3.1.1. PM2.5 Data Output of the PA-II Device
One of the main issues that need to be accounted for is which of the data fields (CF = 1, CF = atm), provided by the PMS5003 and, consequently, the PA-II monitor, should be considered when performing PM
2.5 field calibration, since the two outputs deviate notably in certain concentration ranges [
25]. Examining the CF = 1 versus CF = atm scatterplots, a linear 1:1 relationship is observed for low concentrations. Then, a non-linear adjustment appears to be applied in mid-range CF = 1 values to derive CF = atm. Finally, another linear correction is applied for higher concentrations, with a coefficient that is smaller than one. Based on repeated observations on all datasets, the different PM
2.5(CF = 1) ranges (in μg m
−3) where the three different corrections are applied to obtain PM
2.5(CF = atm), are approximately the following:
Low Range: PM2.5(CF = 1) < 20, where PM2.5(CF = atm) = PM2.5(CF = 1).
Mid-Range: 20 < PM2.5(CF = 1) < 110, where an unknown correction is applied by the sensor manufacturer.
High Range: PM2.5(CF = 1) > 110 where PM2.5(CF = atm) ≈ 0.66 PM2.5(CF = 1).
To illustrate this, three scatterplots corresponding to the above concentration ranges, along with the linear regression equations for the first and third range are depicted in
Figure 2a–c. The winter-time measurements in Ioannina were utilized for the visualization, as the dataset is characterized by a wide concentration range (hourly concentrations from as low as a few μg m
−3 up to several hundreds of μg m
−3). Splitting the reference dataset into 35 equidistant concentration bins that cover the entire ambient concentration range, calculating each bin’s PM
2.5(CF = 1) and PM
2.5(CF = atm) average, and finally plotting those values against the reference PM
2.5 values, leads to the plot of
Figure 2d. It can be seen that for both data fields, while, at low concentrations binned averages follow a 2:1 line, they gradually deviate from it with increasing ambient concentrations.
A similar pattern has been reported for the PA-II monitor when comparing its outputs to a reference (TEOM) instrument [
48], while non-linear fits with a similar curvature, as in
Figure 2d, have been documented to better describe the sensor to reference relationship [
31]. In
Figure 2d, the pattern seems to be more pronounced for the CF = atm output, which, while falling on the 2:1 line for low concentrations, it tends towards the 1:1 line for higher ambient PM
2.5 levels. While this non-linear behavior of CF = atm should be addressed by any proposed calibration scheme, the use of the CF = 1 field may provide a more straightforward approach, also allowing for a more direct assessment of the effects that are related to aerosol physical properties, since there seems to be a more direct linkage between PM
2.5(CF = 1) and measured particle number concentrations. Thus, the CF = 1 field was utilized in the subsequent analysis. It is noted that the proprietary algorithm used for the correction of CF = 1 to CF = atm [
31], to the best of our knowledge, has not been yet documented in the literature, is not available in the sensor manual or any relevant documentation, while it is also not known to the manufacturer of the PA-II monitor either (personal communication with Purple Air LLC).
3.1.3. Coarse Particle and Relative Humidity Effects on Sensor Bias
During the intercomparison campaigns in Athens and Ioannina, the PA-II devices operated under a wide range of ambient conditions, regarding the meteorology, aerosol sources, and chemical composition. Furthermore, several dust episodes were observed, linked principally to air masses originating from northern Africa. Scatterplots of reference measurements conducted in Ioannina versus the PA-II PM
2.5(CF = 1) measurements can be seen in
Figure 3a,b, providing an indication of coarse particle effects. Respectively, data from the 2nd Athens intercomparison campaign (the Grimm 11D OPC that provided the particle fraction ratios was not available during the 3rd intercomparison campaign) are depicted in
Figure 3c,d.
