NO2 sensors in the field (termed “candidate sensors” here) were calibrated using one of three methods: reference site collocation, transfer standard collocation, or remote network calibration method. For reference site collocations, a candidate AQMesh unit was installed alongside a reference monitor from the London Air Quality Network (LAQN) or UK Automatic Urban and Rural Network (AURN) (Fig. S1 shows a picture of an example reference site collocation). Transfer standard collocations relied on nine AQMesh devices that were periodically (every 2–4 months) collocated and calibrated against reference monitors; these calibrated AQMesh units were then used as transfer standards and were collocated with so-called candidate AQMesh units in the field to determine the latter's calibration parameters. The duration of typical calibration collocations was 7–14 d (for both reference and transfer standard methods), though long-term collocations were also conducted for further performance evaluation purposes. Calibration gain and offset parameters were obtained by performing a linear regression on the hourly averaged collocation time series after excluding the 1st and 99th percentile of hours during the collocation based on the ratio of reference / candidate values. Calibration parameters were deemed valid and applied to the candidate sensor if the scaled collocation time series met statistical criteria of normalized root mean square error (nRMSE) <50 % (Eq. 3) and R2>0.7 (Eq. 4), which ensured that sensor performance was sufficient to calculate robust calibration parameters and effectively excluded periods where the NO2 variability was too low to provide a meaningful test of sensor gain and offset.

The remote network calibration method is a novel approach, developed and applied by the University of Cambridge project team, that remotely derives unit-by-unit calibration parameters for the entire sensor network in lieu of physical collocations. The algorithm uses a spatial scale separation methodology described in previous work (Heimann et al., 2015; Popoola et al., 2018) to calibrate sensors in relation to each other when pollution levels are consistent across the network and obtains traceability (connection to reference standard with known uncertainties) from a single calibrated reference monitor (Popoola et al., 2022; Popoola and Jones, 2020). For BL, a single (site-dependent) calibration was performed using the period May–December 2019 and applied to the entire dataset. This paper does not intend to evaluate the network calibration method compared to other approaches. The method and its performance will be addressed in more detail in a separate study (Popoola et al., 2022). However, we describe the method here because it was used to scale a subset of BL sensors which had no physical (reference or transfer) calibration available, and we include data from this subset of sensors to maximize the number of sensor locations in our analysis and comparisons to the reference network.

When multiple valid calibration options were available for a specific AQMesh sensor in the network, a decision tree was used which prioritized (i) reference site collocation (n=11), (ii) transfer standard collocation (n=73), and (iii) network calibration (n=38); the total number of calibrations applied exceeded the number of devices because failed sensors were replaced and re-calibrated.

A long-term upward drift in BL NO2 sensor measurements was observed (Fig. S2), which we hypothesized to be caused by an ozone cross-interference. A correction was applied to the hourly NO2 dataset that subtracted a fraction of the derived site-specific O3 concentration from the scaled NO2 readings. Site-specific O3 was deduced using upwind background reference O3 measurements; under low-NOx conditions (<10 ppb) the site-specific O3 was assumed to be the upwind background O3 concentration; otherwise it was assumed to be the difference between background O3 and the site-specific NO concentration. Because the effect appeared to increase as a sensor aged, the cross-sensitivity correction for ozone was assumed to start at 0 % upon initial sensor deployment and exponentially increase to a maximum of +18 % of estimated site-specific ozone concentrations 6 months later. Figure S2 illustrates the effect of the correction on BL network mean NO2 concentrations throughout the campaign. Figures S3 and S4 show evidence supporting the ozone cross-interference hypothesis and an evaluation of the correction method for an individual sensor. Note that due to a complex set of factors including the combination of factory (AQMesh) and field calibration methods, we could not exclude other possible causes of observed irregularities in sensor measurements. Except for the short-term collocation analysis results (Fig. 2), the results presented throughout this paper use the scaled hourly average ozone-corrected dataset.

Detailed documentation of the static network QA/QC procedures is available in the project QA/QC manual in the Breathe London Technical Report (Breathe London, 2020).

