{18,166,174,201,205}. In some citizen science projects, commercial sensor systems process LCS data before providing it to analysts. Although convenient, this limits long-term quality control and calibration, as manufacturers may alter proprietary processing algorithms without disclosure, hindering full understanding and control.

Economic barriers arise when citizen science projects face challenges related to funding and resource availability². High costs associated with purchasing or maintaining equipment and technologies can restrict participation. Limited financial resources can hinder the scalability and sustainability of initiatives, preventing them from reaching a broader audience²⁰⁶.

Competition barriers arise when multiple citizen science projects with similar themes or geographic focus compete to attract participants, secure resources, and capture attention²⁰⁷. For example, if several initiatives are seeking to monitor air quality within a specific city, citizens may be overwhelmed with options, leading to reduced engagement and fewer data contributions. Additionally, despite the vast amount of data being generated, the siloed operations of stakeholders impede its effective utilization²⁰⁸.

Individual and societal barriers include factors such as limited awareness, motivation, and time constraints^209,210. Lack of awareness about citizen science opportunities may discourage potential participants from engaging in projects²⁰⁷. Additionally, time constraints and competing priorities can limit the level of involvement from citizens. For example, neighborhoods characterized by a high turnover of residents, such as areas with many short-term rental houses or flats (typically for durations of 1–3 years), may show decreased interest in policy co-creation and air quality monitoring initiatives.

Socio-economic disparities may also influence participation rates and the representation of diverse communities, creating challenges in achieving broad inclusivity²¹¹. Neighborhoods with a high concentration of immigrants and refugees may also exhibit lower levels of interest in air quality citizen engagement projects²¹². Citizen science projects are predominantly established in Western, educated, industrialized, rich, and democratic (WEIRD) societies, often overlooking Indigenous and non-Western perspectives²¹³. These initiatives tend to attract participants with formal education, economic stability, and access to technology and infrastructure, typically in politically stable regions^212,214. As a result, marginalized and Indigenous communities, both in the Global South and the Global North, are often excluded or unaware of such opportunities, limiting their engagement in citizen science efforts²¹². We explicitly note that some of the insights presented here are informed by our direct involvement in several citizen engagement initiatives, particularly in the context of air quality monitoring. We acknowledge that a more rigorous, quantitative documentation of such participatory efforts remains an important direction for future research.

Similarly, in a neighborhood characterized by socio-economic disparities, residents from lower-income backgrounds may face financial constraints that limit their ability to actively participate in citizen science initiatives²¹⁵. deSouza and Kinney²¹⁶ analyzed deployment patterns of a commonly used LCS (PurpleAir) based on demographics and pollutant concentrations. They found census tracts with LCS were wealthier, more predominantly white, and better educated than the US average. The costs associated with acquiring sensor equipment, attending meetings or workshops, or even accessing reliable internet connections for data submission may pose barriers for individuals with limited financial resources.

Environmental priorities can also influence participation rates and the representation of diverse communities in citizen science projects²⁰⁷. In Nordic countries, for example, the priorities regarding environmental issues may differ from regions where air quality is considered a critical concern. The Nordic countries are known for their emphasis on sustainability, clean energy, and environmental stewardship. However, air quality concerns may not be at the forefront of environmental consciousness (see e.g., Sääksjärvi²¹⁷) due to factors such as strict environmental regulations, relatively low levels of air pollution compared to other regions, and a strong focus on other environmental issues such as climate change and biodiversity.

Lastly, legal and regulatory barriers add complexities to citizen science projects. Issues related to privacy, confidentiality, fairness, transparency, data ownership, and ethical considerations can impede progress^205,218,219. Compliance with relevant laws and regulations, obtaining permits or approvals, and ensuring data protection can be time-consuming and challenging. For example, a project involving the use of drones for vertical distribution of pollutants monitoring may face restrictions due to aviation regulations or privacy concerns related to aerial imagery. Furthermore, the collection of personal data represents a common concern and barrier to citizen engagement. Fairness in data collection and analysis is another legal and regulatory consideration. It is important to ensure that the data collected from citizen science projects is representative and unbiased, without favoring specific demographics or locations²²⁰. This requires implementing appropriate sampling strategies, considering potential biases in participant recruitment, and applying rigorous data validation and quality control measures. Transparency in data management and reporting is essential for establishing trust and credibility in citizen science projects^218,221,222. Moreover, making project results and findings openly accessible to the public promotes transparency and allows for scrutiny and further analysis. As citizen science projects increasingly use advanced technologies like ML, new ethical considerations emerge that need to be addressed. It is important to address concerns such as algorithmic bias, transparency of decision-making processes, and potential unintended consequences of relying solely on automated systems. Training ML algorithms on diverse and representative datasets and implementing interpretability and explainability techniques are essential to maintain fairness, accountability, and transparency.