Improving Estimates of PM_2.5 Concentration and Chemical Composition by Application of High Spectral Resolution Lidar (HSRL) and Creating Aerosol Types from Chemistry (CATCH) Algorithm

Abstract

Improved characterization of ambient PM_2.5 mass concentration and chemical speciation is a topic of interest in air quality and climate sciences. Over the past decades, considerable efforts have been made to improve ground-level PM_2.5 using remotely sensed data. Here we present two new approaches for estimating atmospheric PM_2.5 and chemical composition based on the High Spectral Resolution Lidar (HSRL)-retrieved aerosol extinction values and types and Creating Aerosol Types from Chemistry (CATCH)-derived aerosol chemical composition. The first methodology (CMAQ-HSRL-CH) improves EPA’s Community Multiscale Air Quality (CMAQ) predictions by applying variable scaling factors derived using remotely-sensed information about aerosol vertical distribution and types and the CATCH algorithm. The second methodology (HSRL-CH) does not require regional model runs and can provide atmospheric PM_2.5 mass concentration and chemical speciation using only the remotely sensed data and the CATCH algorithm. The resulting PM_2.5 concentrations and chemical speciation derived for NASA DISCOVER-AQ (Deriving Information on Surface Conditions from COlumn and VERtically Resolved Observations Relevant to Air Quality) Baltimore-Washington, D.C. Corridor (BWC) Campaign (2011) are compared to surface measurements from EPA’s Air Quality Systems (AQS) network. The analysis shows that the CMAQ-HSRL-CH method leads to considerable improvement of CMAQ’s predicted PM_2.5 concentrations (R² value increased from 0.37 to 0.63, the root mean square error (RMSE) was reduced from 11.9 to 7.2 μg m⁻³, and the normalized mean bias (NMB) was lowered from −46.0 to 4.6%). The HSRL-CH method showed statistics (R²=0.75, RMSE=8.6 μgm⁻³, and NMB=24.0%), which were better than the CMAQ prediction of PM_2.5 alone and analogous to CMAQ-HSRL-CH. In addition to mass concentration, HSRL-CH can also provide aerosol chemical composition without specific model simulations. We expect that the HSRL-CH method will be able to make reliable estimates of PM_2.5 concentration and chemical composition where HSRL data are available.

1. Introduction

Particulate matter is one of the six criteria air pollutants in the United States, which can adversely affect human health and the environment. Fine particles with an aerodynamic diameter of 2.5 μm or smaller (PM_2.5) have the most serious impacts on human health and the environment, due to their small size and long residence time (Shisong et al., 2018). Despite the significance of public exposure, the ground measurement networks for PM_2.5 mass concentration and chemical composition are sparse in many regions, especially in rural areas (Goldberg et al., 2019).

Remote sensing retrievals and chemical transport model (CTM) or air quality model (AQM) simulations are widely used to fill the gaps in ground measurements of PM_2.5. CTMs (e.g., GEOS-Chem) and AQMs (e.g., CMAQ) simulate PM_2.5 concentrations by solving the mass continuity equations for each chemical component given emissions, meteorology, and topography (Jin et al., 2019a). These models are valuable tools for long-term epidemiologic studies on PM_2.5 nationally and globally. However, due to their coarse spatial resolution (often larger than 12 km), they are usually less reliable to provide PM_2.5 predictions at local scales (Wang et al., 2016). In addition to the coarse grid resolution, the accuracy of PM_2.5 simulations from models is affected by uncertainties in emissions, chemical processes, and meteorological data.

Remote sensing has been used for providing observational products of aerosols over the last two decades (Kaufman et al., 2002). Early studies (e.g., Wang & Christopher, 2003) have found a correlation between the satellite-retrieved aerosol optical depth (AOD) and ground-level PM_2.5 concentration. AOD is the vertically integrated light extinction caused by aerosols and is a widely-used remote sensing product. Since the early 2000s, considerable efforts have been made to predict ground-level PM_2.5 using remotely sensed AOD, along with several other parameters. Generally, these methods can be divided into two categories: the statistical methods and the geophysical methods (Jin et al., 2019b). The statistical methods typically include satellite-retrieved AOD along with several meteorological factors (e.g., temperature, relative humidity, and the boundary layer height) and land-use variables. These parameters are used in conjunction with the ground-level PM_2.5 measurements to train the statistical models. Once properly trained, the statistical models can be used to predict the ground-level PM_2.5 (e.g., Kloog et al., 2014). The disadvantage of the statistical methods is that their training process relies on ground measurements, so the predictions are limited to regions with sufficient ground monitoring sites (Jin et al., 2019b).

Traditional geophysical methods combine remotely sensed AOD with model simulations to predict ground PM_2.5 using the following equation (Liu et al., 2004):

$${[PM}_{2.5\_sat}]={[PM}_{2.5\_model}]\times \frac{{AOD}_{sat}}{{AOD}_{model}}$$ (1)

where [PM_{2.5_sat}] represents the predicted ground PM_2.5 using the geophysical method, [PM_{2.5_model}] represents simulated PM_2.5 in the lowest model grid, AOD_sat represents satellite-retrieved columnar AOD, and AOD_model represents model-predicted columnar AOD calculated based on model-simulated PM_2.5 chemical components. The method was further refined by van Donkelaar et al. (2012) through the application of climatological-based regional scaling and spatial smoothing. The geophysical methods do not rely on ground measurements and can provide PM_2.5 predictions with high spatiotemporal coverage. However, since AOD_sat is a column measure of aerosol load, the ability to infer surface PM_2.5 from AOD is affected by both the vertical distribution of aerosols and the satellite overpass time (van Donkelaar et al., 2006). Moreover, the chemical composition and hygroscopic growth of aerosols (i.e., their ability to absorb water) can substantially affect their extinction efficiencies and hence the AOD (Flores et al., 2012).

Unlike passive remote sensing instruments which rely on the sun’s energy as a source of light (thus termed as passive remote sensing), active sensors (such as lidars and radars) emit their own source of electromagnetic radiation which is then scattered back to the receiver (e.g., Hostetler et al., 2018). Lidars can retrieve vertically resolved aerosol optical properties at different wavelengths, allowing them to overcome some of the shortcomings of the geophysical methods. Several studies have applied the lidar-retrieved surface aerosol vertical extinction in their predictions of ground PM_2.5 (e.g., Schaap et al., 2009; Koelemeijer et al., 2006; van Donkelaar et al., 2012; Chu et al., 2013; 2015). Methods utilizing the available insights into aerosol vertical distributions lead to improvements in AOD-PM_2.5 relationships. However, to account for the aerosol hydration effects on extinction, models either prescribe some campaign-averaged value for the hygroscopic growth of aerosols (e.g., Chu et al., 2015) or rely on model-predicted information on aerosol chemical composition.