The BL network generated nearly 1.5 million hourly calibrated NO2 measurements from 100 devices at 112 locations over the course of the 26-month pilot campaign. The number of sensor locations producing valid, calibrated data gradually increased over the first 7 months (Fig. S6). The initial delay in network data capture was caused by logistical challenges faced at the outset of the project, including obtaining permissions for monitor deployment and conducting calibrations for each sensor. By the spring of 2019 the majority of the network was operational and generated valid data for the remainder of the project, though the total number of NO2 sensors producing valid data fluctuated due to redaction of flagged data and the downtime of sensors that failed during the project before replacement and re-calibration were performed. In total, 35 NO2 sensors were replaced due to failure, with most failures occurring during the winter. Additional considerations and lessons learned for stationary sensor network setup and maintenance are discussed in the Breathe London Blueprint (Breathe London, 2021b).

Figures 2 and 3 present measurement uncertainty statistics for calibrated BL LCSs based on short-term (typically 7–14 d) and long-term (>80 weeks) reference collocations. Both analyses quantify uncertainty of sensor measurements that were calibrated based on results of a prior reference site collocation (Sect. 2.3.1). These results allow us to evaluate the effectiveness of the project's QA/QC procedures (including calibration) since each repeat collocation serves as an independent test of the project-long uncertainties in sensors that were calibrated during a discrete time period.

Figure 2 shows evaluation results for 10 calibrated sensors that were collocated for subsequent short-term periods (typically 7–14 d) that began 1–84 weeks after an initial reference site calibration period. These subsequent collocation periods (n=35) were used to estimate calibrated sensor uncertainty compared to reference measurements (e.g., unit 99 was calibrated based on the first reference site collocation in October 2018; uncertainty statistics in Fig. 2 were calculated from the second and third reference collocations, which occurred in April and July of 2019).

A median R2 of 0.79 indicates that calibrated sensors effectively captured changes in NO2 concentrations that were measured by reference instruments. The median MBE was 8.0 µg m-3 (23 % of mean concentration) with a range of -19 to 34 µg m-3 (-37 % to 121 % of mean concentration), and the median normalized RMSE was 35 % (range of 16 % to 189 %). The large range of biases exhibited by individual sensors and the systematically high median bias of the collocated sensors reveal variability in the consistency of sensor response over time (and under different meteorological conditions) and serve to assess the robustness of initial sensor calibrations when applied to a longer time series. However, we note that uncertainty statistics in Fig. 2 are calculated from sensor data that were not corrected using the ozone cross-interference correction (Sect. 2.3.2) and thus represent an upper bound of the BL network uncertainty.

The Fig. 2 results and summary statistics are affected by a group of outlier collocations (unfilled circles in Fig. 2) that started in July 2019, during which most sensors exhibited higher measurement error and poorer correlation to reference measurements. The eight collocations with the highest normalized RMSE (>70 %) all occurred during July 2019 (Fig. S7). Additionally, seven of these July 2019 collocations had R2 values below 0.7, meaning they would have failed the statistical screening criteria used for determining valid collocation calibrations (Sect. 2.3.1). During this month, we observed high-biased sensor measurements when local air temperatures were above 20–25 ∘C, which we discuss further below. With the July 2019 collocations (n=11) excluded, the median nRMSE and MBE of the remaining collocations (n=24) improve to 30 % and 4.5 µg m-3, respectively, and the median R2 increases slightly to 0.81.

