Independent evaluation of point source fossil fuel CO₂ emissions to better than 10%

Significance

The 1,000 largest power plants comprise 22% of total global fossil fuel CO₂ emissions, making them an obvious target for regulating and reducing emissions. The success of existing and upcoming regulations and emission trading schemes requires reliable monitoring and verification of emissions, preferably using independent, objective evaluation to establish trust and transparency. However, such methodology has thus far been elusive, and emissions reporting currently relies solely on self-reported “bottom-up” inventory data. We demonstrate a method using time-integrated atmospheric observations and modeling to reliably quantify fossil fuel CO₂ emissions from point sources to within 10%. This level of uncertainty is a marked improvement over current ∼20% uncertainties for individual power plants and allows independent evaluation of reported emissions.

Abstract

Independent estimates of fossil fuel CO₂ (CO₂ff) emissions are key to ensuring that emission reductions and regulations are effective and provide needed transparency and trust. Point source emissions are a key target because a small number of power plants represent a large portion of total global emissions. Currently, emission rates are known only from self-reported data. Atmospheric observations have the potential to meet the need for independent evaluation, but useful results from this method have been elusive, due to challenges in distinguishing CO₂ff emissions from the large and varying CO₂ background and in relating atmospheric observations to emission flux rates with high accuracy. Here we use time-integrated observations of the radiocarbon content of CO₂ (¹⁴CO₂) to quantify the recently added CO₂ff mole fraction at surface sites surrounding a point source. We demonstrate that both fast-growing plant material (grass) and CO₂ collected by absorption into sodium hydroxide solution provide excellent time-integrated records of atmospheric ¹⁴CO₂. These time-integrated samples allow us to evaluate emissions over a period of days to weeks with only a modest number of measurements. Applying the same time integration in an atmospheric transport model eliminates the need to resolve highly variable short-term turbulence. Together these techniques allow us to independently evaluate point source CO₂ff emission rates from atmospheric observations with uncertainties of better than 10%. This uncertainty represents an improvement by a factor of 2 over current bottom-up inventory estimates and previous atmospheric observation estimates and allows reliable independent evaluation of emissions.

Fossil fuel carbon dioxide (CO₂ff) emissions are the main driver of the increasing atmospheric CO₂ mole fraction (1). Of the ∼10 GtC/y of CO₂ff now emitted globally, the largest 1,000 power plants emit 22% of the total (2). Thus, large power plants are an obvious target for regulating and reducing CO₂ff emissions. They are already subject to regulation or emission trading schemes in some regions (3, 4), and the Paris Agreement requires transparency in emission reporting (5). The success of such regulations requires the ability to reliably monitor and verify emissions, which is currently achieved through “bottom-up” inventory data, in which CO₂ff emission estimates are based on self-reported fuel use and carbon content statistics (6). Uncertainties in this method are on the order of 20% for individual power plants (7). Some studies suggest significant errors in emission reporting may already be occurring (8, 9) and under a regulatory environment, there may be incentive for deliberate misreporting. Independent, objective evaluation of emissions is needed to establish trust and ensure that self-reported bottom-up emission rates are unbiased (5, 10, 11). As power plants represent such a large proportion of total emissions, reducing the current uncertainties for point sources would have a strong impact in overall CO₂ff emission uncertainties (11, 12).

“Top-down” atmospheric observations have long been proposed as a complementary method to independently evaluate CO₂ff emissions (11, 13, 14). Constraining emissions to within 10% or better using top-down methods would substantially improve on the existing uncertainties in many self-reported bottom-up emission estimates and be sufficient for meaningful evaluation of those estimates (11).