The HSRL offers the opportunity to acquire information on the chemical composition of aerosols. The unambiguous retrievals of four intensive aerosol parameters (i.e., lidar ratio, depolarization ratio, backscatter color ratio, and spectral depolarization ratio) by the HSRL allows for qualitative classification of aerosols into eight types: pure dust, dusty mix, maritime, polluted maritime, urban, fresh smoke, smoke, and ice (Burton et al., 2012). Aerosol types are expected to be associated with chemical composition. For example, dust and maritime particles are expected to contain soil minerals and sea salt, respectively, while urban and smoke particles typically contain different fractions of sulfate, nitrate, ammonia, black and organic carbon. Nevertheless, the HSRL-retrieved aerosol types are classified according to their interaction with light and therefore require additional constraints to be linked to aerosol chemical composition. Dawson et al. (2017) developed a novel algorithm (CATCH- Creating Aerosol Types from Chemistry) for transforming model-predicted aerosol microphysical parameters (i.e., size distribution, complex refractive index, density, etc.) and chemical composition (i.e., sulfates, nitrates, and ammonia, organic carbon, black carbon, sea salt, and mineral dust) into aerosol types similar to those derived by HSRL. It was shown that the CATCH aerosol types, derived using a training data set from the coastal outflow of the eastern U.S., matched closely with expected spatial distributions near heavily populated cities and MODIS-observed active fires elsewhere in the U.S. (Dawson et al., 2017). The CATCH algorithm revealed that each aerosol type over the U.S. can be characterized by some distinct composition of the chemical species such as sulfate, nitrate, ammonia, black and organic carbon, sea salt, and mineral dust. The advantage of using CATCH is that once the algorithm is properly trained using the data acquired through different field campaigns over a specific geographic region (e.g., continental U.S.), it can be used for linking aerosol chemical composition with remotely sensed types without specific model runs for different years and seasons. Dawson et al. (2017) also showed that across the U.S. the spatiotemporal variability in inter-type aerosol chemical composition was narrower than its intra-type variability, suggesting that the aerosols in each type can be linked to their respective chemical composition.

In this study, we develop two new methods to calculate the mass concentration and chemical composition of PM_2.5 by combining the products of the CATCH algorithm with HSRL-retrieved aerosol extinction and types. The results of these methods are examined against baseline CMAQ model simulations and the CMAQ simulations corrected with HSRL measured aerosol extinction values. All methods use the data collected during the NASA DISCOVER-AQ (Deriving Information on Surface Conditions from COlumn and VERtically Resolved Observations Relevant to Air Quality) Baltimore-Washington, D.C. Corridor (BWC) Campaign (2011). The results are compared against the ground measurements from EPA’s Air Quality System (AQS) network.

2. Data

2.1. Remote sensing

During the DISCOVER-AQ BWC campaign in July 2011, the airborne HSRL instrument was flown onboard the UC-12 aircraft. The aircraft conducted two flights each day: one in the morning and one in the afternoon. The exception was on 28 July 2011 when it conducted only the morning flight. The flights were conducted at nearly constant altitudes of ~8 km. The HSRL is reporting the data at 30 m vertical resolution with extinction starting at approximately 300 – 400 m above the surface. The detailed discussion on the HSRL measurements during the DISCOVER-AQ BWC campaign can be found in (Chu et al., 2015). Figure 1 shows an example of the flight path on 29 July 2011.

Figure 1.

The study region. The blue line shows the 29 July 2011 afternoon flight path. The teal diamonds represent hourly PM_2.5 sites (A1 - A4 from north to south) and the red squares depict daily PM_2.5 speciation sites (B1 - B4 from north to south). A3, B2 and A4, B3 are the same sites.

2.2. Models

The Community Multiscale Air Quality (CMAQ) version 5.0.2. is used in this study for conducting air quality model simulations. The CMAQ results used here were previously evaluated in publications including Appel et al. (2017) and Simon et al. (2018). To briefly summarize, CMAQ was applied to a 4 km resolved domain (150 × 150 cells) centered over the Baltimore DISCOVER-AQ area. The vertical structure uses 35 terrain-following hydrostatic pressure coordinate extending to 50 hPa with the finest surface layer being approximately 20 meters at the surface. Lateral boundary conditions were derived from a 12 km continental simulation described in more detail in Simon et al. (2018). CMAQ was driven by the Weather Research and Forecasting model (Skamarock et al., 2008) version 3.7. Emissions were from the 2011 National Emission Inventory (NEI) version 2 (U. S. Environmental Protection Agency, 2016) speciated for Carbon Bond 05. Sector-based domain wide NO_x, CO, SO₂, and PM_2.5 and CO emission totals are available in Table S1 in the supplemental material. The evaluation of model-predicted CO against the aircraft measurements showed that the meteorology in CMAQ version 5.02 can credibly represent regional transport in the Baltimore and Washington D.C. areas (Simon et al., 2018). The model-predicted mass concentrations of selected aerosol species (i.e., sulfates, nitrates, and ammonia (SNA), organic carbon (OC), black carbon (BC), sea salt (SS), and mineral dust (Dust)) in both Aitken and accumulation modes (including accumulation mode unspeciated aerosol mass (OTH) (see Sec. 3.3.1)) are used to calculate PM_2.5 chemical speciation. The CMAQ outputs throughout the campaign were extracted along the flight path of the UC-12 aircraft.

The CATCH algorithm is used here for transforming the HSRL retrieved aerosol types into the plausible aerosol chemical composition. A detailed description of the algorithm can be found in Dawson et al. (2017). In general, CATCH is designed to take the Goddard Earth Observing System- with Chemistry (GEOS-Chem) model predicted aerosol extensive (i.e., the mass ratio of total carbonaceous aerosol to total carbonaceous aerosol and sulfate, the mass ratio of organic carbon to black carbon) and intensive (i.e., the effective lidar ratio, the effective real index of refraction, and the effective single-scattering albedo, all at 532 nm) properties and generate aerosol types similar to those retrieved by HSRL (i.e., fresh smoke, smoke, urban, dusty-mix, and maritime). While doing this, CATCH also retains the information about the chemical composition of individual aerosol types over different locations and times, allowing for the aerosol chemical composition to be related to each HSRL retrieved aerosol type. The GEOS-Chem model simulations were carried out for the period of 18 July to 6 August 2014 and are describe in more detail in Dawson et al. (2017).

