Multiscale observations of NH3 around Toronto, Canada

. Ammonia (NH 3 ) is a major source of nitrates in the atmosphere, and a major source of ﬁne particulate matter. As such, there have been increasing efforts to measure the atmospheric abundance of NH 3 and its spatial and temporal variability. In this study, long-term measurements of NH 3 derived from multiscale datasets are examined. These NH 3 datasets include 16 years of total column measurements using Fourier transform infrared (FTIR) spectroscopy, three years of surface in-situ measurements, and 10 years of total column measurements from the Infrared Atmospheric Sounding Interferometer (IASI). The 5 datasets were used to quantify NH 3 temporal variability over Toronto, Canada. The multiscale datasets were also compared to assess the observational footprint of the FTIR measurements. All three time series showed positive trends in NH 3 over Toronto: 3.34 ± 0.46 %/year from 2002 to 2018 in the FTIR columns, 8.88 ± 2.83 %/year from 2013 to 2017 in the surface in-situ data, and 8.38 ± 0.77 %/year from 2008 to 2018 in the IASI columns. To assess the observational footprint of the FTIR NH 3 columns, correlations between the datasets were exam- 10 ined. The best correlation between FTIR and IASI was obtained with coincidence criteria of ≤ 25 km and ≤ 20 minutes, with r = 0.73 and a slope of 1.14 ± 0.06. Additionally, FTIR column and in-situ measurements were standardized and correlated. 2 with an ensemble of surface observations over Europe, and r as high as 0.81 and 0.71 for measurements made at Fyodorovskoye, Russia and the Monte Bondone, Italy, respectively. It should be noted that the comparisons in Van Damme et al. (2015a) were done by converting IASI NH 3 columns to surface concentration by using the same model used in the retrieval process, as opposed to the standardized dataset approach used in this study. r 2 = 0.33. These results suggest that TAO, representative of NH 3 at a city-size scale ( ∼ 50 km), requires higher-resolution model runs for comparison. This is also evident when comparing GEOS-Chem against IASI within the single model grid cell that includes TAO; this comparison led to a poorer correlation with r 2 = 0.13. In addition, GEOS-Chem overestimated NH 3 in the larger domain when compared with IASI. However, in the single grid cell over TAO, the model underestimated NH 3 columns compared to both IASI and TAO. This study showed a positive trend of NH 3 over Toronto derived from ground-based FTIR, satellite, and in-situ measurements. The NH 3 total columns using an FTIR situated in downtown Toronto showed an observational footprint at a city-size scale, although this also highlights the need for models simulating NH 3 to be run at higher resolution than 2° × 2.5° for comparisons with ground-based measurements. the ﬁnal

and other regions of Canada (Griffin et al., 2013;Whaley et al., 2015;Lutsch et al., 2016Lutsch et al., , 2019a. As such, the time series of total column NH 3 measured at TAO exhibits long-term trends and pollution episodes. Toronto also has in-situ (surface) measurements of NH 3 (Hu et al., 2014), made by Environment and Climate Change Canada. A study by Hu et al. (2014) investigating NH 3 in downtown Toronto has shown that greenery within the city is an important source of NH 3 when temperatures are above freezing, and that potential sources at temperatures below freezing have yet to be investigated. Additionally, recent 60 studies have shown the increased capacity for satellite-based instruments to measure spatial and temporal distributions of NH 3 total columns at global Warner et al., 2016;Shephard et al., 2020), regional Warner et al., 2017;Viatte et al., 2020), and point-source scales Clarisse et al., 2019a;Dammers et al., 2019).
In this study, NH 3 variability over Toronto is investigated using ground-based FTIR data, in-situ measurements, and satellite-65 based observations from the Infrared Atmospheric Sounding Interferometer (IASI). This study is a part of the AmmonAQ project, which investigates the role of NH 3 in air quality in urban areas. Trends in the NH 3 time series and their statistical significance are determined, and the GEOS-Chem model is also used to supplement and compare against observations. Additionally, correlations between FTIR and IASI, FTIR and in-situ, in-situ and IASI, FTIR and model data, as well as IASI and model data on a regional scale are analyzed to assess the observational footprint of the FTIR NH 3 measurements. 70 The paper is organized as follows: Section 2 describes the FTIR retrieval methodology, the in-situ and satellite data, the GEOS-Chem model, and the analysis methodologies. Section 3 presents the trend analysis of the FTIR, in-situ, and IASI measurements, the results of the correlation studies, and the FTIR NH 3 observational footprint analysis. Section 4 presents the evaluation of the GEOS-Chem model, and conclusions are provided in Section 5.

