Calibration and assessment of electrochemical low-cost sensors in remote alpine harsh environments

This work presents results from an original opensource low-cost sensor (LCS) system developed to measure tropospheric O3 in a remote high altitude alpine site. Our study was conducted at the Col Margherita Observatory (2543 m above sea level), in the Italian Eastern Alps. The sensor system mounts three commercial low-cost O3/NO2 sensors that have been calibrated before field deployment against a laboratory standard (Thermo Scientific; 49i-PS), calibrated against the standard reference photometer no. 15 calibration scale of the World Meteorological Organization (WMO). Intraand intercomparison between the sensors and a reference instrument (Thermo Scientific; 49c) have been conducted for 7 months from May to December 2018. The sensors required an individual calibration, both in laboratory and in the field. The sensor’s dependence on the environmental meteorological variables has been considered and discussed. We showed that it is possible to reduce the bias of one LCS by using the average coefficient values of another LCS working in tandem, suggesting a way forward for the development of remote field calibration techniques. We showed that it is possible reconstruct the environmental ozone concentration during the loss of reference instrument data in situations caused by power outages. The evaluation of the analytical performances of this sensing system provides a limit of detection (LOD) < 5 ppb (parts per billion), limit of quantification (LOQ) < 17 ppb, linear dynamic range (LDR) up to 250 ppb, intra-Pearson correlation coefficient (PCC) up to 0.96, inter-PCC > 0.8, bias > 3.5 ppb and ±8.5 at 95 % confidence. This first implementation of a LCS system in an alpine remote location demonstrated how to obtain valuable data from a low-cost instrument in a remote environment, opening new perspectives for the adoption of low-cost sensor networks in atmospheric sciences.


Introduction
The troposphere is a very complex system, which is subject to continuous inputs, production and removal processes of ozone 15 from natural phenomena and human activities (lifetime 25 days (Young et al., 2013)). In southern Europe the background tropospheric ozone concentration appears to be significantly affected by three main air mass transport processes: (i) transport of polluted air masses on regional and long-range scales; (ii) downward transport of stratospheric air masses; (iii) transport of mineral dust (Cristofanelli and Bonasoni, 2009). Large gaps remain in the surface observation network, despite many years of research and monitoring of surface ozone on regional and global scales. Especially in terms of areas without monitoring 20 and in terms of regions that have monitoring programs but no public access to the data archive (Schultz et al., 2017). Future improvements in the database would require better data harmonization, enhanced data sharing and monitoring in data-sparse regions to develop and integrate in-situ networks, complementary to satellite instruments, in order to improve measurement accuracy and spatial/temporal sampling (O'Neill et al., 2015). Therefore covering in-situ spatial data gaps to increase the effectiveness of satellite observations, which must be calibrated using ground-base reference measurements (Organization,25 2017), is necessary to achieve a better agreement between observations and models.
Tropospheric ozone was chosen for this pilot study due to its high relevance to the Earth's climate (Tørseth et al., 2012), ecosystems and human health. It is one of the most important atmospheric gases involved in photochemical reactions (Crutzen et al., 1999). Ozone is the precursor of oxidizing substances like OH − and NO − 3 and it is a key agent determining the oxidation capacity of the troposphere (Gauss et al., 2003). Tropospheric ozone influences climate as it plays a central role in the 30 radiative budget of the atmosphere (IPCC, 2013 p.55) and it is the third most important greenhouse gas in the free troposphere (Forster et al., 2007). Furthermore, surface ozone is a dangerous secondary pollutant causing harm to human health and ecosystems (Cooper et al., 2014;Jacobson and Jacobson, 2002).
Earth monitoring is a key aspect in improving our understanding of global processes and Climate. In this framework, remote areas, such as mountain regions, are considered reference background sites due to their sensitivity to Climate Change (Bonasoni 35 et al., 2008;ISAC-CNR;Cristofanelli et al., 2006;Barbante et al., 2004;Gabrieli and Barbante, 2014). Therefore data coming from high-altitude observatories provide valuable insights on the Earth's Climate and are crucial for the climate communities.
At present, the problem of establishing the spatio-temporal representativeness of measurements of ozone remains a difficult task, especially in presence of great spatial variability as in the case of remote regions. Increasing the number of referencegrade observatories devoted to long-term baseline observations in the alpine area is not practicable due to the high costs of 40 construction, maintenance and labour. Moreover, global atmospheric observatories have to be located in remote areas to reduce the influence of local source pollution, thus increasing the logistical costs and discomfort for the personnel. In this context, remote-sensing can not fully solve the spatial problem as satellite systems can only meet the established requirements if they are supported by correlative data of known quality, reliable ground-based observations and quantitative science (Dobber et al., 2006;ESA, 2020). 45 Emerging commercial Low-Cost Sensors (LCSs) present an unique opportunity to overcome the challenge of increasing the spatial density of monitoring sites. The rapid development and continuous improvement of low-cost technologies are demonstrating notable applications (Hertel et al., 2015;Hagan et al., 2018). Nowadays high-quality LCSs are beginning to play a role in areas such as modelling or emissions validation in support of state-of-the-art instrumentation and established networks (Mead et al., 2013;Win;Borrego et al., 2016Borrego et al., , 2018Heimann et al., 2015;Castell et al., 2015). While most Low-Cost 50 Sensor Network applications are designed for the built environment (Mueller et al., 2017;Jiang et al., 2005;Bauman et al., 2013;Andersen and Culler, 2014;Levis et al., 2005;Andersen et al., 2017) (e.g. smart cities, indoor air quality), there is a lack of studies in remote alpine regions where data are crucial for the Climate and meteorological research community. The World Meteorological Organization Global Atmosphere Watch (WMO-GAW) recognizes that the fate of the next generation of monitoring stations could be dramatically modified by the breakthroughs of new LCSs technologies (Lewis et al., 2018), but 55 there are still open issues to be addressed such as: (i) standardization of tools and protocols to ensure growth in the number of low-cost nodes without having to fundamentally change the LCSs network architecture, (ii) compatibility with established observing systems architecture (e.g. Zhang and Director (2010)), (iii) the simplification of the remote system management and (iv) the assurance of data quality.
In this work we aim to assess the reliability of LCSs for monitoring near-surface ozone concentrations in a remote Alpine 60 region, focusing on the precision, accuracy and reliability of LCSs measurements compared with a reference instrument. We carried out two laboratory calibration experiments, in April and July 2018, in order to evaluate the LCSs performances and their stability over time in a controlled environment before field deployment. We conducted our study from March to December 2018 at the Col Margherita Atmospheric Observatory (MRG). Due to the harsh weather conditions recorded at Col Margherita, this site was considered ideal to test the performances of the LCSs in view of the modern challenge of deploying low-cost 65 applications in real world difficult situations.