For all the methods examined in this study to be directly comparable and representative of the ground measurements, the model results and the HSRL retrievals were averaged within the planetary boundary layer (PBL). While CMAQ determines the PBL heights, and the mixed layer heights can also be computed from the HSRL data, we have obtained the PBL heights from the North American Mesoscale Forecast System (NAM) hybrid sigma-pressure coordinate data with 12 km spatial and 1 h temporal resolution to use consistent averaging for all the methodologies examined in this study. Previous studies that compared model-predicted outputs for the PBL height against radiosonde observations have shown that out of different state-of-the-art models with variable PBL schemes, NAM forecasts had the smallest mean absolute error (~ 40m) in diagnosed PBL height (Coniglio et al., 2013). For all the parameters examined, the NAM forecasts were also found to have lower biases compared to the five most widely used PBL schemes (Coniglio et al., 2013).

2.3. Ground measurements

Ground data were acquired from the EPA’s AQS network of surface monitors within the study region (see Fig. 1). The AQS database contains measurements of air pollutant concentrations throughout the U.S. and its territories. Details about these AQS sites are listed in Table 1. Four AQS sites (A1 – A4) provide hourly surface PM_2.5 concentration data, while the other four (B1 – B4) provide daily averaged PM_2.5 speciation. The speciation is measured once every 3 or 6 days (Solomon et al., 2014) and the dates with available speciation measurements are 2, 14, 20, and 26, July 2011. The ground measurement sites were selected to be the closest to the UC-12 aircraft tracks. Note that statistically, the surface PM_2.5/PM₁₀ ratio at the DISCOVER-AQ BWC surface sites was ~ 70% (Chu et al., 2015). Therefore, throughout this study, it is assumed that PM_2.5 particles are primarily responsible for the measured aerosol light extinction measured by the HSRL.

Table 1.

Summary of the AQS sites.

ID	Type	AQS ID	Latitude	Longitude	State
A1	Hourly PM2.5	24-015-0003	39.701444	−75.860051	MD
A2	Hourly PM2.5	24-510-0040	39.297733	−76.604603	MD
A3/B2	Hourly PM2.5, daily PM2.5 Speciation	24-033-0030	39.055277	−76.878333	MD
A4/B3	Hourly PM2.5 Daily PM2.5 Speciation	11-001-0043	38.921847	−77.013178	DC
B1	Daily PM2.5 Speciation	24-005-3001	39.310833	−76.474444	MD
B4	Daily PM2.5 Speciation	51-087-0014	37.55652	−77.40027	VA

3. Methodology

3.1. HSRL data processing

The following data processing procedures were implemented for the HSRL data. First, we derive aerosol extinction near the surface. As mentioned above, the first range bin for the HSRL retrieved aerosol extinction values at 532 nm (σ_{ext_532}) is ~ 300 meters above ground level (AGL). Therefore, to infer σ_{ext_532} near the surface, the HSRL product AOT_hi_col is used. AOT_hi_col is the additive AOD at 532 nm from the measurement height extrapolated to the ground. Therefore, the AOD for each layer including the surface can be determined by subtracting the AOD for the layer above it. The difference $\frac{AOt\_hi\_col\left(z\right)-AOT\_hi\_col(z+dz)}{dz}$ yields σ_{ext_532} (Mm⁻¹) in the first 300 m AGL.

Second, the original HSRL aerosol types are converted to the aerosol types in the CATCH algorithm. In this step, the maritime and polluted maritime types from HSRL are reclassified as the maritime type, and the ice and dusty mix types, along with the outlier, overlap, and unclassified types are removed from the analysis of aerosol chemical composition. Similar to extinction values, aerosol types are not available below 300 m AGL. To derive aerosol types near the ground, aerosols are assumed to be well-mixed and the aerosol types available in the lowest vertical bin are extended to the ground. This assumption is supported by the CMAQ simulations that show PM_2.5 to be well mixed within the PBL. However, under the conditions when the PBL height is less than 300 m, the proposed methodology may not work.

Third, the HSRL-retrieved aerosol extinction and types get regridded to match the vertical resolution of the CMAQ grids. As the vertical resolution of CMAQ grids is typically coarser than HSRL grids, to calculate the average HSRL extinction coefficients within a single CMAQ grid, the HSRL layer AOD values closest to CMAQ grid vertical boundaries were summed and then divided by the cumulative height. The regridding of the aerosol type data is more complex, as various aerosol types can exist in a single CMAQ grid. Since different aerosol types could be associated with different chemical composition, during regridding the fractional contribution of each aerosol type was calculated by weighting the types by their extinction values as following:

$$Typ{e}_i=\frac{\sum _{l=1}^n{{\sigma }_{ext}}_{i,l}Typ{e}_{i,l}}{{\sum }_{i=1}^5{\sum }_{l=1}^n{{\sigma }_{ext}}_{i,l}Typ{e}_{i,l}}$$ (2)

where i stands for the CATCH aerosol types and l for a number of HSRL vertical layers within a given CMAQ grid layer. The initial CATCH types are assigned discretely for whole levels, so Type_i,l are 1 and 0 in the original HSRL levels. Thus, Type_i at CMAQ levels is an extinction weighted fraction of the aerosol type.

3.2. Mass fractions of PM_2.5 chemical components from CATCH

The CATCH algorithm (Dawson et al., 2017) is used to link aerosol types to respective chemical compositions over the continental U.S. and Mexico. This is achieved by prescribing the average mass fractions of chemical components (K_ij) to the regridded HSRL aerosol types. Here i stands for CATCH aerosol types and j for chemical components (i.e., SNA, organic matter (OM), BC, SS, and Dust). Table 2 summarizes the CATCH-derived average chemical compositions for each aerosol type. Table 2 shows that the maritime and dust aerosol types are primarily composed of one chemical component (SS and Dust, respectively). The urban aerosol is largely comprised of SNA (61%) and OM (28%) with smaller contributions of BC (3%). The composition of the smoke aerosol is similar to the urban aerosol, with lower SNA fractions (60%) and higher OM fractions (33%). Compared with the smoke aerosol, the fresh smoke aerosol contains less SNA (40%) and more OM (49%) and BC (7%). CATCH, like HSRL, also has a category for unclassified/outlier aerosol types but they are not included in the current analysis. The optical properties cannot be assigned to unclassified/outlier aerosol types as they do not have a defined chemical composition and aerosol size distribution.

