Assessment of ground-based atmospheric observations for verification of greenhouse gas emissions from an urban region

Previous Studies of Urban CO₂

Many studies (e.g., refs. 8⇓–10) have described near-surface CO₂ concentrations and fluxes in a variety of urban environments and attributed observed variability to both atmospheric dynamics and local emission patterns. City-scale fluxes have been derived with mass-balance approaches using surface data from Krakow (11) and aircraft data from Indianapolis (12). Ratios of carbon monoxide (CO) and CO₂ from a site near Beijing were used to define trends in combustion efficiency (13). None of the approaches taken in these studies are capable of quantifying trends in emissions at the full urban scale and with the accuracy required for verification.

Levin et al. (14) presented the only prior study to accurately assess emission inventories over time at a regional scale, using a multidecadal dataset of atmospheric GHG concentrations, , and radon-222 from Heidelberg. Unfortunately, the key element of this unique study (long-term, high-frequency, radioisotope measurements) are not currently widely reproducible due to cost and technological requirements. The framework described below uses measurements made by readily available sensors combined with open-source meteorological data and models, and is thus scalable to numerous locations, as needed for MRV.

Characterization of CO₂ Observations from Salt Lake City

The SLC CO₂ measurement program was initiated as part of a study of carbon and oxygen isotopes focused on urban source attribution. Pataki et al. (15⇓–17) estimated the proportional contributions of natural gas versus gasoline combustion and biological respiration to observed CO₂ enhancements, and (18) suggested the application of CO₂ as a tracer of atmospheric transport and mixing in complex terrain. Pataki et al. (19) found general agreement between SLC eddy-flux measurements and an emissions inventory compiled for the upwind area. SLC CO₂ was simulated with a multiple box model to understand the relative contributions of meteorology and anthropogenic and biological surface fluxes to observed daily and seasonal cycles (20).

The SLC CO₂ data follow a distinctive diel pattern, in which concentrations are higher at night and lower during the day (Fig. 1A), following the daily cycle of the mixing height, which is shallow at night due to thermal stratification, and deep most days due to solar heating of the surface. The diel cycle of CO₂ concentrations is notably out of phase with emissions (Fig. 1B), implying that thermally forced circulations impose a stronger influence on near-surface concentrations than emission rates (20). Mean hourly enhancements over background in 2006 at the downtown site ranged from approximately 0–20 ppm (1σ ≅ 16 ppm) in the afternoon (12–18 h mountain standard time, MST) and from approximately 20–60 ppm (1σ ≅ 35 ppm) at night (22-04 h MST). Peak concentrations are typically observed in the early morning due to the combined effects of atmospheric stratification and increased emissions from rush-hour traffic; concentrations drop rapidly thereafter with the onset of deeper vertical mixing in midmorning (20) (Fig. 1A).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Average hourly observed CO₂ concentration at the downtown site in 2006. (A) Averaged by hour of the day from winter (Dec.–Feb.) and summer (June–Aug.) months, and from nearly the entire year (April–Dec.) at the background site. (B) Observed CO₂ from the whole year (Left y axis) versus average hourly CO₂ emissions estimated from the Vulcan inventory for a 0.5° × 0.5° area encompassing the Salt Lake Valley (Right y axis). (C and D) Distribution of observed CO₂.

Seasonal averages are lower in summer than winter (Fig. 1A). During the growing season, CO₂ concentrations sometimes fall below background in the afternoon, likely due to uptake by urban trees (21). In the winter, the Salt Lake Valley (SLV) is prone to atmospheric temperature inversions, which suppress vertical mixing and give rise to sustained periods of elevated concentrations. The overall distribution of observed CO₂ is heavily right-skewed (Fig. 1C) and is seemingly comprised of two subpopulations, representing stratified and unstratified conditions (Fig. 1D).