The large and varying CO₂ background makes it difficult to use CO₂ measurements alone to constrain CO₂ff emissions (11, 13). However, the ¹⁴CO₂ content of CO₂ can unambiguously quantify CO₂ff, because all ¹⁴C has long since decayed away from fossil fuels, whereas all other sources contain ¹⁴C levels close to that of the current atmosphere (15). The concept is well established, but only a few studies have actually evaluated CO₂ff emission flux rates with this method. Uncertainties have been 30–100% (16–18), similar to uncertainties in (unbiased) bottom-up statistical data at local and regional scales (7, 19). Further, existing technology limits ¹⁴CO₂ measurement to laboratory-based analysis of individual samples at low sampling resolution. Alternative surrogate tracers such as carbon monoxide that are coemitted or colocated with CO₂ff emissions have been successfully used in top-down evaluations of regional and urban CO₂ff emissions (16, 17, 20–22), but are unlikely to be useful for power plants that, as a consequence of regulation, have very low pollutant gas emissions.

A major source of uncertainty in the top-down method is the atmospheric transport models that translate the observed atmospheric mole fractions to the emission flux rate by describing the movement of air (14). Previous work has shown that these models cannot adequately simulate the short-term turbulent atmospheric variability needed to interpret individual “grab samples,” flasks of air filled over a period of a few minutes (23).

In this study, we use time-integrated ¹⁴CO₂ sampling and atmospheric transport modeling to resolve the twin challenges of measurement technology and model inadequacy. Time-integrated sampling is used to aggregate CO₂ from many hours or days into a single ¹⁴CO₂ measurement. The model is averaged over the same time period, negating the need to resolve short-term turbulent atmospheric variability. Our simple field collection methods allow ¹⁴CO₂ measurements from multiple locations, further reducing potential model biases. We will show that with a modest number of samples collected using these techniques, emissions can be evaluated to within 10%.

Site and Instrumentation

We demonstrate the method and quantify uncertainties using a case study at a small isolated CO₂ff point source in rural Taranaki, New Zealand (Fig. 1). The source has a CO₂ff emission flux of ∼3,000 gC/s (0.1 TgC/y), which is small compared with the world’s largest power plants [e.g., Taichung, the world’s largest power plant, emits ∼300,000 gC/s (2)], but has several advantages as a test case. The CO₂ff source is isolated from most other sources, with the exception of one smaller source located 700 m to the east, which we include in our model analysis; sparse local traffic and farmhouse heating contribute only very minor sources in this region. Average daily emission rates are available from stack monitoring, allowing us to quantify uncertainties in the method. The surrounding terrain is relatively flat dairy pastureland, making atmospheric transport modeling straightforward and also providing easy access for sample collection. The pastureland is, however, highly productive, resulting in large and diurnally varying biospheric CO₂ fluxes (23).

Fig. 1.

Image showing the Ballance (yellow circle) and Kapuni (red circle) sources. Right part of image has been digitally altered to match brightness. Scale is the same as Fig. 2. Image courtesy of Google Earth copyright 2015 Digital Globe.

Time integrated ¹⁴CO₂ samples from two different methods are used. First, we simply collect live grass which faithfully preserves the ¹⁴C content of photosynthesized CO₂. The exact time period over which CO₂ is assimilated into the grass depends on weather, growing period, and plant growth stage (24). In these managed pastures, the grass is grazed once every 18–21 d, and we assume that each sample represents the mean ¹⁴CO₂ during daylight hours over the 10 d preceding collection, for a total of ∼120 h (23). We tested a second technique that absorbs CO₂ from air into sodium hydroxide (NaOH) solution as an alternative for other sites where fast growing plant material may not be available. NaOH samples were collected over a known 15-h period consisting of 3 h in the midafternoon on each of 5 consecutive days. Grass samples were collected over seven sampling periods from 2012 to 2014 and NaOH samples during the last four time periods in October/November 2014. Samples were collected from multiple locations around the emission source, varying for each sampling time period (Fig. 2). For each sample, the average recently added CO₂ff mole fraction [in parts per million (ppm)] was calculated from the measured ¹⁴C content.

Fig. 2.