FTIR Measurements
Ground-based NH 3 total columns used in this study were retrieved from infrared solar absorption spectra recorded using an ABB Bomem DA8 FTIR spectrometer situated at the University of Toronto Atmospheric Observatory in downtown Toronto, Ontario,Canada (43.66°N,79.40°W,174 masl). This instrument has been making measurements since mid-2002, and trace gas measurements are contributed to the Network for Detection of Atmospheric Composition Change (NDACC; http://www. 80 ndsc.ncep.noaa.gov/) (De Mazière et al., 2018). The DA8 has a maximum optical path difference of 250 cm, with a maximum resolution of 0.004 cm −1 , and is equipped with a KBr (700-4300 cm −1 ) beamsplitter. While the FTIR is equipped with both InSb and HgCdTe (MCT) detectors, NH 3 profiles were retrieved using the MCT detector, which is responsive from 500-5000 cm −1 . The DA8 is coupled to an active sun-tracker, which was manufactured by Aim Controls. The tracker is driven by two Shinano stepper motors on elevation and azimuth axes. The active tracking was provided by four photo-diodes from 2002-85 2014. This was upgraded to a camera and solar-disk-fitting system in 2014. Detailed specifications of the system can be found in Wiacek et al. (2007). Due to the nature of solar-pointing FTIR spectroscopy, the measurements are limited to sunny days, resulting in gaps in the time series. Measurements are typically made on 100-150 days per year.
The TAO FTIR uses six filters recommended by the NDACC Infrared Working Group (IRWG), and measures spectra through each filter in sequence. NH 3 profiles were retrieved using two microwindows of 930.32-931.32 cm −1 and 966.97-967.675 90 cm −1 . Interfering species include H 2 O, O 3 , CO 2 , N 2 O and HNO 3 . The solar absorption spectra recorded by the DA8 were processed using the SFIT4 retrieval algorithm (https://wiki.ucar.edu/display/sfit4/). SFIT4 uses the optimal estimation method (OEM) (Rodgers, 2000), and works by iteratively adjusting the target species volume mixing ratio (VMR) profile until the difference between the calculated spectrum and the measured spectrum, and the difference between the retrieved state vector and the a priori profile is minimized. The calculated spectra use spectroscopic parameters from HITRAN 2008 (Rothman et al.,95 2009), and atmospheric information (temperature and pressure profiles for any particular day) provided by the US National Centers for Environmental Prediction (NCEP). A priori VMR profiles were obtained from balloon-based measurements (Toon et al., 1999). The NH 3 retrieval methodology used at TAO is described in detail in Lutsch et al. (2016).
Uncertainties in the retrievals include measurement noise and forward model errors. Smoothing errors that arise due to the discretized vertical resolution were not included, to conform to NDACC standard practice. Measurement noise error includes 100 errors due to uncertainties in instrument line shape, interfering species, and wavelength shifts. Uncertainties in line intensity and line widths were calculated based on HITRAN 2008 errors. Error analysis was performed on all retrievals (following Rodgers, 2000); the resulting errors were grouped into random and systematic uncertainties, and added in quadrature. The resulting mean uncertainties averaged over the entire time series, were 12.9% and 11.8% for random and systematic errors, respectively, for a total average error of 18.8% on the NH 3 total columns. The mean degrees of freedom for signal (DOFS) 105 averaged over the 2002-2018 time series was 1.10.

