1 Introduction

Atmospheric CO2 concentrations have increased more than 50% compared to those pre-industrialization due to increasing fossil fuel consumption (Friedlingstein et al. 2022). Various efforts have been made to reduce global warming induced by CO2 emissions. The Kyoto Protocol of the United Nations Framework Convention on Climate Change) (UNFCCC), adopted on December 11, 1997, aims to reduce the emissions of six types of greenhouse gases, including CO2 (unfccc.int/process-and-meetings/the-kyoto-protocol/what-is-the-kyoto-protocol/). The Paris Agreement, adopted on December 12, 2015, aims to keep the global average temperature increase within 2 °C of that before industrialization by reducing CO2 emissions (unfccc.int/process-and-meetings/the-paris-agreement). Although East Asia is the third-largest source region of CO2 after North America and Europe, the number of surface CO2 observations in East Asia is relatively small compared to that in North America and Europe (Moran et al. 2018), which makes the estimation of surface CO2 fluxes in East Asia highly uncertain (Stephens et al. 2007).

To decrease the uncertainties associated with surface CO2 flux estimation in East Asia, many studies have been conducted using various models. Jing et al. (2018) simulated global CO2 concentrations using the Goddard Earth Observing System Chemistry (GEOS-Chem) model and compared them with observed column-averaged CO2 (XCO2) concentrations of Greenhouse Gases Observing Satellite (GOSAT) and total carbon column observing network (TCCON). CarbonTracker, developed by the National Oceanic and Atmospheric Administration (NOAA) Earth System Research Laboratory (ESRL), has also been used to estimate surface CO2 fluxes over East Asia and globe (Kim et al. 2014a, 2014b, 2017, 2018; Park and Kim 2020; Cho and Kim 2022). Although a nesting domain is used for more detailed simulations of CO2 flux or concentrations over East Asia using global models (i.e., GEOS-Chem and CarbonTracker) (Shim et al. 2013; Kim et al. 2014a), it is difficult to produce high-resolution simulations at a regional scale using global models because of limitations in resolution size and computational resources. Therefore, studies using regional models (e.g., Community Multiscale and Air Quality (CMAQ) model and the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) have been conducted to simulate high-resolution CO2 concentrations. CMAQ is an offline model that performs chemical forecasts using atmospheric fields as an input, whereas WRF-Chem is an online model that simultaneously conducts atmospheric and chemical forecasts. Li et al. (2017) simulated the CO2 concentration in East Asia using the CMAQ and compared the results with observed XCO2 concentrations of GOSAT. Dong et al. (2021) simulated the CO2 concentrations in China using the WRF-Chem and diagnosed the spatial and temporal variations of CO2. Zheng et al. (2019) compared the XCO2 concentrations produced by WRF-Chem with those produced by Orbiting Carbon Observatory 2 (OCO-2) to evaluate CO2 emissions from power plants located in the United States (US). To evaluate the most suitable setup in the WRF-Chem to simulate CO2 concentrations in the US, sensitivity experiments for physical processes and CO2 emission inventory have been performed (Martin et al. 2019; Feng et al. 2016, 2019).

Among the regional models, WRF-Chem considers the interactions between atmospheric variables and chemical components as the interactions occurred in the real atmosphere. To simulate the high-resolution regional CO2 concentrations using WRF-Chem, several components including the emission inventories, atmospheric initial and boundary conditions, initial and boundary conditions of CO2, and physical parameterizations in WRF-Chem need to be considered. For the emission inventory, anthropogenic and oceanic CO2 concentrations can be provided from each emission inventory data. Initial and boundary conditions of CO2 can be provided by the global models such as GEOS-Chem and CarbonTracker. For the physical parameterizations in WRF-Chem, Díaz-Isaac et al. (2018) investigated the impact of physical parameterizations and initial conditions on simulated CO2 concentrations in the US.

For the biogenic CO2 concentration, the Vegetation Photosynthesis and Respiration Model (VPRM) combined with WRF-Chem can be used. The VPRM is a model that calculates gross ecosystem exchange (GEE) and respiration (R) to simulate biogenic CO2 concentrations (Ahmadov et al. 2007). To calculate the GEE and R in the VPRM, four parameter values are required (Mahadevan et al. 2008), which should be optimized for the experimental area (Hilton et al. 2013). The parameters of the VPRM model need to be optimized using eddy-covariance observations for each vegetation type present in the experimental region (Mahadevan et al. 2008). The VPRM parameters optimized for the US, Europe, and tropical regions are provided in WRF-Chem. Several studies have attempted to optimize parameters for their experimental region. Park et al. (2018) performed parameter optimization for the downtown area of California, USA, and Hilton et al. (2013) performed parameter optimization in North America. In East Asia, Dayalu et al. (2018) performed parameter optimization for 2005–2009 using eddy-covariance observations for the same period in China and Korea. Li et al. (2020) suggested using the VPRM parameter table modified from Hilton et al. (2013) for northeastern China and analyzed the uncertainties related to parameters for northeastern China. Park et al. (2020) simulated CO2 with a VPRM model using the parameters optimized for the US in WRF-Chem and showed reasonable CO2 simulation results over Korea. Thus, to estimate CO2 over parts of Asia, Li et al. (2020) used the modified VPRM parameters for the US, while Park et al. (2020) directly used the VPRM parameters for the US. Although Dayalu et al. (2018) used the VPRM parameters optimized for Asia to estimate CO2 over East Asia, the VPRM parameters were optimized using past observations during 2005–2009. Therefore, to simulate recent CO2 concentrations over East Asia, suitable VPRM parameters in WRF-Chem need to be investigated.