Site description
Col Margherita site is characterized by an alpine climate. Considering the 2008 -2018 time window, the annual mean temperature was 3.2 • C, with August as the warmest month with an average temperature of 11.8 • C and January as the colder month 70 with an average temperature of -4.7 • C. The average annual rain precipitation was 1485 mm, with August as the wettest month with 161.9 mm and December as the driest month with 59.5 mm (Trentino). Although not particularly high, the location is representative of the synoptic conditions of the free troposphere as represented on maps at around 700 hPa. This is possible since the site is distant and scarcely influenced, by surrounding orographic barriers (Figure 1). Despite the location could have been affected by pollutants emitted by major cities and transported by local winds (Masiol et al., 2015;Diémoz et al., 2019), recent 75 studies showed that the atmospheric composition of Col Margherita is related to air masses on a regional scale; in particular during winter the observatory is located above the atmospheric boundary layer so that local sources are not significant (Barbaro et al., 2020;Sprovieri et al., 2016).
About 20 thousand people live within 10 km of the MRG observatory area and about 70 thousand within 50 km (ISTAT, 2011;Veneto, 2020;gis, b, a).

MRG Observatory description
The MRG observatory is a GAW Regional Station (WIGOS Id: 0-380-0-MRG) led by the Institute of Polar Sciences of the National Research Council of Italy (ISP-CNR). It is located on the southern-eastern side of the Italian Alps (46.36683 N, 11.79192 E), based at an altitude of 2543 meters above mean sea level (m a.s.l.), it was chosen as representative for the surrounding alpine region (Barbaro et al., 2020).

85
The observatory is a prefabricated insulated shelter with external dimensions of 3.00m x 2.42m x 3.22m. The observatory is equipped with a complete automatic weather station mounted on a 3 m mast (AWS: CS215-L probe, Campbell Scientific, Utah, USA; PTB110 Barometer, Vaisala, Helsinki, Finland; 05103-45 Wind Monitor, R. M. Young Company, Michigan, USA) and an UV-absorption ozone analyser (Thermo Fisher Scientific 49C, SN: 0503110399). The MRG observatory is unmanned and fully automated, it is connected to the main electrical grid and supplied with a backup solar power system with ≈200 Ah 90 batteries in case of grid failure. The observatory is equipped with remote control via GSM / GPRS technology.