Table 2.

Average percent mass fraction matrix (K_ij) of PM_2.5 chemical components for different aerosol types.

	Urban	Smoke	Fresh Smoke	Maritime	Dust
SNA
Median	60.8	59.8	40.4	5.9	3.7
Q1, Q3*	54.1, 66.5	53.6, 66.4	39.0, 45.8	4.4, 7.3	3.5, 3.9
OM
Median	27.7	33.0	48.6	0.2	0.8
Q1, Q3	23.0, 33.8	27.1, 38.6	43.1, 51.8	0.1, 0.3	0.7, 0.8
BC
Median	3.3	3.0	6.9	0.1	0.1
Q1, Q3	3.0, 3.7	2.7, 3.5	6.2, 8.4	0.05, 0.08	0.10, 0.11
Dust
Median	0.6	2.6	1.0	4.6	94.8
Q1, Q3	0.08, 2.7	1.7, 4.0	0.8, 1.3	1.8, 8.9	94.4, 95.0
Sea Salt
Median	1.2	0.5	0.4	88.9	0.7
Q1, Q3	0.5, 8.5	0.3, 0.7	0.3, 3.5	84.3, 93.0	0.6, 0.8

^*Q1 indicates quartile 1 (25^th percentile) and Q3 indicates quartile 3 (75^th percentile).

3.3. Methods for PM_2.5 prediction

The two new methods for improving estimates of PM_2.5 concentration and chemical composition developed here are compared against the original CMAQ simulations and the preexisting approach introduced by Liu et al. (2004).

3.3.1. CMAQ-HSRL method

The CMAQ-HSRL method is similar to one introduced by Liu et al. (2004). CMAQ-predicted and HSRL-measured layer AODs (i.e., AOD_CMAQ, AOD_HSRL) are calculated by multiplying the simulated $\left({{\sigma }_{ext}}_{CMAQ}\right)$ and measured $\left({{\sigma }_{ext}}_{HSRL}\right)$ extinction coefficient by the CMAQ grid heights. The ${{\sigma }_{ext}}_{CMAQ}$ is calculated using CMAQ-predicted aerosol chemical composition as:

$${{\sigma }_{ext}}_{CMAQ}=\sum _{j=1}^6{b}_j{f}_j{[PM}_{CMAQ,j}]$$ (3)

where ${\sigma }_{ex{t}_{HSRL}}$ is the extinction coefficient at 550 nm (in units of Mm⁻¹) calculated from speciated extinction coefficients (b_j), hygroscopic growth factors (f_j), and accumulation mode dry mass concentrations ([PM_CMAQ,j] in μg/m³) where j∈{SNA, OM, BC, Dust, SS, OTH}. Extinction coefficients b_SNA = 3.0, b_OM = 4.0, b_BC = 10.0, b_Dust = 1.0, b_SS = 1.37, and b_OTH = 1.0 are the dry mass extinction coefficients (in m² g⁻¹) (Binkowski & Roselle, 2003; Malm & Hand, 2007; Pitchford et al., 2007). Hygroscopic growth factors are relative humidity (RH)-dependent scattering enhancements caused by the water uptake of the aerosols (hygroscopic growth). The values of f_SNA and f_SS are from the look-up tables of the IMPROVE algorithm (Pitchford et al., 2007) and f_OM is calculated following Brock et al. (2016) as f_OM=1+0.03×RH/(100−RH). RH values over 95% are set equal to f values at RH=95% following Pitchford et al. (2007). The RH-dependent scattering enhancements for Dust, BC, and OTH are assumed to be equal to 1 (i.e., no change in aerosol optical extinction associated with the water uptake). The vertical profiles of RH are taken from the CMAQ output.

$$\left[{PM}_{CMAQ-HSRL,j}\right]=\left[{PM}_{CMAQ,j}\right]\times \frac{AO{D}_{HSRL}}{{AOD}_{CMAQ}}$$ (4)

where j refers to aerosol chemical components (i.e., SNA, BC, OM, Dust, SS, and OTH) and [PM_CMAQ,j] and [PM_{CMAQ−HSRL,j}] stand for CMAQ model and CMAQ-HSRL method predicted chemically speciated PM_2.5 concentrations respectively (in μg m⁻³).

3.3.2. CMAQ-HSRL-CH method

The CMAQ-HSRL-CH method is similar to CMAQ-HSRL, except CMAQ-predicted chemically-speciated PM_2.5 concentrations are iteratively corrected using the HSRL extinction and aerosol types data with the CATCH-derived aerosol chemical composition as:

$$\left[{PM}_{CMA{Q-HSRL-CH}_{n+1,j}}\right]=\left[{PM}_{CMA{Q-HSRL-CH}_{n,j}}\right]\left(1+R_j.\frac{\left({AOD}_{HSRL}{-AOD}_{CMAQ}\right)}{{AOD}_{HSRL}}\right)$$ (5)

where j refers to aerosol chemical components (i.e., SNA, BC, OM, Dust, and SS) and n is an iteration number. At iteration n=1, the initial PM estimate is from CMAQ. R_j represents the extinction weighted fraction of component of chemical species j in the CMAQ grid (unitless) that is the product of the fractional contribution of HSRL-derived aerosol types and the CATCH-derived aerosol chemical components within each CMAQ grid (R_j= Σ_iType_iK_i,j). R_j does not change during the iteration; its role is to introduce the aerosol chemical information, in addition to extinction, into corrected PM_2.5 concentration. Since there is no unspeciated aerosol mass in the CATCH algorithm, no correction is performed for PM_OTH in CMAQ extinction calculations.