Figure 3 presents the collocation time series and monthly error statistics between calibrated BL sensor and reference monitor measurements during two long-term (>18 month) collocations, where the sensor measurements are calibrated based on the collocation results during the 2-week period directly preceding the extended evaluation period. While the aggregate MBE of both collocations is small (<2 µg m-3), BL sensors exhibit biases that vary seasonally relative to reference measurements; MBE of sensors during 14 d periods (Fig. 3b) ranges from -3 to +11 µg m-3 for unit 17 (-8 % to +34 % of the 14 d mean concentration) and from -8 to +19 µg m-3 for unit 83 (-20 % to +91 % of the 14 d mean concentration). The drifting sensor response follows the same seasonal pattern for both long-term collocations, with the highest bias occurring during summer months and peaking during August 2020. Variations in RMSE error are largely driven by sensor bias; nRMSE is highest during summer months, corresponding to peak BL sensor bias. Figure S8 further illustrates the occurrence of high-biased BL sensor measurements during hours when the local air temperature exceeded 20–25 ∘C. While the results presented above quantify uncertainty of sensors calibrated using reference collocations, the data use cases in the following sections also include sensor data calibrated using two additional approaches when sensors could not be collocated at reference sites, as described in the “Methods” section (Sect. 2.3.1): transfer standard calibration and network calibration method. The transfer standard method is more difficult to validate because collocations occur at BL sites in the field instead of at reference sites. The uncertainty of this method is expected to be marginally higher than the direct reference site collocations in Fig. 2 due to the additional step where the calibration is transferred between BL AQMesh units. A high level of precision and consistency in response across BL NO2 sensors (R2=0.94, nRMSE = 0.1; Fig. S9) gives confidence that calibrations would transfer effectively between units. Preliminary evaluations have shown that the estimated uncertainty of BL sensor measurements scaled with the network calibration method is broadly similar to the uncertainty of reference collocation-calibrated sensors (Popoola et al., 2022; Popoola and Jones, 2020). The results in Figs. 2 and 3 demonstrate that the long-term measurement uncertainty of sensors calibrated during a brief, discrete period is influenced by the changes in sensor response during different seasons and environmental conditions. Enhanced QA/QC such as calibration on a near-continuous basis or seasonal bias corrections such as the one shown in Fig. S10 (see Sect. 3.2.1) could minimize variations in measurement uncertainty due to sensor performance.

For a long-term measurement campaign using sensors, evaluation against reference measurements should be performed throughout the course of the project. The evaluation results above point to the ability of BL sensors to accurately reproduce changes in NO2 concentrations captured by the reference monitors (high R2 values) with average uncertainty (nRMSE) of ∼ 35 %. However, our results also show that seasonal biases due to time-varying effects of environmental interferences can lead to larger uncertainties (>100 % nRMSE) during periods when local air temperatures reached above 20–25 ∘C. This characterization of sensor uncertainties can inform how results from the BL LC sensor network are interpreted, ensuring derived insights are robust (e.g., differentiating between high-biased sensor artifacts and elevated NO2 concentrations).

We next present a series of analytical use cases to evaluate the applicability of BL NO2 LCS network results for deriving insights about the local air pollution environment. Results from each use case using BL data are compared against results generated from reference network data. In addition, the collocation sensor evaluations presented above are used to assess BL network uncertainty and interpret differences between BL and reference network results.

We first examine the ability of the BL sensor network to characterize trends in the regional (Greater London) pollution load by comparing monthly mean NO2 concentrations of the BL network with the reference network results (Fig. 4). We compare monthly values here to assess the sensor network's ability to reproduce long-term patterns observed by the reference network on timescales that would be sensitive to effects of seasonal variations in pollutant concentrations or sensor performance as well as long-term ambient pollution changes resulting from major interventions. We note two major events during the measurement campaign which are expected to impact NO2 concentrations: (i) introduction of the Ultra Low Emission Zone (ULEZ), which became effective on 8 April 2019, imposed tolls to discourage entry of older, higher-emitting vehicles into Central London, with increasing fractions of compliant vehicles and fewer vehicles overall observed in the zone through calendar year 2019 (GLA, 2020b), and (ii) Covid-19 pandemic restrictions beginning March 2020, including social distancing measures and stay-at-home orders, disrupted activity patterns throughout Greater London.

The BL network tracked the reference network trend while exhibiting lower mean concentrations for most of the campaign (on average 7 µg m-3 lower throughout campaign; 9 µg m-3 during 2019). We attribute this partially to differences in location, site types, and sampling points (height, distance to road, road traffic volume, etc.) between the networks, and this is confirmed through comparisons of measured and modeled concentrations using the ADMS-Urban air pollution dispersion model (described in Sect. 2.6). Modeled network mean NO2 concentrations for 2019 at reference network monitoring site receptors were 5 µg m-3 higher (∼ 15 %) than the modeled mean concentrations at BL receptor locations. Because the model only predicts 55 % of the difference between the two networks, we examined the model–network comparisons more closely. The model exhibits little systematic bias at reference sites (<1 µg m-3; see Fig. S11). By contrast, the mean of modeled concentrations was higher than that observed at the BL sites by 6 µg m-3, with the difference driven by BL sites with the lowest observed concentrations (Fig. S11). We note that 16 BL sites exhibited lower concentrations than the 20 µg m-3 minimum observed by the reference network, so we cannot rule out the possibility of low sensor bias in a portion of the BL network. In sum, we are unable to fully resolve the cause of the systematic difference between modeled and observed BL concentrations, although it may have contributions from uncertainty in sensor network measurements (and underlying QA/QC) and model uncertainty.