The PMS5003 sensor has been documented to have poor size selectivity for coarse particles [
30], while the design of the flow path—from inlet to optical cavity—forces the sample flow along two consecutive 90
o angles, favoring the deposition of larger particles before reaching the optical detector [
31]. Therefore, the response is susceptible to error when aerosol is dominated by coarse particles, which in Southern European areas is typically observed during Saharan dust transport episodes [
50].
In the present case, when color-coding data-points in the scatterplots according to their corresponding PM
1/PM
2.5 and PM
2.5/PM
10 ratios (calculated from measurements of the multi-channel OPC instruments), it is evident (
Figure 3) that the data-points deviate from the general pattern for increasing coarse PM fractions (lower ratios). Points corresponding to lower ratios follow a steeper, almost 1:1 line, as compared to points corresponding to higher ratios, for which a clear overestimation by the PA-II is observed. Because these deviations become clearer when using PM
1/PM
2.5, this ratio will be henceforth used as a proxy in the related graphs.
On the other hand, important sensor biases can be expected in elevated ambient RH, since no conditioning of the sample takes place in the PMS5003 sensor. When evaluating the PA-II monitor, Magi et al. [
51] reported linearly increasing MAE and RMSE errors with increasing RH, while a positive relationship between RH and MBE was documented by Feenstra et al. [
21].
For the investigation of these effects, the absolute error of the PA-II PM
2.5(CF = 1) relative to the reference values in both datasets was first calculated. Given the fact that wintertime mean PM
2.5 concentrations in Ioannina were high (57.2 μg m
−3), the analysis was initially focused on the springtime measurements (mean 13.4 μg m
−3), in order to achieve comparability with the Athens dataset where the mean PM
2.5 concentration for the 2nd and 3rd campaigns combined was 15.2 μg m
−3. Plotting the absolute error against the PM
2.5 reference concentrations, steeper slopes in the relationships with ambient concentrations were observed for higher RH values (
Figure S5). This was more obvious in Ioannina, while, in Athens, this pattern was much less pronounced, probably due to the drier ambient conditions (RH: 49% in Athens versus 64% in Ioannina, on average, for the respective periods).
Data were divided in 20 equidistant bins according to RH, and the corresponding Mean Absolute Error (MAE) values was plotted, in order to examine the humidity effect. The results are shown in
Figure S6, where an evident dependence of MAE on RH can be observed. The effect is more evident in the Ioannina dataset, with MAE rising from 8.5 μg m
−3 to 35.7 μg m
−3 when RH increases from 30% to 90%. This behavior could be linked to the frequent lake-effect fog events during wintertime in Ioannina [
52], with the error being possibly related to water droplets that are introduced in the sample stream and counted in the PM
2.5 size range. A similar observation, with fog events inducing a positive bias in PM
2.5 measurements performed with the PMS1003 sensor—a previous version of PMS5003—has been also reported in Brisbane, Australia [
53]. A positive but nevertheless more modest relationship can be observed in the Athens dataset, with MAE rising from 4.3 μg m
−3 to 9.3 μg m
−3 along the 30% to 90% RH increment.
It should be noted that higher concentration levels in Ioannina were mostly observed during the evening, coinciding with elevated RH levels. This means that part of the bias could be due to an additive effect of high concentration and RH, given the positive relationship between absolute error and ambient PM
2.5 levels (
Figure S5). Other than that, the observations noted earlier, regarding the sensor behavior under elevated coarse mode concentrations (when the PA-II underestimates real PM
2.5 concentrations), implies that, in this case, the examination of the absolute error hinders the correct attribution of the actual measurement bias. Thus, it is necessary to isolate the influence of different effects and then examine their corresponding errors.