The HSRL layer AOD (i.e., AOD_HSRL) is calculated through data regridding as described in Sec. 3.1. If AOD_CMAQ is differed from AOD_HSRL by more than 3% (i.e., |AOD_HSRL−AOD_CMAQ>0.03∙AOD_HSRL), the concentrations of CMAQ-simulated PM_2.5 components are varied iteratively, until either the solution converges or the number of iterations exceeds 100. In CMAQ grids where the iteration number exceeds the threshold value [$\left[{PM}_{CMA{Q-HSRL-CH}_{n+1,j}}\right]$ is set to equal to $\left[{PM}_{CMA{Q-HSRL-CH}_{1,j}}\right]$, which is the raw model result. The average number of iterations was 3.7 and the total fraction of the gridpoints in which the iteration number exceeded the threshold value was 0.4% on average and was less than 1.3% for any flight.

3.3.3. HSRL-CH method

In the HSRL-CH method, the HSRL retrievals are combined with the CATCH algorithm products to predict surface PM_2.5. The total PM_2.5 concentration in each CMAQ grid ([PM_HSRL−CH) is then calculated as:

$$\left[P{M}_{HSRL-CH}\right]={\sigma }_{ex{t}_{HSRL}}/\left(\sum _{j=1}^5{b}_j{f}_j{R}_j\right)$$ (6)

where j (=1, 2,…,5) refers to aerosol chemical components, ${\sigma }_{ex{t}_{HSRL}}$ is HSRL extinction in the CMAQ grid, b_j and f_j are the same as in Eq. 4., and R_j is the same as in Eq. 5. Note that there is no unspeciated mass in PM_2.5 predicted by Eq. 6. The vertical profiles of RH are taken from HSRL data sets that use the NASA Goddard Earth Observing System version 5 (GEOS-5) forecasts provided by the NASA Global Modeling and Assimilation Office (GMAO). The data is interpolated to each altitude, latitude and longitude for each profile.

For each grid, the concentration of chemical species (j∈{SNA, OM, BC, Dust, SS, OTH}) in μg m⁻³ are derived as:

$$\left[P{M}_{HSRL-CH,j}\right]=R_j\left[P{M}_{HSRL-CH}\right]$$ (7)

where R_j represents the extinction weighted fraction of component j in the CMAQ grid (as in Eq. 5). It should be noted that the application of R_j in Eq. 7 could lead to some uncertainty for variable RH conditions, as it derives aerosol mass in a dry state using the extinction weighted data.

3.4. Error calculations

Here we only quantify the uncertainties introduced by CATCH-derived aerosol chemical composition for individual aerosol types. According to Table 2, for a given aerosol type (i.e., urban, maritime, etc.) CATCH derives slightly different aerosol chemical compositions (and therefore mass fraction ratios) over different regions of the continental U.S. Monte Carlo simulations were conducted to assess how uncertainties in prescribed average mass fractions of individual aerosol types (K_ij in Table 2) affect the results for each data point. To do so, 10,000 samples were taken within the range of chemical components (25^th and 75^th percentiles) for each aerosol type from Table 2. To avoid erroneous results in aerosol chemical composition data, the sum of the fractions was always kept the same as the median values in Table 2. Although by doing so, the composition fractions are not treated as fully independent, the large number of samples assures a wide range of the aerosol chemical composition. Uncertainties in PM_2.5 mass concentration and chemical composition related to the uncertainty in K_ij are then calculated as 10^th and 90^th percentiles of all Monte Carlo simulation.

Measurements at the AQS sites also contain uncertainties with each monitoring network often using its sampling methodology and measurement frequency (Eder & Yu, 2006). However, uncertainties associated with the AQS measurements are not generally reported as part of the data. Additionally, there are some small uncertainties in HSRL extinction values (Hair et al., 2008) and derived aerosol intensive parameters (Burton et al., 2012) associated with systematic errors in measurements. The quantification of these uncertainties is beyond the scope of this study.

3.5. Comparison with the ground measurement

As described in Sec. 2.3, the AQS surface PM_2.5 mass concentration data are averaged hourly while the chemical speciation data are averaged over 24 hrs. To make the best comparison between the HSRL (that provides a snapshot of vertical curtains for extinction values and aerosol types along the aircraft path) and the ground station data, several additional data processing procedures were implemented. Once the calculations for mass concentration and chemical speciation are carried out for all three methods (i.e., CMAQ-HSRL, CMAQ-HSRL-CH, and HSRL-CH), the data were averaged within 8 km of the ground sites and compared with the closest (hourly) surface PM_2.5 measurements. To find the optimal span for the data analysis, we varied the averaging distance from 12 km to 4 km. The data analysis showed that the reduction of the distance increases correlation, but reduces the data points available for the comparison. The previous analysis of DISCOVER-AQ BWC data found 10 km from the AQS stations to be an optimal distance for the acquisition of HSRL aerosol extinction profiles (Chu et al., 2015). In addition to the spatial averaging, the data were vertically averaged within the PBL; as mentioned above, the HSRL retrievals of aerosol types typically started at approximately 300 m above the surface.

The full list of the AQS parameters (with the corresponding codes) used for comparison is provided in Table S2 in the supplemental material. The measured sulfate, nitrate, and ammonium concentrations were summed up at the ground site to be comparable with CMAQ SNA results. The choice of the parameter used for determining BC and OC was made based on the results of Spada and Hyslop (2018). Following the results of El-Zanan et al. (2005) the scaling factor of 1.92 was used to convert measured organic carbon concentration to OM. For the AQS data, unspeciated mass (OTH) is defined as the difference between the total PM_2.5 measured and the sum of all mass-apportioned to specific species. Specifically,

$$\left[{PM}_{A{QS}_{OTH}}\right]=\left[{PM}_{A{QS}_{total}}\right]-\left(\left[{PM}_{A{QS}_{SNA}}\right]+\left[{PM}_{A{QS}_{OM}}\right]+\left[{PM}_{A{QS}_{BC}}\right]+\left[{PM}_{A{QS}_{DUST}}\right]+\left[{PM}_{A{QS}_{SS}}\right]\right)$$ (8)

Dust concentration was determined from elemental speciation as El-Zanan et al. (2005):

$$\left[Dust\right]=2.2\times \left[Al\right]+2.49\times \left[Si\right]+1.63\times \left[Ca\right]+2.42\times \left[Fe\right]+1.94\times \left[Ti\right]$$ (9)

[PM_2.5] sea salt concentration was calculated from chloride concentration for CMAQ and ground data using the scaling factor of 1.8 (Pitchford et al., 2007).

