The Carbosense CO2 sensor network covers the whole of Switzerland with a regional focus on the city of Zurich (Fig. 1). It is formed by three classes of sensors: (i) seven high-precision cavity ringdown spectrometers (Picarro G1301, G2302 and G2401); (ii) 20 temperature-stabilized, mains-powered NDIR medium-cost sensors with active sampling and reference gas supply (Senseair HPP; Hummelgard et al., 2015); and (iii) 300 nodes of battery-powered CO2 low-cost diffusive NDIR sensors (Senseair LP8). The deployment of the first low-cost sensors started in July 2017, and the network has been continuously extended, reaching 230 sensors by September 2019. The CO2 low-cost sensors are deployed at antenna locations of the telecommunication company Swisscom (4–150 m above ground), meteorological measurement sites of the Federal Office of Meteorology and Climatology (MeteoSwiss; 10 m above ground), and sites of the National Air Pollution Monitoring Network NABEL (5 m above ground). Within the city of Zurich, the sensors are also mounted on lamp posts or electricity poles (3–5 m above ground). The low-cost sensor network covers a wide altitude range from 200 to 2390 m a.s.l. and various orographic conditions and landscape types (urban areas, agricultural lands, forests, mountain areas). This implies a wide range of environmental conditions during the operation of the sensors. The deployment of the HPPs started in August 2018, and instruments were operating at 15 locations as of 1 September 2019.

The raw data from the sensor units, after being transmitted via the LPN to a Swisscom server, are forwarded via the Internet to Decentlab where they are stored in an Influx database (InfluxDB, 2019), providing near-real-time access to the data. Decentlab provides web-based dashboards for data visualization as well as APIs for data access in various scripting languages. Information about the sensor network such as deployment history, calibration runs, calibration parameters, observations from reference instruments and processed sensor measurements is stored in a MySQL database hosted by Empa.

The deployment of a large number of LP8 sensors in this study revealed two issues that are important for ambient long-term measurements with this sensor type. First, the response characteristics of the LP8 infrared measurement can change over time, both steadily and abruptly. Sudden changes in the sensor response might be due to mechanical stress of the plastic housing under continuously varying environmental conditions. Second, the infrared measurements are susceptible to humidity exceeding a value of about 85 %. This behaviour is common to all LP8 sensors, but actual thresholds differ among individual sensors. Therefore, additional processing steps subsequent to the application of the calibration function are required to achieve a data set of sufficiently high accuracy and completeness (Sect. 3.3, 3.4, and 3.5).

Several analyses that are presented in the following sections refer to the term deployment. We define deployment as the time period within which a specific sensor unit is placed at a particular outdoor location. A sensor unit can be used in several consecutive deployments.

Each LP8 sensor was individually calibrated. For this purpose, each sensor unit was placed in the climate and pressure chambers for at least one complete calibration. Furthermore, each unit was operated under ambient conditions at site DUE until it was shipped for deployment in the Carbosense network. The sensors were run at DUE under ambient conditions in parallel with a Picarro instrument for a time period of between several weeks and several months.

Thus, an extensive data set of both chamber and ambient measurements was collected for each sensor unit to determine the calibration parameters of Eqs. (4) and (5). Filters that exclude conditions near condensation and large changes in IR measurements or in ambient CO2 were applied to this data set for optimal parameter estimation. The data filtering during calibration is more rigorous than the outlier detection applied to the sensors deployed in the Carbosense network (see Sect. 3.4). A robust estimator (Huber loss function) was used for the parameter estimation to minimize the impact of large residuals (e.g. persons breathing near the sensors). The parameters of the LP8 sensor models are stored in the MySQL database. The sensor unit has to pass a new calibration cycle whenever the LP8 sensor is exchanged.

Measurements from LP8 sensors deployed within the Carbosense network are processed by using Eqs. (4) and (5) with the corresponding coefficients determined during the calibration phase, yielding a first-guess CO2 molar fraction CO2,CAL. Thereby, there are two parallel processing chains, but besides the computation of CO2,CAL, the further processing is performed equally (outlier detection, drift correction). The function f(t) in Eqs. (4) and (5) is replaced by a constant that equals the last value of this step function during calibration.

The first-guess CO2,CAL is subsequently corrected for sensor drifts as described in Sect. 3.5. Equations (4) and (5) require the pressure at the sensor location. This value is derived from 10 min pressure measurements from the meteorological measurement network SwissMetNet operated by MeteoSwiss (Supplement Fig. S2). A procedure was set up that estimates the vertical pressure gradient in Switzerland and horizontally interpolates the pressure reduced to sea level every 10 min. These values allow for the computation of the pressure for any location and height above ground level with an uncertainty of about 1 hPa. The accuracy of the pressure interpolation has been determined from a comparison to measurements at SwissMetNet sites performing a leave-one-out cross validation.

