Source-apportionment methods

6-30-16 DRAFT ONLY do not cite or quote


An example of surface ozone attributed to the emissions from Maryland (left) and Ohio (right) at 2 p.m. on 7 July 2011 using CAMx v6.10. 

1. Motivation and Overview 

The goal of source apportionment in air quality modeling is to identify model inputs and processes that have the largest influence on various air pollution metrics. Model inputs may consist of emissions, initial conditions, and boundary conditions. Typical model responses of interest may include regional pollution episodes, average concentrations in a particular city, or the number of exceedences of an air quality standard. Quantifying how these responses depend upon emissions, chemical processes, and transport is valuable for developing efficient control strategies as well as identifying contributions of exceptional events to poor air quality. 

Ideally, it would be great if we could estimate the relationship between all model inputs and all model responses (i.e., how changes to emissions of any amount, at any single instant, in any particular source site, affects pollutant concentrations of any species at any downwind location at any later time). However, obtaining this amount of information becomes incredibly challenging from a computational perspective, amounting to running millions of air quality model simulations. 

Various source-apportionment methods have thus been developed, which specialize in providing some subset of this information in order to answer particular types of science or policy questions. Here we present a brief categorization of methods around typical air quality management questions, and provide an overview (Table 1) of which air quality models currently (in 2016) employ these methods. Summaries and examples of each method are then presented in Section 2, and in-depth theoretical descriptions can be found in review articles such as Cohan and Napelenok (2011).


What is the spatio-temporal distribution of pollution owing to a few broadly defined or aggregated source sectors or species, such as all transportation versus power plant emissions? 

Source-oriented modeling approaches consider the impact of a small number of model inputs on concentrations throughout the model domain. These include methods such as zero out, tagging (e.g., OSAT, APCA) and decoupled direct methods (DDM). These approaches provide a rich perspective on how a small number of sources (such as the emissions from a few particular power plants) or from a few aggregated sets of sources (such as emissions throughout North America for a few species) impact pollution in any place, at any time.

What is the origin of air pollution occurring in a particular location? 

Receptor-oriented approaches (Lagrangian, adjoint) provide a detailed estimate of the influence of emissions from each model grid cell, species, and time on a limited number of receptor metrics, such as pollution levels in a few urban areas, or the number of exceedences summed throughout the country. 

What would be the impact of implementing control strategies in a different order, or the co-benefits of implementing them together over each implemented in isolation? 

Higher-order decoupled direct methods (HDDM) or response surface models (RSM) can be used to evaluate impacts of new emissions control strategies within the context of an evolving emissions landscape owing to previously implemented policies. 

What are the marginal pollution responses to changes in emission (e.g., ppb per amount emitted)? 

Source-receptor methods (perturbation, DDM, adjoint) are best suited for evaluating the response of pollution to incremental changes in emissions, such as from the addition or removal of a small amount of emissions.

What is the complete breakdown of pollutant sources contributing to total concentration levels? 

Source attribution methods (HDDM, tagging, zero-out) attempt to fully characterize the origin of pollution concentrations in terms of the percent owing each type or location of emission. 

What can be learned from satellite data itself regarding source attribution of ozone and particulate matter? 

A separate article addressing this topic specifically can be found here.

Table 1: Models and source attribution, and source-response methods. Note that all models are capable of performing brute-force zero-out and perturbation calculations, and response-surface models (RSMs) could be built using any of the approaches listed here.


2. Methods and Applications 

Brute force zero-out and perturbation

These approaches entail altering the emissions in the model by either completely removing the source, or set of sources, of interest (aka the zero-out approach) or by perturbing them by a small amount (perturbation, often 10-20%). This approach is very easy to implement, since usually only slight modifications to the air quality model are required to facilitate these perturbations. This approach is thus often seen as easy to implement and reliable. The downside is the computational cost scales directly with the number of source regions to be considered, as a new simulation must be run for each source considered. While such simulations can be run in parallel, this approach is still best suited for questions regarding the influence of a small number of source categories. Whether or not emissions are completely removed or just perturbed depends upon the science or policy goal, with the former usually used to estimate source attribution and the later the response of air quality to near-term emissions changes.

