This preliminary study, conducted in Italy, aims to investigate the potential of growth functions and multi-layer perceptron neural networks (MLP NN) in reducing the impact of humidity on low-cost particulate matter (PM) sensors, with a focus on the sustainability of low-cost sensors compared to reference stations. All over the world, low-cost sensors are increasingly being used for air quality monitoring due to their cost-efectiveness and portability. However, low-cost sensors are susceptible to high humidity, which can lead to inaccurate measurements due to their hygroscopic property. This issue is particularly relevant in Italy, where many cities such as Rome, Milan, Naples, and Turin experience high mean relative humidity levels (>70%) for most months of the year. To improve data quality and gain useful data for quantitative analysis, techniques must be developed to reduce the impact of humidity on the final data. The sensors used in this study were placed in proximity to a reference station, solely for validation purposes in the case of corrective functions and involved in the training phase in the ∗Corresponding author.
The issue of air pollution is becoming increasingly
prevalent due to various factors such as urbanization,
industrial activities, and transportation [
These sensors have the potential to make air
quality monitoring more accessible and widespread,
particsources [3]. The afordability and portability of these
ence projects and community-led monitoring eforts.
Adsensors have opened up new possibilities for citizen sci- cities.
enabling prompt actions to be taken to address air
pollution hotspots. However, the drawback of LCSs is their
tion dificult due to the fact that these sensors are very
sensitive to environmental factors, compared to reference
stations [4], [
nEvelop-O (L. Po) corrective function as input to the MLP NN. Finally, in Section 6, evaluation metrics are presented. This section outlines the various metrics that were used to evaluate how they can be used to assess the accuracy of low-cost ularly in developing countries or areas with limited re- sustainability of low-cost sensors in comparison to refditionally, low-cost sensors can provide real-time data, ing a diferent aspect of the study. In Section 2, laser restricted technology, which makes quantitative evalua- tions as a method for mitigating the efects of humidity threshold, the water in the air can be detected by the sen- ing humidity efects is discussed. Section 5 presents the
Sensor scattering is a technique (shown in Figure 1) commonly utilized to measure the concentration of PM in the air. This technique involves the use of lasers to scatter light of particles in the air, which are then detected by a sensor. The amount of scattered light is proportional to the number of particles in the air, allowing for an estimation of the PM concentration. However, this technique is limited by the sensitivity and accuracy of the sensor, particularly in the case of low-cost sensors.
Low-cost sensors lacking a drying function or installed in humid environments, such as coastal areas, are more susceptible to humidity interference. Particles with hygroscopic properties absorb water from the air [7], resulting in larger particle sizes and an increase in light scattering within the sensor [8]. This leads to overestimated PM concentration levels (Figure 2) for high levels of RH.
It is important to note that pollution and humidity are not directly correlated and that water vapor is not harmful to human health. Therefore, to accurately measure pollution levels, it is essential to clean the data from this artifact. The EU air quality standards [9] and the WHO air quality guidelines [10] and other governmental organizations measure pollution impact based on the dry PM concentration.
Calibrating a low-cost sensor using data from a reference station and humidity levels is possible if a reference station is available. However, it may not work well for PM sensors because the problem is highly localized, with diferent types of pollutants having diferent hygroscopic properties based on surrounding environmental emissions [11], [12], [13]. Placing the sensor near a reference station, even if it is far from the final location of the LCS, is another possibility for calibration, but it may also result in poor data quality. Therefore, the proposed methods aim to clean data in a location-agnostic manner.
In addition to the previously proposed methods that utilize reference stations [14], two potential techniques were proposed for cleaning low-cost sensor data in a location-agnostic manner. The first technique involves using a corrective function to reduce the correlation between PM concentration and humidity [15]. By reducing the correlation, the concentration level detected can be reduced by a particular factor. The second technique involves training a multi-layer perceptron neural network using data from a reference station, taking into account not only the relative humidity but also other atmospheric variables, such as pollutants and meteorological factors. This could enable the NN to learn the relationship between humidity, pollutants, and the growth of PM concentration. Although the network is trained using data from a reference station, the learned function is generic and not location-specific. To make this technique work, the network needs to be exposed to multiple contexts to be used in locations far from the training site.
High relative humidity causes low-cost air particulate matter sensors to measure higher values than professional sensors, due to the condensation efect and hygroscopic properties of the particles. To correct this efect, a growth function that estimates the increase in PM values due to humidity can be applied [15]. In the context of air quality monitoring, a growth function is used to estimate the increase in particulate matter concentration due to the absorption of humidity by the particles. Specifically, the growth function is used to estimate the amount by which the concentration of wet particulate matter exceeds the concentration of dry particulate matter at a given relative humidity level.
The growth function, given the level of the RH, returns the coeficient to use in the correction. Dividing the PM concentration detected by the low-cost sensor, which we define as is returned the PM concentration without the humidity contribution, defined as . See equation 1.
= / ( ) (1)
threshold approach, depending on the sensor and environmental factors. Applying the growth function across = 1 + ⋅ ℎ 2 (2) the entire dataset may result in coeficients being applied (1 − ℎ) to low RH levels, which can negatively impact already The growth functions can be customized to fit meacorrected data. Therefore, a threshold approach is pre- sured data by selecting appropriate values for the paramferred, whereby the growth function is only applied to eters and . A key feature of all growth functions is RH values above the chosen threshold. The threshold that they have a value of 1 when the relative humidity is level can be determined based on the specific sensor be- 0 and a significantly larger value as the relative humidity ing used or by taking into consideration environmental approaches 100%. knowledge. A commonly used threshold is around 70%.