Results and flags of subsequent processing steps are stored in the MySQL database to guarantee full traceability and to support the comparison of different processing options.

A relative humidity threshold RHtrsh was determined for every LP8 sensor based on the measurements from the ambient calibration performed at DUE. The purpose is to review the operation limits specified by the manufacturer and to develop a method for flagging the sensor measurements that may be impacted by humidity.

First, the standard deviation of the CO2 residuals (difference between computed CO2 values of the sensors and CO2,moist measured by the Picarro) is computed in 2 % RH intervals in a range of relatively dry conditions between 40 % RH and 70 % RH (resulting in 15 values in total), and the median of these values denoted as σres is determined. Second, the 95 % quantile of the residuals is computed in 2 % intervals from 0 % RH to 100 % RH. RHtrsh is then selected as the maximum interval for which the 95 % quantile is smaller than 3⋅σres. The CO2 residuals and the computed RHtrsh values are exemplarily depicted for two sensors in Fig. 3a and b. The operation limits for humidity indicated by the manufacturer (0 % RH–85 % RH) concur with our results (Fig. 3c). All the RHtrsh values are stored in the database.

Flagging the measurements of the deployed sensors by applying the criterion RH > RHtrsh results in a data set with very few outliers but, concurrently, a significant number of measurements are falsely rejected. In Sect. 3.4 a more adaptive outlier detection algorithm is presented that does not rely on any reference measurements. The choice of the filtering approach depends on the intended use of the measurements and whether the number of undetected outliers or the number of falsely flagged outliers is more important.

We call an LP8 measurement an outlier when it cannot be related to the ambient CO2 molar fraction by means of the sensor models described by Eqs. (4) and (5). Outliers are primarily caused by relative humidity exceeding about 85 % (see Fig. 3). Under these conditions the light absorption within the measurement cell can be increased due to the presence of water droplets or condensation of water on the mirrors. Such conditions may last for periods of between a few minutes and more than a day. The difficulty in detecting such events is that the signals in the LP8 IR and SHT21 RH time series do not follow a characteristic profile but exhibit significant variation depending on the actual progression of the meteorological conditions. The distinction between small humidity effects and a true increase in CO2 is not a simple task as the sensor measurements do not fully describe the conditions in the measurement cell. In addition, temporary enhancements from closely located emission sources can unusually impact the CO2 measurements and should not be treated as outliers.

The outlier detection algorithm was designed to rely entirely on the measurements from the sensor unit itself and to require no auxiliary information such as measurements from a reference instrument. It analyses and processes quantities derived from sensor observations that, under normal conditions, vary only slowly. The algorithm learns the sensor's usual behaviour at its current location from data obtained during the particular deployment and flags unusual measurements. Prerequisites for the algorithm are that environmental conditions and their changes remain within certain limits and that stable relations exist between specific sensor quantities and environmental conditions. Learning sensor behaviour in the field is an important element for minimizing the required calibration time.

Thus, the LP8 outlier detection algorithm is primarily based on the differences in consecutive log(IR) and temperature values plus statistical measures that are derived from a large number of IR measurements. The algorithm also reviews the relative humidity to enhance the robustness of the algorithm. The absolute values of IR and the corresponding values of CO2,CAL were not directly used as both are not stable over time due to drift or jumps and as they depend on CO2, temperature and pressure, which are variable over time.

First, the outlier detection algorithm requires the computation of several auxiliary quantities. In the following, IR, T and RH denote the infrared measurement, the LP8 temperature and the SHT21 relative humidity. The subscripts M and L refer to the mean and the last single measurement in a 10 min interval. Δt indicates the time between subsequent measurements transmitted to the database (subsequent measurements are only considered if the difference does not exceed 20 min).

Difference in log⁡(IR). ΔIR,M(t)=log⁡(IRM(t))-log⁡(IRM(t-Δt))

Second, based on the samples in relatively dry conditions (RH < 80 %), a quadratic function ΔIR,M=f(ΔT,M) is robustly determined which describes the normal change in log⁡(IR) with a change in temperature. The corresponding residuals r for all samples are computed, and, again based only on the dry samples, the median absolute deviation (MAD) is calculated (Fig. 4).

