## 1. Introduction

Air pollution is currently a global fret, which can be linked to the extensive population growth and urbanisation, together with their aftereffect in traffic, industrialisation and energy consumption [

1]. Human health is closely linked to the air we breathe [

2], as evidence from recent studies for the adverse health effects has shown [

3,

4]. A recent report of the World Health Organization (WHO) suggests that 92% of the world’s population lives in places that exceed the recommended annual mean concentrations of Particulate Matter (

$P{M}_{2.5}$) [

5]. Because of the growing health effects of chronic exposure to ambient air pollution, policy makers and scientists are showing an increased interest in monitoring air pollution at a higher spatial resolution. Various recent studies from spatial epidemiology and public health have set out a specific interest in traffic based pollution [

6,

7,

8], particularly in Stuttgart, Germany [

9]. Generally, air pollution monitoring is done by environmental or governmental organisations using a network of fixed monitoring stations. Typical regulatory decisions are taken based on long duration temporal trends and statistics [

10], considering conditions related to hotspots estimated based on real-time data, if available. Interpreting the pathways from the generation of emission, dispersion and chemical transformation of pollutants in ambient air pollution concentrations is very challenging due to its high spatiotemporal variability [

11]. In the recent years, land use regression (LUR) has been widely used in various health and epidemiological studies to estimate air pollution at a finer spatial scale in the urban areas [

12,

13,

14]. However, due to economic reasons, the number of air quality monitoring stations in cities is usually sparse and limited, and this considerably limits an accurate assessment of the intraurban variability of air pollution.

Citizens and environmental agencies are exploring the potential of small, low-cost air quality monitoring sensors to enable detailed real-time information on air quality in the city [

15,

16,

17]. Several low-cost sensor deployments have been conducted in recent years extending from citizens investigating air quality in the houses and surrounding areas, to networks of sensors to measure community-level air quality, to a vast network of sensors covering the cities [

15,

18]. However, the datasets provided by low-cost sensor networks are arguably of less accurate [

19,

20,

21]. Despite such a limitation, the demand for sensor technology is high, driven by widespread concerns about the air pollution as well as an interest in reducing the personal exposure [

22]. While crowdsourcing approaches for air quality data gathering and related technologies are escalating, research to inform the translation of low-cost air quality sensor data into real applications remains limited. The term “low-cost” might be interpreted differently depending on the end users and the specific purpose of the study. For instance, U.S EPA Tier 3 instruments can be low cost (2000–3000 USD) for regulatory authorities but not for general citizens who are willing to participate in the data collection process [

23]. Therefore, in our study, we refer to low-cost sensors as devices which cost less than 200 Euros (and can thus be used by individuals or communities for air quality monitoring).

To capture the spatial variability in detail, accuracy of data will profoundly be relating to “where” the data are collected. To better understand exposure in microenvironments, it is crucial to take into account the spatial coverage of air pollution monitoring networks. Inappropriate location selection may lead to over- or underestimation of pollution originated from various emission sources in the city. When considering low-cost sensors to gather air quality data, previous studies suggest that, generally, the datasets generated with the help of citizen or community participation approaches inherit serious data gaps and the measurements collected are from irregularly spread sensors [

21]. Since the process of air pollution monitoring to capture spatial variability involves specific cost and time [

24], it is desirable to optimise the monitoring locations. Hence, to overcome the data gaps and irregular spatial spread of sensors to make data collection more efficient, there is a need for methods that can help in extending wide-spread and optimal location identification.

Participatory data approaches can be helpful to enable detailed air quality data collection, but exploiting the datasets generated from these low-cost devices requires tools and techniques for data cleaning and processing. The vast amount of dynamic, varied, detailed, and interrelated datasets from citizen participatory approaches could be enhanced by preparing the protocols and infrastructure that enables scientifically sound data collection [

25]. The systematic deployment of low-cost sensors for urban air quality monitoring can be useful for air quality data collection. With the potential of low-cost air quality monitoring sensors to increase the spatial coverage [

26], along with its application to foster participation [

22], it is desirable to systematically identify the optimal placement of monitoring stations to make the best use of advanced sensor technology and citizen engagement efforts.

