Satellite Formaldehyde to Support Model Evaluation

Abstract

Formaldehyde (HCHO), a known carcinogen classified as a hazardous pollutant by the United States Environmental Protection Agency (U.S. EPA), is measured through monitoring networks across the U.S. Since these data are limited in spatial and temporal extent, model simulations from the U.S. EPA Community Multiscale Air Quality (CMAQ) model are used to estimate ambient HCHO exposure for the EPA National Air Toxics Assessment (NATA). Here, we employ satellite HCHO retrievals from the Ozone Monitoring Instrument (OMI)—the NASA retrieval developed by the Smithsonian Astrophysical Observatory (SAO), and the European Union Quality Assurance for Essential Climate Variables (QA4ECV) retrieval—to evaluate three CMAQ configurations, spanning the summers of 2011 and 2016, with differing biogenic emissions inputs and chemical mechanisms. These CMAQ configurations capture the general spatial and temporal behavior of both satellite retrievals, but underestimate column HCHO, particularly in the western U.S. In the southeastern U.S., the comparison with OMI HCHO highlights differences in modeled meteorology and biogenic emissions even with differences in satellite retrievals. All CMAQ configurations show low daily correlations with OMI HCHO (r = 0.26 – 0.38), however, we find higher monthly correlations (r = 0.52 – 0.73), and the models correlate best with the OMI-QA4ECV product. Compared to surface observations, we find improved agreement over a 24-hour period compared to afternoon-only, suggesting daily HCHO amounts are captured with more accuracy than afternoon amounts. This work highlights the potential for synergistic improvements in modeling and satellite retrievals to support near-surface HCHO estimates for the NATA and other applications.

1. Introduction

Formaldehyde (HCHO) is an air pollutant affecting allergies and lung function (Mendell, 2007) and cancer risk (Salthammer et al., 2010), leading to its classification as a hazardous air pollutant in the 1990 Clean Air Act, and is included in the National Air Toxics Assessment (NATA) of the United States Environmental Protection Agency (U.S. EPA). While exposure to formaldehyde is greater in indoor air than outdoors (e.g. Daisey et al., 2003; Kinney et al., 2002), ambient (outdoor) concentrations account for about 30% of total exposure (Loh et al., 2007). In addition to its direct health effects, formaldehyde serves an indicator of chemical processes forming ground-level ozone, due to its role as a byproduct of volatile organic compound (VOC) oxidation (Sillman, 1995).

As a primary pollutant, HCHO is directly emitted from plants and soils, wildfires and other wood burning, mobile vehicle sources, industrial processes, and other combustion and chemical activities (Salthammer, 2013; Liteplo et al., 2002). Secondary production from VOC oxidation exceeds primary emissions in most areas of the U.S. (e.g. Luecken et al. 2012), with natural VOC sources greater than anthropogenic sources (e.g. Millet et al., 2006). After photolysis and reactions with the hydroxyl radical (OH), formaldehyde has an atmospheric lifetime of a few hours (Atkinson, 2000; Levy, 1972).

To support health assessment and regulatory activities, the EPA monitors ambient HCHO across the U.S. However, the data coverage provided by ground-based monitors is limited, with a high level of spatial and temporal heterogeneity. In the summer of 2011, the main focus period for this study, ground-based formaldehyde observations made with a 1-hour, 3-hour, or 24-hour duration are available from 132 monitors across the U.S., but most monitors report data for a small fraction of the monitoring period. For example, 85% of monitors only report every 6 – 12 days.

Since 1997, satellite data have been used to detect HCHO atmospheric abundance (Burrows et al., 1999). There are four satellite-based observations of formaldehyde currently available: 1) Ozone Monitoring Instrument (OMI); 2) Global Ozone Monitoring Instruments (GOME-2A and GOME-2B); 3) Tropospheric Monitoring Instrument (TROPOMI); and 4) Smithsonian Astrophysical Observatory Ozone Mapping and Profiler Suite (SAO OMPS). OMI (Levelt et al., 2006), launched on July 15, 2004 on the NASA Aura satellite and provides data with a 13 km by 24 km horizontal resolution with global coverage daily (however, since 2008 a row anomaly has limited this to every other day). The European Space Agency GOME-2A and GOME-2B (De Smedt et al., 2012; Munro et al., 2016) launched on October 19, 2006 on the METOP-A satellite, and September 17 2012 on the METOP-B satellite, respectively. GOME-2A and GOME-2B provide data with global coverage every 1.5 days, with a horizontal resolution of 80 km by 40 km before July 15, 2013 and 40 km by 40 km afterwards. TROPOMI (Veefkind et al., 2012), launched on the Sentinel-5P on October 13, 2017 and provides data with a horizontal resolution of 7.2 km by 5.6 km and daily global coverage. The Smithsonian Astrophysical Observatory (SAO) Ozone Mapping and Profiler Suite (OMPS, González-Abad et al., 2016), launched on October 28, 2011 on the SUOMI-NPP satellite, provides data with a 50 km by 50 km horizontal resolution with global coverage daily.

For our analysis of the summers of 2011 and 2016, OMI offers the highest resolution HCHO dataset available. There are two operational OMI HCHO retrievals with data for these periods: 1) the NASA retrieval developed by SAO (González Abad et al., 2015); 2) the European Union Quality Assurance for Essential Climate Variables (QA4ECV) project retrieval (De Smedt et al., 2018; Zara et al., 2018). QA4ECV replaces the earlier BIRA-IASB HCHO retrieval (De Smedt et al., 2015). These retrievals differ in both their calculations of slant column density and computation of the vertical column density (VCD) using the slant column density and a model-derived air mass factor.

Satellite-based HCHO observations have been used to evaluate global and regional atmospheric chemistry models and their emissions inputs (e.g. Baek et al., 2014; Bauwens et al., 2016; Chaliyaknnel et al., 2019; Ring et al., 2019; Stavrakou et al., 2015; Wiedinmyer et al., 2005; Zhu et al., 2014). Palmer et al. (2003) and Millet et al. (2006) found satellite-based observations of HCHO to be a reliable proxy for VOC emissions. As such, satellite-based HCHO has been used to evaluate isoprene emissions (e.g. Kaiser et al., 2018; Zheng et al., 2015; Zhu et al., 2017a), while Liao et al. (2019) investigated both modeled and satellite-based relationships HCHO and organic aerosols. Satellite-based HCHO has also been used to derive column hydroxyl radical amounts in the remote troposphere (Wolfe et al., 2019) and to identify long-range transport of biomass burning plumes over North America (Alvarado et al., 2020). The most direct application of HCHO satellite data to air quality management has been in the area of characterizing ozone production regimes (e.g. Duncan et al., 2010; Jin and Holloway, 2015; Jin et al., 2017; Jin et al., 2020; Martin et al., 2004; Schroeder et al., 2016; Souri et al., 2020; Sun et al., 2018).