Diferent corrective functions have been developed to address the problem of humidity afecting PM concen- 3.1. Parameters optimization tration measurements. Soneja et al. [16] research sug- One approach to choosing the growth function paramegests that there is an overestimation bias that becomes ters is to use reference station values as a ground truth. significant at around 75% relative humidity, while an un- Diferent and parameters are chosen within a certain derestimation bias exists at very low RH levels (below range and the parameters that best improve the PM de30%). To address this issue, they have proposed humidity tected compared to the data detected by the reference adjustment equations that cover the entire RH range. station are selected, as in [8] and [19]. However, this
In this paper [15], the author proposes a compre- approach links the parameters to the specific location of hensive approach to correct nephelometric 1 PM for the reference station and is limited to the period studied. humidity-related bias. The paper starts with an overview This approach is only useful if the low-cost sensor is coof diferent sources that explain the principle behind the located and fixed near the reference station and never hygroscopic growth of particulates [17], [18]. moved. In this way, it is possible to update the param
The author proposes a new approach, named “combo”, eters over time and use the reference station as ground described in 2, to address humidity-related bias in PM truth, [20]. However, this approach involves the optimeasurements. This approach wants to combine some mization of the growth function parameters concerning of the existing methods. the reference station values, which may not be feasible in all situations. 1Nephelometry is a method used to measure the concentration of To address this limitation, an alternative approach is suspended particles in a liquid or gas. to select growth function parameters that result in the lowest correlation between corrected PM concentrations In training the network, data from one or more sensors and RH [15]. This method does not require optimization located near a reference station can be used, including of the growth function parameters concerning reference PM concentration, meteorological and atmospheric varistation values and may be more practical in certain situ- ables, and the additional variables as inputs, with the ations where a reference station is not available or the reference station as the output. It’s critical to ensure low-cost sensor is not co-located with a reference station. that the additional variables used are also available in
In this approach, it is necessary to have data on the PM other locations where low-cost sensors are placed. The concentration as well as the relative humidity detected in elemental analysis is the most crucial additional variable, the same location and at the same time. However, since describing the concentration of each element present. To the accuracy of the sensors used is lower than that of improve the network’s ability to generalize and correct a reference station, it is always better to preprocess the PM concentrations accurately, the network should learn initial data and remove anomalies using typical anomaly the correlation between RH and PM growth in diferent removal methods. It is also important to note that this pollutant contexts, as the hygroscopic properties are deapproach works on the assumption that relative humidity pendent on the specific pollutants present at the time of and PM concentration are not strongly correlated, as detection. shown by the lack of correlation observed in reference It is common in the literature to find studies that aim station data. to calibrate low-cost PM sensors against reference stations. In addition to using an MLP NN, there may be other suggestions in the literature [22]. Some possible 4. Neural network methods include regression analysis, decision tree models, and support vector machines. Each method has its advantages and limitations, and the choice of method depends on the specific application and data available.
We believe that MLP NN has the potential to correct PM concentration levels more efectively than other methods. To achieve this, a wide range of scenarios, including those that the sensors are likely to encounter, and diferent periods should be presented to the network during the training phase. This will improve the network’s ability to generalize and correct PM concentration levels accurately, even for sensors located far from the reference station.
multi-layer perceptron neural network to model and generalize the relationship between relative humidity and PM concentration growth. To achieve this, the neural network is trained on a dataset containing input-output pairs of RH, PM concentrations, and other relevant variables like meteorological and atmospheric variables. Once trained, the neural network can be used to correct PM concentrations. This method has the advantage of being able to capture complex, nonlinear relationships between RH and PM concentrations and has the potential to generalize well to diferent locations and periods. However, it requires a significant amount of high-quality training data and careful tuning of the network architecture. 5. Cooperative techniques
To efectively generalize the relationship between RH and PM hygroscopic growth using a neural network, it is One alternative and potential way to improve the accunecessary to feed the network with additional variables racy of PM concentration measurement is to combine beyond just RH and PM concentration data. Hygroscopic diferent methods. Preprocessing techniques can be apgrowth is a process that occurs as water vapor accumu- plied to eliminate obvious anomalies in the raw data lates on the surface of aerosol particles with increasing before applying the growth function. The resulting correlative humidity. The extent to which this process oc- rected data can then be used as input to train a neural curs depends on the chemical composition of the parti- network, which can further improve the accuracy of the cles, which can vary widely in time and space [21]. PM concentration measurements. This approach has the
Therefore, to improve the neural network’s ability to potential to benefit from the strengths of each method, accurately correct PM concentrations and generalize its leading to more accurate and reliable results. However, predictions to new contexts, additional variables that cap- the success of this approach depends on the quality and ture information about the chemical composition of the consistency of the data, as well as the efectiveness of the particles should be included as inputs to the network. preprocessing.
However, these variables are often not available at the sensor location and must be obtained from online resources like Copernicus 2. 6. Evaluation metrics
program that provides free and open access data (www.copernicus.eu)
In conclusion, it is expected that correction functions based on hygroscopic growth factor can be applied in locations where reference stations are not available. However, a neural network trained in various contexts may still be superior. Therefore, the winning approach may eventually come from a combination of these two techniques. Future work will focus on exploring the potential of these methods in mitigating the efects of humidity on low-cost PM sensors and improving the accuracy of their measurements. data from these devices are often made public and used for monitoring pollutant levels. To address this issue, the U.S. EPA has proposed guidelines [23], containing metrics of Table 1, for evaluating the data quality of low-cost sensors, which can be useful for non-regulatory purposes such as identifying local air quality trends and hotspots, promoting environmental awareness, and providing supplemental monitoring.
The goal of the report is to establish a consistent set of testing protocols, metrics, and target values for evaluating the performance of PM2.5 air sensors in outdoor, ifxed-site environments specifically for non-regulatory supplemental and informational monitoring (NSIM). The metrics proposed can be used to assess the performance of these methods and guide future developments in this ifeld.