A measurement IR(ti) is flagged when |r(ti)| > 3⋅MAD∩|ΔIR,M(ti)|>3⋅σM(ti)∩RH(ti)>70% (value set according to Fig. 3c). The positive flagged residuals are denoted as rflag,pos and the negative flagged residuals as rflag,neg. Starting from ti, consecutive (for all the rflag,pos(ti)) or preceding (for all the rflag,neg(ti)) measurements are also flagged until RH drops below MRH(ti). In general, high relative humidity leads to decreased IR and, concurrently, increased CO2 values for the LP8 sensor. Concurrently, the sign of r(ti) determines the direction of backward or forward flagging of temporally adjacent measurements.

In addition, two more quantities are determined based on rflag,pos: RHQ75 is the 75 % quantile of the RH values, and TDPQ25 is the 25 % quantile of the difference between T and the dew point (Td).

Measurements are also flagged if (i) |γ(ti)|>5⋅IRnoise∩RH(ti)>85%, (ii) |ΔIR,M(t)|>5⋅ΔIRlarge, (iii) RH > RHQ75 or (iv) T–Td < TDPQ25. Under (i), the first criterion is already fulfilled if two of seven |γ(ti)| adjacent to ti indicate increased noise.

IR measurements from LP8 sensors and the corresponding calibrated molar fractions CO2,CAL, corresponding to χCO2 in Eqs. (4) and (5), are not stable in time. For sensors deployed in the field, this drift has to be corrected in order to compute unbiased CO2 molar fractions. Since usually no reference measurement is available at the location of the LP8 sensor to determine the drift, a method was developed that makes use of specific weather conditions during which horizontal gradients in CO2 are small and links the measurements of the LP8 sensor to those of the closest accurate instrument. The criterion of small horizontal gradients and a well-mixed planetary boundary layer is best met during situations of high wind speeds.

The drift correction algorithm involves two consecutive steps: first, the identification of time periods Pslow when the sensor behaviour is slowly evolving and the drift can be corrected and of periods Pfast when the behaviour changes abruptly and, second, the determination of the drift and its correction.

For the first task, the identification of Pslow, the calibrated measurements CO2,CAL from the afternoon are analysed because CO2 molar fractions are most comparable from day to day in the afternoon when the planetary boundary layer is usually well mixed (Fig. S1).

The algorithm computes for each sensor and day td the following quantities from the calibrated measurements CO2,CAL (time in Switzerland refers to CET or CEST):

Qprev7d (td) – 20 % quantile of CO2,CAL(τ) where τϵ [t-7 d… t-1 d] ∩τϵ (13:00–17:00 UTC)

Further, an empiric threshold ΔQTRSH is computed as median (Qn7d-Qp7d)+5⋅MAD(Qn7d-Qp7d). Sensor behaviour is considered unsteady (Pfast) if |Qnext7d(td) – Qprev7d(td)| > ΔQTRSH or if |b15d (td)| > 3 ppm d-1 ∩|Qnext15d(td) – Qprev15d(td)| > 40 ppm. Drift correction is applied separately to the measurements of each sensor deployment and continuous time period Pslow. An example of the working principle of the algorithm is shown in Fig. 5.

The actual drift correction is based on wind measurements from MeteoSwiss sites (Fig. S2) and CO2 measurements from the high-precision instruments deployed in the network (both 10 min averages). The drift correction algorithm is applied to the calibrated measurements CO2,CAL from the sensors deployed in the Carbosense network.

First, all the MeteoSwiss sites within a distance of 40 km from a sensor are selected. Time periods are identified when all the selected sites report for at least 90 min i.

wind speed > 2 m s-1 or

For the data set presented in this study, we use only the high-precision measurements from the sites DUE, PAY and GIMM for the adjustment of the LP8 sensors. This procedure allows for quantifying the accuracy of the concept by means of the remaining reference sites. In fact, measurements from GIMM are only used to adjust LP8 sensors deployed at PAY. Thereby, co-located sensor and reference measurements are independent in this data set (see Sect. 4.2). Obviously, three reference sites are not sufficient to accurately adjust all the LP8 sensors deployed in Switzerland as weather conditions often differ from region to region. Drift correction for a final and optimized LP8 data set will rely on measurements from all the reference sites and also from the HPP instruments (Fig. 1).