In addition to the biogenic emission inventory, atmospheric variables (e.g., wind) affect the simulated atmospheric CO2 concentrations because the distribution and concentration of simulated atmospheric CO2 are more affected by wind transport than by reactions with other chemicals in the atmosphere (Nasrallah et al. 2003). Seo and Kim (2023) showed that enhanced atmospheric variables by meteorological data assimilation have large impact in improving the accuracy of CO2 concentration simulations in East Asia. In previous studies that simulated CO2 concentrations using WRF-Chem, various atmospheric fields were used as the initial and boundary conditions of the model. As WRF-Chem is a regional model, initial atmospheric conditions and atmospheric boundary conditions greatly affect the simulation results and forecast error (Kim and Kim 2021). To simulate CO2 concentrations using WRF-Chem in the US, Hu et al. (2020) used the National Center for Environmental Prediction-Department of Energy (NCEP/DOE) R2 data (Kanamitsu et al. 2002); Chen et al. (2019) and Feng et al. (2019) used the European Center for Medium-Range Weather Forecasts Interim Reanalysis (ERA-Interim; Dee et al. 2011); and Martin et al. (2019) conducted an experiment using NCEP North American Regional Reanalysis (NARR; Mesinger et al. 2006). To simulate CO2 concentrations using WRF-Chem in China, Li et al. (2019, 2020) used NCEP/DOE R2 data, and Liu et al. (2018) used ERA-Interim data. Ballav et al. (2012) and Park et al. (2020) simulated and verified CO2 concentrations in Tokyo and Korea using WRF-Chem, respectively, using the final analysis (FNL) of NCEP.

Although various atmospheric reanalysis fields have been used as the initial and boundary conditions in WRF-Chem for various experimental areas, no previous studies have investigated the sensitivity of simulated CO2 concentrations with respect to atmospheric reanalysis data, especially focusing on East Asia. In addition, the sensitivity of simulated CO2 concentrations with respect to VPRM parameters has not been investigated in East Asia. Therefore, to appropriately simulate high-resolution CO2 concentrations in East Asia, the effects of atmospheric conditions and VPRM parameters on simulating CO2 concentrations over East Asia need to be evaluated using WRF-Chem. Therefore, in this study, sensitivity studies using WRF-Chem were conducted to find the most appropriate experimental framework for simulating high-resolution CO2 concentrations in East Asia.

Section 2 presents the model description and observations, Section 3 presents the results, and Section 4 provides a summary and conclusions.

2 Methods

2.1 Model

WRF-Chem is a chemical transport model based on the WRF developed by NCAR (Grell et al. 2005). The WRF-Chem version 4.1.5 was used in this study. WRF-Chem is a fully compressible non-hydrostatic model with dynamic and chemical parts integrated together in each time step (Powers et al. 2017). Because the atmospheric and chemical parts are fully coupled (“online model”), both parts are transported on the same grid, with the same physics and transport system. Because the chemical and dynamic parts affect each other, the “online” calculations can suitably simulate chemicals in the atmosphere (Grell et al. 2005).

The physical schemes used in WRF-Chem are the shortwave and longwave scheme (RRTMG (Iacono et al. 2008)), microphysics scheme (WRF Single-moment 6-class Scheme; WSM 6-class (Hong and Lim 2006)), cumulus parameterization scheme (Grell 3D Ensemble Scheme (Grell and Dévényi 2002)), planetary boundary layer physics scheme (Yonsei University (YSU) scheme (Hong et al. 2006)), surface layer scheme (Revised MM5 (Jiménez et al. 2012)), and land surface scheme (Unified Noah Land Surface model (Tewari et al. 2004)).

In WRF-Chem, CO2 is subdivided into four components:

$${CO}_{2}\_TOTAL={CO}_{2}\_ANT+{CO}_{2}\_BIO+{CO}_{2}\_OCE+{CO}_{2}\_FIRE,$$
(1)

where CO2_ANT is anthropogenic CO2, CO2_BIO is biogenic CO2, CO2_OCE is oceanic CO2, and CO2_FIRE is CO2 due to fire. Because CO2 is treated as an inert gas in WRF-Chem, each component does not affect the other components during integration (Zheng et al. 2019).

For CO2 simulations in WRF-Chem, the emission input data of anthropogenic, biogenic, oceanic, and fire emissions, and background CO2 are required. Anthropogenic and oceanic emission input data were generated from inventory data, as described in Section 2.1.1. The background CO2 data was from that specified in CarbonTracker 2019 (CT2019). For biogenic emission, VPRM was used as described in Section 2.1.2. In this study, fire emission was not considered because the fire inventory showed few fire events during the experimental period over East Asia. The CO2 concentrations were predicted by integrating the WRF-Chem with emission input data and atmospheric and chemical initial and boundary conditions.

2.1.1 Anthropogenic and ocean emission inventory

The Emission Database for Global Atmospheric Research (EDGAR) and the open-source data inventory for anthropogenic CO2 (ODIAC) are widely used as anthropogenic emission inventories (Zheng et al. 2020). The EDGAR inventory generally overestimates observations around large urban area and ODIAC shows better agreement with observations (Hu et al. 2020). For this study, both EDGAR and ODIAC anthropogenic emission inventory were tested with WRF-VPRM. By using the ODIAC inventory, the simulated CO2 concentrations become more similar to the observed CO2 concentrations in most validation sites compared to those with EDGAR inventory, which implies that the ODIAC inventory is appropriate for CO2 simulations in East Asia (not shown).

Therefore, the ODIAC was used as an anthropogenic CO2 emission inventory. ODIAC is generated based on GOSAT satellite data from the National Institute for Environmental Studies (NIES) in Japan (Oda and Maksyutov 2015). Fossil fuel (i.e., anthropogenic) emissions in ODIAC are calculated using space-based nighttime light data of GOSAT, the emissions from each plant, and the latitude and longitude for each plant. ODIAC version 2019 (ODIAC 2019) was downloaded from the Center for Global Environmental Research (CGER) and NIES (http://db.cger.nies.go.jp/dataset/ODIAC/, https://doi.org/10.17595/20170411.001.). Monthly average data at a resolution of 1 × 1 km during 2000–2018 in GeoTIFF format in ODIAC 2019 were provided and monthly average data in March 2018 of ODIAC 2019 was used in this study. The monthly average data of the ODIAC emission inventory have been used as an anthropogenic emission inventory in multiple studies (Li et al. 2019, 2020; Hu et al. 2020).