The present study aimed to develop a method which can help in systematically identifying the optimal locations of citizen sensors for air pollution monitoring. The method was tested using citizen-contributed data from the city of Stuttgart, Germany as a scenario. The primary objective of the optimisation method included the identification of the most advantageous spread and optimal monitoring site locations that minimise mean prediction error for land use regression (LUR) estimations of air quality parameters for the study area. LUR is a method for spatial exposure assessment. It helps to model pollutant concentrations (e.g., particulate matter and nitrogen dioxide) at any location using various environmental characteristics of the surrounding area (e.g., traffic intensity, elevation, and land use type). The spatial simulated annealing (SSA) algorithm was used to run the objective function for finding the optimal monitoring network design.

The main contributions of this study can be summarised as follows:

We extended the optimisation method proposed by Gupta et al. [

27] by incorporating wide-spread distribution aspects (in addition to LUR’s predictor error aspects) into the placement of low-cost citizen sensors for air quality monitoring.

We demonstrated the applicability of the proposed optimisation method in two practical scenarios: starting a new volunteered geographic information (VGI)-based air quality monitoring campaign; and finding out where to place new sensors to extend the existing VGI-based air quality monitoring network from the city of Stuttgart, Germany.

While existing works on analysing the quality of VGI mostly aim at examining the degree to which a fact contributed by a volunteer is likely to be true (see, e.g., Goodchild and Li [

28]), this work approached the question of quality of VGI from a slightly different angle. By trying to find the spatial distribution of volunteers which can minimise the global prediction error of the air pollution monitoring network, this work intended to inform the coordination of VGI efforts for air pollution monitoring at the city level. As such, the method proposed can be classified as belonging to the fourth type of VGI validation process identified in [

29], namely “measure of fitness by way of completeness” (not the amount of points, but the promise of detail or spatial extent). There are a couple of methods in the literature to tackle the VGI quality aspects of positional accuracy, thematic accuracy, and topological consistency, but a general lack of methods addressing other aspects such completeness, temporal accuracy, or vagueness, as a recent review by Senaratne et al. [

30] reminded. Criteria such as road density or errors of omission/commission are used as surrogates for completeness in some studies [

31,

32], but completeness in this study was approximated using the combination of two criteria: spatial spread and minimal prediction error of the air pollution monitoring network.

The rest of the paper is organised as follows. In

Section 2, we present a brief overview of the previously done work on the topic.

Section 3 describes the study area and the data used in the study. A new methodology for optimisation is described in

Section 4. In

Section 5, we present the results and discuss the objective function used in the proposed optimisation method.

Section 6 presents the discussion regarding the developed optimisation method. We draw the conclusion in

Section 7.

## 4. Method

As mentioned in

Section 1, the present study sought to develop an optimisation method that can take into account fitness-of-purpose as objective for VGI data collection. As Clements et al. [

22] suggested, the identification of the research question before planning the deployment of the VGI based air quality sensor network can help in useful data collection. The optimisation method presented below takes the question “

what are the set of locations in the city where measurements are required for estimating air pollution with minimal LUR prediction error?” as the research question. It used Spatial Simulated Annealing (SSA) to run the objective function.

SSA is a random search algorithm that explicit deals with spatial vicinity. It is the spatial version of the probabilistic techniques simulated annealing, which was developed by van Groenigen [

67] for spatial soil sampling design optimisation. The SSA technique mimics the cooling of metal phenomenon to reach global optima, i.e., simulated annealing. In the starting phase of the annealing process, the locations for sensors can change greatly, with low probability even at not so optimal locations. As the process cools down with time, changes in locations become smaller, and acceptance of worse designs of monitoring network becomes less likely. During the optimisation process, the algorithm takes several hundred or thousands of iterations to identify the optimal configurations. The SSA algorithm is widely used in sampling design for mapping [

67,

68]. The SSA algorithm requires an objective function, whose output value acts as “energy” in the optimisation process. The optimal design identification is made based on the set of the location which represents the minimal energy of all iterations in SSA. Hence, the objective functions should be formalised as a single objective optimisation function, pointing at discrete-valued variables which calculate the energy value.

