Evaluating Anthropogenic CO2 Bottom-Up Emission Inventories Using Satellite Observations from GOSAT and OCO-2

Abstract

Anthropogenic carbon dioxide (CO₂) emissions from bottom-up inventories have high uncertainties due to the usage of proxy data in creating these inventories. To evaluate bottom-up inventories, satellite observations of atmospheric CO₂ with continuously improved accuracies have shown great potential. In this study, we evaluate the consistency and uncertainty of four gridded CO₂ emission inventories, including CHRED, PKU, ODIAC, and EDGAR that have been commonly used to study emissions in China, using GOSAT and OCO-2 satellite observations of atmospheric column-averaged dry-air mole fraction of CO₂ (XCO₂). The evaluation is carried out using two data-driven approaches: (1) quantifying the correlations of the four inventories with XCO₂ anomalies derived from the satellite observations; (2) comparing emission inventories with emissions predicted by a machine learning-based model that considers the nonlinearity between emissions and XCO₂. The model is trained using long-term datasets of XCO₂ and emission inventories from 2010 to 2019. The result shows that the inconsistencies among these four emission inventories are significant, especially in areas of high emissions associated with large XCO₂ values. In particular, EDGAR shows a larger difference to CHRED over super-emitting sources in China. The differences for ODIAC and EDGAR, when compared with the machine learning-based model, are higher in Asia than those in the USA and Europe. The predicted emissions in China are generally lower than the inventories, especially in megacities. The biases depend on the magnitude of inventory emissions with strong positive correlations with emissions (R² is larger than 0.8). On the contrary, the predicted emissions in the USA are slightly higher than the inventories and the biases tend to be random (R² is from 0.01 to 0.5). These results indicate that the uncertainties of gridded emission inventories of ODIAC and EDGAR are higher in Asian countries than those in European and the USA. This study demonstrates that the top-down approach using satellite observations could be applied to quantify the uncertainty of emission inventories and therefore improve the accuracy in spatially and temporally attributing national/regional totals inventories.

1. Introduction

Controlling anthropogenic carbon dioxide (CO₂) emissions, which lead to the increase in atmospheric CO₂ concentration, is important to mitigate global warming for human sustainable development. Therefore, an accurate investigation of anthropogenic CO₂ emissions is critical to evaluate the effects of emission reduction and control measures. Current gridded data of anthropogenic CO₂ emissions are estimated based on the spatial and temporal attribution of national or regional totals in inventories by proxy data, such as population density, transportation network and enterprise emitting sources. These bottom-up gridded emission inventories, such as the Emissions Database for Global Atmosphere Research (EDGAR), Open-source Data Inventory for Anthropogenic CO₂ (ODIAC) [1,2,3], a high-resolution inventory of CO₂ emissions in China produced by The Ministry of Ecology and Environment of China (called CHRED hereafter) [4,5], and the emission inventory with various proxy data developed by Peking University [5] (called PKU hereafter), have been widely used. The used national or regional emissions are generally accurate in the counties/regions with reliable data sources and standardized statistical methods [6]. The gridded emissions, however, remain highly uncertain in space and time due to the imperfections of proxy data, which can be irregular and insufficient, used to disaggregate the national/regional emissions in space and time [7]. For example, the uncertainty (2σ) of gridded ODIAC is up to 120% on average and ranges from 4% to 190% [8]. The emissions from different inventories, moreover, are inconsistent at the same location and time. It is therefore important to quantify the uncertainties of these gridded inventories for the assessment of anthropogenic CO₂ emissions and application in modeling global carbon cycle.