The role of satellite data in evaluating regional photochemical grid models for air quality management has grown in recent years. Evaluation of regional photochemical grid models supports air quality management and public health analysis, as these models are widely used for regulatory agencies to inform human exposure to pollutants. Favorable comparisons are described between these models and satellite retrievals of chemical species such as nitrogen dioxide (Canty et al., 2015; Harkey et al., 2015; Uno et al., 2007), and sulfur dioxide (Wang et al., 2011; Zhao et al., 2013).

This evaluation of regional models with satellite HCHO builds on a large body of work comparing satellite-based HCHO and global models (e.g. Barkley et al., 2011; Bauwens et al., 2016; Chan Miller et al., 2017; Flemming et al., 2015; Huijnen et al., 2010; Liao et al., 2019; Pfister et al., 2008; Stavrakou et al., 2015; Vigouroux et al., 2009; Zeng et al., 2015). Wiedinmyer et al. (2005) found good agreement between HCHO simulated with the Comprehensive Air Quality model with extensions (CAMx) and boundary layer HCHO amounts inferred from GOME-2 over the eastern United States. Wang et al. (2015) found moderate correlations between HCHO observed by the SCIAMACHY satellite instrument and a suite of air quality models over North America. Zhang et al. (2016) found the Weather Research and Forecasting Model with Chemistry with the Model of Aerosol Dynamics, Reaction, Ionization and Dissolution (WRF/Chem-MADRID) consistently overpredicted column HCHO amounts over the southeastern U.S. for three consecutive ozone seasons relative to monthly averaged OMI HCHO. Wang et al. (2017) found the EPA Community Multiscale Air Quality (CMAQ) model captured seasonal and spatial trends in OMI-BIRA HCHO, though overestimated column amounts in the southeastern U.S.

We extend this line of research, using multiple satellite retrievals to evaluate the performance of CMAQ. By evaluating multiple CMAQ configurations against two OMI HCHO retrievals, we highlight uncertainty in model emissions as well as satellite data products. We evaluate the CMAQ model using ground-based observations of HCHO, as well as HCHO VCD from the OMI instrument. CMAQ data are also used to scale satellite HCHO columns to assess the utility and challenges of using satellite-based HCHO to estimate near-surface amounts. The potential of surface HCHO fields derived from a combination of model and satellite data may be used in support of health assessments, including the NATA, and complement ground-based monitoring.

2. Data and Methods

Model, satellite, and ground-based data used in this work cover summer (June-August) in 2011 2016. During the warm months, biogenic precursors, and hence HCHO amounts, over the U.S. are greatest.

2.1. Model configurations

We evaluate three simulations of the EPA CMAQ model (Byun and Schere, 2006; Nolte et al., 2015): 1) 2011 simulation from EPA for the 2011 NATA (referred to here as the NATA simulation); 2) 2016 simulation from EPA using updated model chemistry, and emissions from the 2016 National Emissions Inventory Collaborative (referred to here as the 2016 simulation); 3) 2011 simulation from University of Wisconsin-Madison (referred to here as the UW simulation, Abel et al., 2019).

The NATA configuration used CMAQ version 5.0.2 (doi:10.5281/zenodo.1079898), and was run with 25 vertical layers with a top of approximately 50 hPa, with meteorology input from the Weather Research and Forecasting (WRF) model version 3.4 (Skamarock et al., 2008), with temperature, humidity, and wind constrained to the 12 km North American Model (NAM) analysis via analysis nudging at all levels above the boundary layer. The WRF configuration used has been shown to compare well with measurements and capture daytime vertical mixing (Baker et al., 2013) and more specific evaluation is available elsewhere (U.S. EPA, 2014). The model was run for the whole of 2011 although we focus on summer results. Anthropogenic emissions were based on version 2 of the 2011 National Emission Inventory (NEI; U.S. EPA, 2015a), with day-specific wild and prescribed fire emissions based on satellite and ground-based information about location and size (Baker et al., 2016). Biogenic emissions were calculated using the Biogenic Emission Inventory System (BEIS) version 3.14 with the Biogenic Emissions Landuse Database, version 3 (BELD3; Carlton and Baker, 2011). A GEOS-Chem simulation of 2011 was used to provide lateral boundary inflow of chemical species which vary in space and time (Henderson et al., 2014).

The 2016 simulation used CMAQ version 5.2.1 (doi:10.5281/zenodo.1212601). The model was run with 35 vertical layers with a top of approximately 50 hPa, and meteorology input from WRF version 3.8, configured the same as for input to NATA with analysis nudging to NAM above the boundary layer. Anthropogenic emissions were taken from the 2016 beta version of the National Emissions Inventory Collaborative (NEIC, 2019), and biogenic emissions were calculated using the BEIS version 3.61 with BELD version 4.1. A hemispheric CMAQ simulation with a similar configuration was used to provide time-varying lateral boundary conditions (U.S. EPA, 2019). As with the NATA simulation, the 2016 simulation was run for the whole of 2016 and we focus on the summer months.

The UW simulation employed CMAQ version 5.0.1, and was run with 25 vertical layers, with the top layer at approximately 100hPa. Meteorological input came from WRF version 3.2.1, which was run with temperature, humidity and wind constrained to the North American Regional Reanalysis dataset (Mesinger et al., 2006) via analysis nudging at all levels, as described in Harkey and Holloway (2013). The model was run from May 20 through August 31, 2011, to include eleven days of spin-up. Anthropogenic emissions, as well as emissions from prescribed fires and wildfires, are also based on the 2011 NEI version 2 (U.S. EPA, 2015b). Unlike the NATA and 2016 configurations, the UW configuration also included in-line estimates of NO and NO₂ produced by lightning, and used boundary conditions from the Model for Ozone and Related Chemical Tracers version 4 (Emmons et al., 2010).