Emission sources (K, Kapuni plant; B, Ballance plant) and sampling locations (+) for all sampling times and methods, superimposed on CO₂ff contours created from an aggregate of all samples over all time periods. The impact of the dominant westerly winds is clearly seen, resulting in the highest CO₂ff to the east of the Kapuni plant. The high contours in the northeast are likely spurious due to edge effects in the aggregate contouring. (Inset) Wind rose aggregated over all sampling time periods, with wind speeds indicated by color from 0 (navy blue) to 12 m/s (orange).

We forced a small-scale plume dispersion model with locally measured 10-min meteorological data for each sampling time period (all daylight hours for the preceding 10 d for grass samples, ∼120 h, and the 15-h sampling time for the NaOH samples) and the reported emissions from both sources (Fig. 1). The 10-min model output CO₂ff was sampled at each receptor location and averaged over the full sampling period (25). At this small spatial scale, emissions are rapidly dispersed and do not re-enter the domain, and emissions from other sources are not considered because they are included in the locally determined background.

Results and Discussion

The observed CO₂ff mole fractions ranged from 0 to 20 ppm (Fig. S1 and Dataset S1) with uncertainties of ∼0.7 ppm. The prevailing westerly wind results in highest CO₂ff values east of the emission source, reducing as the plume disperses downwind. Different spatial patterns were observed for each time period, with some time periods strongly dominated by the prevailing westerly winds and others exhibiting more mixed wind patterns (Fig. S2).

Fig. S1.

Plots for each sampling date showing CO₂ff values at each location from Δ¹⁴C observations (black) and Windtrax modeling with reported emission rates (red). For grass samples, the date of collection is indicated; for NaOH samples, the dates of midafternoon sampling are shown.

Fig. S2.

Wind roses for each sampling period, generated using local meteorological observations as described in Supporting Information. Wind speeds (m/s) are indicated by the inset color bar.

To evaluate the reported emissions, we determine a scaling factor that relates the observed CO₂ff to the simulated CO₂ff. The scaling factor is then used to adjust the reported bottom-up emission rate to determine the estimated top-down emission rate (26). Because the measurement and model are both imperfect, individual samples give a wide range of scaling factors, particularly when CO₂ff is small and measurement uncertainties dominate. Thus, we determine the scaling factor from the slope (and uncertainty) of the correlation between simulated and observed CO₂ff (Fig. 3 and Methods). We also evaluate the median model:observed ratio (20), giving a second indication of the scaling factor.

Fig. 3.

Observed vs. modeled CO₂ff for each integrated sampling period and method. One-to-one line is shown in black and the actual best-fit slope is in gray. Error bars are omitted for clarity and 89 data points are included.

For the entire dataset of 89 samples collected by the two different methods for seven different sampling periods over 3 y, we find a scaling factor (based on the slope method) of 1.06 ± 0.08. That is, the top-down method determines that the reported emissions are correct, with an uncertainty of 8%. The median method gives a scaling factor of 0.95, consistent with the slope method. This result is robust across various subsets of the data (Fig. 4 and Table 1). The grass samples alone (n = 68) give a scaling factor even closer to unity and with slightly more certainty (1.02 ± 0.07), whereas the NaOH samples have more scatter (1.08 ± 0.13). The scaling factor is more variable when calculated for individual time periods, which is unsurprising because there are only 4–14 observations for each individual date. A Student t test shows that despite the scatter from these small subsets, all scaling factors are statistically indistinguishable from unity. For subsets of 25 or more samples (i.e., averaging across multiple sampling dates), uncertainty in the scaling factor derived from the slope method is better than 10% and as good as 7%.

Fig. 4.

Scaling factors determined for subsets of observations, determined by either the slope method (filled symbols) or median method (open symbols). The four leftmost datasets each include multiple sampling dates and more than 20 observations; the remaining points are individual datasets with 4–14 observations each. The 1-σ error bars are shown for the slope method. The interquartile range tends to be large, so for clarity, no uncertainty is given for the median method.

Table 1.