Several AQAST projects have used brute-force methods for policy relevant applications. One common application is estimation of background versus the U.S. or North American anthropogenic contributions to O3. This application entails a small number of perturbation regions and is thus readily implemented using this approach. For example, the second column of Figure 1 shows the ozone resulting from of U.S. anthropogenic emissions to the maximum daily 8-hour O3 (MDA8) on several high-ozone days of June, 2012, as estimated with the GEOS-Chem model. Also shown are the contributions from anthropogenic emissions in a single state, and the contributions from Canada plus Mexico. Simulations of this type were used to define the North American Background O3 levels in the EPA O3 NAAQS Policy Assessment (U.S. EPA, 2014). AQAST scientists have also used perturbation runs to evaluate sources of aerosol in central Asia (Kulkarni et al., 2014) and the contribution of global long-range transport of ozone and aerosol precursors through the HTAP2 project.

Figure 1. Surface O3 concentrations on three days in June, 2012, owing to anthropogenic emissions from the following regions: the entire U.S. (Column 2), just Tennessee (Column 3), and Canada plus Mexico. 

Tagging 

Tagging consists of adding auxiliary variables to the air quality model itself to track pollution emanating from a pre-defined set of sources. The results allow for a complete source attribution of the pollution to this set of sources (as opposed to quantifying marginal responses, i.e., those in response to small changes in emissions). Tagging for inert primary species is a fairly straightforward process; methods for addressing secondary pollutants use sophisticated analysis of the limiting reagents to tag secondary constituents. 

Unlike perturbation simulations, multiple model runs are not required, and the model state does not need to be perturbed. The limitation is that for each source region considered, the computer memory requirements increase considerably. Tagging is thus typically used to identify the contribution for dozens of regions or less. 

Some air quality models have generic tagging capabilities (AM3, GEOS-Chem, WRF-Chem), and tagging is one way in which stratospheric influences are often identified (e.g., Lin et al., 2012; Zhang et al., 2014). CAMx has specific tagging capabilities for O3 (OSAT, APCA) and particulate matter, or PM (PSAT, OPSA), while CMAQ uses TSSA. Differences between these tagging methodologies (see Cohan and Napelenok, 2011) lie in how secondary species are attributed to different sources (e.g., Figure 3), and whether or not the attribution is performed online or offline, which can be important when considering tagging for species formed nonlinearly. 

Tagging and zero-out runs across different models have been found to provide similar estimates of North American background O3 in the western U.S. (Dolwick et al., 2015); other comparisons note that tagging is more appropriate for source attribution than for estimating responses to emissions changes (Collet et al., 2014; Koo et al., 2009). 

AQAST research projects have used tagging methods to evaluation long-range contributions to pollution on regional to global scales. Pfister et al. (2013) used a tagging algorithm implemented in the WRF-Chem model to evaluate contributions of continental inflow to surface O3 and CO, as shown in Figure 2. Saide et al. (2015) used source tagging to evaluate the contribution of biomass burning events to smoke and haze. The CAMx model has a tagging scheme (OSAT) that is commonly used for policy applications (Ramboll Environ, 2014). AQAST members applied the CAMx OSAT source apportionment add-on to evaluate contributions from regional boundary conditions to O3 in Baltimore, MD (Goldberg et al., 2015; 2016), work that contributed to the State Implementation Plan (SIP) for Maryland (Figure 3).

Figure 2: Modeled mean and standard deviation of O3 and O3INFLOW and COINFLOW as well as percent contribution of O3INFLOW % (local afternoon), from Pfister et al. (2013). 



Figure 3: Diurnal pattern of ozone source attribution at the Edgewood, MD site for the 5 July 2018 projected scenario using (top) OSAT and (bottom) APCA. The latter approach attributes anthropogenic/biogenic interactions to the controllable, anthropogenic source, resulting in more ozone formation attributed to anthropogenic sources.

Decoupled Direct Methods 

Decoupled direct methods (DDM) use a linearized form of the air quality model to propagate perturbations from a set of inputs to the complete space of concentrations. The benefit over brute-force perturbation simulations is that the computational cost of evaluating these perturbations is much lower than performing unique air quality model simulations for each perturbed input. First-order DDM methods are best suited for estimating the marginal response of pollution metrics with respect to emission changes. For ozone and aerosols, this may be applicable for emissions perturbations of up to 20-50%, depending upon the conditions. To estimate the response from larger perturbations, or perturbations of multiple precursors implemented simultaneously, high-order DDM (HDD) methods may be employed. The advantage of this approach over zero-out methods is that the actual forward model state is never perturbed, hence the model conditions best reflect the air quality model base state. The downside of DDM and HDDM methods for source-oriented modeling is cost of construction of the DDM form of the model itself, which takes time and code development. 