The ocean CO2 map from the Japan Meteorological Agency (JMA), with a spatial resolution of 1° × 1° (Iida et al. 2021; Takatani et al. 2014), was used as the oceanic CO2 emission inventory. The ocean CO2 map provides the air–sea CO2 flux, pH, carbon dioxide partial pressure (pCO2), dissolved inorganic carbon (DIC) concentration, and total alkalinity (TA). Among these, pCO2 is calculated from sea surface temperature (SST), chlorophyll-a, and salinity observations from a satellite (Takatani et al. 2014).

2.1.2 Vegetation Photosynthesis and Respiration Model (VPRM)

VPRM (Mahadevan et al. 2008) is a model for calculating biogenic CO2 in WRF-Chem, and was combined with WRF-Chem starting from version 3.1.1 (Xiao et al. 2004; Ahmadov et al. 2007). Before conducting VPRM, a pretreatment process called the VPRM preprocessor needs to be performed. MODerate resolution Imaging Spectroradiometer (MODIS) satellite observations were used in the VPRM preprocessor. The MODIS is operated on two spacecraft, Terra and Aqua. In this study, the MOD09A1 version 6 of Terra was used. The enhanced vegetation index (EVI) and the land surface water index (LSWI) were calculated in the VPRM preprocessor using the surface reflectance values of MOD09A1 for each land use type of the synergetic land cover product (SYNMAP) proposed by Jung et al. (2006).

The VPRM was calculated simultaneously with model integration. In VPRM, the sum of GEE and R are calculated for each land use type using EVI and LSWI from the VPRM preprocessor and 2 m temperature and downward shortwave radiation from WRF. The calculation formula is as follows (Mahadevan et al. 2008):

$$GEE=\lambda \times {T}_{scale}\times {P}_{scale}\times {W}_{scale}\times EVI\times \frac{1}{\left(1+\frac{PAR}{{PAR}_{0}}\right)}\times PAR$$
(2)
$$R=\alpha \times T+\beta$$
(3)

where \(PAR\) is photosynthetically activate radiation and calculated using downward shortwave radiation from WRF; \(\alpha\), \(\beta\), \(\lambda\), and \({PAR}_{0}\) are empirical parameters for each land use type; T is 2 m temperature from WRF; \({T}_{\mathrm{scale}}\) denotes the relationship between photosynthesis and temperature, \({P}_{\mathrm{scale}}\) denotes the effect of leaf expansion, and \({W}_{\mathrm{scale}}\) denotes canopy moisture calculated from LSWI of the MODIS satellites, which are dimensionless variables with values between 0 and 1 (Hilton et al. 2013) and calculated using the following equations.

$${T}_{scale}=\frac{\left(T-{T}_{min}\right)\left(T-{T}_{max}\right)}{\left(T-{T}_{min}\right)\left(T-{T}_{max}\right)-{\left(T-{T}_{opt}\right)}^{2}}$$
(4)
$$P_{scale} = \frac{1 + LSWI}{2}$$
(5)
$$W_{scale} = \frac{1 + LSWI}{{1 + LSWI_{max} }}$$
(6)

where \({{T}_{\mathrm{max}}, {T}_{\mathrm{min}},\mathrm{ and} T}_{\mathrm{opt}}\) represent the maximum, minimum, and optimum temperatures during photosynthesis, respectively, and are provided as tables for each land use type in VPRM, and \(LSWI_{max}\) denotes the maximum LSWI in the growing season.

In VPRM simulations of biogenic CO2 in WRF-Chem, α, β, λ, and PAR0 should be optimized for the experimental region (Hilton et al. 2013). In this study, three tables previously used for East Asia were used to investigate the VPRM parameter table that is the most appropriate over East Asia. The three tables are the US table, Li table (used by Li et al. 2020), and Dayalu table (used by Dayalu et al. 2018) (Table 1).

Table 1 VPRM parameter values for different vegetation types in each VPRM table
Full size table

2.1.3 Initial and boundary conditions of WRF-Chem

As WRF-Chem is a regional model that combines meteorology and chemistry, the chemical initial and boundary conditions and the atmospheric initial and boundary conditions are required to run the WRF-Chem. Since the atmospheric CO2 concentration is primarily affected by the transport of CO2 rather than chemical reactions (Nasrallah et al. 2003), only chemical initial and boundary conditions for CO2 were used as the chemical initial and boundary conditions.

In accordance with previous studies (Li et al. 2019; Li et al. 2020; Liu et al. 2018; Park et al. 2020), CT2019 data from the ESRL of NOAA (Jacobson et al. 2020) were used as the chemical initial and boundary conditions for CO2. The global CO2 concentrations of CT2019 are provided at a spatial resolution of 3° × 2°. As in WRF-Chem, CO2 in CT2019 is subdivided into four components: CO2_ANT, CO2_BIO, CO2_OCE, and CO2_FIRE. As in WRF-Chem, the fire emission in CT2019 was not considered.

The fifth generation atmospheric reanalysis of the European Centre for Medium-range Weather Forecasts (ECMWF; ERA5) (Hersbach et al. 2018) and FNL of NCEP (NCEP/NOAA 2000) were used as atmospheric initial and boundary conditions.

2.2 Experimental design

To investigate the most appropriate atmospheric initial and boundary conditions and VPRM tables for simulating CO2 over East Asia, several experiments were conducted for the one-month period of March 2018.

Table 2 shows the configuration of WRF-Chem used in this study. The horizontal resolution of WRF-Chem was set to 9 km with 393 × 336 grid points over the experimental region, as shown in Fig. 1. The model’s vertical layers were 51 vertical layers with the top of the model as 50 hPa.