Where the NATA and 2016 configurations employed BEIS, biogenic emissions for the UW modeling were calculated with WRF output using the Model of Emissions of Gases and Aerosols from Nature version 2.1 (Guenther et al., 2012). Both MEGAN and BEIS have been shown to have positive biases in isoprene emissions (e.g. Bash et al., 2016; Bauwens et al., 2016; Kaiser et al., 2018; Wang et al., 2017; Zhang et al., 2017). Using the same versions of BEIS and MEGAN employed in this work, Zhang et al. (2017) found MEGAN overestimated isoprene by a factor of 2.6, and BEIS overestimated isoprene by a factor of 1.1—despite the application of satellite-based photosynthetic radiation (PAR). The modeled isoprene overestimates in turn affect ozone (Zhang et al., 2017) and formaldehyde (Wang et al., 2017), to a degree dependent on the model chemical mechanism.

The NATA and UW model simulations utilize the Carbon Bond 5 chemical mechanism with updated toluene and chlorine chemistry (Whitten et al., 2010; Sarwar et al., 2011), while the 2016 configuration used the Carbon Bond 6 mechanism (Emery et al., 2016; Luecken et al., 2019). All three configurations used “AERO6” aerosol chemistry (Nolte et al., 2015), and in-line photolysis; the 2016 configuration used the non-volatiles primary organic carbon version of AERO6. The models are configured with 396 × 246 grid points, with a 12 km × 12 km horizontal resolution. A comparison table for the three simulations is provided in Supplemental Table 1.

2.2. Observational Datasets

Two satellite retrievals of column HCHO are compared with CMAQ simulations, with both data products based on measurements from OMI onboard the Aura satellite. As a polar-orbiting instrument, OMI provides daily global coverage with an equatorial overpass time of approximately 1:40 pm local standard time. OMI provides observations of a number of trace gases, including sulfur dioxide, for which there is one product (Krotkov et al., 2006), nitrogen dioxide, for which there are two products with global coverage and one product with coverage over the U.S. (Boersma et al., 2011; Bucsela et al., 2013; Russell et al., 2011), and HCHO, for which there are two global products that are currently operational (De Smedt et al., 2018; González Abad et al., 2015; Zara et al., 2018). The HCHO retrievals are developed by first calculating a slant column density (SCD), which depends on the viewing geometry, surface albedo, and attenuating properties of the atmosphere. Model-simulated a priori HCHO columns, as well as scattering weights, are integrated through the column to calculate an air mass factor (AMF). The VCD is related to the SCD and AMF by

$$\text{VCD}=\text{SCD}/\text{AMF}$$

For our analysis we employ each OMI HCHO products VCD.

The NASA Level 2, version 3 HCHO product developed by SAO (González Abad et al., 2015), referred to here as OMI-SAO, is publicly available through NASA’s Earth Observing System Data and Information Systems (EOSDIS, https://earthdata.nasa.gov/). OMI-SAO SCD is calculated with a high-resolution solar reference spectrum and a fitting window of 328.5–356.5 nm (Gonzalez Abad et al., 2015). The GEOS-Chem model is used to provide a priori HCHO profiles in the derivation of OMI-SAO VCD, and to provide a climatological reference sector background correction (González Abad et al., 2015). In keeping with applications of OMI-SAO discussed by Zhu et al. (2017a) and Zhu et al. (2017b), we employ the reference sector corrected HCHO, with the understanding that these data have biases and the product is currently under revision (Zhu et al., 2020).

The QA4ECV Level 2 HCHO product (De Smedt et al., 2018; Zara et al., 2018), referred to here as OMI-QA4ECV, is publicly available online (http://www.qa4ecv.eu/ecv/hcho-p/data). In calculations of HCHO SCD, OMI-QA4ECV use Earth radiance spectra over the equatorial Pacific as reference. The OMI-QA4ECV uses a fitting window of 328.5–359.0 nm (Zara et al., 2018). The TM5-MP model is used to provide daily a priori HCHO profiles in the derivation of the OMI-QA4ECV product (De Smedt et al., 2018).

Level-2 OMI-SAO and OMI-QA4ECV data were regridded to a 12 km × 12 km grid over the continental U.S. using the Wisconsin Horizontal Interpolation Program for Satellites (WHIPS, https://nelson.wisc.edu/sage/data-and-models/software.php). As part of the WHIPS gridding, pixels were removed due to quality flags, row anomalies,where the solar zenith angle was greater than 70 degrees, and where the cloud fraction was greater than 0.4.

Measurements from ground-based monitors were used to evaluate estimates of surface layer HCHO. All available formaldehyde data was utilized from the EPA Ambient Monitoring Archive (AMA, https://www3.epa.gov/ttnamti1/toxdat.html#data), which comprises data from a number of monitoring networks, including the Interagency Monitoring of Protected Visual Environments (IMPROVE) network, the National Air Toxics Trends Stations (NATTS), and the Photochemical Assessment Monitoring Stations (PAMS). During the summer of 2011, 4919 HCHO measurements were made at 130 monitor locations, 109 of which are in urban areas. There were 1942 24-hour average observations made at 127 monitor locations, and of those 120 were in urban areas as defined by the 2010 US Census. There were 509 3-hour average observations encompassing the OMI overpass at 1 pm LST at 19 of those monitors, and of those, 18 monitors were located in urban areas, and 158 observations coincided with available OMI HCHO. Monitor locations for data used in this study are given in Supplemental Figure 1.

Measurements used the 2,4-dinitrophenylhydrazine (DNPH) cartridge sampling method, standardized for local conditions as described by Oommen et al. (2020). The DNPH sampling of formaldehyde may have a negative bias (as much as −25%; Hak et al., 2005), and measurements may be affected by interference with water vapor, ozone, or nitrogen dioxide (e.g. Achatz et al., 1999; Arnts and Tejada, 1989; Karst et al., 1993; Rodler et al., 1993). The low bias contributes to an uncertainty of about 22% for mean HCHO measured with DNPH (Dunne et al., 2018). The AMA contains the best publicly available, quality-checked and assured HCHO dataset for this time period, and for consistency and reproducibility we did not alter or adjust these data.

3. Results

In this section, we evaluate the NATA, UW, and 2016 CMAQ model configurations against satellite data, and evaluate near-surface estimates based on scaled OMI-SAO and OMI-QA4ECV HCHO against ground-based monitoring data.