Simulation of CO₂ Observations

We used the Weather Research and Forecasting—Stochastic Time-Inverted Lagrangian Transport (WRF-STILT) atmospheric transport model (http://stilt-model.org) and antropogenic and biogenic flux inventories to simulate the SLC CO₂ data, as described in Materials and Methods and references therein. The observation-model framework was tested for four monthly time periods from different seasons in 2006 (Table 1) and three sampling locations in SLC. We assessed two horizontal resolutions of the ATDM, 4 km (“baseline”) and 1.3 km (“high-resolution”) (Fig. S1), the latter of which was tested for only a 2-wk subset of the autumn time period. The high-resolution ATDM included parameterization of an urban canopy model (UCM) (22), which allows for greater heterogeneity in surface properties related to the urban environment than is available in standard WRF configurations. Fig. 2 shows hourly observed and simulated CO₂ for baseline and high-resolution models. The time series demonstrates the model’s general capability for capturing the typical diel pattern of near-surface CO₂ concentrations, albeit with a systematic underestimation using a priori emissions (Fig. 2B).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Hourly observed and modeled CO₂ concentrations for two weeks in October 2006 (A) at the downtown site, and (B) averaged by hour of the day at the downtown, neighborhood, and junior high sites.

Elsewhere (23), we evaluated the performance of the two WRF configurations (baseline and high-resolution with an UCM) by comparing observed and modeled meteorological parameters from the SLV. The high-resolution meteorological configuration with the UCM led to improved representation of the daily evolution of the surface (2 m) temperature and boundary-layer height, especially when the flow was driven by local circulations. Likewise, the high-resolution WRF-STILT model resulted in changes in simulated CO₂, in particular related to the timing of the nocturnal boundary-layer formation and breakup (Fig. 2B).

The model captures many weather-related events such as the multiday persistence of low concentrations around October 16 and August 30 (Fig. 2A and Fig. S2B). Some, generally short, time periods are poorly simulated, such as on June 24–25, when the model substantially underestimates observed CO₂ (Fig. S2A). The model also captures seasonal variability in the magnitude and variance of CO₂ enhancements (Table 1). In December, the model is often unable to simulate CO₂ concentrations at hourly resolution (Fig. S2D), although this result was not unexpected because meteorological conditions during strong winter temperature inversions in valleys are difficult to simulate (24). But the model does capture the very high variances in CO₂ and the general amplitude of enhancements over background during December (Table 1 and Fig. S2D).

The model biosphere sometimes draws down simulated CO₂ in SLC below background during midday, in the spring and summer months only, in agreement with the observations (Fig. 1A) and known irrigation practices, and contributes small enhancements from plant and soil respiration during all other times. Overall, the model suggests that the biosphere has a relatively minor influence on CO₂ concentrations in SLC, which is not surprising given the semiarid ecosystems of the region and previous findings (19⇓–21).

Quantification of the Data-Model Relationship

Two approaches were employed to quantify the relationships between hourly observed and simulated CO₂ concentrations for the four time periods and three observation sites. We first applied a type II, standard major axis regression (25) to fit a line to observed versus simulated values (Fig. 3). The inverse of the regression line slope provides an estimate of the optimum factor by which to scale the emissions inventory to best match the observed data. Confidence intervals (CIs) on the slope define our ability to detect changes in emissions over time at the 95% confidence level, assuming spatial and temporal biases in the model are unchanging over time.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Hourly modeled versus observed CO₂ at three sites for a two-week time period in October 2006 resulting from (A) high-resolution and (B) baseline model configurations. Solid lines are standard major axis regression lines and dashed lines are one-to-one shown for reference.

Because the regression is based on modeled and observed values paired in time, it may be susceptible to sporadic failures in the transport model, and thus may lead to CIs which are overly pessimistic in terms of trend detection capability. We therefore adopted an alternative approach which gives less weight to poorly timed events by selecting scaling factors that minimize the differences in observed and simulated population means from each site and time period. Ninety-five percent CIs for this optimization procedure were calculated using a percentile bootstrap.

Table 2 gives scaling factors generated by both optimization methods for the high-resolution and baseline models for October. Table S1 gives scaling factors for the other time periods. Mean scaling factors for October from the high-resolution model are between 1.5 and 1.8, with 95% CIs that are ± 7–8% of the mean (Table 2). Mean scaling factors for other seasons from the baseline model mostly range between 1 and 2 and have 95% CIs that are ± 6–13% of the means (Table S1).