DDM methods have been used in several AQAST projects aimed at air quality forecasting and source attribution. The Hi-Res2 Forecasting System uses source-receptor relationships calculated with CMAQ-DDM as part of its data assimilation algorithm, wherein model emissions are adjusted to minimize the air quality forecast error. The source impacts calculated with CMAQ-DDM can be used for dynamic air quality management. CMAQ-DDM has also been used for a developing hybrid source apportionment methods (Hu et al., 2014; Ivey et al., 2015). 

Adjoint Methods 

Adjoint methods are receptor oriented source-apportionment tools, which means they are used to evaluate contributions of parameters such as emissions, boundary conditions, and initial conditions to receptor metrics such as ozone or particulate concentrations in particular cities, or aggregated metrics such as the number of exceedances in a region. A linearized, transposed, auxiliary version of the air quality model (the adjoint) is constructed to calculate how variations in the response metric of interest depend upon perturbations to each model input, such as the emission from any particular species or sector, at any hour, in any grid location. Their advantage is thus in providing detailed information about the different types of locations of sources that contribute to pollution in a single location, or to a single metric of air pollution (e.g., national PM2.5 exposure). Like DDM or brute-force perturbation methods, the results of an adjoint calculation estimate the influence of small changes to emissions. The downside of this approach is the considerable cost of construction of the adjoint model itself. Adjoint methods are available in GEOS-Chem (Henze et al., 2007) and certain version of CMAQ (e.g., Hakami et al., 2005; Turner et al., 2015; Bastien et al., 2015).

Adjoint source sensitivity analysis was used in AQAST projects to evaluate the long-range contributions of emissions to vegetative exposure and crop loss to ozone under present and future conditions (Lapina et al., 2014; 2015; 2016), as well as contributions of emissions to reactive nitrogen deposition (Lee et al., 2016), see Figure 4. The values in each grid cell of this figure show the contribution of the total nitrogen emissions (both oxidized and reduced) in that grid cell to a single national park. Total contributions from different sectors are provided by the inset pie charts. This type of analysis helps identify the scales at which emissions controls would be required to mitigate nitrogen deposition, and the variability in the important of different sectors. 


Figure 4: Source attribution of reactive nitrogen deposition to Great Smoky Mountains (SD) and Big Bend (BB) national parks, whose locations are indicated by the black circles. The sector definitions are ls: livestock, fe: fertilizer, na: natural, sf: surface inventory, eg: electric generating units, ne: non-eg industrial stacks, ac: aircraft, li: lightning, and so: soil. Adapted from Lee et al. (2015). 

Lagrangian Methods 

Lagrangian particle dispersion (LPD) models are another type of receptor oriented source-apportionment tools. These tools track the trajectory of air parcels arriving at a particular receptor site backwards in time to reveal their origin. These methods are often quite efficient for considering thousands of such trajectories. The resulting trajectories, or footprints, can be combined with an emission inventory to calculate the source contribution of emissions in each emission grid-cell or time to the receptor location. Lagrangian tools are currently limited to species that are chemically inert or conservative on short time scales, such as primary aerosols, CO or CH4. 

AQAST collaborators have used the WRF-STILT LPD model to evaluate greenhouse-gas sources contributions (e.g., "Modeling the carbon balance in Arctic ecosystems"). Figure 5 provides an animation of the track of particles over a 10-day period prior to arriving at the receptor location shown as a yellow pin. The different files correspond to different times and altitudes at which the particles were released from the receptor location. Note that in STILT the particles were released from the receptor and the calculation proceeds backward in time. However in the creating the movie the particles are shown moving forward in time. Each arrow represents a particle and the arrow length shows its movement over the previous 4 hours. The red arrows are within the boundary layer and the white ones are above the boundary layer, and the colors on the ground represent the accumulated footprints over the 10-day period. The footprints represent the accumulated contribution to the mixing ratio of a measured gas from surface emissions from that location and generally have units of ppmv/umole/m2/s. 

Figure 5: [VIDEO]

Response Surface Methods 

The methods discussed previously describe how to estimate the contribution of emissions to pollution within a single snapshot in time, considered in the context of static atmospheric conditions. However, formation of pollutants such as O3 and PM is an inherently nonlinear process, and air quality managers are often faced with consideration of multiple air quality control policies. Source-receptor methods that consider how pollutant concentrations respond to incremental changes in emissions (perturbation, adjoint, DDM) may provide results that can accurately estimate the response to emissions changes of more than 20-50%. Zero-out methods may provide different results if one estimates the impacts of, for example, removing traffic emissions while power-plant emissions are left unchanged compared to removing power-plant emissions while traffic emissions are unchanged. 