The uncertainty of inventories can be evaluated by cross-validation with more detailed regional emission inventory data [9,10] or the model simulation at the specific time and regions [11]. The column-averaged dry air mole fractions (XCO₂) derived from satellite observations, such as the Greenhouse gases Observing Satellite (GOSAT) and Orbiting Carbon Observatories (OCO-2, OCO-3), have been developed to detect the magnitude of anthropogenic CO₂ emissions from regional emissions and point sources. Satellite observations provide important ‘top-down’ information for evaluating bottom-up inventories and reducing uncertainty [12,13]. Many studies indicated that the emissions by human activities could lead to 1–3 ppm enhancements of XCO₂ regionally [14,15,16,17] and up to 6 ppm enhancement over large power plants [18], as detected by the XCO₂ retrievals derived from GOSAT, OCO-2, and OCO-3 observations. The spatial pattern of XCO₂ derived from GOSAT is consistent with the spatial patterns of the inventories of ODIAC and CHRED [19,20]. The XCO₂ variation explains more than 40% of ODIAC in the grids and up to more than 90% in the binned emissions based on clustering analysis of gridded XCO₂ [21,22,23]. The anthropogenic CO₂ emissions could be estimated using gridded XCO₂ retrievals from GOSAT and OCO-2 observations by the generalized regression neural networks (GRNN) using training datasets combined with ODIAC data [22,23]. The estimated emissions were generally consistent with inventory emissions but showed a large difference with ODIAC in the urban areas, likely because of the uncertainty of spatial distribution of the emissions in ODIAC. Recently, Chevallier et al. [12] indicated that the large fossil fuel CO₂ emissions derived from OCO-2 and OCO-3 explain more than one third of the inventory variance from EDGAR at the corresponding cells and hours, and the signal-to-noise ratio for the XCO₂ retrieval tend to increase with the emission level. These studies suggested that the satellite observations, as a top-down measure, can be potentially applied to evaluate the uncertainty of bottom-up inventories [24]. As an attendant measure, the guidelines for the national emission survey improvement program in the IPCC report (2019) [25] suggest applying the satellite observations to verify and rectify the emission inventories. However, a comparison of the emission inventories, especially for the top two national emitters (China and the US), has been less explored.

Here, we use satellite observations of XCO₂ from GOSAT and OCO-2 to evaluate the differences and uncertainties of bottom-up emission inventories in space and time, aiming to develop data-driven approaches for the verification of emission inventories using satellite observations. In the next section, we describe the datasets used, including satellite XCO₂ observations and four widely used inventories (CHRED, PKU, ODIAC, and EDGAR) and the developed data-driven approaches. In Section 3, we demonstrate the results of how the emission inventories are correlated with the magnitude of observed XCO₂ and the uncertainties of the emission inventories in space and time by comparing with the emissions predicted by a machine learning-based model. The discussion and conclusion are presented in Section 4 and Section 5, respectively.

2. Materials and Methods

2.1. Materials

2.1.1. Satellite Observations

• Globally mapping XCO₂ dataset based on GOSAT and OCO-2

GOSAT and OCO-2 have been observing and retrieving global XCO₂ since 2009. Original GOSAT and OCO-2 level 2 retrieval data are available from the Goddard Earth Science Data Information and Services Center (GES DISC) [26]. The XCO₂ retrievals are the latest ACOS level 2 Lite data product (v9r) for GOSAT and the latest level 2 lite data product (v10r) for OCO-2 [27,28,29]. In this paper, we used the global land monthly mapping XCO₂ dataset (Mapping-XCO₂) with a spatial resolution of 1° in latitude and longitude from April 2009 to December 2020. Mapping-XCO₂ data are generated based on a spatiotemporal geostatistical method by using the integrated XCO₂ retrievals obtained from GOSAT and OCO-2 through the corrections of the systemic errors between GOSAT (from April 2009 to August 2014) and OCO-2 (from September 2014 to December 2020), including sensor sensitivity, and the differences in observing time and space [17,30,31]. This Mapping-XCO₂ dataset has shown its validation and effectiveness in detecting the anthropogenic CO₂ emissions, variations of XCO₂ in the global and regional scales by comparing with Total Column Carbon Observing Network (TCCON), and CarbonTracker model simulations [17,30,31,32,33].