The assumption of spatially homogeneous CO2 mole fractions during strong wind events was tested by treating measurements from reference instruments in the same way as those from the LP8 sensors. Whenever an LP8 sensor was corrected at sites BRM, LAEG, HAE and RIG relative to DUE or at site PAY relative to GIMM, the 10 min CO2 molar fractions measured by the Picarro instruments at the two sites were compared. Not considered are RH and the measurement completeness of the LP8 sensor. Figure 7 shows CO2 differences in measurements from sites LAEG (distance d=19 km; height difference Δh=423 m), RIG (d=39 km; Δh=598 m), BRM (d=41 km; Δh=366 m) and HAE (d=61 km; Δh=-2 m) with respect to DUE as well as CO2 differences at PAY (d=35 km; Δh=57 m) with respect to GIMM. All these sites are located in or adjacent to the Swiss Plateau (Figs. 1 and 7f) and therefore have mostly similar weather conditions. The CO2 differences are depicted in two histograms placed on top of each other. The histograms in light grey shows all 10 min CO2 differences, while the histograms in dark grey only present those differences during windy conditions. The concept works well for background sites (LAEG, RIG, BRM) but has limitations for sites that are locally impacted by emissions (HAE is located next to a motorway). For all the site pairs, the differences in the CO2 measurements show a small bias (-2.1 to 0.8 ppm) and a scatter component (2.2–2.8 ppm at background sites, 6.0 ppm at the traffic site HAE). The root mean square error (RMSE) of the differences amounts to 2.3 to 3.6 ppm (background site) and 6.2 ppm (traffic site). The situation for HAE can be improved if the effect of local emissions is reduced and only measurements between 22:00 and 04:00 UTC (LT = UTC+01:00 or 02:00) and/or wind directions upward of the motorway are selected (Fig. S5). Obviously, that coincides with a reduction in the number of adjustment periods. For three sites in the city of Zurich that are located next to a busy road, sensor corrections are performed only during windy conditions at night-time in order to reduce the effect of local traffic emissions.

There are instances when the LP8 sensor drift cannot be corrected as frequently as required. This can be caused by an extended meteorological situation with low wind speeds or by sensor-related issues (e.g. unstable behaviour, simultaneous wind and high relative humidity). Consequently, the difference between the computed and the true CO2 molar fraction may increase over time. In addition, the outlier detection algorithm can be less effective during prolonged time periods with no dry conditions.

In order to identify such periods of suspicious or less accurate data, the measurements of individual sensors were checked for consistency with the more accurate measurements from HPP and Picarro instruments in a similar geographic setting. Although the true CO2 mole fractions at a given site are unknown, CO2 time series of sites within a particular region are expected to exhibit similarities, e.g. similar daily CO2 minima in the afternoon when the boundary layer is usually well mixed.

For this purpose, all the locations of the Carbosense network were divided into three groups based on their region and the surrounding topography.

All the sites in the Canton of Ticino (Fig. 1) are part of group one as only two HPPs operate in this region. The sites in the other regions of Switzerland are divided into two groups depending on whether they are located on a hilltop (group 3) or not (group 2). A hilltop location is defined by the following criteria: (i) the difference in altitude, i.e. the topography in a 2.5 km perimeter including the actual altitude of the mounted sensor, is larger than 300 m and (ii) more than 90 % of the topography in a 2.5 km perimeter is at a lower altitude than it is at the sensor location. The second criterion is omitted if the difference in altitude exceeds 400 m.

The CO2 molar fraction of the reference instruments and the HPPs are analysed group by group. The 10 % quantile of the preceding 24 h is computed for each instrument and HPP every sixth hour (CO2,Q10%). Afterwards, a band is derived (CO2,limits= median (CO2,Q10%) ±2.0⋅range (CO2,Q10%)) that indicates plausible daily minimum CO2 molar fractions. The preceding 24 h of measurements from a sensor are flagged in case the sensor's daily CO2 minimum is outside the computed band.

The employed sensor model that is based on the Beer–Lambert law and is extended by an empirical parametrization that can relate the sensor IR measurements and the ambient CO2 molar fraction in all relevant CO2, temperature and pressure conditions (Fig. 8a, b, d, e). The sensor's factory calibration is intended for using the sensor in a narrower temperature range, like that encountered indoors, and does not include pressure information. Measurements based on the factory calibration are not as accurate as they can be under outdoor conditions when using an extended model such as those described by Eqs. (4) and (5) (Fig. 8c). The 1 / IR terms in Eq. (4) require that the sensor's IR measurement does not heavily drift or jump because in this case the error cannot be compensated for by a simple offset. The data quality of sensors whose IR values significantly jumped during deployment (∼300 ppm in CO2) is usually better when the simplified sensor model (Eq. 5) without 1/IR-terms is applied (see Fig. 12). Equation (5) provides a less optimal fit under particular operating conditions (e.g. for high CO2 molar fractions; compare with Fig. 8a and b). However, this is of minor importance for most locations. The CO2 molar fraction at locations which are not impacted by nearby emissions is usually within 380–550 ppm. Temperature effects cause by far the largest deviation of the sensor response from the ideal gas law (Fig. 8c, d, f). Pressure effects are of a much smaller magnitude (∼0.1 ppm hPa-1).