Both networks show a downward year-on-year trend in NO2 concentrations and seasonal variability with peak concentrations in the winter. However, BL NO2 means exhibit local maxima in July and August, when reference network measurements are lowest. This effect is the most pronounced in summer 2020, which is the only time when the BL network average exceeds that of the reference network. This bias of the BL network compared to reference network trends during summer months is likely due in part to a systematic high bias in the BL network's NO2 sensors coinciding with local air temperatures above 20–25 ∘C, an effect which was evident during collocations with reference monitors (Figs. 3, S7, and S8). However, spatially varying NO2 pollution trends (e.g., Covid-19 restrictions having a larger impact on emissions at specific monitoring sites or city neighborhoods) may have also affected the two networks differently and contributed to the converging network means towards the end of the BL campaign.

The long-term collocations (Sect. 3.1.2) were used to quantify seasonal changes in sensor bias and could serve as a basis for an empirical correction to the Fig. 4 BL network time series to improve the accuracy of the LCS results. This correction relies on the performance results being consistent across the network; the high precision between AQMesh units in our transfer standard collocations (median R2=0.94; Fig. S9) supports this assumption for the BL project. In Fig. S10 we show the Fig. 4 BL time series with a monthly bias correction based on the long-term collocations that would largely mitigate the seasonal irregularities in the BL time series compared to the reference network.

We next compare the recurring temporal patterns in NO2 concentrations measured by the BL and reference networks (Fig. 5).

The BL network captures diurnal and day-of-week patterns with three key differences from the reference network. First, BL network mean concentrations are ∼ 10 µg m-3 (23 %) lower than the reference network result. Most of this difference was predicted in the modeling exercise discussed in Sect. 3.2.1, with additional contribution from uncertainty in sensor measurements. Second, BL network mean concentrations show a reduced diurnal range compared to the regulatory network (i.e., though daytime average BL concentrations are lower, nighttime values are similar to the reference network). This behavior may be due to previously discussed differences in site characteristics (e.g., higher sensor placement and lower traffic volume at near-road sites) yielding reduced heterogeneity in site types across the BL network, which as a whole appears to be measuring diurnal pollution patterns that are more in line with urban background reference sites (Fig. 6). A similar effect is observed in a comparison of near-road (kerbside and roadside) and urban background reference sites, where the concentration difference was smallest during late-night/early-morning hours (Fig. S12). A third key difference in diurnal day-of-week concentration patterns is the magnitude of the evening peak, which is consistently lower than the morning peak in the BL network. On Wednesdays, for example, the reference network evening peak reached 57 µg m-3 at 18:00 LT, while the BL network reached 40 µg m-3 at the same time; other weekdays similarly have the largest difference in network mean concentrations during the evening rush hour peak. We have not identified a mechanism to explain this difference, which is evident, to a varying degree, throughout the year (Fig. S13).

The BL network was able to accurately characterize timing of peaks and troughs in diurnal variability as well as capture differences in weekday and weekend pollution levels. Uncertainties in the precise magnitudes of some features remain, with the evening peak registering relatively lower in the BL network.

We next examine the ability of the BL network to detect differences in diurnal NO2 concentration patterns at different monitoring site types. Figure 6 shows the weekday diurnal averages for the BL and reference network at near-road (kerbside and roadside) sites compared to urban background sites.

In the morning, near-road concentrations peaked at 08:00–09:00 LT in both the BL and reference networks, reaching 60 µg m-3 in the reference network and 50 µg m-3 in the BL network. The time of the evening peak was also consistent between networks, occurring at 18:00–19:00 LT and reaching 60 µg m-3 in the reference network compared to a lower peak of 44 µg m-3 in the BL network at the same time. In both networks, the evening peak in concentrations occurred 1 h later (19:00–20:00 LT) at background sites than near-road sites. The greatest difference between BL and reference means at both near-road and urban background sites occurred during the evening peak in NO2 concentrations; this feature is identified in the network-wide trends in the prior section.