• SIF

We collected the solar-induced chlorophyll fluorescence (SIF) data, the Global SIF Dataset (GOSIF), for interpreting the amount of vegetation ecological CO₂ uptake in XCO₂. This dataset was developed by Xiao et al. [34] based on discrete OCO-2 SIF soundings, moderate Resolution Imaging Spectro radiometer (MODIS) data, and meteorological reanalysis data based on a data-driven approach. GOSIF covers the period 2000–2020 with 0.05° grid and 8 days’ intervals [34]. Previous studies have demonstrated that the SIF can be used to derive gross primary productivity (GPP) that is closely related to the magnitude of CO₂ uptake [35,36]. We resampled the GOSIF (8 days/0.05°) to match the Mapping-XCO₂ dataset (month/1°) by using the cubic convolution.

• Other Testing Data

To obtain the optimal combination for predicting the anthropogenic emissions with the minimum deviation by using GRNN, we also used NDVI, GPP, wind speed, and night-time light for testing. We collected the Normalized Difference Vegetation Index (NDVI) and Gross Primary Productivity (GPP) from the Moderate Resolution Imaging Spectroradiometer (MODIS) observations. NDVI data from the MOD13C2 product have temporal and spatial resolutions of 0.05° and 2 days, respectively. The GPP data is from the MOD17A2H product with temporal and spatial resolutions of 500 m and 8 days [37,38]. The wind speed data are the horizontal speed of air moving at a height of ten meters above the surface of the Earth in meters per second, which is provided by ERA5 (the fifth generation ECMWF (European Centre for Medium-Range Weather Forecasts) atmospheric reanalysis of the global climate) [39]. The night-time light data were sourced from Visible Infrared Imaging Radiometer Suite (VIIRS) day/night band (DNB) stray light characterization and correction product [40].

2.1.2. Anthropogenic CO₂ Emission Inventories

We collected two global emission inventories (ODIAC and EDGAR) and two regional emission inventories (CHRED and PKU) focusing on China. Table 1 presents the specifications of these inventories, and Figure 1 shows the spatial distribution of their total emissions in 2012 in China (Hong Kong, Macao, and Taiwan in Figure 1a show as “no data” as CHRED data were only produced over the mainland).

ODIAC is a gridded CO₂ emission dataset generated by the spatial and temporal disaggregation of national emissions from the CO₂ Information Analysis Centre (CDIAC) and British Petroleum (BP), using proxy data including satellite observed night-time lights and power plant emissions [1,2]. Similarly, EDGAR is a gridded CO₂ emission dataset generated by disaggregating the national emission inventories. The national emissions are from the energy balance statistics of the International Energy Agency (IEA), and country-specific activity datasets from BP, United States Geological Survey (USGS), World Steel Association, Global Gas Flaring Reduction Partnership (GGFR)/U.S. National Oceanic and Atmospheric Administration (NOAA), and the International Fertilizer Association (IFA). The spatial disaggregation of national emissions was implemented using proxy data including population density, traffic networks, night-time lights, and CO₂-emitting point sources [3]. EDGAR has multiple versions released, such as EDGARv5.0, EDGARv6.0, and EDGARv4.3.2_bp shown in Table 1. EDGARv5.0 provides monthly emissions only for the year 2015. EDGARv6.0 provides monthly emissions data from 2000–2018. EDGARv4.3.2_bp used in this study, which was produced based on EDGARv4.3 and BP statistics 2019 (Version 2.0) [41], is the monthly emissions updated to 2019 which facilitates a comparison with ODIAC from 2009 to 2019. The temporal changes in emissions were allocated based on the datasets from the monitoring atmospheric composition and climate—the Netherlands Organization (MACC-TNO) and the CO₂ release and oxygen uptake from Fossil Fuel Emission Estimate (COFFEE) [42,43,44,45]. It has been shown that bottom-up emission inventories have a large uncertainty in the Chinese region due to a lack of reliable and promptly updated statistical data, inappropriate emission factors (EF), and accurate proxy data [46]. In order to evaluate the uncertainty of these inventories, we also collected CHRED and PKU, which were produced by Chinese institutions with more accurate emissions statistics. CHRED is the annual anthropogenic CO₂ emission gridded data in 10 km resolution generated by the Ministry of Ecology and Environment of China based on the statistics of regional CO₂ emissions in 2012, and field investigations of point source emissions which involved 15 million factories and enterprises in the Chinese regions (excluding ocean islands) in 2012 [4]. PKU is a global CO₂ emission inventory generated by Peking University using local emissions from fossil fuel combustions, populations, vegetations, and geographical data. The large database of subnational fuel consumption was used for 45 major countries in the PKU data. The fossil, biomass, and solid waste fuels were included and categorized into 64 types in 6 economic sectors, and the uncertainties of the CO₂ emission maps were quantified. The relative uncertainty range of CO₂ emissions could be reduced from 364% to 63.2% by using the sub-national disaggregation [5].