As shown in Fig. 9, the RMSE of the LP8 CO2 measurements with respect to the Picarro during chamber and ambient calibration is between 6.8 and 12.5 ppm when applying Eq. (4) and between 8.0 and 13.9 ppm when applying Eq. (5) for the deployed sensors. Data filtering during calibration is chosen to be very selective in order to optimally determine the sensor model parameters.

The performances of the outlier detection and drift correction algorithms are presented together as both processing steps have to be applied to obtain accurate CO2 measurements. The results shown in this section refer to sensor measurements in the period 1 July 2017 to 1 September 2019.

Several sensor units are operated at sites equipped with a CO2 reference instrument (HAE - five sensor units; PAY – five sensor units; RIG – five sensor units; LAEG – two sensor units; BRM – one sensor unit) in order to test different calibration and processing options. Drift correction for the sensors at PAY relies on the CO2 measurements from GIMM and for the sensors at RIG, HAE, LAEG and BRM on the CO2 measurements from DUE (Fig. 7). Thus, the sensor and reference instrument measurements are independent at these sites.

A slightly modified data processing scheme was applied to the data from the 141 sensor units that were operated at DUE beyond 1 December 2017. This additional data treatment provides the opportunity to assess the data quality for a larger set of sensors. The calibration data set for these sensors contains all data before 1 December 2017 and is applied to the measurements thereafter. The sensor data are processed as described in Sect. 3, but drift is corrected by referring to measurements from sites LAEG and BRM instead of DUE (site LAEG, being located closer to DUE, is used when both instruments provide data). The accuracy of the CO2 molar fraction from these sensors located at DUE can therefore be compared to that from sensors deployed in the Carbosense network. Among the sensors at DUE there are also those with a performance that is not sufficient for deployment, and therefore they are held back at DUE.

The comparison of the median difference between CO2 measurements from the sensors and from the reference instruments reveals that sensor drift can be adjusted over the long term when the sensor measurements can regularly be referred to CO2 predictions (Fig. 10). The frequency of the required adjustments depends on the individual sensor as the change in sensor behaviour and the corresponding drift do not evolve constantly.

By means of the sensors which operate co-located with reference instruments the effect of different processing options can be assessed. This includes the employed sensor model (Eq. 4 or 5), the applied outlier detection (no outlier detection, outlier detection based on RHtrsh or the algorithm presented in Sect. 3.4) and the use of additional consistency checks. The sensor and reference measurements are compared for weekly periods by means of the root mean square error (RMSE) and the correlation (Fig. 11a and b). In addition, the fraction of valid measurements with respect to the total number of measurements in the database is indicated (Fig. 11c). It shows the effect of data filtering on the number of usable measurements. Scatter plots of the comparisons between the LP8 measurements and the measurements from the reference instruments at HAE, PAY, RIG, LAEG and BRM are shown in Figs. S6 to S9.

The median of the weekly RMSE of the sensor measurements with respect to the reference measurements at BRM, HAE, LAEG, PAY and RIG amounts to 10 ppm (25 % and 75 % quantiles – 6.8 and 14.3 ppm, respectively). The accuracy of the measurements is not constant over time but has a dependency on the effectiveness of the outlier detection and drift correction algorithms and thereby also on the prevailing weather conditions. The two described outlier detection algorithms differ in terms of the resulting RMSE values. Rigorous data filtering using RHtrsh leads to the best RMSE values. The outlier detection algorithm performs slightly worse in terms of the RMSE. Overall, it classifies a slightly larger number of measurements as valid than the filtering using RHtrsh. Differences in performance between the sensor models described by Eqs. (4) and (5) are small for this set of sensors. The accuracy of the measurements can be further improved when they are validated against measurements from high-precision instruments operated in the Carbosense network. This is shown for the combination of the outlier detection algorithm and the consistency check. Correlation between sensor and reference is about 0.9 on average. At sites RIG, LAEG and BRM the correlation coefficients are smaller due to smaller CO2 variations encountered at these locations (Fig. S1).