Table 2 Configuration of WRF-Chem simulation
Full size table
Fig. 1
figure 1

a Meteorological observation sites (SYNOP: black dot, SONDE: red triangle) and b CO2 observation sites (surface CO2 observation sites: blue dot, OCO-2 XCO2 observation sites: grey dot) in the model domain

Full size image

Figure 2 shows the schematic diagram to simulate CO2 using WRF_Chem. From 1 March 2018, a 30-h prediction of WRF-Chem was conducted every 18 UTC, and the previous 24 h CO2 prediction field was used as the initial condition for the next run as in Pillai et al. (2011) and Zhao et al. (2019). This was to simulate long-distance transport by allowing CO2 transported in the previous run to be reflected in the next run (Ballav et al. 2012; Liu et al. 2018; Li et al. 2019, 2020). The emission inventory, atmospheric initial and boundary conditions (i.e., FNL and ERA5), and chemical boundary conditions (i.e., CT2019) were updated every 18 UTC. To run the WRF-Chem for the one-month period of March 2018, 7 days of model spin up was performed from February 22 to 28, 2018, as Ballav et al. (2012) and Ballav et al. (2020) have used 5 days of model spin up to simulate CO2 concentrations for 1–2 years over Asia. The validation used only 24-h forecasts from 6 to 30 h.

Fig. 2
figure 2

The schematic diagram of the CO2 simulation using WRF-Chem

Full size image

Table 3 shows the experimental names depending on the atmospheric conditions and the VPRM table used.

Table 3 Experiment names depending on meteorological initial and boundary conditions and VPRM tables
Full size table

2.3 Validation

Validation was performed for the 24 h forecast field from 6 to 30 h forecasts, to avoid possible discontinuities caused by initial and boundary condition updates.

For validation, the bias and root mean square error (RMSE) were used and calculated as:

$$Bias = R_{i} - O_{i}$$
(7)
$$RMSE = [\frac{1}{n}\sum\limits_{i = 1}^{n} {(R_{i} - O_{i} )^{2} ]^{1/2} }$$
(8)

where R is the model simulation value, O is the observed value, and n is the number of observations.

2.3.1 Meteorological observations for validation

For the atmospheric field, the NCEP PrepBUFR conventional observations were used to validate the surface and upper-air simulation results. For surface observations, the wind speed and direction of land surface synoptic weather observations (SYNOP) were used every 6 h (00, 06, 12, and 18 UTC). For upper-air observations, upper-air wind profiles from radiosonde (SONDE) observation data at 925, 700, 500, 300, and 200 hPa were used every 12 h (00, 12 UTC). Figure 1a shows the locations of the SYNOP and SONDE observations used in this study.

2.3.2 CO2 observations and model output for validation

Various CO2 observations were used to examine whether the CO2 concentrations were accurately simulated in WRF-Chem. Table 4 provides information on surface CO2 observation sites used for validation. CO2 observation data for Anmyeon-do (AMY, Republic of Korea), Mt. Dodaira (DDR, Japan), Kisai (KIS, Japan), Lulin (LLN, Taiwan, Province of China), Ryori (RYO, Japan), Tae-ahn Peninsula (TAP, Republic of Korea), Ulaan Uul (UUM, Mongolia), and Yonagunijima (YON, Japan) are provided by the World Data Centre for Greenhouse Gases (WDCGG, https://ds.data.jma.go.jp/wdcgg). These data are observed by NOAA ESRL, Center for Environmental Science in Saitama (SAIPF, Japan), JMA, and Korea Meteorological Administration (KMA; Republic of Korea). The Gosan (GSN, Republic of Korea) and Ulleung-do (UL, Republic of Korea) observations are provided by KMA (https://data.kma.go.kr/data/gaw/selectGHGsRltmList.do?pgmNo=587). The observation data for AMY, DDR, KIS, RYO, and YON are at 1 h intervals, GSN and UL data are at 1 day intervals, and LLN, TAP, and UUM provide data discontinuously.

Table 4 Information on surface CO2 observation sites
Full size table

Satellite-based XCO2 observations were used to compensate for the lack of surface CO2 observations over East Asia. OCO-2 is the National Aeronautics and Space Administration (NASA)’s first Earth remote sensing satellite for atmospheric CO2 observations, launched after GOSAT. OCO-2 provides a space-based global measurement for the absorption and emission of local CO2 and carries out observations at 13:30 LST along a solar synchronous orbit. The OCO-2 observation data used were ACOS L2 Lite Output Filtered with oco2-lite_file_prefilter_b9 converted from Level 1 radiance to Level 2 data using the ACOS retrieval algorithm (O’Dell et al. 2012), produced by the Jet Propulsion Laboratory (JPL) (https://co2.jpl.nasa.gov/download/?dataset=OCO2LtCO2v9&product=LITE). The data quality of the OCO-2 observations can be checked by the values of xco2_quality_flag and warn_level as described in the OCO-2 Data Product User’s Guide (Osterman et al. 2018). The xco2_quality_flag value is 0 or 1, where 0 means “good” and 1 means “bad”. In this study, OCO-2 data with ‘0’ xco2_quality_flag value were used for validation.

In WRF-Chem, CO2 concentrations are simulated at each pressure level, while OCO-2 observes the column-averaged CO2 mole fraction (XCO2). Because the data types of the simulated CO2 and satellite observed XCO2 are different, they need to be converted into the same data type for comparison. Thus, CO2 concentrations simulated at each pressure level in WRF-Chem were converted to XCO2 concentrations. First, the simulated CO2 concentrations in WRF-Chem were interpolated to the latitude and longitude of OCO-2 data. Then, the XCO2 concentrations of WRF-Chem were calculated as in Connor et al. (2008) and O’Dell et al. (2012):

$${XCO}_{2}^{model}={XCO}_{2a}+\sum_{i}{w}_{i}^{T}{A}_{i}{\left({CO}_{2}^{interp}-{CO}_{2a}\right)}_{i}$$
(9)

where \({XCO}_{2a}\) is a priori XCO2, \({w}_{i}^{T}\) is the pressure weighting function, \({A}_{i}\) is the column averaging kernel, \(C{O}_{2}^{interp}\) is the interpolated simulated CO2 concentrations of WRF-Chem, and \(C{O}_{2a}\) is a priori CO2.