To overcome these limitations, one needs to consider the response of pollution to emissions changes within the context of emissions that themselves are already changing. This type of second order information is the goal of methods such as HDDM and response surface modeling. HDDM estimates the second derivative of pollutant concentrations with respect to changes in emissions (e.g., Hakami et al., 2003). Alternatively, response surface models attempt to map out the entire range of possible pollution levels that may correspond to any particular emissions levels. These maps can be built by evaluating air quality models across a range of emissions settings. This requires ensembles of simulations (typically hundreds) but can be performed using simple perturbation approaches (e.g., Fann et al., 2009).

References 

Bastien, L. A., B. C. McDonald, N. J. Brown, and R. A. Harley, High-resolution mapping of sources contributing to urban air pollution using adjoint sensitivity analysis: benzene and diesel black carbon, Environ Sci Technol. 2015 Jun 16;49(12):7276-84. doi: 10.1021/acs.est.5b00686 

Cohan, D., and S. Napelenok (2011), Air quality response modeling for decision support, Atmosphere, 2, 407-425, doi:10.3390/atmos2030407.

Collet, S., H. Minoura, T. Kidokoro, Y. Sonoda, Y. Kinugasa, P. Karamchandani, J. Johnson, T. Shah, J. Jung, and A. DenBleyker (2014), Future year ozone source attribution modeling studies for the eastern and western United States, J. Air Waste Manage. Assoc., 64(10), 1174–1185, doi:10.1080/10962247.2014.936629. 

Dolwick, P., Akhtar, F., Baker, K.R., Possiel, N., Simon, H., Tonnesen, G., 2015. Comparison of background ozone estimates over the western United States based on two separate model methodologies. Atmos. Environ., http://dx.doi.org/10.1016/j.atmosenv.2015.01.005. 

Fann, N., C. M. Fulcher, and B. J. Hubbell (2009), The influence of location, source, and emission type in estimates of the human health benefits of reducing a ton of air pollution, Air Qual. Atmos. Health, 2, 169-176.

Goldberg, D. L., T. P. Vinciguerra, K. M. Hosley, C. P. Loughner, T. P. Canty, R. J. Salawitch, and R. R. Dickerson (2015), Evidence for an increase in the ozone photochemical lifetime in the eastern United States using a regional air quality model, J. Geophys. Res. Atmos., 120, 12778–12793, doi:10.1002/2015JD023930. 

Goldberg, D. L., T. P. Vinciguerra, D. C. Anderson, L. Hembeck, T. P. Canty, S. H. Ehrman, D. K. Martins, R. M. Stauffer, A. M. Thompson,. J. Salawitch, and R. R. Dickerson. (2016), CAMx ozone source attribution in the eastern United States using guidance from observations during DISCOVER-AQ Maryland, Geophys. Res. Lett., 43, doi:10.1002/2015GL067332. 

Hakami, A.; Odman, M.T.; Russell, A.G. High-order, direct sensitivity analysis of multidimensional air quality models. Environ. Sci. Technol. 2003, 37, 2442–2452. 

Hakami, A., D. K. Henze, J. H. Seinfeld, K. Singh, A. Sandu, S. Kim, D. Byun, and Q. Li (2007), The adjoint of CMAQ, Environ. Sci. Technol., 41(22), 7807-7818, doi:10.1021/es070944p. 

Henze, D. K., A. Hakami and J. H. Seinfeld (2007), Development of the adjoint of GEOS-Chem, Atmos. Chem. Phys., 7, 2413-2433.

Hu, Y., Balachandran, S., Pachon, J. E., Baek, J., Ivey, C., Holmes, H., Odman, M. T., Mulholland, J. A., and Russell, A. G.: Fine particulate matter source apportionment using a hybrid chemical transport and receptor model approach, Atmos. Chem. Phys., 14, 5415-5431, doi:10.5194/acp-14-5415-2014, 2014. 

Ivey, C. E., Holmes, H. A., Hu, Y. T., Mulholland, J. A., and Russell, A. G.: Development of PM2.5 source impact spatial fields using a hybrid source apportionment air quality model, Geosci. Model Dev., 8, 2153-2165, doi:10.5194/gmd-8-2153-2015, 2015. 

Koo, B., G. M. Wilson, R. E Morris, A. M. Dunker, and G. Yarwood, Comparison of source apportionment and sensitivity analysis in a particulate matter air quality model, Environ. Sci. Technol. 2009, 43, 6669–6675, 2009. 