Calculated slopes and estimated emissions for aggregated datasets

Sample subset	No. of samples	Median scaling factor	Slope observation vs. model	χ2ϖ
All grass and NaOH	89	0.95	1.06 ± 0.08	9
NaOH 2014	21	1.22	1.08 ± 0.13	27
All grass 2012–2014	68	0.92	1.02 ± 0.07	3

In our experiment, the observation-model agreement tends to be better for the grass samples than for the NaOH samples (Fig. 3 and Table 1), most likely because the grass samples have an ∼10 times longer time-integration period of ∼120 h per sample. At any given site, CO₂ff varies dramatically through time as the plume wanders with atmospheric flow (Fig. 5 and Supporting Information). Thus, a longer averaging time reduces biases due to short-term atmospheric variability in the model.

Fig. 5.

Example of modeled variability in CO₂ff over the 720 10-min time steps for NZA-55832 collected in December 2013.

The observed CO₂ff values are strongly dependent on the ¹⁴CO₂ background selected. Our ¹⁴CO₂ background is the long-term monthly mean ¹⁴CO₂ observed at Baring Head (41.41°S, 174.87°E) (27), 225 km south of Kapuni. Baring Head is expected to be a good background choice for our location because the upwind region is the well-mixed Southern Ocean for both sites. To confirm this, we compared ¹⁴C in annual tree rings from an oak tree 2 km upwind of the Kapuni plant with the growing season (September–April) average ¹⁴CO₂ content at Baring Head and found no significant difference between the two records (mean offset, 0.2 ± 1.2 ‰ for five samples during the 2005–2012 period) (28). For each sampling time period, we also measured a single background grass sample collected ∼20 km from Kapuni. However, the uncertainty in these single measurements translates to 0.5 ppm (1-σ) in CO₂ff, introducing noise into the overall dataset. With this alternative choice of background, the scaling factor for the overall dataset was slightly but not significantly higher at 1.13 ± 0.10. Thus, background measurements should be carefully considered and preferably more than one background sample collected for each sampled period.

From the scatter of individual observed and modeled CO₂ff values (Fig. 3) and the reduced χ² statistics (χ²_ν) (Table 1), it is clear that the measurement uncertainty and stochastic model uncertainty we included do not entirely explain the scatter of the results. The additional scatter is presumably due to errors in the meteorological data, model transport errors, and sampling biases. However, as we have shown, these additional errors average out with sufficient observations. Multiple sampling sites at varied locations around the emission source likely aid in the good agreement, because biases in the modeled transport at individual sites from, for example, uneven terrain or vegetation, will be reduced. Samples from different time periods experience different meteorological conditions, some of which may be better represented by the model than others, also improving agreement. Unsurprisingly, a sufficient number of samples are needed to obtain statistically meaningful results, and our results show clearly that at least 20 samples are needed to obtain repeatable results with uncertainty in the emission rate of <10%.

Sites very close to the emission source tend to be more difficult to model well (high CO₂ff points in Fig. 3), because short-term and small-scale turbulent flow are larger effects and the ability of the model to correctly simulate vertical motion is more important. Conversely, observations too far from the source will sample only a diluted plume so that measurement uncertainties dominate.

The large point sources for which these top-down measurements would be of most interest are one or two orders of magnitude larger than the source used in our study site (2). These larger emissions imply easier CO₂ff quantification, because signals will be large, although the emission plume may be more dilute by the time it reaches the surface from taller emission stacks (often 100–200 m high). The particular atmospheric transport model that we used (WindTrax) is valid only for small scales of up to a 1,000-m range, yet many other models with physics appropriate for larger scales are available, from larger-scale plume dispersion models to meso-scale models (29, 30), and the principle of averaging across long time periods is equally valid for those models. A more significant challenge is that only a few large point sources are isolated from other CO₂ff emissions, although the plume may be positively identified by other methods (12).