Figure 1b shows the locations of the surface CO2 and satellite XCO2 observations used in this study. In addition to surface CO2 and satellite XCO2 observations, the CarbonTracker output (CT2019) was used to validate the reliability of the simulated CO2 concentrations.

3 Results

3.1 Distribution of surface biogenic CO2 concentrations

Figure 3a shows the average surface biogenic CO2 concentrations simulated in the six experiments, and Fig. 3b shows the surface biogenic CO2 concentrations in CT2019. In the surface CO2 concentrations averaged for six experiments (Fig. 3a), the CO2 absorption by vegetation is weaker over central China and the Korean Peninsula than that in other regions. These regional patterns of the averaged simulation results are similar to those of the biogenic CO2 concentrations in CT2019 in Fig. 3b. Because the horizontal resolution of the experiments was denser than that in CT2019, more detailed distributions were simulated in the experiments using WRF-Chem. However, the amplitude of biogenic CO2 absorption in the averaged simulation results in WRF-Chem is greater than that in CT2019 (compare Fig. 3a and b). This difference in biogenic CO2 may be due to different model framework between WRF-Chem and CT2019.

Fig. 3
figure 3

Distribution of a average surface biogenic CO2 concentration (ppm) simulated in six experiments and b surface biogenic CO2 concentration (ppm) in CT2019. Anomaly distributions for average surface biogenic CO2 concentration (ppm), simulated in c FNL_US, d ERA_US, e FNL_Li, f ERA_Li, g FNL_Da, and h ERA_Da from the average of six experiments shown in (a)

Full size image

Figures 3c–h show the difference between the surface biogenic CO2 concentrations of each experiment and the average biogenic CO2 concentrations over six experiments. FNL_US and ERA_US show very similar distributions and amount of biogenic CO2 concentrations (Fig. 3c and d), indicating that the difference in atmospheric initial and boundary conditions for WRF-Chem simulations did not seem to significantly affect the simulated biogenic CO2 concentrations. Compared to the average biogenic CO2 concentrations, FNL_US (Fig. 3c) and ERA_US (Fig. 3d) show lower biogenic CO2 absorption over central China and the Korean Peninsula. This underestimated biogenic CO2 absorption in both FNL_US and ERA_US compared to the average biogenic CO2 absorption results in greater differences in Fig. 3c and d.

FNL_Li and ERA_Li show very similar distributions and amount of biogenic CO2 concentrations (Fig. 3e and f). Compared to the average biogenic CO2 concentrations, FNL_Li (Fig. 3e) and ERA_Li (Fig. 3f) show lower biogenic CO2 absorption over central China and the Korean Peninsula and greater biogenic CO2 absorption in southern China. However, the magnitude of the differences between the simulated biogenic CO2 concentrations (FNL_Li and ERA_Li) and the average is small.

FNL_Da and ERA_Da also show similar distributions of biogenic CO2 concentrations (Fig. 3g and h). Compared to the average biogenic CO2 concentrations, FNL_Da (Fig. 3g) and ERA_Da (Fig. 3h) show greater biogenic CO2 absorption over central China and the Korean Peninsula. This overestimated biogenic CO2 absorption in both FNL_Da and ERA_Da compared to the average biogenic CO2 absorption results in greater differences (Fig. 3g and h).

In contrast to the similar distribution and magnitude of biogenic CO2 concentrations between the experiments using different atmospheric initial and boundary conditions and the same VPRM table, there were substantial differences between the experiments using the same atmospheric initial and boundary conditions and different VPRM tables. Therefore, the simulated surface biogenic CO2 concentrations were more sensitive to differences in the VPRM tables than those in the atmospheric initial and boundary conditions. In terms of region, the differences in biogenic CO2 concentrations in the experiments were the greatest over central China and the Korean peninsula.

The distributions of the simulated total CO2 concentrations (not shown), which are the sum of the biogenic, anthropogenic, oceanic, and background CO2 concentrations, showed similar distributions as in Fig. 3.

3.2 Validation with observations

As the simulated CO2 concentrations may be affected by the simulated transport, the simulated wind speed and direction were validated against the observed wind speed and direction. In addition, the simulated CO2 concentrations of each experiment were compared with the observed CO2 concentrations to validate whether the simulation results were appropriate and to investigate the experiment that led to the most accurate simulation results.

3.2.1 Validation of wind speed and direction

In WRF-Chem, the atmospheric and chemical fields interact with each other. According to Baklanov et al. (2014), in WRF-Chem, various atmospheric variables such as temperature, precipitation, wind direction, and wind speed can affect the chemical species. In addition, the physical characteristics of aerosols and the concentrations of radiatively active gases can affect atmospheric variables. However, in this study, only CO2 was simulated without considering the reaction with aerosols in the atmosphere. Therefore, there was no change in the atmospheric field with changes in the CO2 concentration.

Among the six experiments in Table 3, the experiments with the same atmospheric initial and boundary conditions simulated the same atmospheric fields. This implies that the atmospheric fields of FNL_US, FNL_Li, and FNL_Da were simulated identically. Thus, when verifying the atmospheric field, the six experiments can be divided into two groups: experiments using FNL (FNL_exp) and experiments using ERA5 (ERA_exp).

Figure 4 shows the time series of bias and RMSE for each experimental result (i.e., FNL_exp and ERA_exp) with respect to surface SYNOP observations for wind speed and direction. The bias and RMSE for each experiment are summarized in Table 5. For both FNL_exp and ERA_exp, the biases of the surface wind speed show high fluctuations centered around 0 (Fig. 4a), leading to small bias values (0.05 and 0.01 m s−1 in FNL_exp and ERA_exp, respectively) for both experiments (Table 5) despite high fluctuations. The RMSEs of the surface wind speed in both experiments showed high fluctuations (Fig. 4b) with an approximate value of 3.2 m s−1 (Table 5). In contrast to wind speed, the biases of the surface wind direction were mostly positive (Fig. 4c), implying that the surface wind direction in both experiments was overestimated compared to the observation values. The bias for FNL_exp (ERA_exp) was 22.84° (24.05°) (Table 5). The RMSEs of the surface wind direction in both experiments showed large values (Fig. 4d) with 82.81° for FNL_exp and 84.21° for ERA_exp (Table 5). Compared to the surface wind speed with smaller bias and RMSE, the surface wind direction showed a large bias and RMSE compared to the observed values for both experiments. Although the difference between FNL_exp and ERA_exp was small for both surface wind speed and direction, FNL_exp showed a slightly smaller bias and RMSE than ERA_exp.