#### 4.1. Optimisation Objective Function

The optimisation was performed based on some rules and objectives that are used as a function. The optimisation objective function is usually composed of one or many constraints, which were calculated using the explanatory variables of a given LUR model in our case. The objective function was implemented using SSA, where it estimated the objective function value (also called the energy of annealing) to identify the set of locations which fulfil the given optimisation objectives. The objective function used in our study considered two aspects:

#### 4.1.1. Prediction Error Aspect

The first aspect of the objective function was adopted from the previous work done by Gupta et al. [

27] to identify the set of locations for crowdsourcing sensors which collectively can help in modelling

$PM$ concentration with the less spatial mean of prediction error for the study area. The prediction error is the measure of the accuracy of the model to predict the value of the variable of interest by using various independent variables. The prediction error aspect of the objective function considered the covariates of the selected LUR model that can help in estimating the

$PM$ concentration in the study area. The evaluation of the prediction error was done considering the matrix approach for the least squares estimate of linear regression:

where

${\widehat{y}}_{o}$ and

${y}_{o}$ represent the predicted and measured value of the dependent variable;

${x}_{o}$, …,

${x}_{n}$ represents the vector of the set of predictor variables values at the prediction site;

${x}_{o}^{\prime}$ represents the transpose vector; and

X and

${X}^{\prime}$ are the matrix and transpose matrix of the predictor variables for the randomly selected monitoring site.

${\sigma}^{2}$ is the residual variance of the regression, which is constant [

69] and independent of sampling locations. Hence, it can be left out of the objective function. The average predicted error of the LUR model is thus proportional to:

and that is what we use as objective function.

For a two-dimensional study area

A (represented by the number of its grid cells), the prediction error aspect is computed using

n observation sites for a network design

D.

D is the design of the set of monitoring locations identified at each iteration of the optimisation process. The optimisation process starts using a network configuration fed in as input or by randomly selecting monitoring design

${D}_{o}$, consisting of observation points

${s}_{o},\dots ,{s}_{n}$ with corresponding predictor variable vectors

${x}_{o}$, …,

${x}_{n}$. During the optimisation process, the monitoring sites are transformed into a random vector with only one element different from the initial one, yielding a new monitoring design

${D}_{1}$. The optimisation process computes the prediction error for each

${D}_{x}$ (

x ∈ [0, …,

n]) utilising each node of the rasterised study area

A, until the minimum value is achieved using

${x}_{o}$, …,

${x}_{n}$, which represents the set of predictor variables values at prediction location in

A, with

X and

${X}^{\prime}$ being the matrix and transpose matrix of the predictor variables for the randomly selected monitoring network design

${D}_{x}$ (

x ∈ [0, …,

n]) in the optimisation process. For the objective function, the manipulation of the set of locations leads to the modification of

X matrix values. For further details about the above-mentioned objective function, we suggest referring to the work done by Gupta et al. [

27].

#### 4.1.2. Widespread Distribution Aspect

Along with the requirement of the objective function to decrease the mean prediction error for the study area, the second aspect in the objective function focuses on enforcing the wide-spread distribution of sensor network in the study area. A wide-spread deployment is necessary because it can help in providing higher spatial resolution to air pollution data, which in turn better informs the identification of pollution sources and supports more conclusive studies on the effect of air pollution on the quality of life in cities [

70,

71]. A widespread deployment also helps in reducing the uncertainties associated with the modelled forecasting results.