At this aggregate level, the lower-cost network captures similar diurnal features and effectively differentiates between pollution levels and time-of-day trends at urban background and near-road sites.

Here we discuss the application of BL LCS data for identifying hotspots and characterizing spatial heterogeneity in NO2 concentrations, using a case study where BL sensor measurements led to identification of an air pollution hotspot. During the first winter of the project (December 2018 through February 2019), a BL sensor deployed at Holloway Bus Garage measured mean weekday NO2 concentrations of 77 µg m-3, 89 % higher than the BL network weekday mean of 41 µg m-3 (Fig. S14). Though the concentration gradient (between Holloway Bus Garage and the BL network mean) was larger than the typical sensor uncertainties (∼ 35 % nRMSE) and occurred during winter months when large positive biases were not observed during collocation evaluations, additional steps were taken to establish confidence that the local pollution levels were accurately characterized and not sensor artifacts. Two additional BL sensors were deployed in the area, and a follow-up transfer standard collocation was performed which verified the accuracy of the deployed pod's calibration factors.

The BL monitoring at Holloway Bus Garage ultimately led to corrective action by local authorities, and this successful example demonstrates the potential value of LCSs for identifying air pollution hotspots. The case study also emphasizes the need for rigorous verification of measurements from an individual sensor. The collocation analyses quantified a wide range in the bias of BL sensors over the course of the project as well as uncertainty in the consistency of sensor performance over time (Figs. 2 and 3). Therefore, especially for concentration gradients of similar magnitude to the estimated uncertainty of the sensors, there is a need for caution when analyzing site-specific data; we established confidence in the LC sensor hotspot characterization through the deployment of additional LC sensors to verify results.

Here we test the viability of the BL network to detect short- to medium-term (hours to days) episodes of elevated NO2 concentrations using a well-characterized historical air pollution event in December 2019 (LAQN, 2019). Weather conditions in Greater London resulted in the formation of a strong temperature inversion that caused a build-up of primary pollutants, including NO2, in the layer of colder air close to the ground, with pollution peaking at morning rush hour on 4 December (LAQN, 2019). Figure 7 compares the hourly mean NO2 concentrations as measured by the BL and reference networks for the week of the pollution episode.

The BL network detected a short-term regional build-up of pollution with a temporal profile that provides excellent comparability with the reference network result (R2=0.96) and corresponds to the London Air Quality Network's published report about the event. The highest peak occurred during late rush hour (09:00–10:00 LT) on the morning of 4 December, with the BL network registering a peak of 87 µg m-3 compared to 103 µg m-3 for the reference network during the same time period. Another smaller peak occurred when evening emissions were trapped on 4 December, and the event subsided when the inversion broke near midnight on 4 December. The BL lower-cost network captures the basic features of the event, although there is a low bias compared to the reference sites (nRMSE = 23 %, MBE = -11 µg m-3), partially explained by the different site types (see Sect. 3.2.1).

However, we note that the network was less effective in characterizing pollution events during periods of poorer sensor performance. In Fig. S15 we present a more cautionary case study during July 2019. We demonstrated in Sect. 3.1.2 that collocated BL sensors produced high-biased measurements during periods when local air temperatures reached above 20–25 ∘C, with worst-case nRMSE exceeding 100 % (dominated by positive bias). Figure S15 shows an instance where this effect leads to an overestimation of regional NO2 pollution levels using the BL network (for example, the BL network mean during daytime hours on 25 July 2019 is 91 µg m-3 compared to the reference network mean of 65 µg m-3, a 40 % positive bias across the network). Due to the extensive reference network in London and frequent short-term as well as ongoing long-term BL sensor collocations, we were able to identify the apparently anomalous BL sensor behavior under these environmental conditions which resulted in a systematic positive bias across the network. However, in cases with limited reference monitoring infrastructure, the BL measurements could have led to an overestimation of the magnitude of the pollution event in question. While technological and methodological solutions to address this sensor issue are viable, another project with different technologies or environmental characteristics may experience different effects, illustrating the importance of rigorous data validation and uncertainty evaluation in the context of each new application of LCS technology.