Fig. 4
figure 4

Time series of a bias and b RMSE of simulated 10 m wind speed (m s−1) with respect to the observed 10 m wind speed at SYNOP sites during March 2018. Time series of c bias and d RMSE of simulated 10 m wind direction (°) with respect to the observed 10 m wind direction at SYNOP sites in March 2018. FNL_exp (ERA_exp) is denoted by red (blue) line

Full size image
Table 5 Bias and RMSE of the simulated 10 m wind speed and direction for FNL_exp and ERA_exp with respect to the observed 10 m wind speed and direction at SYNOP sites
Full size table

Figure 5 shows the time series of wind speed, bias, and RMSE for each experimental result (i.e., FNL_exp and ERA_exp) with respect to the upper-air SONDE observations at each pressure level. The bias and RMSE for each experiment are summarized in Table 6. For both FNL_exp and ERA_exp, the wind speed and bias increased as go up into the upper atmosphere (Fig. 5a–c). The average biases of wind speed below 700 hPa (at 500 hPa) were 0.20 and 0.22 m s−1 (− 0.63 and − 0.71 m s−1) in FNL_exp and ERA_exp, respectively (Table 6). In contrast to the negative biases in other layers, the biases of the wind speed at 925 hPa were positive in both experiments (Fig. 5c and Table 6), which implies an overestimation of the wind speed at 925 hPa in both experiments. As the wind speed increased in the upper atmosphere, the biases in the upper atmosphere also increased (Fig. 5d–f). Similar to the bias, for both experiments, the RMSE increased as go up into the upper atmosphere (Fig. 5g–i). The average RMSEs of wind speed below 700 hPa (at 500 hPa) are 4.00 and 4.04 m s−1 (6.45 and 6.50 m s−1) in FNL_exp and ERA_exp, respectively (Table 6).

Fig. 5
figure 5

Time series of a–c simulated wind speed (m s−1), d–f bias (m s−1) of simulated wind speed, and g–i RMSE (m s−1) of simulated wind speed with respect to the observed wind speed at SONDE sites at each pressure level in March 2018. FNL_exp (ERA_exp) is denoted by red (blue) line

Full size image
Table 6 Bias and RMSE of the simulated wind speed and direction for FNL_exp and ERA_exp with respect to the observed wind speed and direction at SONDE sites at each pressure level
Full size table

Figure 6 shows the time series of wind direction, bias, and RMSE for each experimental result (i.e., FNL_exp and ERA_exp) with respect to the upper-air SONDE observations for wind direction at each pressure level. The bias and RMSE for each experiment are summarized in Table 6. For both FNL_exp and ERA_exp, the wind direction was approximately less than 240° in the lower atmosphere at 925 hPa and greater than 240° above 700 hPa (Fig. 6a–c), implying that the wind is veering towards the east in the lower atmosphere. In contrast to the wind direction, the fluctuations of wind direction and bias decreased as go up into the upper atmosphere (Fig. 6a–c). The high fluctuation in wind direction in the lower atmosphere is due to the complex topography. The average biases of the wind direction below 700 hPa (at 500 hPa) were 8.93° and 7.99° (5.20° and 4.58°) in FNL_exp and ERA_exp, respectively (Table 6). For both experiments, the wind direction was mostly overestimated at all levels (Fig. 6d–f and Table 6). Similar to the bias, for both experiments, the RMSE increased as go down into the lower atmosphere (Fig. 6g–i). The average RMSEs of the wind direction below 700 hPa (at 500 hPa) were 53.01° and 53.15° (32.12° and 32.25°) in FNL_exp and ERA_exp, respectively (Table 6). In both wind direction and wind speed, the difference between FNL_exp and ERA_exp was not large, as for the surface wind field validation. However, the mean RMSE of FNL_exp was smaller in the lower atmosphere (below 700 hPa) as well as in the upper atmosphere (at 500 hPa) (Table 6). In other words, the wind field of the entire atmosphere was slightly better simulated by FNL_exp.

Fig. 6
figure 6

Time series of a–c simulated wind direction (°), d–f bias (°) of simulated wind direction, and g–i RMSE (°) of simulated wind direction with respect to the observed wind direction at SONDE sites at each pressure level in March 2018. FNL_exp (ERA_exp) is denoted by red (blue) line

Full size image

Throughout the atmosphere from 925 to 500 hPa, the average RMSE of wind speed in FNL_exp was 4.81 m s−1 and that in ERA_exp was 4.86 m s−1. For the wind direction, the average RMSE of FNL_exp was 46.05° and that of ERA_exp was 46.19°. Therefore, based on the surface and pressure level validations, FNL_exp showed slightly better results for wind forecasts than ERA_exp in East Asia.

3.2.2 Validation of simulated surface CO2 concentrations with observed surface CO2 concentrations

The simulated surface CO2 concentrations in the six experiments were validated with respect to the observed surface CO2 concentrations. In addition to the observed surface CO2 concentrations, a comparison with surface CO2 concentrations simulated in CT2019 was conducted to validate the reliability of the surface CO2 concentrations simulated in this study.