3.1. Column HCHO estimates

CMAQ-modeled HCHO is calculated as mixing ratio at every model layer, and must be converted to column density for comparison with satellite retrievals. To calculate model-simulated VCD, column weighting factors (averaging kernel) from OMI-SAO and from OMI-QA4ECV CMAQ HCHO are applied to CMAQ HCHO layer mixing rations, taken at at 1 pm. The averaging kernel associated with each retrieval, was gridded to the CMAQ model grid using WHIPS. By applying the averaging kernels, we account for the vertical sensitivity of the satellite observation as estimated by each retrieval, as well as the distribution of HCHO in the model column. For consistency in our comparisons between model column and OMI HCHO, we also filter all model data at gridpoints corresponding to where satellite-based HCHO are missing or have been removed (e.g. where satellite observations experience a row anomaly, or the satellite-observed cloud fraction was greater than 0.4).

Results discussed here represent summer 2011 (June-August; two model configurations) and summer 2016 (one model configuration), and are shown in Figures 1 and 2 and Tables 1 and 2. The pixelated and noisy nature of the 12-week average illustrates the impact of random errors in space-based HCHO observations. Both OMI products show maximum summer-average values over the southeastern U.S., with amounts over 12 × 10¹⁵ molecules/cm². In this region, OMI-SAO HCHO shows column amounts of 12–15 × 10¹⁵ molecules/cm² with embedded areas of column amounts > 15 × 10¹⁵ molecules/cm² (top left, Figures 1 and 2). OMI-QA4ECV show wider extents of column amounts between 15–20 × 10¹⁵ molecules/cm², with areas of column amounts > 20 × 10¹⁵ molecules/cm² (top right, Figures 1 and 2). Satellite HCHO patterns are largely consistent over both years.

Figure 1.

Summer (June-August) 2011 average column HCHO in units of molecules/cm², from OMI-SAO and OMI-QA4ECV (top row), as calculated using the NATA configuration using the OMI-SAO and QA4ECV averaging kernels (middle row), and as calculated using the UW configuration using the OMI-SAO and OMI-QA4ECV averaging kernels (bottom row).

Figure 2.

Summer (June-August) 2016 average column HCHO in units of molecules/cm², from OMI-SAO and OMI-QA4ECV (top row), as calculated using the 2016 configuration using the OMI-SAO and QA4ECV averaging kernels (bottom row).

Table 1.

Summer 2011 average formaldehyde from the NATA and UW CMAQ simulations and associated error metrics compared to observations. Unless otherwise noted, model column HCHO amounts and error metrics are given in units of 10¹⁵ molecules/cm²; model surface layer and scaled OMI-SAO HCHO amounts and error metrics are given in micrograms per cubic meter (μg/m³). Ground-based HCHO observations from the Ambient Monitoring Archive (AMA) given for 1 pm LST include all available measurements occurring within one hour of 1 pm LST when coincident OMI observations were available (n = 158).

	Simulated	Observed	RMSE	Mean Error	MFE (%)	Mean Bias	MFB (%)	r
HCHO (1 pm LST)	CMAQ column	OMI-SAO
NATA	6.81	8.79	11.66	4.64	58.89	−1.97	−2.80	0.26
UW	10.15		12.73	6.03	63.66	1.36	20.71	0.27
2016	7.00	9.00	17.10	4.80	57.17	−1.99	0.004	0.13
		OMI-QA4ECV
NATA	8.06	12.68	9.79	6.71	63.08	−4.62	−26.70	0.38
UW	12.23		10.73	7.57	62.98	−0.45	−1.10	0.38
2016	8.43	12.16	9.29	6.35	59.61	−3.74	−14.86	0.32
	CMAQ surface layer	AMA
NATA	3.35	5.15	3.21	2.68	64.60	−1.81	−34.50	0.32
UW	4.66		3.34	2.47	52.54	−0.50	−6.46	0.14
	scaled OMI-SAO	AMA
NATA	4.13	5.15	4.04	3.22	73.79	−0.98	−28.63	0.10
UW	4.09		4.05	3.14	73.29	−1.03	−28.02	0.15
	scaled OMI-QA4ECV	AMA
NATA	4.98	5.15	3.64	2.83	60.72	−0.17	−6.10	0.17
UW	4.95		3.71	2.87	61.34	−0.21	−5.56	0.17
HCHO (24-hour)	CMAQ surface layer	AMA
NATA	2.82	3.82	2.01	1.36	40.98	−1.00	−28.06	0.64
UW	3.22		2.20	1.47	42.72	−0.60	−18.91	0.53

OMI-SAO shows lower HCHO column amounts than OMI-QA4ECV in the western U.S. OMI-QA4ECV average column amounts in the west range from 4–25 × 10¹⁵ molecules/cm², with much of the land area west of 100°W showing amounts > 8 × 10¹⁵ molecules/cm² (top right). OMI-SAO column amounts range from 4–14 × 10¹⁵ molecules/cm², and peak amounts (> 12 × 10¹⁵ molecules/cm²) collocated with elevated terrain, such as the Sierra Nevada in California and Rocky Mountains in Colorado and New Mexico (top left). Differences between the two OMI HCHO products in HCHO abundance over elevated terrain may be due to differing calculations of scattering column density and AMF. As a result of higher HCHO overall, the OMI-QA4ECV product has a somewhat higher signal-to-noise ratio than does OMI-SAO (Supplemental Figure 2).

In both 2011 and 2016 summer-average maximum HCHO occurs over the southeastern U.S. in all datasets (Figures 1 and 2), and the satellite retrievals show elevated “hot spots” of summer-average column amounts in central Colorado, parts of New Mexico, Arizona, and Utah. Because the OMI-QA4ECV product shows higher background amounts overall, these localized high values are less pronounced relative to OMI-SAO (top right, Figures 1 and 2). Observations of elevated summer HCHO column amounts over California vary by retrieval: OMI-SAO HCHO shows amounts over 12 × 10¹⁵ molecules/cm² near the California-Nevada border and along the Pacific coast south of San Francisco (top left, Figures 1 and 2); QA4ECV HCHO shows amounts over 15 × 10¹⁵ molecules/cm²in the north and central parts of the state (top right, Figure 1 and 2).