All of these four collected inventories described above are converted to unified Mt CO₂ units, and integrated into same spatial (1 degree) and temporal (a month) scales.

2.2. Methods

2.2.1. Correlating Emission Inventories and Satellite Observed XCO₂

The change of XCO₂ signals, as obtained by satellite observations over a specific target, include the background CO₂ that is globally mixed driven by the atmospheric circulation (C_BK), the anthropogenic CO₂ emissions (C_AE), biological CO₂ fluxes balanced by CO₂ uptake and release from the biospheres (C_Bio), as well as CO₂ fluxes of incoming and outgoing CO₂ driven by wind field (C_Tran) and the error term (ε) [47,48], as shown in Equation (1):

XCO₂ = C_BK + C_AE + C_Bio + C_Tran + ε(1)

Accordingly, XCO₂ is bound to correlate with the anthropogenic CO₂ emissions. We can therefore use the XCO₂ as a base to compare the emissions in the four inventories (CHRED, PKU, ODIAC and EDGAR) created based on different data sources and methods. We defined the spatial XCO₂ patterns by the clustering analysis of the annual mean Mapping-XCO₂ in 2012. The clustering analysis was implemented by slicing XCO₂ from the minimum (390.4 ppm) to the maximum XCO₂ (394.4 ppm) with 0.2 ppm step into 20 segments in the land region (excluding the ocean islands). The averages of XCO₂ and emissions in CHRED, PKU, EGDAR and ODIAC are computed for each XCO₂ segments, respectively. The XCO₂ enhancement, moreover, is calculated by subtracting the minimum XCO₂ from the XCO₂ average of each segment. The corresponding differences of emissions in the four inventories corresponding to the same XCO₂ level can indicate the consistencies of the four inventories.

2.2.2. Comparing Emission Inventories with Predicted Emissions by a Machine Learning-Based Model Based on Satellite Observations

Based on Equation (1), the anthropogenic CO₂ emissions could be predicted by a neural network-based machine learning model trained using the data of XCO₂, C_BK, C_Bio, C_Tran, and the emission inventory [1,2,3,22,23]. Specifically, for the training datasets, we used: (1) the time series from satellite observations (from 2010 to 2019) for XCO₂; (2) NDVI, GPP, and SIF data to represent variable C_Bio; and (3) wind speed data to represent variable C_Tran. We introduced the generalized regression neural network (GRNN) as a machine-learning simulator to predict the emissions using satellite observation data and emission inventories. The GRNN has the advantage of effectively avoiding the effects of artificial subjective assumptions on the prediction because the training of GRNN generally depends on the samples, less artificial adjustments for the operating parameters, with only one smoothing factor [49]. GRNN, moreover, has a strong nonlinear mapping capability and quick learning speed. These advantages of GRNN make it provide better predictions using less training data, such as the time series [22].

Firstly, we tested the performance of the multiple combinations of variables (XCO₂, dXCO₂, NDVI, GPP, SIF, wind speed, night-time light, emissions in EDGAR) to obtain the optimal combination for predicting the anthropogenic emissions with the minimum deviation from EDGAR by using GRNN built using temporal data for each grid. The dXCO₂ is the background removed XCO₂ by subtracting the latitudinal average of XCO₂ at the same time which could enhance the signals induced by the anthropogenic emissions. As a result, the combination of XCO₂ and SIF shows an optimal prediction among the multiple combinations of variables. This demonstrates that using SIF to represent the ecological flux associated with XCO₂ can better predict the anthropogenic CO₂ emissions. We, moreover, tested the GRNN that was built based on the training of all grids spatially using the same combination of variables for both the global grids and the regional grids in China, respectively. The results indicated that the predictions from the same GRNN for all of the grids showed large deviations.