Because the RH and coarse particle effects tend to bias the performance in opposite directions, the bias error was examined as a metric of the sensor’s performance, for both Ioannina springtime and Athens datasets (
Figure S7). The effect of coarse mode particles can be now observed clearly (
Figure S7), with the majority of negative errors being calculated for low PM
1/PM
2.5 ratios, while a linear relationship with increasing concentration can be suggested for any PM
1/PM
2.5 ratio, starting with apparently negative slopes for low ratios and gradually moving towards positive slopes for higher ratios. A positive relationship with RH, even at mid-range concentrations is again suggested (
Figure S6), while negative errors seem to be linked to lower RH, possibly related to drier and dust-laden southern air masses affecting continental Greece [
54]. A boxplot of ΜBEs in 10% RH bins is also presented (
Figure S7), where the RH dependence is obvious.
It is necessary to exclude that this increase was introduced by the positive error-concentration relationship to examine whether the increased bias linked to RH can be actually attributed to rising RH values. In this scope, the ratio of PA-II PM
2.5(CF = 1) to PM
2.5 reference concentrations (PA-II/Ref) was calculated and its association with the observed reference concentration was explored (
Figure 4). In
Figure 4a,b, data-points are color-coded by the PM
1/PM
2.5 ratio and RH respectively, and it is indicated that lower PM
1/PM
2.5 ratios correspond to low PA-II/Ref ratios (ranging from zero to roughly 1.5). We used the PM
2.5/PM
10 ratio (according to the analysis in
Figure S8) in order to remove the points affected by the elevated coarse mode concentrations and consequently focus on the RH effect, given that, as an external parameter, it should be more readily available by regulatory monitoring stations, where, in most cases, PM
1 will not be routinely monitored. The colored scatterplot in
Figure 4c, were coarse-related data-points (PM
2.5/PM
10 ratios less than 0.5) have been removed, clearly illustrates the RH effect on the sensor data deviation from reference dried measurements, while the resulting boxplot of
Figure 4e points towards a linear pattern with the lower bin median PA-II/Ref at 1.41 and the higher at 2.10. A line was fitted to the calculated medians of each bin versus RH yielding an excellent correlation (R
2 = 0.96).
Through the error analysis, it is suggested that, in order to correct the PA-II PM
2.5 values, a multivariable approach should be followed, incorporating both humidity and coarse particle effects. The PA-II devices include a RH sensor, characterized by high repeatability (
Section 3.1.2), which can be calibrated ahead of field deployment (
Figure S4). However, the representative PM ratios required for corrections related to the presence of coarse mode particles should be externally provided. Such a solution could lie in the integration of PM
2.5, PM
10 data from a central or the nearest available regulatory air quality station (AQS), and it should be necessary when coarse-particle episodes are frequent and effective over wide urban areas. Similar calibration approaches have been already documented [
20,
55], albeit directly using reference PM
2.5 measurements coming from nearby regulatory AQS. Nevertheless, because lower PM ratios are not only related to regional phenomena, such as Saharan dust events in Southern Europe, this approach could face limitations regarding locally re-suspended dust. Therefore, its use in traffic sites, for example, where traffic induced resuspension can be substantial [
56], may result in unsolicited errors.
3.1.4. Models for Correction of PA-II PM2.5 Measurements
The Ioannina dataset was randomly divided in two subsets, selecting 60% as the base dataset (
n = 2556 hourly observations) and the remaining 40% as the evaluation dataset (
n = 1706) (
Figure S9). The base dataset was used to fit different regression models. Reference PM
2.5 concentration was used as the dependent variable and PA-II PM
2.5 (CF = 1) concentration, ratios related to the presence of coarse particles (PM
1/PM
2.5 or PM
2.5/PM
10) and RH, were used as predictors. In total, ten different models were fitted, using all or a combination of the aforementioned predictors, starting with simple linear and quadratic regression models involving only the sensor and reference PM
2.5 data (iModel 1, iModel 2), then incorporating the PM ratios (iModel 3 to iModel6), and finally RH (iModel 7 to iModel 10). Motivation for including the PA-II PM
2.5 (CF = 1) as a squared term comes from the convex patterns observed in
Figure 3, as well as the PA-II versus reference relationships documented in other studies [
21,
26,
31]. A detailed description of the developed models can be found in
Table S7. Generally, the larger part of the variability in the models can be explained by the PA-II PM
2.5 concentration variable. Positive increments with added predictors were somewhat small in terms of adjusted R
2, however coefficients of all added predictors were statistically significant at the 0.05 level, for all the models.