All three CMAQ configurations (center and bottom rows, Figure 1; bottom row, Figure 2) capture the observed spatial patterns, including elevated HCHO in the southeastern U.S. relative to other parts of the domain. However, the relative bias of the model to the satellite depends on the model configuration and the HCHO satellite product.

The greatest differences among the datasets is observed in low HCHO environments such as over water bodies and much of the western U.S., i.e. “background” HCHO. The OMI-QA4ECV product estimates higher amounts in these areas relative to OMI-SAO (top rows, Figures 1 and 2). The CMAQ configurations’ HCHO amounts are lower than both satellite retrievals, but the the differences are more pronounced when comparing CMAQ with OMI-QA4ECV. The low background column HCHO amounts in CMAQ, relative to both retrievals, may be due to an underestimation of sources impacting the West Coast as well as local sources in the western part of the U.S. Differences in background amounts between OMI HCHO retrievals also reflect uncertainty in retrieval background HCHO estimates. As González Abad et al. (2016) describe, the choice of model used for the reference sector correction has a significant impact on the retrieval, one that may be temporal and regional in scope: they estimate biases in GOES-Chem (used in the derivation of OMI-SAO HCHO) contribute 16–40% of the uncertainty in AMF calculations, and the GOES-Chem climatology used to calculate the reference sector column HCHO directly accounts for ~10% of the total error in VCD.

The NATA and 2016 simulations agree well with the OMI-SAO product in most regions (average bias of −2.8% and 0%, respectively, Table 1), but exhibit a pronounced low bias when OMI-QA4ECV (NATA average bias −26.7%, 2016 average bias −15%; Table 1). The NATA and 2016 configurations (when processed with retrieval-appropriate averaging kernels) show lower column HCHO over the southeastern U.S. than either OMI-SAO or OMI-QA4ECV. The OMI-QA4ECV product estimates somewhat higher column HCHO amounts in the Southeast U.S. relative to OMI-SAO, so this difference is more pronounced in comparing NATA and 2016 CMAQ with OMI-QA4ECV. NATA and 2016 CMAQ HCHO column amounts are lower than both satellite retrievals, but the difference is more pronounced when comparing NATA CMAQ with OMI-QA4ECV (Figure 1, center right).

The UW CMAQ simulation for 2011 agrees well with the OMI-QA4ECV product on average across the domain (average bias −1.1%), but exhibits a pronounced high bias when OMI-SAO (average bias +20.7%; Table 1). However, the low domain average bias in the comparison of UW CMAQ versus OMI-QA4ECV arises from an underestimation of HCHO VCD in background areas and an overestimation over the Southeast. The overestimate of HCHO in the Southeast leads to a large extent of simulated formaldehyde column amounts above 15 × 10¹⁵ molecules/cm² over the southeastern U.S. in Figure 1 (bottom left). The underestimation of background column HCHO is consistent with the NATA CMAQ simulation, and consistent with comparisons of both OMI-SAO and OMI-QA4ECV (Figure 1, bottom row).

The different biases in CMAQ-modeled column HCHO for 2011 relate to differing temperatures, as well as differing biogenic emissions from models using differing emission factors. Compared with near-surface temperatures from Real-Time Mesoscale Analysis (RTMA; De Pondeca et al., 2011), the UW configuration temperatures have a slight positive bias (~4% mean biases in daily average temperature), while the NATA configuration temperatures exhibit a weak negative bias (< 1% mean biases in daily average temperature; Supplemental Figure 3). Across the southeastern U.S., isoprene emissions from MEGAN used in the UW simulation are about twice as great as those from the BEIS estimates used in the NATA simulation; at isoprene monitor locations, UW isoprene amounts are over 3 times greater than those estimated by the NATA. The UW configuration exhibits a positive bias in daily average isoprene concentrations relative to daily isoprene surface observations of 98%, where the NATA configuration has a negative mean bias of −17% (Supplemental Figure 4). Isoprene is the most abundant biogenic precursor of HCHO, and the large differences between modeled column HCHO amounts between the two CMAQ configurations may be attributed to the higher isoprene emissions and hotter temperatures in the UW CMAQ simulation, relative to the NATA CMAQ and available observations.

Both the NATA and UW 2011 configurations perform similarly in the representation of the daily variability of column formaldehyde relative to satellite retrievals: the correlations (r) with OMI-SAO are 0.26 and 0.27 (NATA and UW, respectively), and the correlations with OMI-QA4ECV are 0.38 for both simulations (Table 1). Correlations between month-averaged observations and simulated column amounts are significantly higher: 0.53 (NATA) and 0.54 (UW) when compared to OMI-SAO; 0.73 (NATA) and 0.71 (UW) when compared with OMI-QA4ECV. The 2016 CMAQ simulation also shows a higher daily r when compared with OMI-QA4ECV (r= 0.32) than OMI-SAO (r = 0.13), as well as higher correlations between month-averaged observations and OMI-QA4ECV (r = 0.70) and OMI-SAO (r = 0.35). These results agree with Duncan et al. (2014), who suggest OMI HCHO should be averaged over at least 6 weeks in order to minimize the effects of random measurement errors.

3.2. Near-surface HCHO estimates

We also use CMAQ-simulated concentrations of HCHO to scale OMI-SAO and OMI-QA4ECV HCHO to near-surface amounts for 2011 conditions, using the same technique as Liu et al. (2004), van Donkelaar et al. (2006), and Lamsal et al. (2008). To calculate scaled-surface amounts, we take the ratio of CMAQ model column amounts to surface layer amounts, both at 1 pm LST to correspond with the OMI overpass. Though model surface layer amounts are taken directly from CMAQ output, as noted above, the CMAQ column HCHO amounts are calculated by applying the OMI-SAO or OMI-QA4ECV averaging kernel as appropriate.

Across all 24-hour duration (daily average) observations (n = 1942), the average observed HCHO amount is 3.82 μg/m³. Across all available 3-hour observations coincident with OMI HCHO (n = 158), monitored HCHO is ~35% higher, at 5.15 μg/m³ (Table 1). The increase in concentrations is consistent when all available (n = 509) observations at overpass time are considered ~38% higher, Supplemental Table 2). The higher values during the mid-day overpass arise from photochemically produced HCHO. At monitor locations, both CMAQ simulations also report higher average values at 1 pm compared to daily averages.