Our results indicate that emissions from a single point source can be determined to within 8% with a modest number of time-integrated ¹⁴CO₂ samples. The method would be further improved if the horizontal winds were always measured during the experimental time period and within the sampling region, and if the stability conditions assumed in the model were better constrained by vertical wind speed measurements. We expect that our results will scale to larger emission sources and thus that this top-down time-integrated ¹⁴CO₂ method has wide applicability to point sources around the world, providing a relatively inexpensive, robust method for independently evaluating point source emissions with quantifiable uncertainties.

Methods

Our study site is a small isolated CO₂ff point source in rural Taranaki, New Zealand (Fig. 1). The Kapuni Vector processing plant scrubs CO₂ from locally produced natural gas (with ∼40% CO₂ content) and vents that CO₂ff to the atmosphere from two 35-m-high stacks. Average daily emission rates from stack monitoring vary from 3,095 to 3,789 gC/s over our study period; hour-to-hour variability in emissions is less than 5% (Table S1). A smaller nearby source, the Ballance Agri-nutrients plant 700 m to the west, produces ∼1,000 gC/s; only annual average emission rates are available, but this source contributes only ∼10% of the CO₂ff at our observational sites, which are for the most part closer to the Kapuni source (Table S1). The surrounding terrain is relatively flat dairy pastureland with very small additional CO₂ff sources from farmhouse heating and sparse traffic, although there are large and varying biospheric CO₂ fluxes (23). A stream and trees limit sampling south of the emission source.

Table S1.

Reported emission rates used in the model simulations

Sample subset	Start date	End date	Vector Kapuni emission rate, gC/s (SD)	Ballance emission rate, gC/s
NaOH	10/11/14	10/15/14	3,735 (596)	1,476
	10/18/14	10/22/14	3,286 (346)	1,476
	10/25/14	10/29/14	3,736 (127)	1,476
	11/1/14	11/5/14	3,789 (106)	1,476
Grass	10/10/14	10/16/14	3,648 (615)	1,476
	10/17/14	10/23/14	3,326 (392)	1,476
	10/24/14	10/30/14	3,762 (124)	1,476
	10/31/14	11/6/14	3,753 (106)	1,476
	11/30/13	12/6/13	3,449 (62)	1,476
	10/18/12	10/24/12	3,095 (278)	884
	8/8/12	8/14/12	3,561 (130)	884

The Vector Kapuni operator provided daily average emission rates for each day we sampled and we report the average rate for each sampling period along with the SD across the days averaged. Ballance provided only annual averages. All emission rates were converted to gC/s.

Single blades of grass were pretreated with acid, combusted to CO₂ gas, and reduced to graphite, and the ¹⁴C content was measured by accelerator mass spectrometry (AMS) (31, 32). NaOH samples were collected by pumping air from an inlet at 2.5 m above ground level through 50 mL 1 M NaOH to absorb CO₂. The pump was switched on from 1:00 PM to 4:00 PM local time each day for 5 consecutive days to collect a total of 15 h of CO₂. Poppet valves up- and downstream of the NaOH solution automatically closed when the pump was off, ensuring that CO₂ was only absorbed during the pumping period. NaOH solutions were prepared by dissolving NaOH pellets in distilled water. Any CO₂ initially present was precipitated with barium chloride, and each 40-mL aliquot of the remaining nominally CO₂-free solution was decanted into 100-mL Nalgene bottles under a stream of nitrogen (33). After sampling, CO₂ was evolved in the laboratory by acidification with phosphoric acid, reduced to graphite, and measured by AMS.

Side-by-side tests were performed at Baring Head, Wellington, comparing pumped NaOH samples with routinely performed passive collection of CO₂ into open bottles of NaOH. Two independent pairs showed no significant differences in Δ¹⁴C. δ¹³C measurements show depletion in ¹³C relative to ambient air by 2–4‰, indicating that CO₂ is not quantitatively stripped from air passing through the NaOH solution, but this fractionation is accounted for in the ¹⁴C determination (32). Contamination of the NaOH solution due to incomplete removal of CO₂ during solution preparation and absorption of CO₂ during handling contributes a larger ¹⁴C background than for other sample types. Contaminating CO₂ was determined to have Δ¹⁴C of –40‰ and is 2% of the total CO₂ collected, based on test solutions from the same NaOH batch; this large blank changes Δ¹⁴C by <0.1 ‰.