The obtained equations were then applied in the evaluation dataset and the model-corrected PA-II PM
2.5cor concentrations were compared to the reference measurements, calculating R
2, MAE, and normalized RMSE (nRMSE) as performance metrics.
Table 1 summarizes the results of the evaluation. The fitting ability of the models increased by progressively adding as predictors the variables that have been shown to directly influence the sensor’s measurement. Specifically, the nRMSE decreased by approximately 17% when the PM ratios were incorporated (iModel 3 and iModel 5) and approximately 21% when the polynomial model included also RH (iModel 7 and iModel 9), in comparison to the simple linear model (iModel 1). An important decrease (19%) in nRMSE, was also found by the simple quadratic models (iModel2), with no other external predictors. The model with the best descriptive power was the one incorporating PM
1/PM
2.5, RH, and a PA-II PM
2.5(CF = 1) quadratic term (iModel 8), yielding a 40% improvement in nRMSE, relative to the simple linear model. It is noted here that the PM
2.5(CF = 1) MAE, calculated for the evaluation dataset (before application of a correction model), was 25.4 μg m
−3, much higher than the MAE of models, for which the lowest value was for iModel 8 (MAE = 2.2 μg m
−3, 41% lower as compared to the simple linear iModel 1). Including PM ratios in the PA-II PM
2.5 calibration scheme appeared to drastically improve the behavior of the post-processed sensor signal during dust events.
Trying to discern which model incorporating the two coarse-related ratios (iModel 8 or iModel 10) performs better during a dust event, we focused on a severe dust episode recorded during 20–22 December 2019 and a milder event during 11–21 May 2020 in Ioannina. During these periods, air masses mostly originated from the southern sectors, with PSCF analysis indicating dust transport from north Africa (
Figure S11), in agreement with the low PM
2.5/PM
10 ratios calculated. In
Figure 5, the time-series of coarse PM as measured by the reference instrument is depicted, with the shaded areas corresponding to the respective dust events.
The mean PM
10-2.5 concentration was 38.7 μg m
−3 during the December and 27.3 μg m
−3 during the May event, in contrast to the 6.7 μg m
−3 average that was registered for the entire period. In the bottom panels of
Figure 5 the behavior of the two examined models along with the linear model (iModel 1) is shown. iModel 8 seems to generally perform better (
Tables S10 and S11), yielding average PM
2.5 values that were closer to the recorded for each period (34.0 μg m
−3 for December and 13.6 μg m
−3 for May, against 37.7 and 14.2 μg m
−3 that were measured by the reference instrument, respectively).
The same approach was followed for the investigation of the best corrections applicable to the Athens intercomparison campaigns, using the beta attenuation monitor measurements as reference. As opposed to Ioannina, in Athens the models were not found to benefit from the inclusion of PA-II PM
2.5(CF = 1) as a quadratic term, since maximum ambient concentrations were relatively low (below 80 μg m
−3 throughout the entire period;
Figure 3). Because the PM
1/PM
2.5 and PM
2.5/PM
10 ratios data availability was limited to the 2nd Athens intercomparison campaign, multiple linear regression models incorporating those ratios were tested only for this period. Additionally, in Athens, models using only PA-II PM
2.5(CF = 1) and RH as predictors were tested, with data spanning both the 2nd and 3rd campaigns, after excluding data-points evidently affected by dust events. As criteria for this exclusion, we used: (i) a threshold of 0.5 for the PM
2.5/PM
10 ratio for the 2nd campaign (considering the analysis presented in
Figure S8) and (ii) inspection of backward air mass trajectories when the PM
2.5/PM
10 ratio was not available.