The extended sensor model described in Eq. (4) is applicable for a wider range of environmental conditions (CO2, T, P) than the reduced version (Eq. 5). However, when the IR signal shows large changes (> |300| ppm expressed in molar fraction) as in the case of sensor unit (SU) 1314 deployed at HAE, the application of the simplified sensor model provides more accurate results (Fig. 12).

Co-located sensor units are an additional option to assess the sensor performance. They reveal how similarly two sensors behave when they encounter comparable environmental conditions. There are 12 locations where two sensor units operate in parallel but where no reference instrument is available (Figs. 13 and S6). The horizontal distance between the sensor unit pairs does not exceed 45 m. There are no close emission sources for cases with distance ≠0. The indicated RMSE refers to the difference in simultaneous measurements. The sensor pairs exhibit fairly good correlations at most locations. For the sensor pairs operated at Hallau (HLL) and Birmensdorf (BSCR) there is better agreement when processing the measurements using the sensor model given by Eq. (5) instead of the model given by Eq. (4). The IR measurement of sensor unit 1012 changed significantly in January 2018; those of sensor unit 1120 changed significantly in March 2018. The difference between the processing models is small for the other sensor pairs.

Eight sensor units were deployed in the Carbosense network, and they were brought back to site DUE to review their performance due to sensor malfunctioning (e.g. LP8 sensor dropped out of the board) or suspicious CO2 measurements. For completeness, the comparison between the measurements from these sensors and from the Picarro instrument is shown in the Supplement.

The Carbosense network consists of 230 LP8 sensors as of 1 September 2019. In total, there were 262 deployments in the period 1 July 2017 to 1 September 2019. Over 75 % of the deployments lasted longer than 1 year, and five lasted less than 30 d.

The data transmission over Swisscom's Low Power Network (LPN) works reliably. The 25 %, 50 % and 75 % quantiles of the fraction of transmitted data for individual deployments at MeteoSwiss and NABEL locations and at locations within the city of Zurich amount to 88 %, 95 % and 98 %, respectively (Fig. 14a). Performance is even better at Swisscom's transmitter locations (25 % quantile – 98 %). However, these are usually equipped with an LPN gateway and built at elevated locations. We cannot assess to which part of the data transmission process the data loss is attributed (transmission module used in the sensor unit, LPN infrastructure, LPN network coverage). The transmission module (Microchip RN2483) of several sensor units was found to have a reduced reliability at high temperatures (above about 30 ∘C).

A small number (∼1 %) of the transmitted LP8 measurements had a nonzero status flag, for instance, when temperature was below -8.5 ∘C (LP8-specific threshold) or the sensor was malfunctioning. For a minor fraction of measurements a drift adjustment could not be performed as the sensor was assessed to be in an unstable phase. The outlier detection algorithm flags 23 % of the measurements that were drift corrected. In combination with the consistency check, 29 % of the measurements are flagged. There is considerable variability in these fractions related to the individual sensor performance and the location. A clear relationship is evident between the fraction of outliers and the humidity conditions encountered at the deployment location (Fig. 14b). Overall, the median of usable measurements from all individual deployments amounts to 67 %. There is a diurnal variation in the fraction of flagged measurements closely related to the diurnal variation in relative humidity (Fig. 15a). The outlier detection algorithm has the advantage of retaining a larger number of measurements in conditions of high relative humidity compared to the method using RHtrsh (Fig. 15b).

The conversion of wet CO2 to dry CO2 requires the water molar fraction χH2O. This value is computed for the sensor units based on the SHT21 T and RH measurements and the pressure that is interpolated for the specific location. The uncertainty in the estimation of χH2O and the corresponding uncertainty in the dry-air mole fraction of CO2 can be assessed for a total of 55 sensor units operated at MeteoSwiss SwissMetNet sites that are equipped with more accurate meteorological instruments. At those sites, χH2O has been computed from the sensor units and from reference T, RH and p measurements (Figs. S3 and S4). The agreement is best (±0.07 %) when global radiance is low (< 50 W m-2). In this case T and RH measured inside the box are representative for the outside conditions. Deviation is slightly worse (±0.15 %) for higher global radiance. For the majority of the measurements, the conversion of wet-CO2 to dry-CO2 molar fraction is associated with an error below 1.2 ppm (assessment of deviation for an error ϵ=0.2% and χCO2,wet=600 is as follows: χCO2,dry=χCO2,wet/(1-ϵ/100)=600ppm/(1-0.2/100))=601.2 ppm).