4. Results and discussion

4.1. HSRL retrievals

Figure 2 shows two examples of HSRL retrievals of aerosol extinction coefficient at 532 nm and derived aerosol types. The HSRL curtains for aerosol extinction and types for each flight are provided in Fig. S1. Extinction coefficient values were generally high in the first couple of km AGL, suggesting that aerosols were largely confined within the PBL. A good example is 1 July 2011 (Fig. 2a) when the atmosphere above the PBL was so clean, that HSRL was not able to infer the aerosol types. Near the top of the PBL, increased values of aerosol extinction coefficient were associated with the high RH (often manifested by the presence of clouds) and the resulting hydration of the particulate matter. The HSRL aerosol type classification algorithm is using a flight-by-flight threshold for separating the clouds from aerosols in each lidar profile (Burton et al., 2012). The dominant aerosol type during the BWC campaign was urban, which accounted for the majority of retrievals. However, several days exhibited a complex mixture of different aerosol types. Smoke aerosols were frequently present near the top of the PBL and in the free troposphere. Several days have shown the existence of dust particles both near the surface (e.g., 14 and 27 July) and between 3 to 4 km AGL (e.g., 11 and 29 July) (see Figs. 2 and S1).

Figure 2.

HSRL-retrieved aerosol extinction coefficient at 532 nm and aerosol types for afternoon flights on 1 July 2011 (top) and 29 July 2011 (bottom). The vertical arrows show times when the aircraft came within 8 km of the AQS sites.

4.2. PM_2.5 mass concentrations

Figure 3 shows PM_2.5 mass concentrations predicted by CMAQ, derived using the three methods discussed above, and the surface measurements made at the AQS sites. Table 3 provides the performance metrics commonly used for model evaluations (e.g., Eder and Yu, 2006). Although, A2 and A4 sites are located in large metropolitan areas, i.e., Baltimore, MD and Washington, DC (see Fig. 1), Fig. 3 does not demonstrate considerable regional differences in hourly averaged PM_2.5. According to Fig. 3a, the original CMAQ simulations tend to underestimate ground-level PM_2.5, especially under high PM_2.5 conditions. The PM_2.5 performance evaluation statistics for the CMAQ output applied in this study against the surface observations given in Table 3 are comparable to those reported by Kelly et al. (2019). Kelly et al. (2019) Supplemental Table S1 shows the mean R² value of 0.44 (with the range of 0.05 to 0.5) and NME of 36.9% (35.2 to 54.6 %) for the Northeast summer are close to ones reported in Table 3. However, the NMB and RMSE reported in Table 3 are on the high end compared to the values reported by Kelly et al. (2019) (NMB −18.2%, −35.4 to 15.0% and RMSE 6.07 μg m⁻³, 4.06 to 8.39 μg m⁻³). The high bias is likely due to a combination of issues including (1) the method of evaluation used here and (2) the local nature of evaluation. First, the method of evaluation reported here uses the nearest satellite track prediction as a surrogate for the monitor location. Second, the statistics reported by Kelly et al. (2019) are for the Northeast as a whole, while these results are for a small select area on a limited number of days.

Figure 3.

Scatter plots of PM_2.5 observations vs. predictions a) from original CMAQ simulations, b) from CMAQ-HSRL, c) from CMAQ-HSRL-CH, and d) HSRL-CH. The points are colored by sites. The error bars correspond to the uncertainty in predicted PM_2.5 due to the range in CATCH-derived chemical composition for HSRL-retrieved aerosol types.

Table 3.

Summary statistics associated with CMAQ simulations and three methodologies for predicting PM_2.5 mass concentration.

Method	R2	RMSE (μg m−3)	MB (μg m−3)	NMB (%)	NME (%)
CMAQ	0.37	11.91	−8.58	−46.00	49.49
CMAQ-HSRL	0.58	7.39	1.48	7.92	31.38
CMAQ-HSRL-CH	0.63	7.16	0.84	4.55	29.30
HSRL-CH	0.75	8.58	4.42	24.02	33.89

R²= coefficient of determination, RMSE= root mean-square error, MB= mean bias, NMB= normalized mean bias, normalized mean error. All definitions can be found in Eder and Yu (2006).

The assimilation of HSRL retrieved extinction data (i.e., CMAQ-HSRL) leads to an improved agreement between AQS measured and CMAQ simulated PM_2.5 (see Fig. 3b and Table 3). The results of the CMAQ-HSRL-CH method (i.e., correction of CMAQ results with HSRL retrieved extinction and hygroscopic particle growth corrections) in Fig. 3c are comparable to the ones obtained using the simpler CMAQ-HSRL method. This result suggests that corrections to aerosol vertical extinction profiles were more important than corrections to their chemical composition related changes in particle hydration. This is largely due to the fact that the additional chemical information obtained through HSRL/CATCH derived aerosol types did not lead to significant differences in the CMAQ-predicted fractional composition of the aerosols. Our data analysis further shows that for 1.5 km AGL CMAQ-HSRL predicted PM_2.5 concentration is comparable (with slightly higher values) to CMAQ-HSRL-CH predictions (see Fig S2). Figure S2 shows that there is no discernable vertical bias between the two methods.

Figure 3d and Table 3 show that when compared with the AQS data, HSRL-CH results (i.e., the ones without CMAQ model simulations) were significantly better than original CMAQ simulations and analogous to the other two methods (i.e., CMAQ-HSRL and CMAQ-HSRL-CH). Table 3 also shows that despite the good correlation, HSRL-CH demonstrates some positive bias (NMB = 24%) compared to surface measurements. Overall, the data analysis suggests that the calculations based on the HSRL-retrieved data combined with the CATCH algorithm could lead to accurate results for PM_2.5 mass concentration in the campaign domain.

The error bars for CMAQ-HSRL-CH and HSRL-CH methods show the uncertainty associated with the assumption of average chemical composition for specific aerosol types across the U.S. The uncertainties shown in Figs. 3c and 3d are rather small. This is largely caused by the fact that the Monte Carlo simulation-derived uncertainty in the extinction weighted fraction of component chemical species (i.e., uncertainty in R_j) gets weighted by the AOD ratio that is commonly less than 3%. Small error bars in Figs. 3c and 3d suggest that the errors in CATCH-derived chemical composition are not the primary source of the uncertainty in calculated PM_2.5 mass concentrations. This result is not surprising, as it was found that aerosol vertical extinction was the primary driver of aerosol mass concentration corrections. Nevertheless, it suggests that HSRL-CATCH derived aerosol chemical composition was a realistic representation of the true chemical makeup of the ambient particles.