In general, errors and biases for the two 2011 CMAQ configurations are lower when daily averages are considered than when calculated for 1 pm only. For 1 pm the model RMSE is 3.21 μg/m³ (NATA) and 3.34 μg/m³ (UW); the mean error is 2.68 μg/m³ (NATA) and 2.47 μg/m³ (UW). These metrics are reduced by over 20% when all times are considered (Table 1). When comparing with daily average observations, the mean bias is reduced by over a third for the UW simulation and is reduced by almost half in the NATA simulation (Table 1). Temporal agreement between NATA and UW CMAQ configurations and monitor data is also higher for daily averages (NATA r = 0.64, and UW r = 0.53) and lower when considering only 1 pm values (NATA r = 0.32, UW r = 0.14; Table 1).

In aggregate, the improved agreement over a 24-hour period compared to afternoon-only suggests that the model configurations capture daily HCHO amounts with more accuracy than afternoon amounts. This behavior may be due to a number of factors that impede model resolution of the diurnal cycle. For example, photochemical production of HCHO from biogenic precursors is highest during the daytime hours, so the lower afternoon agreement may suggest the model captures primary HCHO (higher amounts at night relative to other sources) better than secondary HCHO (higher amounts in daytime). For example, we find while observed HCHO amounts peak in the late morning, the NATA configuration shows peak near-surface amounts in the late afternoon, and the UW configuration shows peak amounts near the early afternoon OMI overpass (Supplemental Figure 5). Additionally, the model configurations show improved performance overnight, for the 3 am – 6 am average (Supplemental Table 2), albeit with a (smaller) negative bias. The persistent low model bias across the diurnal cycle (mean fractional biases between −18 and −80%; Supplemental Table 2, Supplemental Figure 5) is likely an artifact of comparing model-simulated HCHO amounts at a 12 km × 12 km horizontal resolution to point-based observations (e.g. Blond et al., 2007; Tan et al., 2015). Though the aggregated results presented above do not factor in potential biases caused by differences in monitor data availability, we find similar results in an analysis using only co-located and coincident (on a daily basis) observations (Supplemental Figure 6).

The location of ground-based observations used in model evaluation--whether the observation is made in an urban or high-HCHO emitting location, or not—can also inform sources of bias. Of ground-based observations at the OMI overpass time, only one monitor is located in a non-urban area, which does not allow for a robust analysis of near-surface, diurnal model performance across urban-rural gradients. However, for monitors in urban areas, we find the 1pm NATA and UW mean biases are double that of the 24-hour average (Supplemental Table 3), similar to model performance when all monitor data are aggregated (Table 1). Additionally, NATA and UW show smaller errors and biases in 24-hour average HCHO across non-urban locations than urban (Supplemental Table 3), which suggests the model may perform better (relative to point-based observations) where HCHO emissions are lower. However, when near-surface model performance is compared to local, direct HCHO emissions estimates on the model grid, we find a consistent negative bias in 24-hour average HCHO with no trend relative to changing emissions; at mid-day, model biases tend to decrease (become less negative) with increasing local emissions (Supplemental Figure 7). These results reinforce that model performance is greater on a daily basis than in the afternoon, consistently so across locations, although with confounding effects of model resolution.

An evaluation with ground-based observations allows us to compare the model-simulated surface HCHO with scaled satellite data. As discussed above, the OMI-SAO and OMI-QA4ECV HCHO columns are used to estimate near-surface values, using the ratio of column-to-surface HCHO from CMAQ. Because each model configuration provides a different ratio, for each OMI HCHO product we have two alternate versions of scaled HCHO, one scaled with NATA and one scaled with UW (Figure 3). As shown in Figure 3, scaled satellite HCHO provides the same broad spatial coverage as the model, and shows similar spatial patterns as near-surface model estimates, model column estimates, and OMI HCHO observations (Figures 1 and 2).

Figure 3.

Summer 2011 average near-surface formaldehyde amounts, in micrograms per cubic meter, from the NATA configuration (top left) from the UW configuration (top right), from OMI-SAO scaled with NATA HCHO (center left), from OMI-SAO scaled with UW HCHO (center right), from OMI-QA4ECV scaled with NATA HCHO (bottom left), and from OMI-QA4ECV scaled with UW HCHO (bottom right). Model values used correspond to ~1 pm LST.

Summer-average scaled OMI-SAO HCHO calculated with the NATA CMAQ configuration (Figure 3, center left) ranges from ~0.5 – 24 μg/m³ and averages 2.35 μg/m³ over land. The data show average HCHO amounts over 4 μg/m³ in the southeastern U.S., the central valley of California, and isolated locations in other western states such as Wyoming and Utah. Summer-average scaled OMI-SAO HCHO calculated with the UW CMAQ configuration (Figure 3, center right) shows a similar spatial pattern, with values ranging from ~0.4 – 21 μg/m³ and an average value of 2.21 μg/m³ over land. These data also show average HCHO amounts over 4 μg/m³ in the southeastern U.S., the central valley of California, although over a lesser spatial extent than OMI-SAO scaled with NATA (Figure 3, center row).

Summer-average scaled OMI-QA4ECV HCHO calculated with the NATA and UW configurations shows a similar spatial pattern as scaled OMI-SAO HCHO, but with broader areas of higher concentrations. Scaled OMI-QA4ECV shows amounts over 4 μg/m³ extending into the Midwest and Mid-Atlantic (Figure 3). In the western U.S., average scaled OMI-QA4ECV HCHO amounts are > 1 μg/m³ (Figure 3). OMI-QA4ECV scaled with NATA ranges from 0.55 to ~11 μg/m³and averages 3.05 μg/m³ over land; OMI-QA4ECV scaled with the UW configuration ranges ~0.5 – 22 μg/m³ and averages 2.88 μg/m³.

In the western U.S., both model configurations show a large area of low near-surface HCHO (< 1 μg/m³, Figure 3, top row). Scaled-to-surface OMI HCHO amounts are greater than 1 μg/m³ over the same region (Figure 3, middle and bottom rows). These patterns are consistent with the higher HCHO column values from OMI versus the models in the western U.S. shown in Figure 1.