In-Situ Measurements
To complement the FTIR total column NH 3 measurements, the publicly available in-situ data obtained by Environment and Climate Change Canada (ECCC) as a part of the National Air Pollution Surveillance Program (NAPS) were used (http: //maps-cartes.ec.gc.ca/rnspa-naps/data.aspx) (National Air Pollution Surveillance Program). The data span December 2013 to 110 April 2017, with a sampling frequency of one in three days. The sampling interval is 24 hours, from 00:00 to 24:00 local time, and samples were collected with a Met One SuperSASS-Plus Sequential Speciation Sampler. The detection limit is 0.6 ppb (Yao and Zhang, 2013). The integrated samples were brought back to the lab for analysis (Yao and Zhang, 2016). While errors are not reported in the dataset, the uncertainty is 10% when the NH 3 VMR is between 3 to 20 ppb (Hu et al., 2014). The instrument is situated less than 500 m away from the TAO FTIR, at 43.66°N, 79.40°W, 63 masl. with apodization. IASI can make off-nadir measurements up to 48.3°on either side of the track, leading to a swath of about 120 2×1100 km. At nadir, the field-of-view is 2 × 2 circular pixels, each at 12 km in diameter (Clerbaux et al., 2009). 4 https://doi.org/10.5194/amt-2020-319 Preprint. Discussion started: 31 August 2020 c Author(s) 2020. CC BY 4.0 License. The IASI NH 3 total columns (IASI ANNI-NH3-v3) are retrieved using an artificial neural network retrieval algorithm, with ERA5 meteorological reanalysis input data (Van Damme et al., 2017;Franco et al., 2018). Due to this retrieval scheme, there are no averaging kernels nor vertical sensitivity information for the retrieved columns . Details of the retrieval scheme and error analysis can be found in Whitburn et al. (2016) and Van Damme et al. (2017). IASI-A and IASI-B 125 NH 3 were combined and used in this study, as this allows for a longer time series and more data points for robust analysis. The retrieved columns of NH 3 from both satellites have been shown to be consistent with each other Viatte et al., 2020).

GEOS-Chem
The GEOS-Chem (v11-01) global chemical transport model (CTM) (geos-chem.org) was used in this study to supplement 130 and compare against observational data. The model was run at 2°× 2.5°resolution (latitude × longitude) using MERRA2 (Modern-Era Retrospective analysis for Research and Applications, Version 2) meteorological fields (Molod et al., 2015), the EDGAR emissions database (Janssens-Maenhout et al., 2019) for anthropogenic emissions, and NH 3 emissions (natural and anthropogenic) provided by Bouwman et al. (1997) and Croft et al. (2016). The GEOS-Chem model includes a detailed tropospheric oxidant chemistry, as well as aerosol simulation (e.g., H 2 SO 4 -HNO 3 -NH 3 simulation) (Park et al., 2004). NH 3 135 gas-aerosol partitioning is calculated using the ISORROPIA II model (Fountoukis and Nenes, 2007). Chemistry and transport are calculated with 20 and 10 minute timesteps, respectively. The model was spun up for one year, and output was saved every hour.

TAO and IASI Comparison
To assess the representative spatial and temporal scale of TAO NH 3 columns, the NH 3 total column measurements around 140 Toronto made by IASI were compared against TAO total columns and modeled NH 3 columns (see Section 3.3). As NH 3 shows high spatiotemporal variability, several definitions of coincident measurements were used in this study, with spatiotemporal criteria of varying strictness. As NH 3 concentrations can vary significantly during the day, the temporal coincidence criterion was chosen to be ≤ 90 minutes . In addition, values of ≤ 60, 45, 30 and 20 minutes were also tested.
For spatial coincidence criteria, ≤ 25 km , 30 km, 50 km, and 100 km were tested. For each criterion, 145 correlations (both r and slope) were calculated. This analysis was used to evaluate the spatial and temporal scales represented by the TAO NH 3 columns.