Many low-cost sensors involved in the optimisation process along with uncertainties that can be caused by the spatial autocorrelation of the predictor variables used for optimisation may lead to clustered results. Furthermore, the selection of locations with extreme values of predictors in the optimisation process, while only considering the prediction error aspect, can also lead to clustered results. It is essential to have a constraint which can limit the clustering and enforce the selection of disparate locations. Hence, we extended the objective function developed by Gupta et al. [

27] to take into account the wide-spread distribution aspects of sensors in VGI-based monitoring network design.

To integrate the wide-spread aspect in the optimisation objective function, we calculated the inverse mean shortest distance (IMSD) for the set of locations selected in each iteration of annealing after calculating the mean prediction error value considering Equation (

3) in the optimisation process. The spread aspect of the objective function can be written as:

where

N is the number of points in the configuration considered for optimisation and

$mi{n}_{j-1}({D}_{ij})$ is second minimum distance between the

ith point and other points of configuration (as the minimum value will be 0 for each point distance to itself).

The algorithm for computing the IMSD (Equation (

4)) to enforce wide-spread distribution of sensor locations (as points) can be summarised as:

Input of a number of points (N) with a different spatial configuration as selected in each iteration of SSA.

Compute the distance matrix for all points.

Identify the second minimum value in each row of the matrix, as the distance matrix will contain the first minimum value as 0.

Compute the mean of the minimum values from each row and column of the distance matrix.

Compute the inverse of the mean value.

After the computation both aspect values using Equations (

3) and (

4), the values are then added to get a single objective function value which is further characterised as energy state in the SSA optimisation process. Furthermore, the optimisation function was made flexible to consider the weight function to prioritise one of the two aspects (prediction error or spread function) when identifying optimal locations during the optimisation process. The weights must be equal to or larger than 0 and sum to 1. The overall equation of the objective function that identifies the set of locations honouring both aspects of the objective function for the study area can be expressed as:

where

${W}_{1}$ and

${W}_{2}$ are the weights which can be assigned to each aspect the objective function based on the aim and fitness aspect of the VGI based air pollution monitoring initiatives. LUR prediction error and spatial spread are both critical for air pollution monitoring. The main idea behind the discussed objective function with the flexibility to consider weights is to give policymakers (e.g., coordinators of VGI initiatives) some control over prioritising their goal considering two crucial aspects of air pollution monitoring campaigns. Minimising the prediction error of the LUR implies confidence in the estimated values of air pollution at locations that were not observed. On the other side, maximising spatial spread leads, as mentioned above, to an air quality monitoring network potentially more informative as to the identification of various unidentified pollution sources in the city.

The overall steps for the optimisation algorithm can be summarised (also see

Figure 3) as follows:

A LUR model is selected/developed (using the air pollution ground data from low-cost sensors and predictor variables). If ground data are not available for LUR creation, already existing LUR models can be selected (arbitrarily or by considering models containing specific predictor variables which are significant for the study area).

Initial monitoring station locations are defined as the input, consisting of N observations, which can also be feed in as a whole number.

The study area A is discretised and the candidate locations are defined based on the resolution expected for the study area.

Random point selection in each iteration starts and calculates the objective function values using SSA.

The design of each previously selected configuration during the optimisation is modified until the network design is accepted based on the objective function value.

A design will be accepted if it reduces the prediction error as well as distribute the sensor in a wide-spread fashion, depending on the weight assigned to each objective as per Equation (

5).

The optimisation continues to iterate and find the set of optimal locations until the new energy value reaches the minimum and is not changing in further iterations based on the energy transition and other annealing parameters.

All geospatial and statistical operations for the study were carried out in the R statistical environment [

72], using packages sp [

73], sf [

74] and SpatialTools [

75]. For running SSA, we used the R package spsann [

76]. The source code of the optimisation method developed in this study can be accessed from GitHub [

77].

#### 4.2. Optimisation Process

The monitoring of air pollution is highly location-dependent. To tackle the challenges of acquiring spatially fine-grained air pollution data for cities using VGI based approach, it is crucial to pay considerable attention to

where the air pollution data must be collected by participants. We tested the performance of the proposed optimisation method for the city of Stuttgart where a large number of citizens are collecting air pollution data using low-cost sensors developed by OK labs [

59].