All six estimates of surface HCHO capture HCHO maxima in the southeastern U.S., with values over 4 μg/m³. Even though the UW CMAQ configuration shows higher column HCHO in the Southeast (versus NATA CMAQ and OMI, Figure 1) and higher surface HCHO in the Southeast (versus NATA, top row of Figure 3), scaled OMI with UW ratios (Figure 3, middle and bottom right) shows lower values in the Southeast than scaled OMI with NATA ratios (Figure 3, middle and bottom left). This suggests that the UW CMAQ configuration has a lower surface-to-column HCHO ratio than NATA, even as it has higher column and higher surface HCHO values. While differences between the NATA and UW simulations’ near-surface HCHO amounts are strongly influenced by their differing biogenics emissions calculations (BEIS in the NATA; MEGAN in the UW; Supplemental Figure 4), differences in modeled surface-to-column HCHO ratios are more affected by differing meteorology, such as temperature (e.g. Supplemental Figure 3) and cloudiness, which impact vertical mixing and photolysis.

The skill of these six spatially continuous HCHO datasets may be assessed by comparison with ground-based observations in Table 1. Both ground-based and satellite-based HCHO observations are non-continuous: ground-based HCHO observations are typically conducted every 3 or 6 days; for any given day OMI-SAO HCHO at a monitor location may be screened for cloudiness, row anomaly, etc. For our assessment, we calculate error metrics based on all available coincident observations across monitor locations, which varies from zero available locations up to 15 monitor locations on a given day (Supplemental Figure 8). Error metrics available for each monitor location are provided in Supplemental Tables 4 – 9.

Over the summer period, at the 1 pm overpass time, ground-based data average 5.15 μg/m³ (Table 1). The six mid-day estimates of surface HCHO have mean fractional biases between −5% and −35%. OMI-SAO scaled with UW HCHO has a larger bias ((−1.03 μg/m³) than either near-surface UW HCHO (−0.50 μg/m³) or OMI-QA4ECV scaled with UW HCHO (−0.21) μg/m³. OMI scaled with NATA HCHO shows a smaller bias than near-surface NATA HCHO (−0.98 μg/m³ and −1.12 μg/m³ in scaled OMI-SAO and OMI-QA4ECV data versus −1.81 μg/m³ modeled). However, the mean errors in scaled OMI HCHO are slightly greater than the model configurations’ surface layers (Table 1). These results suggest that the magnitude of the differences between scaled OMI HCHO and observations are larger than the differences between model simulations and observations.

The correlations of each scaled OMI HCHO product are smaller than correlations for model-simulated column HCHO: the correlations between modeled column HCHO and OMI-SAO HCHO are ~0.26, correlations between scaled OMI-SAO and observations are 0.10 and 0.15 (scaled with NATA and UW, respectively); the correlations between modeled column HCHO and OMI-QA4ECV HCHO are 0.38, while the correlations between scaled OMI-QA4ECV HCHO and observations are 0.17 (Table 1). Correlations among all early-afternoon datasets and observations are low, less than 0.4. These low correlations may reflect both the noisiness of satellite-based HCHO observations, as well as their availability. Combined, these results suggest that model-simulated HCHO better captures daily variances, and use of OMI HCHO, including scaling, is best suited for estimating exposures on seasonal or longer timescales.

4. Discussion

While all datasets used in our analysis have uncertainties, discrepancies among them have the potential to not only improve models, but strengthen the application of satellite data to air quality health. Health impacts assessments of HCHO and other air toxics are primarily informed by ground-based observations and by chemical transport modeling, which are limited by geographical and temporal sampling and by uncertainties in inputs and parameterizations. Satellite HCHO data has a wide spatial coverage that complements the sparse spatial and temporal coverage of ground-based measurements used to evaluate modeling. These near-global, daily space-based observations offer a valuable opportunity for both model evaluation and for informing exposure to a significant cancer risk (e.g. Strum and Scheffe, 2016).

The application of satellite-based observations benefits from the availability of multiple retrievals from the same instrument, as well as observations from multiple instruments. In this study, we evaluated the agreement of three CMAQ simulations over two study years against two retrievals from OMI: the OMI-SAO HCHO and QA4ECV OMI HCHO products. We find all three CMAQ configurations agree in many respects with satellite-derived column HCHO for the summers of 2011 and 2016. Agreement with OMI HCHO varies by retrieval and model configuration, especially meteorology and biogenic emissions. All models showed higher levels of agreement with OMI-QA4ECV daily and monthly variability, and up to 6 × higher agreement in capturing monthly variability versus daily. Monthly average correlations against OMI-SAO in this study are lower than correlations reported between OMI-SAO HCHO and WRF/Chem by Zhang et al. (2016; r = ~0.7–0.8), which were over three ozone seasons.

In this and past work, model-simulated column HCHO amounts show biases of opposite signs: compared to OMI-SAO, NATA CMAQ has a negative bias, while the UW CMAQ has a positive bias; all three model configurations (NATA, UW, 2016) have a negative bias relative to OMI-QA4ECV. The negative biases in our CMAQ column HCHO is similar to the negative bias Chan Miller et al. (2017) found in comparing GEOS-Chem to OMI-SAO HCHO in the southeastern U.S., and that Chen et al. (2019) found in GEOS-Chem HCHO in the free troposphere. The positive bias in UW CMAQ column HCHO is similar to the positive bias Wang et al. (2017) found in comparing CMAQ to OMI-BIRA HCHO in summer months—where the model underestimates column amounts in the western U.S., but greatly overestimates amounts in the southeast.

Our differing CMAQ biases indicate the effects of differing model estimates of biogenic isoprene, e.g. from MEGAN or BEIS, which in turn affect formaldehyde amounts. Carlton and Baker (2011) found model simulations using MEGAN tended to have a positive bias in isoprene, while using BEIS resulted in a negative bias in isoprene. Wang et al., (2017) and Zhang et al. (2017) found both BEIS and MEGAN had positive biases in biogenic isoprene emissions, with BEIS having a smaller positive bias than MEGAN. Using OMI HCHO to constrain isoprene emissions, Bauwens et al. (2016), and Kaiser et al. (2018) found MEGAN has a high bias in the southeastern U.S. Our results generally agree with previous work: over the southeastern U.S. the UW simulation with MEGAN has isoprene emissions much greater than those from the NATA simulation with BEIS, and overall, the use of MEGAN in the UW configuration contributes to a large positive bias in isoprene, while the use of BEIS in the NATA configuration contributes to a smaller, and negative bias in isoprene.