Through the above testing and analysis, we used XCO₂ and SIF to predict the anthropogenic CO₂ emissions in 2012 and 2019, respectively. The processing flow is shown in Figure A1. The biases of the predicted emissions and inventory emissions could demonstrate the uncertainty of inventory emissions, because the predicting emissions’ base on GRNN are from the fitting of the long-time changing trend of emission inventory under the constraint of the satellite observations (Figure A2 and Figure A3).

The processing (Figure A1) includes the following two key points:

• (1). Predicting area and year

We selected the Northern Hemisphere (NH) with 10,312 grids as the study area. The NH included China, India, North America, and Europe, which accounted for most of the global anthropogenic CO₂ emissions. The dynamic GRNN models were built for each grid using the time series data (2010–2019), counting 10,312 GRNN models in the NH, respectively. The predicting year was focused on the years 2012 and 2019. The prediction for the year 2012 was mainly for evaluating the inventories under different scenarios in China where these four inventories are available. The prediction was also implemented in the year 2019, which was the last year in the used time series, in order to compare with the predictions in 2012 and verify the validity of GRNN model;

• (2). GRNN structure

The neural network module of GRNN included generalized regression neural network in MATLAB. The GRNN network consisted of four layers [50], which were input layer, pattern layer, summation layer, and output layer. The number of neurons in the input layer was the same as the dimensions of sample vectors. The pattern layer was also called the hidden regression layer, the number of neurons in the pattern layer was the same as the number of training samples. The Euclidean distance square exponent between the training samples and the ith neural unit is calculated for the summation layer. The summation layer used two types of pattern summation, one was an arithmetic summation of the outputs of the pattern layer, and the other was a weighted summation by the transfer function of the link between the neurons in the pattern layer and the total neurons in the summation layer. In the final output layer, the predicting value was obtained by calculating the ratio of two neurons in the summing layer.

3. Results

3.1. Correlations between Emission Inventories and Satellite-Observed XCO₂

The emission inventories, CHRED, PKU, ODIAC, and EDGAR, show differences in space which can be seen from Figure 1, as they used different emissions coefficients, proxy data sources, and spatially and temporally disaggregating methods (Table 1). Figure 2 demonstrates their correlations with the variations of XCO₂ spatially which are described in Section 2.2.

The largest inconsistence between the four inventories is with the XCO₂ value of 392 ppm corresponding to high emission levels where ODIAC and EDGAR are 60% and 50% higher than CHRED, respectively. ODIAC and EDGAR generally showed larger emissions than CHRED and PKU. The previous investigations indicated that ODIAC and EDGAR likely also overestimated the regional emissions in China [7]. Figure 1a demonstrates that the overestimations of emissions are also with the XCO₂ value of 392 ppm and the emissions of 7–12 Mt CO₂. EDGAR is much larger than the other inventories for the regions with XCO₂ values of 392.6 ppm and 394.2 ppm, which are, respectively, the results from the abnormal magnitudes of emissions in Shenyang city (662 Mt CO₂) and Shanghai city (833 Mt CO₂). The emissions from the cities in EDGAR are mostly allocated by using the Carbon Monitoring for Action (CARMA) power plant database, which was a global database that gathered and presented the best available estimates of CO₂ emissions for 50,000 power plants [51].

It can be seen from Figure 2b that all of the high emissions in the four inventories showed 3–4 ppm enhancements of XCO₂. CHRED and PKU presented the strongest correlation to the spatial variations of XCO₂ (R² = 0.78) whereas EDGAR showed the lowest correlation (R² = 0.59), and the largest RMSE (9.01 Mt CO₂). This result indicates the spatial distributions of emissions in CHRED and PKU tend to be more accurate than the other inventories. CHRED is suggested to be more reasonable spatially as it used more detailed local data, and the survey of CO₂ emissions from the point sources including the power plants burning fossil fuel and the air pollution monitoring network (Table 1).