4.3. PM_2.5 chemical speciation

Figure 4 compares the chemical speciation of PM_2.5 derived from CMAQ simulations, CMAQ-HSRL, CMAQ-HSRL-CH, and HSRL-CH to the AQS measurements. Figure 4 is a composite of the days when the AQS chemical composition data were available and the UC-12 aircraft came within 8 km of the site (i.e., 2, 14, 20, and 26 July 2011) for the morning and afternoon flight legs. One major difference between the AQS site measurements and predictions is that the predictions are based on 3 to 4 instances either in the morning or afternoon (see vertical arrows on Fig. 2), while the monitor observes concentrations all day long. Figure S3 shows the daily speciation data for individual flights at all four sites. At the first look, Fig. 4 shows lower PM_2.5 values for the B4 site, located near Richmond, VA, the southernmost point of the study domain. However, Fig. S3 reveals that the results for the B4 site shown in Fig. 4f are based on a single day 14 July 2011 (when the AQS data were available and the aircraft flew within 8 km of the B4 site). According to Fig. S3, all the sites reported low PM_2.5 concentration for that day. Therefore, the low aerosol mass concentration for the B4 site shown in Fig. 4f is not an indication of a sizable spatial gradient over the study domain. In agreement with Fig. 3, Fig. 4 shows that in general, CMAQ underestimates the total aerosol mass. The agreement between the measured and the predicted values gets improved when the CMAQ outputs get corrected using HSRL extinction measurements. It can be noticed that Figs. 4a and 4b show different AQS data for morning and afternoon flights at the B1 site. This happens because Fig. 4 is a composite of different days and the calculated AQS average for the morning flights does not include 20 July 2011, as the aircraft did not pass within 8 km of the site on that flight.

Figure 4.

Average surface PM_2.5 chemical speciation from AQS, CMAQ, CMAQ-HSRL, CMAQ-HSRL-CH, and HSRL-CH. Morning flights on the left and afternoons on the right. For the morning flights at station B4, there are no data with both the AQS measurements and the airplane flying within an 8 km radius of the ground station. The error bars correspond to the uncertainty in predicted PM_2.5 speciation due to the range in CATCH-derived chemical composition for HSRL-retrieved aerosol types.

The chemical speciation data shown in Fig. 4 gives some insight into processes responsible for the differences in PM_2.5 mass concentration. According to Fig. 4 there is a large difference between the measured and predicted amounts of unspeciated aerosol mass, i.e., OTH. As a reminder, for the AQS sites, the OTH concentration is derived as a difference between the total PM_2.5 measured and the sum of all mass apportioned to chemical species (i.e., SNA, OC, BC, SS, Dust). For the CMAQ results, OTH represents accumulation mode unspeciated aerosol mass. The CATCH algorithm does not contain unspeciated aerosols, therefore there is no OTH in HSRL-CH. Out of the four sites, B2 and B3 show a considerable fraction of AQS PM_2.5 being attributed to OTH. Further analysis shows that both sites could have been influenced by the local sources of mineral dust. B2 site is located less than 150 m away from an agricultural farm, while the B3 site is located near the historic McMillan Reservoir Sand Filtration Site, a twenty-five-acre decommissioned water treatment plant. According to Fig. 4, such local sources of fugitive dust are not well resolved by either CMAQ simulations or HSRL retrievals. Figure 4 also reveals large differences between measured and derived OM. Previous studies have shown that in Southeast U.S., both the CMAQ and the GEOS-Chem models (upon which the CATCH chemical speciation of the aerosol is based) underestimate organic aerosol concentrations during summer in the southeastern U.S. as compared to observations (e.g., Baek et al., 2011; Koo et al., 2014; Wang et al., 2018).

Figure 4 shows that SNA and OM account for the major fraction of total PM_2.5 concentration. This is not surprising, since the urban aerosols were the primary aerosol types over the study domain. However, SS, Dust, and BC also comprise a non-trivial fraction of aerosol mass. For example, HSRL-CH method predicts considerable amounts of sea salt at B1 and B2 sites. Site B1 is located at about 1 km away from the Back River, an estuarine inlet of the Chesapeake Bay, while Site B2 is located at Howard University campus, ~35 km away from the Seven River of the Chesapeake Bay. The chemical speciation data for individual flight legs (see Fig. S3) shows that sea salt was predicted by CMAQ-HSRL-CH and HSRL-CH during the afternoon flight on 2 and 14 July 2011 for the B1 site and 14 July 2011 for the B2 site. Fig. S1 shows that near B1 and B2 stations, the HSRL frequently retrieved polluted maritime aerosol types at different elevations (from 300 m to 2 km). Figure 4 shows that for the dates examined in this study, no discernible amount of sea salt has been measured at the AQS stations or predicted by CMAQ. Currently, it is not possible to explain the differences between HSRL-CATCH predicted and AQS measured sea salt mass concentrations, but points to the need for a comprehensive comparison of HSRL-CATCH predictions at different altitudes over different study domains.

Results of Fig. 4 also help to better understand the differences between CMAQ-HSRL and CMAQ-HSRL-CH methodologies. For example, for the afternoon flight leg at B1 (Fig. 4b), both methodologies predict comparable mass concentrations. However, based on the HSRL retrieved aerosol types and CATCH-derived chemical composition, CMAQ-HSRL-CH predicts the presence of SS and apportions slightly more SNA and slightly less OM, BC, Dust, and OTH compared to CMAQ-HSRL. Because SNA and SS are more hygroscopic, to attain the same extinction values, both CMAQ-HSRL and CMAQ-HSRL-CH methodologies predict comparable PM_2.5 mass concentrations. As stated above, Fig. 4b is a composite of different days, so this effect can be better seen for individual flights. For the afternoon flight leg at B4, CMAQ extinction values were significantly lower compared to HSRL and the CMAQ-HSRL method corrected it by increasing initial compositions with the fixed value to match the measured extinction, i.e., SNA, OM, BC, SS, Dust, and OTH were all multiplied by the same scaling value. On the contrary, CMAQ-HSRL-CH used retrieved aerosol types and CATCH-derived chemical composition to calculate specific R_j values (in Eq. 5) for different chemical components. For this particular case, CMAQ-HSRL-CH predicts significantly more SNA and less OM compared to CMAQ-HSRL. Because SNA is more hygroscopic compared to OM, Dust, and OTH, for a given extinction value CMAQ-HSRL-CH apportions less aerosol mass compared to CMAQ-HSRL. Although in this particular case (i.e., the afternoon flight leg at B4) CMAQ-HSRL compares better with surface AQS measurements than CMAQ-HSRL-CH, overall (see Fig. 4 and Fig. S3), there is no clear favorite for prediction of aerosol chemical composition among different methodologies examined in this study. Finally, the current dataset does not allow us to assess the uncertainty in CMAQ-HSRL methodology introduced by R_j values in Eq. 7. Detailed models (based on aerosol optical and microphysical properties) will be needed to accurately calculate HSRL-derived extinction to dry mass conversion factors for different aerosol types under variable chemical component values and RH conditions.