Low-cost sensors
We used the Alphasense OX-B431 commercial passive sensor which belongs to the class of electrochemical sensors that operate in amperometric mode. These oxidising sensors generate a current that is linearly proportional to the fractional volume of the target gases. The gas sensor provides three electrodes: the working electrode (WE) where the ozone reacts and a current 95 proportional to the ozone concentration is formed; the counter electrode (AUX) that balances the reaction of the working electrode to generate an equivalent current in the opposite direction and the reference electrode that anchors the working electrode potential to ensure that it is maintained at a fixed potential, independently from the current generated.
Each sensor was purchased with its individual sensor board that provides a dual channel voltage output, one for the working electrode and one for the auxiliary electrode. These output pins can be interfaced with a commercial Analog to Digital 100 Converter.
Alphasense provides a calibration certificate and reference regression coefficients for each sensor. The manufacturer declares that the calibration was conducted in a controlled environment (temperature 22 ± 3 • C, relative humidity 40 ± 15 %) using the ozone generator -Thermo Scientific Model 49i-PS (flow 0.5 l · min −1 ) and it was performed considering two points, zero and span (1 part per million (ppm)). The technical sheet (www.alphasense.com) reports that the sensor limit of detection is in 105 the range of units of parts per billion (ppb), a sensitivity that is required for detecting environmental ozone concentration. The manifacturer declares that the sensor voltage output is linear up to 20 ppm of the target gas.
Three equivalent sensors were installed on the sensing system in order to evaluate intra-compatibility between measurements and calibration stability as a function of time. Laboratory calibration was performed at the CNR-ISAC headquarters (Bologna, Italy) before field deployment. Field calibration was evaluated through inter-comparison with a reference UV-absorption O3

Sensor node
The sensing system ( Figure 2) was designed to be easily built. It consists of parts that may be purchased online and they are 115 easy to replace in case of failure. The sensing system was composed of an IP56 enclosure with three holes for the working low-cost sensors and two additional holes for waterproof power and Ethernet connectors. An additional LCS was placed inside the box as a spare sensor. It was not necessary to use it throughout the experiment. The enclosure was neither heated nor regulated nor insulated. Sensor's holes, located at the bottom side of the enclosure, were watertight with rubber O-rings seals.
A bag of silica gel was placed inside the box to keep the environment dry. 120 The sensing system was connected to the main power grid (AC) and to the LTE router of the MRG Observatory to allow data communication and remote control. The sensing system was equipped with an uninterruptible power supply (lead acid battery and battery charge regulator) in case of AC failure. The processing unit of the sensing system was a Raspberry Pi 3b+ and the analog signal of the sensors were digitized using a ± 5V (16 bit) ADC converter (Texas Instruments ADS1115). More detailed description of the hardware components of the sensing system and the approximate cost of the system's part are reported in 125 Supplementary Materials S1.

Sensors calibration
Before field installation, LCSs were calibrated by comparing the analog voltage output of the low-cost sensors with the ozone concentration generated by an ozone calibrator (thermo scientific 49i-PS) with traceability to the WCC-Empa SRP#15 (cal- ibration in 2017). Each sensor was installed on a specific gassing hood equipped with 1/4" barbed Swagelok fitting through which the calibration gas was fluxed using a digital mass flow meter. The calibration procedure was designed for the calibration of reference-grade instruments used in the WMO network. It consists of 21 steps in which the reference ozone (flow 0.9 l · min −1 ) span from 0 to 250 ppb with a declared precision of 1 ppb and zero noise of 0.25 ppb. Each step lasts for 20 minutes and the sensor total calibration time is 7 hr. The reference concentration step sequence simulates a pseudo-random variation of ozone concentration. This procedure allows to evaluate the sensor's tolerance, ability to respond to sudden and un- the 11 th and the 12 th of July under the same conditions of the first calibration. The sensing system was finally re-installed at the Col Margherita Observatory on the 17 th of July. During both calibration period, laboratory temperature and humidity were controlled and maintained at ≈22.5 • C and ≈50% respectively. During laboratory calibration the analog voltage output of the LCSs was recorded every second. Values were then aggregated to 1-minute averages and time-matched with the ozone concentration dataset generated (1-minute) by the reference instrument.

Field installation and data acquisition parameters
The installation of the sensing system was carried out the 25 th of May 2018 at the Col Margherita Observatory. The LCSs instrument was mounted in the AWS mast, 1 m above the ground. The gas inlet through which the reference UV-absorption ozone analyser sucks in the air (≈ 0.6 l · min −1 ) was located on the roof of the observatory at 5 m above the ground. The comparison between the reference instrument in Col Margherita and the LCSs of ozone was performed when both the systems 150 were running, considering the time window from May 2018 to December 2018. The analog voltage output of the LCSs was recorded every 5 seconds during the field experiment. The values were treated for the data processing as described in Sect. 2.7, then aggregated to hourly averages and finally time-matched with the reference ozone concentration dataset validated through manual and automatic data preprocessing (Naitza et al., 2020).