Figure 7 shows the time series of the simulated surface CO2 concentrations in this study and in CT2019 for each surface CO2 observation site during March 2018. The simulated surface CO2 concentrations were mostly similar to the observed surface CO2 concentrations at DDR, KIS, RYO, YON, GSN, and UL (Fig. 7a, b, c, d, f, and g). Except for GSN, the simulated surface CO2 concentrations averaged over the six experiments were more similar to the observed surface CO2 concentrations than those in CT2019. In the case of UUM, TAP, and LLN, surface CO2 concentrations are rarely observed (i.e., approximately once a week), which makes comparisons of simulated surface CO2 concentrations difficult. Nevertheless, the simulated surface CO2 concentrations were mostly similar to the observations at UUM, TAP, and LLN (Fig. 7h, i, and j), indicating the reliability of the simulated surface CO2 concentrations. Compared to the six experiments, CT2019 overestimated the observations at every site (Fig. 7). This overestimation is caused by the anthropogenic emission inventory used in CT2019, which are both Miller emission dataset based on EDGAR v4.2 (European Commission 2011) and ODIAC 2018 (Oda and Maksyutov 2015; Oda et al. 2018). As mentioned in Section 2.1.1, EDGAR anthropogenic emission inventory generally overestimates the observations around local anthropogenic sources (e.g., urban areas).

Fig. 7
figure 7

Time series of simulated and observed surface CO2 concentrations (ppm) for each surface CO2 observation site in March 2018 (FNL_US: red solid, FNL_Li: orange solid, FNL_Da: green solid, ERA_US: blue solid, ERA_Li: purple solid, ERA_Da: light purple solid, CT2019: grey dashed, surface CO2 observation: black star)

Full size image

Figure 8 shows the biases and RMSEs for each experimental result with respect to the observed surface CO2 concentrations at each site. The bias and RMSE for each experiment at each site are shown in Table 7. For rarely observed sites (i.e., UUM, TAP, and LLN), bias and RMSE may not be accurately calculated. Therefore, bias and RMSE were calculated for only seven sites, excluding UUM, TAP, and LLN. The biases were mostly negative except for some experiments at the KIS and YON sites (Fig. 8a and Table 7), which implies that the simulated surface CO2 concentrations mostly underestimated the observed surface CO2 concentrations. Except for AMY with a bias of − 4.71 ppm, the biases at other sites were smaller than 3 ppm (Fig. 8a and Table 7). This is because the simulated CO2 concentrations at AMY were more underestimated than those at other sites, as shown in Figs. 7e and 8a. Among the observation sites, the bias was the smallest at YON (0.01 ppm averaged over six experiments) (Table 7). Among six experiments, FNL_US showed the lowest bias of − 1.18 ppm, followed by ERA_US (− 1.26 ppm) and FNL_Li (− 1.49 ppm) (Table 7). The average biases of all six experiments were less than the bias of CT2019 (Table 7). The RMSEs of KIS, RYO, YON, GSN, and UL were lower than 5 ppm, while the RMSEs of DDR and AMY were greater than 5 ppm (Fig. 8b and Table 7). Among the observation sites, the RMSE at YON was the smallest (1.62 ppm averaged over six experiments) and was much smaller than that of CT2019 (Table 7). On average, the RMSEs of the six experiments were smaller than the RMSE of CT2019 (Table 7). This implies that the surface CO2 concentrations can be simulated more appropriately using high-resolution WRF-Chem compared to a low-resolution global model (e.g., CarbonTracker). Among the six experiments, on average, ERA_Li showed the lowest RMSE (3.68 ppm), followed by FNL_Li (3.71 ppm) (Table 7).

Fig. 8
figure 8

a Bias (ppm) and b RMSE (ppm) of simulated surface CO2 concentrations for each experiment and CT2019 with respect to the observed surface CO2 concentrations at surface CO2 observation sites (FNL_US: red, FNL_Li: orange, FNL_Da: green, ERA_US: blue, ERA_Li: purple, ERA_Da: light purple, CT2019: grey)

Full size image
Table 7 Bias and RMSE of the simulated surface CO2 concentrations for each experiment and CT2019 with respect to the observed surface CO2 concentrations at surface CO2 observation sites
Full size table

Overall, owing to the comparable surface wind fields, FNL_exp showed a similar bias and RMSE for surface CO2 concentrations compared to ERA_exp. For the VPRM tables, the experiments with the Li tables showed smaller biases and RMSEs compared to those with other tables. ERA_Li and FNL_Li showed smaller biases and RMSEs than the other four experiments and much smaller biases and RMSEs than CT2019. Even for the highly underestimated site as AMY, the biases and RMSEs of FNL_Li were the smallest among the six experiments and CT2019. Therefore, ERA_Li and FNL_Li showed the most similar simulated surface CO2 concentrations to the observed surface CO2 concentrations among the six experiments.

3.2.3 Validation of simulated XCO2 concentrations with observed OCO-2 XCO2 concentrations

The distributions of surface CO2 observation sites are limited, and there are few surface CO2 observation sites available in central China. For a more reliable validation, it is necessary to validate the simulated surface CO2 observations in the regions with few surface CO2 observation sites. Therefore, for the regions covered by the OCO-2 satellite, validation was conducted by comparing the XCO2 concentrations deduced from the WRF-Chem results with those of OCO-2.

Figure 9a shows the time series of the simulated and observed XCO2 concentrations. Compared to the OCO-2 XCO2 concentrations, the simulated XCO2 concentrations in all experiments showed similar trends but slightly overestimated values at most times. Figure 9b shows the bias of the simulated XCO2 concentrations with respect to the OCO-2 XCO2 concentrations. Due to the overestimated simulated XCO2 concentrations (Fig. 9a), all six experiments showed mostly positive biases during March 2018 (Fig. 9b), with an average bias of 0.14 ppm (Table 8). Among the six experiments, ERA_Da showed the smallest bias (0.05 ppm) followed by ERA_Li (0.14 ppm) and FNL_Li (0.16 ppm) (Table 8). Similar to the biases smaller than 1 ppm (i.e., average 0.14 ppm), the average RMSE of the simulated XCO2 concentrations with respect to OCO-2 XCO2 concentrations for the six experiments was smaller than 1 ppm (i.e., average 0.61 ppm) (Table 8), indicating that all experiments simulated XCO2 concentrations similar to those observed by OCO-2. FNL_Li showed the smallest RMSE of 0.59 ppm, followed by ERA_US (0.60 ppm), ERA_Li (0.60 ppm), and FNL_US (0.61 ppm) (Table 8). The slightly smaller RMSE of FNL_Li compared to that of ERA_Li may be associated with a slightly smaller RMSE of wind speed and direction in FNL_exp compared to that in ERA_exp in the entire atmosphere, as shown in Table 6. Because the column-averaged XCO2 concentrations are mainly affected by transport in the whole atmosphere, the slightly smaller RMSE of the simulated wind fields in the whole atmosphere in the FNL_exp seems to affect the simulated XCO2 concentrations.