In our study, we tested the application of the developed optimisation method for following different practical scenarios:

#### 4.2.1. Optimisation for Starting a New VGI Campaign

Considering the advantages of new low-cost miniature sensor devices that are capable of monitoring air pollution, we first tested the application of the developed optimisation method for the aim of initiating a VGI campaign. Initiating a new campaign would mean that no crowdsourced air pollution data are available, which leads to either relying on the official monitoring station data for LUR development or starting the process from scratch. Since, for the study area of Stuttgart, only three monitoring stations (see

Figure 1) are measuring the air pollution data (which are not enough to develop the LUR model), we are of the opinion that it would be wise to start the procedure from scratch, assuming no air pollution data availability in the study area for developing a LUR model for the first test case. To initiate the process of identifying optimal locations for the deployment of new sensor network, we need to follow the steps as discussed in the previous section (

Section 4.1). As suggested, the optimisation method requires the input of explanatory variables of a given LUR model for identifying the optimal locations. We selected a LUR model of Austria from the ESCAPE project for

$P{M}_{2.5}$ [

78] as the underlying model for the

$P{M}_{2.5}$ concentration distribution for the study area. The selected model can be presented as:

The selection of the Austria LUR model was based on two underlying assumptions. First, the model utilises square-root of altitude (SQRALT) as one of the explanatory variables. Stuttgart is characterised by uneven altitudes and has a valley around it. We assumed that the SQRALT could help in explaining the dynamics of air pollution. The second factor is the availability of data. The building and altitude data were easily accessible; hence, we decided to use this model for testing the proposed optimisation method for the city of Stuttgart. It is also important to point out that we have only used the number (N = 116) and location of the existing crowdsourcing network’s configuration as the initial monitoring network for the this particular test case. However, it is not mandatory to provide a configuration; the optimisation method can also select random locations as the initial configuration for a certain number of sensors if given as input.

#### Which Locations Are Significant?

Another important factor while starting any air pollution monitoring campaign is to identify locations which are of great significance to the overall process of air pollution monitoring. We also utilised the proposed optimisation method to identify the locations which can be significant for initiating a low-cost sensor deployment in the study area given the selected LUR model.

#### 4.2.2. Optimisation While Placing New VGI Sensors to Extend an Existing Network?

The ideas from the previous sections are useful while planning a new VGI campaign (e.g., a two-day citizen science project to gather some values about pollutant concentrations in the city), and can help VGI coordinators decide where to best channel the available resources. This section considers another scenario, namely that of extending an existing VGI network by adding few new sensors using a systematic approach.

For this new scenario, we used the already existing VGI sensor data (i.e., the 116 sensors). Since the initiation of the optimisation process in this case also requires a LUR model, we developed a new LUR model using data gathered from the existing VGI sensor network (in contrast to what we did in the previous scenario, where we selected a LUR model arbitrarily). The advantage of developing a new LUR model is that it provides a more realistic explanation of the air pollution in the city than any arbitrarily selected model (as we did in the previous test).

In our study, we created a LUR model using the low-cost sensor data by following the steps suggested in the ESCAPE study [

78]. The model uses

$P{M}_{10}$ concentration as the dependent variable and the following explanatory variables: square root of altitude (SQRALT), buildings in 500 m buffer (BUILDINGS_500), industries in 300 m buffer (INDUSTRY_300), major roads length in 1000 m buffer (MAJORROADLENGTH_1000) and low density residential land in 1000 m buffer (LDRES_1000). The final model can be represented as:

Optimisation using the objective function from Equation (

5) can change the overall design of the sensor network compared to the already existing network’s design. However, it is not practically feasible to move existing monitoring sensors from their current location. Thus, we also investigated the applicability of the proposed optimisation method to identify a set of new locations relevant to the objectives of a VGI campaign (e.g., if the VGI campaign decides to extend the existing monitoring network with 20, 40, 60, 80 and 100 more sensors).