LCSs data processing 155
The LCSs raw data were preprocessed to discover possible outliers during the laboratory calibration and the field evaluation.
The filter used was based on the computation of a local polynomial (R function LOESS, R Core Team; Cleveland et al. (1992)) and the median absolute deviation (MAD) between this polynomial and the measurements within a moving window. We define the outliers as measurements deviating more than 5 times the MAD from the local polynomial (Mueller et al., 2017). If the MAD was smaller than the 50% quantile of all differences (| local polynomial -measurement |) it was substituted by this value. This 160 approach prevents the exclusion of measurements during time periods with almost no variation on the ozone concentration. We Team; Brockwell et al. (1991)) for the laboratory calibration LCS dataset. The resulting laboratory calibration outliers were less than the 0.1% of the total LCS measurements (see Supplementary Figure S2.1).
A time window of 1 hr was considered before averaging the data to hourly means for the field LCS dataset. This procedure 165 excluded less than 0.5% of 5-second raw data, mainly generated during the turning on phase of the LCS system (see Supplementary Figure S2.2). Minutes containing less than 9 valid observations (75 %) were excluded and hourly averages were considered if there were data for at least 45 minutes (75 %). Besides, 82 hours for the Mrg2 sensor were manually excluded from the dataset. From 1AM of the 5 th of November to 11PM of the 8 th of November, the WE electrode of the Mrg2 sensor showed a completely different behavior compared to the other two sensors (see Supplementary Figure S2.3).

170
Harsh weather condition caused many periods of absence of main AC current at MRG observatory. This made the reference ozone dataset discontinuous during the summer. On September 14 th , a problem on the pump of the reference instrument was diagnosed and the reference instrument was dismantled for maintenance until the 25 th of October. The 29 th of October we faced a power outage due to a severe storm ("Vaia", 29 th of October 2018, see Supplementary Figure  Evaluation of LCSs accuracy was considered skimming data if threshold value for relative humidity (RH) was overcome.
We considered a multivariate ANOVA model to evaluate the effect of the meteorological variables on the LCSs measurements.

180
Limit of Detection (LOD) was calculated as the average zero signal plus three times the standard deviation. Limit of Quantitation (LOQ) was calculated as the average zero signal plus ten times the standard deviation (Committee et al., 1987;for Standardization, 1994;Harris, 2010).

185
Laboratory calibration was performed through a linear model: where is the slope. Linear model agreement between the reference and the LCS was evaluated using the Pearson correlation coefficient (PCC). Evaluation of bias was performed using the Mean Absolute Error (MAE): where e i was the difference between the prediction (y i ) and the true value (x i , e i = y i − x i ) and n was the number of observations.
Summary of the calibration experiments and comparison with the calibration values declared by Alphasense are reported in Table 1. As an example of the laboratory experiment, the Mrg1 calibrated ozone measurements and analog output signal 195 registered against the reference ozone in July are shown in Figure 3 using the main and secondary y-axis respectively. The linear regressions of the laboratory calibrations are shown in Figure 4. We did not see a change in the analytical performances of the LCSs between the two calibration experiments (Two-sample t-test, p-value > 0.7).
The mean voltage response of Mrg1 was 18.1 ± 0.6 mV, when reference ozone concentration was 0.4 ± 0.1 ppb, and reached 100.2 ± 0.6 mV when reference ozone concentration was 249.75 ± 0.04 ppb. The precision of the Mrg1 sensor, calculated 200 as the relative standard deviation (RSD), was ≈ 3.2% close to the LOD, and decreased to ≈ 0.6% for ozone concentrations higher than 200 ppb (Thompson (1988); Horwitz and Albert (1997), see Supplementary Figure S4.1). MAE for Mrg1 was 3.6 ppb. LOD for Mrg1 was 4 ppb. LOQ for Mrg1 was 14 ppb. The instrumental response to ozone concentration was linear in the interval considered (R 2 = 0.998). Through our laboratory experiment we can confirm the linear dynamic range (LDR) for Mrg1 between the LOD (4 ppb) and 250 ppb. Alphasense datasheet reports that the instrumental response of the OX-B431 205 sensor was linear up to 2 ppm.
The mean voltage response of Mrg2 was 0.8 ± 0.7 mV, when reference ozone concentration was 0.4 ± 0.1 ppb, and reached 89.3 ± 0.9 mV when reference ozone concentration was 250.2 ± 0.3 ppb. Precision of the Mrg2 sensor was 3.5% for values close to the LOD and decreased to 0.9% for ozone concentrations higher than 200 ppb. Mrg2's MAE was 2.5 ppb, LOD was 5 ppb, LOQ was 17 ppb and LDR was 5-250 ppb (R 2 = 0.999).