Small error bars in Fig. 4 suggest that in CATCH-derived chemical composition is not the primary source of the uncertainty in calculated PM_2.5 chemical speciation. CMAQ-HSRL-CH has no error bars for the OTH, as the concentration of the unspeciated aerosol mass did not change in Eq. 5 during the iteration. Overall, Fig. 4 suggests that both CMAQ-HSRL-CH and HSRL-CH yielded results comparable to other methodologies designed for the prediction of PM_2.5 using remotely sensed data. Based on this result, we propose that over the continental U. S. with available HSRL-retrieved vertical extinction and aerosol types, the HSRL-CH method can be used to derived PM_2.5 mass concentrations and chemical composition.

5. Conclusions

Two new methodologies based on a combination of regional air quality model (CMAQ), active remote sensing (HSRL), and model algorithm (CATCH) were evaluated against surface AQS data for the assessment of aerosol PM_2.5 and chemical composition. The analysis showed that the CMAQ-HSRL-CH method of iteratively correcting the air quality model results using the HSRL-retrieved aerosol extinction and types and CATCH-derived chemical composition lead to considerable improvement of CMAQ’s predicted PM_2.5 concentrations (R² value increased from 0.37 to 0.63, the root mean square error RMSE got reduced from 11.9 to 7.2 μgm⁻³, and the normalized mean bias NMB got lowered from −46.0 to 4.6%). This is in line with previously reported results that show that the assimilation of remotely sensed aerosol extinction values into models leads to improvements in model-predicted PM_2.5 concentrations. This study has also shown that HSRL-retrieved extinction and aerosol types and CATCH-derived chemical components can be used to derive PM_2.5 concentration and chemical composition in the atmospheric column without individual model runs. The data analysis showed that the HSRL-CH derived values for surface concentrations (R²=0.75, RMSE =8.58 μgm⁻³, and NMB < 25%) and aerosol chemical composition were comparable to other methodologies in which specific regional air quality model results were corrected with the remotely sensed data. Once accurately validated over different parts of the world, the HSRL-CH method developed here can be successfully used for getting useful information about the vertical distribution of aerosol mass and chemical composition, especially over the regions where neither air quality model results nor surface measurement data are available. The HSRL-CH can also be used to further explore the links between column AOD and PM_2.5, a question that has been discussed by the research community for over two decades. The extinction calculations used in this study were based on dry PM_2.5 mass concentrations. However, the methodology can be revised under conditions when PM₁₀ (i.e., mineral dust or sea salt) comprises the major fraction of the aerosol mass. For such specialized conditions, extinction values can be calculated using empirical relationships between PM₁₀ mass concentrations and AOD described in the literature (Shao et al., 2003; Raut and Chazette, 2009).

Uncertainty calculations were carried out to evaluate the errors introduced by the assumption that the same aerosol types over the continental U.S. have similar chemical composition. The analysis suggests that the possible error in inter-type variability in aerosol chemical composition has a minor effect on derived PM_2.5 concentration and chemical speciation. This result was reached because the vast majority of near-surface aerosols encountered during the measurement campaign were classified as urban and didn’t demonstrate large variability in chemical composition. Therefore, the additional information on the chemical composition of different aerosol types (through CATCH) did not lead to considerable improvements to CMAQ-HSRL-CH compared to CMAQ-HSRL. Future studies should explore the application of the proposed methodologies over different parts of the U.S. as well as outside the U.S., in regions characterized by variable aerosol types. It should be noted that the CATCH algorithm was developed using the Ship-Aircraft Bio-Optical Research (SABOR) campaign that conducting flights between the East Coast of the U.S. and Bermuda. Although the SABOR campaign took place in a different year (18 July to 6 August 2014), it sampled the coastal outflow from the eastern U.S. with the aerosol optical signature likely comparable to that of the DISCOVER-AQ BWC campaign. If the chemical composition of aerosol types happens to be considerably different from that used for the CATCH algorithm development, retraining of the CATCH algorithm for a specific region may be needed. Such studies should also consider the implementation of optical models within HSRL-CH methodology that will use specified microphysical and optical properties within each aerosol type to relate HSRL-measured extinction values to aerosol dry mass chemical composition under variable ambient RH conditions. Future studies should also focus on exploring the other sources of uncertainties such as atmospheric RH profiles and the PBL height, spatial and temporal differences between HSRL retrievals and the ground monitoring sites, and the uncertainties associated with HSRL retrieved aerosol types.

The current study shows that airborne HSRL data can be successfully used for providing information on the vertical profiles of PM_2.5 concentration and chemical composition, information that is crucial for advances in air quality and climate models. Currently, measurements of HSRL extinction and aerosol types do not exist globally. However, such HSRL measurement capabilities are being considered in the next-generation of satellite-based remote sensing systems as part of the NASA Aerosol, Clouds, Convection, and Precipitation (ACCP) Study. Despite having coarser horizontal and vertical resolution than the airborne HSRL measurements, the spaceborne HSRL, in conjunction with passive imagers, will be able to provide superior abilities for satellite-based air quality research and monitoring. We hope the algorithm developed here will be beneficial to the larger scientific community by providing a step towards reducing uncertainties in aerosol chemical composition worldwide.

Highlights.

• New methods for PM_2.5 chemical composition retrievals using HSRL and CATCH

• Improves air quality model predicted PM_2.5 and chemical composition

• Estimates PM_2.5 and chemical composition without air quality model simulations

Data Availability

Supplementary information is available in the online version of this paper. Raw HSRL and model data, processed data, and scripts to generate the figures are accessible through zenodo.org (doi available upon publication).