Trend Analysis and Identifying Pollution Events
With 16 years of data, relatively long-term trends of TAO FTIR column time series can be examined. While a trend analysis simply using monthly averages is possible (Angelbratt et al., 2011), a more sophisticated method of fitting Fourier series of 150 several orders was utilized in this study (Weatherhead et al., 1998). Bootstrap resampling was utilized to derive the confidence interval of the trends (Gardiner et al., 2008). A Q value (the number of bootstrap resampling ensemble members generated for 5 https://doi.org/10.5194/amt-2020-319 Preprint. Discussion started: 31 August 2020 c Author(s) 2020. CC BY 4.0 License. statistical analysis) of 5000 was used (Gardiner et al., 2008). An additional analysis to determine the number of years of measurements needed to give the derived trend statistical significance (2σ confidence) was also conducted, following Weatherhead et al. (1998). This analysis takes into account the need for longer time series to identify trends in data that are autocorrelated 155 (as are atmospheric observations). It should be noted that a major limitation of this analysis is that it assumes that data are collected at regular intervals, while TAO measurements are made at irregular intervals (due to the need for sunny conditions). For this reason, the confidence intervals derived from bootstrap resampling is a more robust method of error analysis, in the case of TAO data. However, as pointed out by Weatherhead et al. (1998), failing to take into account autocorrelation of the noise can lead to underestimations of actual uncertainty, and for this reason, both bootstrap resampling and the Weatherhead method 160 were used in this study. These techniques were combined to assess the intra-and inter-annual trends of NH 3 derived from TAO measurements, including a linear trend of the NH 3 total column along with its uncertainties and statistical significance. This analysis was also applied to the NAPS in-situ and IASI data.
The Fourier fit was used to identify NH 3 enhancements, following Zellweger et al. (2009). This analysis is done by taking the negative residuals of the fit (i.e., measured values smaller than fitted values), mirroring them, and calculating the standard 165 deviation (σ) of the mirrored residuals. Any measurements that are 2σ above the fit are considered enhancements. This analysis reduces biases in the spread due to enhancements by mirroring the negative residuals.
In this study, Fourier series of order 3 were utilized for all analyses. An analysis was done by comparing Fourier series fits of order 1 to 7, and checking for overfitting by running the residuals of the fit through a normality test (the Kolmogorov-Smirnov test). While overfitting was not observed at higher orders, higher orders did not give more statistically-significant trends, so 170 order 3 was chosen.

FTIR Measurements
The FTIR total column time series of NH 3 is shown in Figure 1. The purple points indicate enhancements, and the trends (with and without outliers) are shown as red and cyan lines, respectively. The trend from 2002 to 2018 was found to be 3.34 ± 175 0.46 %/year and 2.23 ± 0.79 %/year (2σ confidence interval from bootstrap resampling), with and without outliers, respectively (see Table 2). The number of years of measurements needed for the trend to be statistically (2σ) significant was found to be 33.8 years and 29.3 years, with and without enhancement events, respectively. Due to the irregular FTIR measurement intervals, these numbers may not represent the true significance of the trends, and should be regarded as best estimates of the significance of the observed trend. The lower magnitude of the upward trend in the analysis without enhancement values indicates that 180 the intra-annual variability of NH 3 is increasing. This is also evident when comparing the mean total column and standard deviations from, for example, the periods 2002-2005 and 2015-2018. In the former period, the mean NH 3 total column and standard deviation (1σ) were 5.94 ± 5.14 ×10 15 molecules/cm 2 , while in the latter time frame, they were 8.13 ± 7.88 ×10 15 molecules/cm 2 . The observed trend at TAO is comparable to a study by Warner et al. (2017), who observed an increasing NH 3 trend of 2.61 %/year over the United States from 2002 to 2016 using data from the Atmospheric Infrared Sounder (AIRS)  satellite-based instrument. Figure 2 shows the annual cycle of the FTIR NH 3 total columns, color coded by year, along with the monthly averages and ± 2σ. TAO NH 3 columns have a maximum in May with a monthly total column average of 13.14 ± 11.69 ×10 15 molecules/cm 2 , due to agricultural emissions increasing in spring/summer (Hu et al., 2014;Dammers et al., 2016). The TAO seasonal cycle is consistent with findings by Van Damme et al. (2015b), who observed maximum NH 3 columns over the central United States 190 during March-April-May (MAM). The mean NH 3 total column across the entire FTIR time series was 7.53 ± 7.10 ×10 15 molecules/cm 2 . These values are higher than remote areas, such as Eureka (located at 80.05°N, 86.42°W), where the highest monthly average was 0.279 ×10 15 molecules/cm 2 , in July . However, TAO NH 3 total columns are far below values observed by the FTIR in Bremen (located at 53.10°N, 8.85°E), which saw values in the range of ∼100 ×10 15 molecules/cm 2 . Monthly mean NH 3 columns are listed in Table 1.