Scaling factors for the high-resolution model are similar to those for the baseline model, but are more consistent between sites and optimization methods, and have narrower CIs (Table 2). We infer that the high-resolution model is better for trend assessment, although this was not immediately apparent by visual inspection of Fig. 2. Improvements by the high-resolution model are especially noticeable in the decreased persistence of very low model values when the data indicate elevated concentrations (Fig. 3). When inventory fluxes are multiplied by the optimal scaling factors, the distributions of simulated CO₂ from the high-resolution model are remarkably close to observed distributions (Fig. 4), although variance was not included in the optimization procedure.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 4.

Quantile–quantile plots of hourly modeled versus observed CO₂ at the downtown site for two-weeks in October 2006 from (A) baseline and (B) high-resolution models. Model values unscaled and scaled by the two optimization methods are shown. For the high-resolution model, scaling factors by the two optimization methods are near identical, so the two scaled model distributions are nearly indistinguishable.

These results support the application of high-resolution modeling, and optimization of sample distributions and regressions, for determining trends in urban emissions. We infer from the results reported in Table 2 for the high-resolution model that 15% is a conservative estimate of the minimum increase or decrease in monthly emissions detectable by our observation-model framework, although caution must be exercised in generalizing these results due to the short time period (approximately 2 wk) for which the high-resolution model was tested. By applying the framework to several years’ worth of data, changes in scaling factors, and thus relative changes in emissions, could likely be estimated with greater precision.

The scaling factors generated in this study are significantly greater than the expected value of 1.0 (Table 2 and Table S1), implying that emissions were underreported for the SLC urban core and/or that modeled meteorology was too well-mixed. However, absolute emissions cannot be evaluated with the same level of accuracy as can a change in emissions over time because our statistical procedures do not account for systematic model errors, such as possible over- or underestimation of the mean boundary-layer height or biases in the presumed spatial distribution of emissions. To evaluate absolute emissions, rather than a change in emissions, a fiducial tracer or a sustained release experiment would be necessary.

Approaches for Improving Emission Trend Detection

A key limitation to further constraining emissions is the inability of current models to simulate small-scale atmospheric processes. Examples of processes which affect concentrations over short timescales at individual urban sites, but which models cannot explicitly represent, include circulations at building, street, and neighborhood scales, and intermittent turbulence in the nocturnal boundary layer. Improved parameterization of these processes could significantly improve the capability of atmospheric models to simulate urban GHG concentrations. Representation of proximal emission processes at an enhanced spatial resolution similar to that of the meteorology (1.3 km) could also lead to improved simulations.

Contrary to the expectations of some (cf. ref. 6), the SLC case suggests that increasing the number of surface measurement stations across the city would be ineffective at substantially improving the observational approach for detecting a change in emissions. Simulations indicate that individual measurements sites are sensitive to emissions across the full urban region (Fig. S3). Observed CO₂ concentrations at the five measurement stations in SLC are strongly correlated on a daily basis (Fig. S4) because within day variance is dominated by the diel cycles in atmospheric stability (Fig. 1) and forcing of these cycles occurs on the scale of the whole valley. This finding suggests that the current network of five stations in SLC is more than adequate for characterizing the daily cycle of urban-scale CO₂. If we consider just the afternoon hours, the opposite problem occurs: Fluctuations at stations quite close together (< 5 km) are not significantly correlated (Fig. S4), suggesting that small-scale processes are responsible for CO₂ concentration variations in the afternoon and are not directly tied to region-wide emissions. Hence, denser measurements of such variations would not help to significantly improve determination of regional trends.