The corresponding datasets were again randomly divided in base and evaluation subsets (60% and 40% respectively), creating in total four subsets (
Figure S10). Six different regression models were tested and they are summarized in
Tables S12 (configuration) and S13 (performance). The results for the evaluation datasets are summarized in
Table 2. Once more, the models incorporating the coarse-related PM ratios (aModel 2 through aModel 5), exhibited the lowest nRMSE (0.133–0.147) and MAE (1.8 μg m
−3) values and they were better correlated with the reference PM
2.5 measurements. Overall, the implementation of multiple regression models, resulted in an improvement of performance metrics, with aModel 4—the model incorporating the PM
1/PM
2.5 ratio and RH—corresponding to the smallest nRMSE (0.142) and MAE (1.8 μg m
−3) values.
It is evident that correcting the raw sensor signal drastically improved the PA-II performance metrics for both the Ioannina and Athens datasets, effectively reducing MAE from 25.4 μg m
−3 to below 3.7 μg m
−3 and from 3.1 μg m
−3 to below 2.2 μg m
−3, respectively. Similar improvements have been observed by the majority of studies applying statistical models for correction. For example, Magi et al. [
51] reported a 45% improvement in MAE after implementing a multilinear regression correction scheme, incorporating RH and reference PM
2.5 measurements as predictors, while Malings et al. [
49], using more complex corrections by calculating an RH-related hygroscopic growth factor, documented a reduced MAE by 40%. Pawar et al. [
57] applied a particle density adjustment followed by growth factor and aspiration efficiency corrections, reporting a 10–15% improvement in RMSE. Tryner et al. [
48] reported that their correction (taking RH into account in a simple regression approach) reduced the absolute value of the mean bias from 2.1 to 0.6 μg m
−3. Significant improvements in PA-II signal corrections have also been shown by the application of non-linear supervised machine learning approaches, such as artificial neural networks [
58,
59].
3.1.5. Comparison to Reference Instrumentation in Different Seasons
The corrected (aModel 6) PA-II output was examined against reference measurements at Thissio, in three different seasons during the 2nd and 3rd intercomparison campaigns, in order to assess the potential influence of seasonal factors on the performance of the PA-II monitor (that could be for example related to the seasonally variable chemical composition of urban aerosols).
The model choice was mainly dictated by the fact that the base dataset used in model aModel 6 was more representative, spanning all seasons in Athens, with the model displaying an overall good performance. The results are presented in
Figure 6. Temperature and relative humidity diurnal patterns and basic statistics during those three seasons are presented in
Table S14 and Figure S12. For the warm season (3 July 2019–3 September 2019,
Figure 6a), the averaged PA-II PM
2.5cor and PM
2.5(CF = 1) data from the five installed PA-II devices at Thissio were used as independent variables in the displayed linear regressions. For the other two seasons, the cold (26 February 2020–7 April 2020) and intermediate (8 April 2020–19 May 2020), data from one PA-II deployed at Thissio were used as the independent variable. The cold and intermediate seasons were categorized according to long-term climatology studies for Southern Greece [
60]. It is noted that, for the warm season data-points affected by coarse particles, according to criterion (i) described previously, were excluded, while for the intermediate and cold seasons, where no concurrent PM
1 or PM
10 measurements were available at Thissio, the affected data were excluded according to criterion (ii).
The comparisons indicate a relatively uniform response of the PA-II monitor, regardless of the measurement season in Athens, which is characteristic of consistent behavior for monitoring short-term PM
2.5 concentrations in the absence of elevated coarse mode concentrations. Furthermore, it is indicative that no significant drift in the sensor’s response was observed in the span of almost one year. The slopes of the linear regression of PM
2.5 reference versus PM
2.5(CF = 1), range between 0.44 and 0.48, while the correlations remain excellent with R
2 varying in the range of 0.86 to 0.93. It must be noted that a statistically significant intercept was observed for the Thissio dataset, probably owing to the different operation principle of the compared instruments, with PMS5003 being limited to the direct detection of particles only larger than 0.3 μm [
30].