210
The mean voltage response of Mrg3 was 38.3 ± 0.9 mV, when reference ozone concentration was 0.4 ± 0.1 ppb, and reached 123.5 ± 0.9 mV when reference ozone concentration was 249.7 ± 0.2 ppb. Precision of the Mrg3 sensor was 2.6% for values close to the LOD and decreased to 0.7% for ozone concentration higher than 200 ppb. MAE for Mrg3 was 2.6 ppb.
on the intercept and an average difference of about 21.6% on the slope.

Field experiment
Field measurements were conducted from the 30 th of May 2018 to the 14 th of December 2018. We measured about 4800 hourly data that were collected in a wide range of environmental temperature ( • C), pressure (hPa), wind speed (m · s −1 ) and RH (%) as summarized in Table 2.

Correlation and bias between LCSs. Intra-comparison
A first estimate to verify that the LCSs were in agreement with each other during field measuring was done calculating their (PCC) and bias, referred as MAE, throughout the period considered, and their variation trends over time. A summary of the inter-PCC and inter-MAE trends are reported in (Table 3). The comparison is consistent except for the month of October. Out of the 744 hourly observations collected during the month, only 59 were considered for the analysis. From the 25 th to the 28 th 225 of October, when both the LCS system and the reference instrument were working (see subsection 2.7). Since there were no signs of malfunctions and given the low environmental variability of the ozone concentration during the three days analysed (29 ppb < O 3 < 47 ppb calculated by the reference instrument and 25 ppb < O 3 < 46 ppb calculated averaging the LCSs measurements), we hypothesized that the low correlation value between LCSs could have been due to the inherent variability of the sensors' measurements. Indeed, if all the 744 LCSs hourly observations of October were considered, we could have 230 calculated: P CC 1,2 = 0.86,P CC 1,3 = 0.85, P CC 2,3 = 0.86, a result consistent with the other periods described in Table 3.

Correlation and bias between low-cost sensors and reference. Inter-comparison
We measured the correlation and the bias between the LCSs and the reference instrument to evaluate the performances of the 235 LCSs in a real case scenario where there could be no possibility to improve the laboratory calibration model. In addition to the calculation of the MAE we consider also the Root Mean Square Error (RMSE) and, to highlight the sign of the bias, the Mean Bias Error (MBE), defined as follows: In this contest the true value is the ozone measured by the reference instrument in Col Margherita. The statistical analysis, considering both the whole dataset and the trend of each sensor, is reported in Table 4. On average, the PCC between the sensors and the reference was ≈ 0.8, with the smallest values registered in December. The average MAE was ≈ 5 ppb and RMSE was ≈ 7 ppb. The bias was not constant through the period. It was larger during summer and it decreased during autumn, with 245 the MBE showing a change in the sign. Probable causes that might affect the accuracy of the LCS measurements could be the environmental temperature and relative humidity, whose dependence is also described in the sensor datasheet. We observed cases with poor agreement. Perhaps lower ozone concentrations and/or low environmental variation of ozone concentrations encountered in the mid winter periods influenced the data quality from the sensors or even the reference instrument. This could explain the low correlation observed during December. Also, as discussed in subsection 3.3, we hypothesized that the low 250 correlation value between LCSs and reference during October could have been due to the short amount of period where both systems were running. The case of June for Mrg2 and for all the LCSs during July were peculiar. Perhaps a role played by the temperature difference between the inside and outside of the box could explain the lower correlation.

Relative humidity threshold
The Alphasense Application Note 106 (AAN 106) reports that the low-cost sensors must operate in the RH range from 15% to 255 90%. We evaluated that the exclusion of the observations collected outside the RH interval does not improve the correlation and accuracy metrics between the LCSs and the reference instrument. Thus, considering the poor improvement obtained through skimming, we did not exclude further LCSs measurements from the dataset.  Table 4. Inter-correlation and bias between the LCSs calibrated in the laboratory and the reference. The sum of the hourly observations in the monthly analysis differ with the total 1793 observations by the data from the 31 th of May to the 14 th of June.