Alternative measurement strategies that are less sensitive to the details of atmospheric circulation and emissions may lead to improved trend detection capability. Shallow circulations rearrange CO₂ between sublayers of the atmosphere on hourly timescales, but for the duration that emitted CO₂ remains in the urban region, total column amounts are directly linked to total emissions. Fig. 5 shows simulations of the characteristic pattern of CO₂ enhancement, vertically integrated through the partial atmospheric column, which comprises the “urban dome” over the SLV. Observations throughout the column are not available for validation, but simulations appear to have sufficient fidelity at the surface (Fig. 2) to justify exploration of the character of the urban dome through modeling. The position of the urban dome shifts with the wind, but, due to the valley topography (Fig. 5C and Fig. S5), its core generally lies on a predictable NW–SE axis. During the day, the dome extends vertically up to 2 km, but at night, excess CO₂ is contained within a thin layer less than 100-m deep (Fig. 5 F and G).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 5.

Simulated partial column-averaged XCO₂ (ppm) enhancements above background up to 3 km above SLC and the surrounding area on October 18, 2006 at 15 h (A) and 23 h (B) MST. (D and E) Simulated CO₂ enhancement near the surface, 50 m above the ground, for the same times and locations as in A and B. (F and G) Vertical slices through the areas of maximum XCO₂ enhancement in the urban domes. (C) Topography in kilometers above sea level. The downtown and rural measurement sites are marked with Xs for reference. Lines in A–E show the positions of the transects plotted in F and G. Note that the two upper-left panels have the same scale, but the four lower panels do not.

The differences between enhancements at the surface and those integrated through the column are striking. The areal extent and magnitude of column enhancements are larger in the day than the night (Fig. 5 A and B), directly reflecting the higher daytime emissions (Fig. 1B) that we wish to measure. We infer that urban enhancements in column amounts are more sensitive to regional-scale meteorology, especially mean wind speeds, and to emissions integrated through the whole urban region. Conversely, surface values are more sensitive to boundary-layer height, shallow circulations, and local traffic emissions. Mean winds are much easier to model than boundary-layer heights, and can be validated with hourly observations from airports and weather stations. Broad-scale emission inventories are better defined than fine-scale, day-to-day traffic patterns.

The magnitude of the anthropogenic CO₂ enhancement in the partial column integral is notably smaller than at the surface, by factors of 2 (daytime, Fig. 5 A and D) to 20 (nighttime, Fig. 5 B and E), suggesting that increased accuracy may be required to both measure and simulate the column enhancement. To our knowledge, ground-based measurements of CO₂ column amounts with the accuracy required for verification have been demonstrated just once, in Los Angeles, where peak concentrations in the column were indeed observed in midday (26). Total column measurements could be made from space, obviating the need for many surface stations and eliminating intrusive measurements on the territory of a treaty signatory. The widespread, spatially heterogeneous, and shifting nature of the simulated SLC urban dome suggests that remote sensing of the dome may offer the best route for its full characterization. Unfortunately, no presently planned satellite has the necessary orbit or targeting capability.

Summary and Future Directions

We have demonstrated an observation-model framework capable of detecting a change in anthropogenic CO₂ emissions of 15% or more from an urban region on a monthly basis. The model framework consists of an atmospheric transport model (STILT) driven by a high-resolution (1.3 km) mesoscale meteorological model (WRF) and coupled to moderately high-resolution models of the spatial and temporal distribution of anthropogenic emissions (Vulcan) and biogenic fluxes. We compared simulations to observations from four time periods and three locations in SLC. Constraints on emission rates were obtained by optimizing the emission model two ways, both of which gave similar central values and confidence intervals.

The observation-model framework is readily scalable to other sites using commercial sensors and open-source models. Measurements are needed to define background (upwind) values, especially for urban areas downwind of other major source regions. For heavily vegetated cities, it will be necessary to distinguish anthropogenic from biogenic emissions, possibly with tracer measurements of fossil fuel combustion (e.g., CO, ¹⁴C). Analysis of the statistical correlations among the SLC measurement sites indicates that five was an ample number, although this result likely depends on SLC’s relatively small size and topographic confinement.

We argue that measurements of vertically integrated column amounts would provide more new information than would additional surface sites. In our estimation, column measurements offer a promising route for improved detection of CO₂ emissions from major source regions, complementing or possibly obviating the need for extensive surface measurements near these areas. Remote sensing of the column-integrated urban dome appears to offer the best route for accurate verification of emission inventories of CO₂ and other GHGs.