NAPS Measurements
The NAPS in-situ NH 3 time series is shown in Figure 3. The purple points indicate enhancements, and trendlines with and without these outliers are shown as the red and cyan lines, respectively. The trendline was found to have a slope of 8.88 ± 2.83 %/year and 6.40 ± 0.18 %/year (2σ confidence interval from bootstrap resampling), with and without outliers, respectively.
where X is the dataset, indexed by i, µ is the mean, and σ is the standard deviation of the dataset. The standardized dataset is centered around zero, and normalized by the standard deviation of the measurements. As the standardized dataset is unitless, it allows for comparison between different measurements in different units. In this study, the TAO NH 3 total columns were used, 210 because the DOFS for the retrieval was around 1 (mean DOFS of the entire time series was 1.10), meaning there is only about one piece of vertical information in these measurements.  and NAPS surface NH3 VMR (blue) monthly averages lines with their respective ± 1σ (shading).

225
The time series of IASI NH 3 total columns (2008 to 2018) within 50 km of TAO is shown in Figure 6. The trend of these IASI measurements is 8.38 ± 0.77 %/year, where the error indicates the 2σ confidence interval obtained by bootstrap resampling analysis. The Weatherhead et al. (1998) method for finding the statistical significance of this trend was not utilized here, as the analysis requires calculating the autocorrelation of data, which is not possible given the spatially scattered dataset. For comparison, the TAO FTIR trend over the same period is 4.02 ± 0.74 %/year.

230
The correlations between IASI and TAO NH 3 columns for the various coincidence criteria listed in Section 2.3 are shown in 10 https://doi.org/10.5194/amt-2020-319 Preprint. Discussion started: 31 August 2020 c Author(s) 2020. CC BY 4.0 License.   showed a correlation of r = 0.79 and slope = 0.73. The same study also found that the FTIR in Paris is capable of providing information about NH 3 variability at a "regional" scale (∼ 120 km) Tournadre et al. (2020). Although not quantified in this study, the line-of-sight through the atmosphere (which changes throughout the day) may also affect the representative scale of ground-based solar-pointing FTIR observations. Additionally, the number of observations is relatively large for each criterion 250 (e.g., N = 923 for 90 minutes, 25 km, while N = 679 for 45 minutes, 50 km), suggesting that the differences in correlation are not simply due to the differences in the number of data points. The correlation plots for 20 min/25 km, 90 min/25 km, 20 min/50 km and 45 min/50 km are shown in Figures 7a, 7b, 7c and 7d, respectively. It should be noted that the slope was calculated through a simple linear regression. For comparison, an additional analysis was done propagating measurement uncertainty using the unified least squares procedure outlined by York et al. (2004) and yielded similar results, with a smaller slope for all 255 cases due to the larger relative uncertainty on IASI measurements (∼68 % for IASI compared to ∼19 % for TAO).
IASI column and NAPS surface NH 3 were also compared in this study by converting to standardized data (see Equation 1). Comparing