Environmental low-cost sensors model
We evaluated the relationship between the bias and the temperature, relative humidity, atmospheric pressure, solar radiation, 260 wind speed and wind direction to further investigate which meteorological variables were contributing to the bias (e i ) between the LCSs (O lcs 3 ) and the reference ozone concentration (O R 3 ). Figure 5 shows the correlation plots for temperature, relative humidity and LCSs signal while the remaining plots for the non correlating variables are reported in the Supplementary Figure   S5.5. The bias showed a correlation trend with the air temperature (PCC ≈ 0.53, p-value < 2e − 16) and relative humidity (PCC ≈ 0.45, p-value < 2e − 16), while no evidence of correlation was shown with the incident solar radiation (PCC ≈ 0.05, 265 p-value ≈ 0.1), the atmospheric pressure (PCC ≈ 0.24, p-value ≈ 0.3), the wind speed (PCC ≈ −0.22, p-value ≈ 0.1) or the wind direction (PCC ≈ −0.10, p-value ≈ 0.4). We finally observed that the bias was dependent on the sensor signal itself (PCC ≈ 0.55, p < 2e − 16).
We considered the following statistical multivariate linear model to evaluate the sensors bias under specific meteorological conditions and LCS signal, implementing in the model only the explanatory variables as previously described: 270 e j = a 0,j + a 1,j T + a 2,j RH + a 3,j O lcs,j 3 (5) where bias (e j , j=1,2,3) is the difference in the ozone concentration measurement between the j-esimo LCS and the reference, a i,j denote the model coefficients, T is the ambient air temperature, RH is the relative humidity and O lcs,j 3 are the j-esimo LCS ozone reading obtained from Equation 1.
concentration measured by the j-esimo low-cost sensors (O F,j 3 ):

Performance of the model
We considered three case scenarios to evaluate the performance of the model. A summary of the values obtained are reported in Table 5. The first scenario considered the whole reference dataset to model the bias and correct the LCSs measurements. The intercorrelation between the LCSs and the reference improved by ≈ 1%. The accuracy between the LCSs measurements and the reference improved by ≈ 60% lowering the average bias (MAE) from ≈ 5.5 to ≈ 3.2 ppb, with the 50% of the bias distribution between ≈ ±2.5 ppb and the 95% between ≈ ±8.5 ppb ( Figure 6). The corrected LCSs dataset, obtained by the continuous 285 calibration of the LCSs in Col Margherita, could be used to reconstruct the environmental ozone concentration in case of loss of reference data, a situation that may occur due to power outages or during the instrumental calibration when the reference instrument is not present at the observatory. During our experiment we obtained 1556 1-hr additional ozone measurements, considering the time periods where the LCSs and the AWS were collecting data (Figure 7 and Supplementary Materials S5).

Scenario 2 290
A second scenario still considered the whole reference dataset, but it aims to evaluate the intra-compatibility of the LCSs bias model. This might be useful if considering the use of the bias model of one low-cost sensor for the calibration of another low-cost sensor. This approach opens the possibility of performing a remote calibration in the surrounding area of the Col Margherita. A remote calibration allows the deployment of a local sensor network where remote standalone sensor's signal is corrected using the bias model studied in a location where a reference instrument is always present. We compared the 295 coefficients of the bias model of each low-cost sensor. Subsequently, we evaluated the accuracy between one of the LCSs and the reference, correcting its dataset using the averaged model coefficients of all the three LCSs (scenario 2a). Next, we evaluated the accuracy between a LCS and the reference, correcting its LCS dataset using the averaged value of the other two LCSs coefficients (scenario 2b). The average value of the intercept coefficient (a 0 ) is -27.76 ± 8.94%. The average value of the temperature coefficient (a 1 ) is 0.26 ± 8.11%. The average value of the RH coefficient (a 2 ) is 0.11 ± 5.41% and the average 300 values of the LCS coefficient (a 3 ) is 0.45 ± 11.32%.

Scenario 2a
We considered the average coefficient values and their relative standard deviations to model the bias of each sensor and to calculate the corrected LCS ozone dataset. In this scenario the inter-correlation between the LCSs and the reference improved by ≈ 1%. The accuracy between the LCSs measurements and the reference improved, lowering the average bias (MAE) to 305 ≈ 3.3 ppb. The 50% of the bias distribution confidence interval (CI) was ≈ ±2.6 ppb and the 95% CI was ≈ ±8.7 ppb.

Scenario 2b
We corrected the ozone measurements of each low-cost sensor using the average model coefficients of the other two low-cost sensors. The inter-correlation metrics between the LCSs and the reference improved by ≈ 1%. The accuracy between the LCSs measurements and the reference improved, lowering the average bias (MAE) to ≈ 3.5 ppb. The 50% of the bias distribution 310 CI was ≈ ±3 ppb and the 95% CI was ≈ ±9 ppb.