During the October/November 2014 sampling campaign, NaOH and grass samples were colocated. Four NaOH samplers experienced leakage problems, and data were discarded.

CO₂ff was determined from the Δ¹⁴C values (23) such that

$$\text{C}{\text{O}}_2\text{ff}=\frac{\text{C}{\text{O}}_2\text{bg}\left({\mathrm{\Delta }}_{obs}-{\mathrm{\Delta }}_{bg}\right)}{{\mathrm{\Delta }}_{ff}-{\mathrm{\Delta }}_{obs}}.$$ [1]

Δ_ff is −1,000‰, indicating zero ¹⁴C content. Δ_obs is our observed Δ¹⁴C value for each time-integrated sample, and Δ_bg is the background Δ¹⁴C value. We assume that at this site, Δ_bg incorporates all other CO₂ sources, so no other corrections need be applied (23). Δ_bg was taken from the monthly mean at Baring Head, Wellington, except for the December 2013 samples, for which the Baring Head Δ¹⁴C was 5 ‰ lower than several of the measured background samples. We hypothesize that strong heterotrophic respiration during this midsummer period resulted in higher background Δ¹⁴C for the grass samples. For this sample set, we used the highest Δ¹⁴C in the sample set; using the Baring Head value did not significantly change the overall result. Uncertainty in the background value chosen for CO₂ itself (CO₂bg) has little impact on CO₂ff, and is taken from Baring Head, Wellington, monthly means.

Simulations were performed with the WindTrax Lagrangian plume dispersion model (34) using the forward mode in which emissions are assumed known and concentrations (or mole fractions) at specified locations around the source are to be determined. WindTrax simulates the transport of trace gases by releasing a set number of particles from the source at each time step, and individual particle trajectories are computed based on model parameters and input meteorology. Valid in the atmospheric surface layer and horizontal distances up to ∼1 km from the source, the model is based on Monin–Obukhov similarity theory and assumes that meteorological observations are averaged over a time interval representing a stable, mean atmospheric state. A few of our samples (Fig. 2) are just beyond the nominal 1-km model limit at ∼1.2 km from the source and thus have low CO₂ff mole fractions in both observation and model. Measurement and transport errors are proportionally larger in these samples, but omitting them does not change the calculated scaling factors and therefore we retain them in the dataset.

For the 2012 samples, a meteorological station was installed 1,000 m to the west of the Kapuni source from mid-August to November 2012, providing 10-min averaged horizontal wind speed and wind direction data (23). No local meteorological data were collected for the August 2012 sampling period, and instead we used meteorological data from the following week. We previously demonstrated that meteorological observations from Hawera, 20 km to the south, showed consistent conditions for the sampling week and the following week for which we have meteorological data (23). In 2013 and 2014, a different meteorological station operated by STOS (Shell Todd Oil Services, Taranaki) located 300 m directly north of the Kapuni source provided 10-min averaged data. We corrected the STOS wind speeds to account for a bias in the raw data identified by comparison with two other meteorological datasets (Fig. S3).

Fig. S3.

Comparison of wind speed and wind direction of the STOS and Smith farm anemometer data during October 2014 (Top), located 200 m northeast and 400 m northwest of the Kapuni source, respectively. (Middle) Comparison of the NIWA anemometer installed 1,000 m east of the Kapuni source in August–October 2012 with the New Zealand MetService Hawera (CliFlo, NIWA's National Climate Database on the Web: cliflo.niwa.co.nz/) observations for the same time period. (Bottom) Comparison of the raw STOS wind speed and wind direction with Hawera data for all of the 2013–2014 sampling periods. The Hawera site is located 20 km to the southeast of the Kapuni source.