On the other hand, the performance of selected models is considered to be satisfactory and consistent, yielding, in all seasons, small intercepts, slopes very close to unity, and excellent correlations to reference measurements, despite changes in chemical composition and aerosol source intensity year-round. The chemical composition and sources of aerosols have been extensively studied in the area of Athens. The main characteristics in urban/suburban background areas (such as those where devices were installed in the local network) are dominated by secondary aerosol throughout the year [
38,
56,
61]. Ammonium sulfate is more abundant during summer, while secondary organics (especially the less oxidized fraction) are enhanced in the cold period [
43]. Moreover, levels of primary organics and black carbon are largely enhanced during winter, mainly due to residential wood-burning emissions, while they diminish in summer, also due to decreased traffic [
38,
41,
43,
62]. The contribution of larger mineral dust particles in PM
2.5 is generally limited in the long-term (typically less than 10%). At background sites, where road-dust resuspension is limited, mineral dust is expected to contribute to PM
2.5 mostly during regional dust transport events [
56].
It is possible that the content in primary organics and black carbon, which are typically emitted in the ultrafine range, is an additional error-inducing factor in the PA-II output, since particles with nominal diameters lower than 0.3 μm are not directly counted by the sensor. Actually, the mass fraction gathered in this size-range can be substantial. For example, Pennanen et al. [
63] in central Athens has found 26% of PM
2.5 mass to be concentrated in sizes below 0.2 μm. These size ranges are dominated by organic and elemental carbon particles, while upper ranges (0.3–0.5 μm, 0.5–1 μm) are characterized by the increased presence of nitrate and sulfate particles [
63,
64,
65]. However, aerosol ions are susceptible to water uptake which alters the scattering characteristics of sampled particles. Moreover, differences on apparent densities and shape factors of particles are related to their chemical characterization [
66] and they can indirectly bias the sensor’s estimation algorithm that assumes fixed values [
67].
The associations between measurement error and concentrations of source-specific aerosol components was examined in order to assess the potential effects of the seasonal variability of sources and chemical composition on the monitor’s performance. Small differences in error indicators have been found when evaluating the PA-II device in laboratory conditions for specific polydispersed particle sources with largely variable density and size distribution characteristics [
25,
27]. However, for ambient aerosols, where particle density is less variable and submicron aerosol dominates the PM
2.5 fraction [
68], there is not much evidence on the chemical composition effects.
In the present case, the absolute error of PA-II PM
2.5cor (corrected with aModel6) relative to reference measurements was correlated against components that function as tracers of specific sources and atmospheric processes. The two aethalometer BC components (BC
ff and BC
bb) were used as indicators of traffic-related and biomass burning emissions (residential wood burning during the cold period and regional forest and agricultural fires during the warm) [
69,
70]. The sulfate concentrations that were determined using the ACSM were used to represent regional transported and processed secondary aerosol [
43]. Comparisons were performed for the 2nd and 3rd Athens intercomparison campaigns; therefore, covering all three examined seasons (cold, warm, intermediate).
Figure S13 shows the season-specific fine aerosol chemical composition measured by on-line instruments and
Figure S14 the results of the comparison.
The results provide a first indication that the PA-II corrected signal is not directly affected by changes in chemical composition and intensity of examined sources. Correlations with aerosol type indicators in all cases were very weak (R
2 < 0.15). Additionally, the performance of the monitor remained stable when shifting from the warm to the cold period, which are characterized by distinct fine aerosol speciation. However, more work is needed in order to find an optimal way of investigating and addressing chemical composition effects and, furthermore, to explore links between the chemical composition and aerosol hygroscopic properties affecting the sensor’s performance in ambient conditions [
49].