## 6. Discussion

This section reflects on the significance of the study as well as its limitations, and points at future work.

#### 6.1. Significance

The study demonstrates the application of the optimisation method which can aid in the systematic deployment of low-cost sensors for detailed air quality monitoring while considering the scientific models such as LUR. Low-cost sensors can provide data with very high spatial and temporal resolution, which is not feasible with conventional measurement approaches. The study provided means to combine low-sensor datasets to a scientifically recognised air pollution modelling approach to facilitate a better air pollution data collection in the city. The developed optimisation method builds upon ideas suggested by Gupta et al. [

27], to optimise air quality monitoring networks for VGI campaigns. The significant performance of the proposed optimisation method to decrease the mean prediction error by 52% along with wide-spread sensor network, demonstrate its applicability to enable systematic planning for (purposeful) VGI campaigns for air pollution monitoring.

The wide-spread VGI campaign sources can be useful for overcoming the issues connected to data quality, such as field duplication, data duplication and irregular spread of sensors, as pointed out by Clements et al. [

22] and Budde et al. [

37]. The optimisation method also helps answer the research question that need to be considered for planning deployment of sensor network (LUR in our case) to drive the data collection process. By defining the objectives before data collection, the method can be useful for reducing the cost of deployment by limiting the number of sensor nodes required. The method can also be beneficial to identify locations which are easily accessible for sensor maintenance and calibration, for example by using population-weighted optimisation [

27]. Such extensions can assist in decreasing sensor failure and replacement costs for successful long-term deployment. If the population weights are considered, the optimisation method can foster construction of LUR models with network design incorporating area close to population and roads, which can better characterise the full range of pollutant concentrations close to population [

79].

Since the currently available sophisticated monitoring stations are not capable of expressing the air pollution variability at a detailed spatial scale, the wide-spread and lower prediction error based low-cost monitoring network can be an alternative for gathering measurements, which can be detailed and informative. Using alternative data sources also helps in overcoming the sparsity and scarcity challenges existing in the literature. The resilience of the developed optimisation objective function to prioritise the wide-spread and prediction-error aspect could be advantageous for developing a systematic crowdsourcing sensor network whose measurements can be used in versatile air quality modelling approaches. The spatial spread aspect of the proposed optimisation method helps in shrinking the effects caused by spatial correlation in LUR residuals (which usually exist, see Beelen et al. [

80]). However, using weighted least squares (WLS) instead of ordinary least square (OLS) for Equation (

3) or considering the kriging prediction error based optimisation as suggested by van Groenigen [

67] would have presented an analogous effect on the spatial spread of optimal configuration as the spread aspect of the proposed optimisation objective function achieved for declustering points. The method also inherits the flexibility offered by LUR and SSA, making it more implementable even in cases where availability of data is limited. The outcome of the optimisation objective function considering the wide-spread distribution aspect can also help in distributing the points in different land use type, which can be constructive for developing robust LUR models as suggested by Wu et al. [

79].

The current state of sensor technologies with relatively large measurement uncertainties lead to concerns regarding engaging the citizens in the data collection process. Observing spikes while collecting data using VGI approaches may promote behavioural change which can help in preventing exposure to bad air quality. On the other side, this may also lead to panic situations, possibly negating any health benefit. Nevertheless, relating the spikes to the geographical variables such as road counts, traffic, and other emission sources by using LUR models may help to better inform citizens about their actions and local area contributions. Furthermore, the low-cost sensors data might not be monetised for proper air quality applications, but the LUR approach used in the study can act as a tool to process and visualise the data; the resulting analysis and the corresponding information generated can be easily monetised.