Scenario 3
The third scenario examined the execution of consequential field calibrations. It represents the situation where there is no chance to lay on local or remote calibration. The low-cost sensing system has to be installed alone, except for scheduled calibration periods during which a reference instrument is placed aside the low-cost sensor to model the bias. The calibrating 315 periods must be chosen to depict as much as possible all the seasonal meteorological conditions of the site.
We modeled the bias considering only the June and the December reference dataset. These periods represent the annual extremes of the meteorological conditions at Col Margherita (Table 2) In June the ambient temperature ranged from 2 • C to 14 • C with a mean temperature of 7 ± 2 • C, ambient relative humidity 320 (RH) ranged from 48% to 100% with a mean RH of 85% ± 10%, atmospheric pressure ranged from 747 hPa to 755 hPa with a mean pressure of 750 ± 2 hPa and wind speed ranged from 0.20 m/s to 12 m/s with a mean wind speed of 3.5 ± 2 m/s. The range of 1-hour validated O 3 concentration was 31.3 ppb with a maximum of 66.3 ppb measured by the Thermo 49c Ozone analyzer.
In December the ambient temperature ranged from -16 • C to 3 • C with a mean temperature of -6.5 ± 5 • C, RH ranged from 325 38% to 100% with a mean RH of 75% ± 15%, atmospheric pressure ranged from 734 hPa to 751 hPa with a mean pressure of 742 ± 5 hPa and wind speed ranged from 1 m/s to 23 m/s with a mean wind speed of 6 ± 4 m/s. The range of 1-hour validated O 3 concentration was 26.6 ppb with a maximum of 43.1 ppb measured by the Thermo 49c Ozone analyzer.
In this scenario the inter-correlation between the LCSs and the reference improved by ≈ 1%.  We summarized the LCSs analytical performance results obtained through two laboratory calibration experiments and a ≈ 7 months field experiment performed at the Col Margherita observatory.

Detection limits 335
The LCSs were capable of detecting ozone in the typical environmental ppb range. During the laboratory calibration experiments we observed that the LOD of the LCSs was about 5 ppb and the LOQ was about 15 ppb. We noticed that the LCSs response to ozone concentration was linear (P CC > 0.99) up to 250 ppb and these analytical performances of the LCSs did not change after two laboratory calibration experiments, conducted three months apart one from another. When used in the field, the linear regression coefficients obtained through laboratory calibration, at fixed temperature and humidity, are not capable to 340 fully describe the behavior of the LCSs, P CC ≈ 0.8.

LCSs calibration
We observed that all the sensors require an individual, both in laboratory and in the field, calibration and caution should be used if it is decided to rely on the declared calibration values, which could be insufficiently accurate for an environmental study, even if variation in ozone concentrations are clearly detectable. Careful should be used if deciding to perform a single field 345 calibration since this can be insufficient as it is very unlikely that, at mid-latitudes, one could choose a representative period of the whole year. A solution might be to perform many field calibrations, covering at least the extremes of the meteorological year condition, in order to depict as much as possible the behavior of the sensors in the environment. It may be considered to perform a more exhaustive laboratory calibration in a climatic chamber simulating the environmental conditions where the sensor will be deployed, to improve the understanding of the LCSs.

Precision
The evaluation of the LCSs intra-correlation during the field experiment revealed that the sensors behave in a similar but not identical way (intra − P CC ≈ 0.9), thus it is important to increase the reliability and reproducibility of the measurements considering LCSs redundancy. This is particularly important when using the sensors in the field without any reference. We detected indeed a case during which one of the three LCSs was not working properly (Supplementary Figure S2.3). These 355 anomalies could not be detectable if redundancy is not considered. The precision of the LCSs, described by the RSD, was evaluated in the laboratory with a constant reference ozone concentration. We observed that near the LOD the LCSs RSD is ≈ 3.5%, decreasing to ≈ 1.6% at ≈ 50 ppb and reach the asymptotic value of ≈ 0.8% for larger ozone concentrations (> 200 ppb). We did not notice a deterioration of the LCSs precision between the two laboratory calibration experiments. During the field campaign it is not trivial to evaluate the precision in terms of RSD, because there is neither a constant reference 360 ozone concentration nor a constant meteorological situation. Thus, since it is not possible to evaluate the precision in field, we assumed that the precision during the field experiments was compatible with the precision studied in the laboratory.

Bias
The LCSs average bias, measured through the MAE between the LCS and the reference instrument during the field experiment, was ≈ 3.5 ppb (≈ 6 ppb without performing the bias correction). The LCSs electrical noise measured in the laboratory was 365 ≈ 1.5 ppb (see Supplementary Figure S3.1) and bias > 4 ppb. During the field experiment we observed that the LCSs bias is dependent from the air temperature, RH and from the LCS electrodes voltage. It is possible to build a multivariate linear model able to describe about 60% the bias variance during the period in which the LCSs operates aside the reference instrument. It is of interest that the environmental model coefficients do not differ much one from another (≈ 8.5%). We noticed that it is possible to reduce the bias of one LCS using the average coefficient values of another LCS, suggesting the development of a 370 remote field calibration of the sensors in the surrounding area of the observatory.