Fig. 9
figure 9

Time series of a simulated XCO2 and OCO-2 XCO2 concentration (ppm) for each experiment and b bias (ppm) of the simulated XCO2 concentration for each experiment with respect to the observed OCO-2 XCO2 concentration during March 2018 (FNL_US: red solid, FNL_Li: orange solid, FNL_Da: green solid, ERA_US: blue solid, ERA_Li: purple solid, ERA_Da: light purple solid, CT2019: grey dashed, OCO-2 XCO2: black solid)

Full size image
Table 8 Bias and RMSE of the simulated XCO2 concentration for each experiment and CT2019 with respect to the observed OCO-2 XCO2 concentrations
Full size table

Figure 10 shows the spatial distribution of the RMSE over 1° × 1° bins for March 2018. The RMSE was calculated only for the bins with 20 or more observations. The RMSEs of all six experiments were similar in northern China and Japan. The greatest RMSE differences among the six experiments were in central China, where the differences in surface biogenic CO2 concentrations among the experiments were the greatest, as shown in Fig. 3. The RMSEs in central China were relatively small in FNL_Li, ERA_US, ERA_Li, and FNL_US (Fig. 10c, b, d, and a), where the surface biogenic CO2 absorption in these three experiments was underestimated compared to the average biogenic CO2 absorption of all experiments (Fig. 3e, d, f, and c). Therefore, the smaller biogenic CO2 absorption in central China in FNL_Li, ERA_Li, FNL_US, and ERA_US compared to that in other experiments resulted in a smaller RMSE over the region.

Fig. 10
figure 10

Distribution of RMSE (ppm) of simulated XCO2 concentration over 1° × 1° bins in a FNL_US, b ERA_US, c FNL_Li, d ERA_Li, e FNL_Da, and f ERA_Da with respect to the observed OCO-2 XCO2 concentration

Full size image

The smallest RMSE of FNL_Li implies that FNL_Li can simulate XCO2 concentrations similar to OCO-2 XCO2 concentrations.

4 Summary and conclusions

In this study, a high-resolution regional WRF-Chem model was used to simulate atmospheric CO2 concentrations in East Asia, where there is high uncertainty in estimating atmospheric CO2 concentrations. To estimate atmospheric CO2 concentrations over East Asia appropriately, the effects of atmospheric conditions and the VPRM parameters used for simulating biogenic CO2 concentrations were evaluated using high-resolution WRF-Chem. Various experiments were performed to evaluate the effects of experimental settings on estimating atmospheric CO2 concentration.

The atmospheric CO2 concentration is more affected by wind than other meteorological variables. Thus, the wind speed and direction need to be accurately simulated to simulate appropriate CO2 concentrations. To examine the atmospheric field that simulates the wind field more accurately, FNL and ERA5 were considered as the initial and boundary conditions of WRF-Chem. In addition, the VPRM parameters that simulate biogenic CO2 concentrations need to be appropriate for estimating atmospheric CO2 concentrations.

To evaluate the effects of the atmospheric field and VPRM parameters on simulating surface CO2 concentrations, six experiments were performed by using two atmospheric reanalysis fields (FNL and ERA5) and three VPRM tables (US, Li, and Dayalu tables) for March 2018 over East Asia.

The simulated surface biogenic and total CO2 concentrations were more affected by differences in the VPRM tables than those in atmospheric initial and boundary conditions. Similar spatial distributions and magnitudes of surface biogenic CO2 concentrations were observed between experiments using different atmospheric initial and boundary conditions but the same VPRM table, whereas experiments using the same atmospheric initial and boundary conditions but different VPRM tables showed distinctly different spatial distributions and magnitudes. In terms of region, the differences in surface biogenic CO2 concentrations among the experiments were large over central China and the Korean peninsula. Since the vertical mixing also affects CO2 concentrations, the effect of physical parameterizations on the vertical mixing and simulation of CO2 concentrations over Asia would be a future work.

To verify the accuracy of the simulated wind and CO2 concentrations, they were compared with observed values. From surface and pressure level validations, all experiments using FNL as the initial and boundary conditions (FNL_exp) were slightly more accurate in wind speed and direction forecasts than those using ERA5 as the initial and boundary conditions (ERA_exp) for the experimental period over East Asia. From the validation of surface CO2 concentrations, on average, the experiments that used either ERA or FNL as the initial and boundary conditions with the Li table as the VPRM table in WRF-Chem showed smaller biases and RMSEs than the other four experiments and also showed much smaller biases and RMSEs compared to CT2019. Therefore, among the six experiments, ERA_Li and FNL_Li simulated surface CO2 concentrations closest to the observed values. From the validation of XCO2 concentrations, FNL_Li using FNL as the initial and boundary conditions and the Li table as the VPRM table in WRF-Chem showed smaller biases and RMSEs than other experiments. Based on all validations of wind and CO2 concentrations, the combination of FNL as the atmospheric initial and boundary conditions and Li table as the VPRM table showed the overall best performance and was thus most suitable for simulating atmospheric CO2 concentrations using WRF-Chem during the experimental period for East Asia.

In future studies, using the WRF-Chem configurations based on the FNL and Li table, high-resolution atmospheric CO2 concentrations over East Asia will be simulated for longer periods, and the characteristics of the high-resolution regional CO2 concentrations will be evaluated.