Overall, the optimisation method can help in defining the locations for systematic VGI campaign planning, which anticipates the wise use of the participation efforts along with reducing the error for air pollution modelling. The use of open and easily accessible data for VGI campaign systematic deployment, make this approach more implementable. Another major benefit of deploying VGI sensors is their ability to measure real-time data and provide immediate feedback that helps in improving the air pollution monitoring strategy systematically with the help of the proposed optimisation method. This also gives the opportunity to serve as a tool to help in building the capacity of participants to understand air pollution and the influence of geographical variables in the proximity, which can also explain air pollutant’s variability. The wide-spread distribution aspect in the proposed optimisation method could also help to identify potential sources of air pollution otherwise unknown to regulatory authorities.

#### 6.2. Limitation and Outlook

Along with the advantages, the proposed optimisation method also brings some challenges and limitations. One of the critical limitations for the application of the low-cost sensor data for air pollution monitoring is the reliability of the measured data. Further challenges include short working time and calibration challenge [

22]. In the study, the quality of the LUR model developed was low (

${R}^{2}$ = 0.1442) which may be due to the quality of data produced by the low-cost devices, and the locations from where they were collected [

79]. Developing new LUR models using inputs from improved VGI sensors could help better estimate the impact of sensor type on LUR model estimations.

In addition to these limitations related to the use of low-cost sensor data, there are limitations concerning the proposed optimisation method. To begin with, the selection of LUR model is the first step to find the optimal location, which means that, if we do not have a LUR model for the study area, we have to select one from the previous studies by specifying some assumption based rules for model selection. The selection of a LUR model based on some assumptions may not involve variables that are convincing enough to explain air pollution in the study area. Another limitation of the approach concerns the use of a LUR model and the underlying assumptions of multiple linear regression (e.g., linearity between dependent and independent variables, independent and normal distribution of error terms may create biases in interpreting the outcomes, which are the typical limitations for any simplistic regression-based studies). Limitations also exist for the SSA approach; as it is a stochastic method, every different run of optimisation method may yield different monitoring network designs. The process of optimisation is also very time-consuming, depending on the input parameters of annealing, variables used for computing the objective function and the study area size. While running the optimisation for the study, the process took 6–8 h for one optimisation outcome.

As can be seen from the results, the output of the optimisation ended up being clustered. This clustering can sometimes be caused by the spatial auto-correlation of the predictor variables, which lead to all points being close to each other. The reliability of the LUR used for the optimisation may also contribute to the clustered results. Devising the methods that address these limitations by taking into account robust LUR, and information on the spatial correlation and interpolation based constraints can be helpful in improving the design objectives of the study. We have not considered such factors in our study but future work could consider integrating it. It would also be interesting to investigate a combination of our method and active learning (see [

81]) for the purpose of optimal air quality network monitoring (e.g., our method helps to identify key locations during the monitoring process, and these could inform the labelling phase of an active learning approach). Extending the developed optimisation method to consider the population distribution weights proposed by Gupta et al. [

27] can also be useful in identifying the locations close to living spaces. A population-based weight can be useful in two ways. Firstly, identifying locations where the citizens live can make the initiation of VGI campaign easier. Secondly, it promotes the gathering of air pollution data that represent the real exposure of the population in the living spaces of the city. For the practical implementation of the proposed optimisation method for VGI approaches, future work can focus on integrating the optimal location identification method with citizen observatory based projects (e.g., FLAMENCO Project [

82]). Integration with citizen observatory based projects can be fruitful because the optimisation method can identify the locations and citizen observatory can identify the participants at the optimal location, making the overall flow of VGI-campaign initiation easy.

As discussed in previous studies related to low-cost sensors deployment [

22,

37], the field of low-cost sensors for environmental monitoring is in transition, and more work is needed to continue exploring the potential of low-cost sensors for air pollution monitoring. With the help of low-cost sensor systematic deployment initiatives by using citizen participation approaches, it is possible to bring forward a whole new system which anticipates the development of open data platforms (e.g., OK Labs [

60]). These initiatives also help in connecting other systems that utilise air quality data such as health informatics, housing companies, and sustainable urban planning, thereby helping in enabling the development of tools and techniques that can improve Quality of Life (QoL) in cities.