Kulkarni, S., Sobhani, N., Miller-Schulze, J. P., Shafer, M. M., Schauer, J. J., Solomon, P. A., Saide, P. E., Spak, S. N., Cheng, Y. F., Denier van der Gon, H. A. C., Lu, Z., Streets, D. G., Janssens-Maenhout, G., Wiedinmyer, C., Lantz, J., Artamonova, M., Chen, B., Imashev, S., Sverdlik, L., Deminter, J. T., Adhikary, B., D'Allura, A., Wei, C., and Carmichael, G. R.: Source sector and region contributions to BC and PM2.5 in Central Asia, Atmos. Chem. Phys., 15, 1683-1705, doi:10.5194/acp-15-1683-2015, 2015. 

Lapina, K., D. K. Henze, J. B. Milford, M. Huang, M. Lin, A. M. Fiore, G. Carmichael, G. G. Pfister, and K. W. Bowman, Assessment of source contributions to seasonal vegetative exposure to ozone in the U.S., J. Geophys. Res., 119, 324-340, doi:10.1002/2013JD020905, 2014. 

Lapina, K., D. K. Henze, J. B. Milford, C. Cuvelier, and M. Seltzer, Implications of RCP Scenarios for future changes in vegetative exposure to ozone in the Western U.S., Geophys. Res. Let., 42, 4190--4198, doi:10.1002/2015GL063529, 2015. 

Lapina, K., D. K. Henze, J. B. Milford and K. Travis, Impacts of foreign, domestic and state-level emissions on ozone-induced vegetation loss in the U.S., Environ. Sci. Technol., doi:10.1021/acs.est.5b04887, 2016. 

Lee, H.-M., F. Paulot, D. K. Henze, K. Travis, D. J. Jacob, L. H. Pardo, B. Schichtel, Sources of nitrogen deposition in Federal Class I areas in the US, Atmos. Chem. Phys, 16, 525-540, doi:10.5194/acp-16-1-2016, 2016. 

Lin Meiyun, A. M. Fiore , O. R. Cooper , L. W. Horowitz , A. O. Langford , Hiram Levy II , B. J. Johnson , V. Naik , S. J. Oltmans , C. Senff (2012): Springtime high surface ozone events over the western United States: Quantifying the role of stratospheric intrusions, Journal of Geophysical Research, 117, D00V22, doi:10.1029/2012JD018151 

Pfister, G. G., S. Walters, L. K. Emmons, D. P. Edwards, and J. Avise, Quantifying the contribution of inflow on surface ozone over California during summer 2008, J. Geophys. Res. Atmos., 118, 12,282–12,299, doi:10.1002/2013JD020336, 2013. 

Ramboll Envrion (2014), CAMx version 6.10 User’s Guide. [Available here]

Saide, P. E., Peterson, D., da Silva, A., Anderson, B., Ziemba, L. D., Diskin, G., Sachse, G., Hair, J., Butler, C., Fenn, M., Jimenez, J. L., Campuzano-Jost, P., Perring, A. E., Schwarz, J. P., Markovic, M. Z., Russell, P., Redemann, J., Shinozuka, Y., Streets, D. G., Yan, F., Dibb, J., Yokelson, R., Toon, O. B., Hyer, E., and Carmichael, G. R.: Revealing important nocturnal and day-to-day variations in fire smoke emissions through a multiplatform inversion, Geophys. Res. Lett., 2015GL063737, 10.1002/2015gl063737, 2015. 

Turner, M., D. K. Henze, A. Hakami, S. Zhao, J. Resler, Jaroslav, G. Carmichael, C. Stanier, J. Baek, A. Sandu, A. Russell, A. Nenes, G.-R. Jeong, S. Capps, P. Percell, R. 

Pinder, S. Napelenok, J. Bash, T. Chai, Tianfeng (2015), Differences Between Magnitudes and Health Impacts of BC Emissions Across the US Using 12km Scale Seasonal Source Apportionment, Environ. Sci. Technol., 49 (7), pp 4362--4371, doi:10.1021/es505968b.

U.S. Environmental Protection Agency, 2014. Policy Assessment for the Review of the O3 National Ambient Air Quality Standards, Research Triangle Park, NC, EPA-452/R-14-006. [Available here]

Zhang, L., Jacob, D. J., Yue, X., Downey, N. V., Wood, D. A., and Blewitt, D.: Sources contributing to background surface ozone in the US Intermountain West, Atmos. Chem. Phys., 14, 5295-5309, doi:10.5194/acp-14-5295-2014, 2014.

Editor: Daven Henze

Partner: Mark Estes