In the absence of detailed vertical wind information, we assume moderately unstable atmospheric conditions for all sampling time periods. The choice of stability condition is an important constraint and source of uncertainty in modeled CO₂ff. Less stable conditions would lead to lower modeled CO₂ff at our sites, as increased vertical mixing dilutes the plume and vice versa. When an alternative stability condition is applied to all samples, the result is a different scaling factor but no change in the correlation between the modeled and observed CO₂ff. For the December 2013 (summer) dataset, the warmer temperatures and higher solar radiation suggest a less stable atmosphere. Thus, we also simulated this dataset with very unstable conditions, which marginally increased the overall scaling factor from 1.06 to 1.07. The expected CO₂ff in each sample was determined by averaging the modeled CO₂ff at each sampling location over all 10-min time steps simulated during the sampling period designated for that sample. For NaOH samples, this is the simple mean of all simulated 10-min time steps. Grass samples require a more complex averaging method to account for the roughly constant rate of CO₂ assimilation even though the CO₂ mole fraction varies strongly through time due to the presence or absence of the emission plume. That is, because we convolve Δ¹⁴C (the relative ¹⁴C content) with the CO₂ mole fraction to determine CO₂ff, the time-averaged sample should ideally collect proportionally more CO₂ when the CO₂ mole fraction is higher (as occurs for the NaOH samples). Because the CO₂ assimilation rate does not increase proportionally with the CO₂ mole fraction for the grass samples, they effectively deweight the periods when the plume is present. We apply a mathematical weighting function to the 10-min modeled CO₂ff to account for this and predict effective CO₂ff in the grass leaves (25).

To determine the scaling factor, we fit a regression slope to the modeled versus observed CO₂ff (Fig. 3), using the ordinary least squares bisector (OLSBI) method (35). OLSBI is appropriate in this case where neither variable is independent and the scatter is largely due to poorly constrained errors in the model.

Evaluation of Meteorological Data

We used the NIWA meteorological station (operational August–October 2012) for the model simulations of the 2012 grass samples. The STOS meteorological station (Davis Instrument Anemometer 7911) provided wind data for the 2013 and 2014 model simulations. We compared these two datasets with other available meteorological data to evaluate and correct any biases: the New Zealand MetService Automatic Weather Station at Hawera, 20 km to the southeast, and another anemometer installation 400 m northwest of the Kapuni source on the Smith farm obtained during part of the October/November 2014 campaign.

The wind directions are comparable among all of the sites, except for a slight skew in the Hawera wind directions relative to the other sites (not surprising for a site 20 km distant in a region with a large mountain only 50 km away, but suggesting that the wind directions from the various Kapuni stations are reasonable).

The wind speeds are comparable between the Hawera, NIWA, and Smith stations (Figure S3, Left) but the raw STOS data appears contracted relative to the three other stations. The STOS station is the only one with continuous data for December 2013 and throughout the October/November 2014 campaign and we therefore apply a correction to the STOS wind speed data. We apply a linear fit to the comparison of STOS and Smith data (because the Smith station has the largest data overlap in time and is closest in space), for which we find r² = 0.87, and Smith_corr = −1.25 + 1.91 × STOS. We also considered a power fit that gives slightly better agreement between the two records at high wind speeds, but the impact of this is small, with less than 5% of time steps having corrected wind speeds above 6 m/s. The highest wind speed time steps have the least impact on the final modeled CO₂ff because CO₂ff is small at high wind speeds, and we estimate that using the linear fit would induce a bias (overestimate) of less than 0.2 ppm in any given result. This bias is small relative to the measurement uncertainties and unquantified model transport errors and therefore we ignore it. We note that over the range of measured wind speeds, the mean of the corrected data are only slightly higher than that of the raw data. The scaling factor for our full dataset derived using the raw wind data is 1.00 ± 0.07, which is within the range of that derived using the corrected data (1.06 ± 0.08).