Reliability
We have not measured differences in the analytical performances between the two LCSs calibration experiments. We showed that during the field experiment the LCSs response to the reference ozone concentration was affected by the local meteorology and their dependence can be analytically described. We observed that a time dependent model for correcting the bias was not 375 necessary for our field experiment, but would be reasonably needed for long-term monitoring, when LCSs aging will not be negligible. The rapid response (≈ 30 ppb · s −1 ) and the small memory effect, measured both during the laboratory experiments and in the field experiment, allowed the measurement of rapid variations of the ozone concentration.
The sensor's datasheet reports that NO 2 interferes with the ozone measurements. Since at the MRG observatory there was not a NO 2 analyzer during our field experiment, we were not able to measure the concentration of this gaseous pollutant. 380 We referred to a local study (Costantino, 2016) to deduce the amount of NO 2 that could have interfered with our experiment.
This work was based on the data of the regional environmental agency: a survey on ozone and nitrogen dioxide at Passo Valles (2.032 m a.s.l., about 3.2 km away from the MRG and about 500 m of lower altitude than MRG) was conducted, from the 1 st of January 2007 to the 1 st of January 2011. The average annual concentration of NO 2 at Passo Valles was 4 ± 1 ppb and the average annual concentration of ozone was ≈ 94 ± 5 ppb. As a comparison, despite referring to different years, the ozone 385 concentration at Col Margherita was ≈ 40 ± 10 ppb during 2018, ≈ 57% lower than what was recorded at Passo Valles. It is worth to notice that the monitoring site at Passo Valles is next to a road, which could be a local source of NO 2 pollution.
Therefore, even if it was not possible to know the NO 2 concentration at MRG, the mean values measured at Passo Valles suggest that a few ppb of NO 2 may have interfered with our experiment in MRG. Nevertheless, since it has been observed that the NO 2 concentrations are decreasing all over Europe (Jamali et al., 2020;Castellanos and Boersma, 2012), it is unlikely 390 that NO 2 interference could have been detected by our LCSs system in MRG and thus explain some of the bias we observed between the LCSs and the reference instrument.
We showed that the LCSs was subjected to many considerable "stresses" throughout the field experiment. During the summer period we faced some severe power outages which caused data losses. These events were due to heavy thunderstorms and bad weather conditions that are characteristic of the alpine summer season. Moreover, the LCSs faced a severe storm ("Vaia", 29 th October 2018, see Supplementary Figure S5.1) which caused several damages all through the North of Italy and caused a general black out in the Col Margherita area. It is worth noticing that the LCSs system was the only one that remained on during the Vaia storm blackout. When the power returned, the system showed no damage.
confidence) throughout the period of the field operation.
We remarked that all the sensors require an individual calibration. We observed that the laboratory calibration is not sufficient to explain the behavior of the sensors during the field experiment. As a consequence, performing a sensor field calibration near a reference site is necessary and this requires infrastructure. Since the quality of the sensor calibration depends on the description of the environmental conditions (i.e. pollutants concentrations, meteorology), we showed that the reference instrumentation is 405 relevant to perform periodic field calibrations.
In this work we discussed three procedures for the LCSs field calibration which can be used for the design of a lowcost sensors network in an alpine region. We showed how to improve the LCSs analytical performances when the reference instrument is always present, when it is present only for scheduled calibration periods and finally how to improve the O 3 measurements of a standalone LCSs using a remote calibration.

410
The comparison with a reference grade instrument revealed that the sensor's bias is impacted by changes in environmental temperature and relative humidity. These effects are demonstrated to be reduced by the application of a correction function and we revealed that a multivariate linear model is capable of describing up to 60% of the bias variability. We noticed that the bias model coefficients were comparable between each sensor (≈ 8.5 % of difference). This opens the chance to perform a remote calibration of low-cost sensors which are not in the proximity of the reference instrument. Future studies could focus on the 415 improvement of the mathematical description of the LCSs working principle and on its environmental dependence, in order to evaluate to which extent a single bias model could be used in a sensors network in an alpine area.
Achievable technical improvements for the enhancement of the analytical performances of the LCSs system are still open as the low-cost technology is getting better. Considering an improved thermal insulation of the sensing system, while ensuring the circulation of the ambient air without impacting the energy efficiency of the instrument, might be taken into consideration for 420 future works. This study reveals how to obtain valuable data from a low-cost instrument in a remote harsh alpine environment and shows procedures for the design of adequate monitoring strategies in the study of tropospheric gases in remote areas.
Foscari University, and Roberto Marin, from the Area Servizi Informatici e Telecomunicazioni of the Ca' Foscari University for their technical support. We thank Meteotrentino for providing the solar radiation and precipitation data from the weather station of the Passo Valles (Trento, Italy) (https://www.meteotrentino.it). Finally, a special thank to Renzo Minella, Loris Scola and the "Ski area San Pellegrino" people (https://www.ski area sanpellegrino.it) for their fundamental cooperation and support during the field activities at the Col Margherita Observatory.