In past studies comparing MEGAN and BEIS, meteorological conditions are typically controlled to isolate the differences in biogenic emissions modeling (e.g. Carlton and Baker, 2011; Wang et al., 2017; Zhang et al., 2017). In this work, differences in simulated formaldehyde amounts, and biogenic precursor emissions, are entangled with differing meteorological inputs. We find the NATA and UW configurations have similar biases in near-surface temperature as they do for isoprene concentrations compared to observations. Other meteorological variables that will impact the distribution and amount of simulated HCHO (as well as precursor biogenic emissions) include vertical mixing, humidity, and insolation (cloudiness).

For the summer of 2011, model-simulated near-surface HCHO shows low biases relative to ground-based observations across all hours, including during the early afternoon OMI overpass. Our results are consistent with work by Marvin et al. (2017), who found the CB05 mechanism underestimated HCHO production from isoprene, and with work by Luecken et al. (2019), who found July 2011 daily average HCHO was underestimated in CMAQ with the newer CB6 chemical mechanism. Both summer 2011 model configurations presented here have a low bias that persists across the diurnal cycle, consistent with measurements and modeling by Cleary et al. (2015). The negative model bias at night, when primary sources dominate, may be an artifact of model resolution, or of unresolved long-range transport (e.g. Alvarado et al, 2020). We find mean errors in model-simulated HCHO are smaller on a daily basis than for the OMI overpass time alone. This suggests further work is needed to clarify the diurnal cycle of HCHO and HCHO-forming reactions, particularly as Schroeder et al. (2016) found ~80% of diurnal HCHO variability is caused by precursors. We note our near-surface analysis is limited by monitor locations, which are primarily urban, and as such model performance with respect to ground-based monitors may improve with better characterization of urban emissions.

As we consider different OMI HCHO datasets, we also note differences and uncertainties among retrieval estimates of background amounts and of amounts over elevated terrain. For example, Gonzalez Abad et al. (2015) note the uncorrected OMI-SAO VCD has uncertainties ranging 45–105%; the reference-sector corrected VCD used in this work has a bias correction, but the formulation of the reference sector column HCHO introduces additional uncertainties. Zhu et al. (2020) found OMI-SAO HCHO biased as low as −30.9% when column amounts are large, and as biased as high as 194.6% when column amounts are low. As such, the OMI-SAO product is being revised to reduce biases related to the reference sector correction and a priori vertical profiles used in the retrieval (Zhu et al., 2020).

The notable differences in some parts of the central and western U.S. between OMI HCHO products highlight the need for satellite-based isoprene emissions approaches (e.g. Kaiser et al., 2018; Millet et al., 2008; Palmer et al., 2006; Stavrakou et al., 2015) that consider temporal aggregation, and potentially coarser scales than used here, to avoid physically unrealistic “hot-spots” of biogenic VOC emissions.

Both OMI HCHO retrievals discussed in this work capture a regional maximum in the southeastern U.S., similar to that found by Zhu et al. (2016) in their evaluation of satellite-based observations in the summer of 2013. Zhu et al. (2016) also found OMI-SAO is biased low relative to aircraft observations in the southeastern U.S., and suggest improvements to the air mass factors used in the retrieval. For satellite retrievals of trace gases, the calculated air mass factor, which is sensitive to surface albedo, atmospheric scattering, viewing geometry, and an a priori vertical gas distribution, strongly affect estimated VCDs (e.g. Millet et al., 2006; Palmer et al., 2001). Particular attention may be given to the a priori HCHO profile, as Wolfe et al. (2016) found models underestimated background HCHO. Additional uncertainty in the utilization of OMI HCHO is introduced by instrument degradation, as discussed by Marais et al. (2012).

Satellite-derived, scaled concentrations offer the potential to complement ground-based observations and model data. Scaled-surface OMI HCHO may be a useful supplement for surface air toxics assessments. However, these scaled estimates inherit errors and biases from the models—for the summer of 2011 we find errors in scaled-surface HCHO are less than those seen in model column estimates, but greater than those seen in near-surface model estimates (e.g., scaled-surface RMSE of 3.64 – 4.05 μg/m³; CMAQ column RMSE of 9.79 – 12.73 × 10¹⁵ molecules/cm²; CMAQ surface RMSE of 3.21 – 3.31 μg/m³). Scaled estimates also inherit the spectral uncertainty of satellite observations of HCHO, as well as limitations in temporal coverage. Our results suggest the use of scaled satellite HCHO has the strongest utility in long-term averages, such as seasonally, as shown here, or on an annual basis. These longer time periods are consistent with cumulative exposure assessment; while shorter-term (~daily) “peak” HCHO exposure contributes to accumulated exposure, it is more often associated with indoor amounts (e.g. Checkoway et al., 2019; Swenberg et al., 2013). Scaled-surface HCHO may be improved with higher resolution HCHO retrievals that better capture the effects of urban-rural gradients in biogenic precursors.

Future work will benefit from observations with a higher spatial and/or temporal resolution, including more ground-based observations at non-urban locations in the western U.S. This may also include model comparison with the SAO OMPS HCHO product from the Suomi National Polar-orbiting Partnership (Suomi NPP) satellite, and with HCHO observations from TROPOMI, onboard the Sentinel-5 Precursor satellite, which have overpasses close to OMI. With higher spatial resolution and less noise compared to OMI, TROPOMI HCHO is well suited for isolating emissions sources and transport (e.g. Alvarado et al., 2020). With the planned launch of the TEMPO instrument in 2022, from which hourly HCHO over the United States will be available, there is potential to greatly increase the utility of satellite observations for HCHO model evaluation and inform air toxics assessments. Given the discrepancies between satellite-based HCHO products, ground-based spectral measurements of HCHO VCD such as those made by PANDORA instruments (Spinei et al, 2018) would provide a critically needed validation for urban and rural environments.

Other satellite-based measurements may benefit from knowledge gleaned in the use of satellite-based HCHO: tropospheric methanol (Beers et al., 2008; Cady-Pereira et al., 2012), also classified as a hazardous air pollutant, and glyoxal, which may be used to better characterize VOCs and related chemistry, including HCHO (e.g. Chan Miller et al., 2017; Kaiser et al., 2015).

Key Points:

• all OMI HCHO satellite products show higher background HCHO than any of the CMAQ model configurations

• biogenic emissions estimates are an important factor in model agreement with satellite-based HCHO observations

• expanded ground-based observations, especially in remote areas of the western US, would aid in evaluating models and satellite retrievals