Comparison with GEOS-Chem
The NH 3 total column from the GEOS-Chem CTM model grid cell containing Toronto (grid center at 44°N, 80°W) is shown in Figure 8a, along with TAO FTIR data. The correlation was obtained by comparing the hourly model data for each FTIR observation. Comparison with the FTIR was done with and without smoothing the model data with the FTIR averaging kernel and a priori profile (Rodgers and Connor, 2003). As smoothing the model data only resulted in differences of less 270 than 1%, the discussion here will focus on the unsmoothed dataset to be consistent with the comparison with IASI. While GEOS-Chem is able to capture the seasonal cycle seen at TAO, the correlation is not strong, with r = 0.51 and the coefficient of determination, r 2 , at 0.26 (see Figure 9a). The calculated slope was 1.16. Both of these values are without smoothing the model data.
Smoothing the data resulted in r 2 = 0.28, and slope = 1.01 It is likely that the model is too coarse (the 2°× 2.5°g rid box corresponds to approximately 220 km × 200 km), and TAO, while able to capture larger-scale variability in NH 3 than 275 in-situ observations, is not sensitive to observations at spatial scales of 100 km or larger. Given the short lifetime of NH 3 , it is 13 https://doi.org/10.5194/amt-2020-319 Preprint. Discussion started: 31 August 2020 c Author(s) 2020. CC BY 4.0 License. unsurprising to see large variability at these spatial scales.
For comparison with IASI, a larger domain was chosen to assess the correlation of the model and satellite observations at a larger regional scale. Model grids spanning 35°N to 53°N, and 93.75°W to 63.75°W were used for the analysis, as these grids capture Toronto, the Great Lakes, and the Atlantic Ocean coastline. The spatial coincidence was calculated by binning 280 the IASI data into the grids of GEOS-Chem, and temporal coincidence was determined by calculating the mean overpass time in the domain and averaging the model data between one hour before and one hour after the mean overpass time. The time series (both GEOS-Chem and IASI were averaged over the domain) and correlation plots are shown in Figures 8b and 9b, respectively. Correlation of GEOS-Chem against IASI is higher than GEOS-Chem against TAO FTIR, with r 2 = 0.33. This is 14 https://doi.org/10.5194/amt-2020-319 Preprint. Discussion started: 31 August 2020 c Author(s) 2020. CC BY 4.0 License. States Great Plains and the Midwest during the summer. The slope was 0.85, meaning NH 3 is overestimated in GEOS-Chem when compared to IASI at this scale. Comparing GEOS-Chem and IASI for one grid cell over Toronto (same cell as the one used for comparison with TAO FTIR) resulted in a lower correlation, at r 2 = 0.13. These results suggest GEOS-Chem is able to model NH 3 on larger regional scales, but a finer resolution is needed for better comparison with smaller regions. In addition, while the modeled NH 3 was overestimated in comparison with IASI over a larger regional domain, the comparison for the

Conclusions
The TAO FTIR spectrometer situated in downtown Toronto, Ontario, Canada has been used to obtain a 16-year time series of total columns of NH 3 . These columns were compared against other NH 3 observations (IASI column and NAPS in-situ surface VMR) and GEOS-Chem model data. This study showed a positive trend of NH 3 over Toronto derived from ground-based FTIR, satellite, and in-situ measurements. The NH 3 total columns using an FTIR situated in downtown Toronto showed an observational footprint at a city-size scale, although this also highlights the need for models simulating NH 3 to be run at higher resolution than 2°× 2.5°for comparisons with ground-based measurements. 17 https://doi.org/10.5194/amt-2020-319 Preprint. Discussion started: 31 August 2020 c Author(s) 2020. CC BY 4.0 License.