Additionally, we implemented the same data processing as in Figure 2 for the ten years’ data (2010–2019) of XCO₂ for ODIAC and EDGAR. The results, as shown in Figure 3, demonstrate that both ODIAC and EDGAR present similar responses to XCO₂ with a slope (4.1–4.3 Mt CO₂/ppm) from 2010–2019, and similar R² of 0.55 and 0.61 for ODIAC and EDGAR, respectively. The high emissions induce the largest CO₂ enhancements of 3.7 ppm in both ODIAC and EDGAR [17].

3.2. Spatial and Temporal Uncertainties of ODIAC and EDGAR

3.2.1. Spatial Distribution of Inventory Emissions

Figure 4 shows the average emissions in inventory, the predictions, and the biases (prediction minus inventory) for ODIAC and EDGAR in the year 2012. The predictions for ODIAC and EDGAR are derived using the satellite datasets of monthly XCO₂ and SIF in 2012 by GRNN built by the training datasets of ODIAC, EDGAR, XCO₂, and SIF during the period from 2010 to 2019 as described in Section 2.2.2. Similarly, the predicted results for the year 2019 are shown in Figure A4.

It can be clearly seen from Figure 4 that, in addition to Figure A4, both of the biases in ODIAC and EDGAR are opposite between China (negative) and the USA (positive). The magnitudes of bias are larger in China than USA either in 2012 or in 2019 (Figure A5 and Table A1). The mean biases in 2012 and 2019 in China are −0.28 ± 0.78 Mt CO₂ and −0.35 ± 3.10 Mt CO₂ for ODIAC and −0.19 ± 0.74 Mt CO₂ and −0.39 ± 1.04 Mt CO₂ for EDGAR, respectively. The large biases are mostly located around the big cities such as Shanghai, Wuhan, and Shenyang. The biases in the USA are much less than those in China and are positive either in 2012 or 2019, which are 0.15 ± 0.27 Mt CO₂ and 0.05 ± 0.16 Mt CO₂ for ODIAC and 0.16 ± 0.41 Mt CO₂ and 0.02 ± 0.47 Mt CO₂ for EDGAR, respectively. These results indicate that both inventories probably overestimated emissions in China and slightly underestimated emissions in the USA. This implies that the uncertainty for both inventories tend to be higher in China than the USA. The biases in India, moreover, are negative for ODIAC but positive for EDGAR in 2012.

Figure 5 illustrates the biases corresponding to the emissions in ODIAC and EDGAR in 2012 by the scatter plot of all grids in China and the USA, respectively. The biases of both ODIAC and EDGAR, as shown in Figure 5 and Figure A6, present much stronger linear correlations with the emissions in China (R² = 0.83–0.91) than the USA (R² = 0.01–0.51), either in 2012 or in 2019, which imply that the uncertainty of emissions in inventories systematically depend on the emissions in China but randomly in the USA. The two grids, located at Shanghai and Shenyang, with abnormal high emissions in China shown in Figure 5c (and Figure A5c) for EDGAR, showed extremely large biases, where the magnitude of emissions in the year 2012 and 2019 are 833 Mt CO₂ and 1064 Mt CO₂ in Shanghai, and 662 Mt CO₂ and 873 Mt CO₂ in Shenyang, respectively. EDGAR allocates abnormally large emissions to these two cities using the source-emitting data of coal-fired power plants from CARMA. Our field investigations and verifications for CARMA data found that nearly half of points’ sources were geographically dislocated or non-existent in CARMA, which is consistent with other investigations [5]. Another investigation found that CARMA has ignored about 1300 small power plants in China [52]. Our investigations found as well that the locations of ground heat sources detected by the satellite observations (Figure A6b) mostly corresponded with the power plants and the chemical industry better than CARMA (Figure A6a). The CARMA data are likely one of the key reasons inducing the large uncertainties of emissions in inventories spatially as the point source emissions accounted for a large proportion of the total emissions [53]. ODIAC allocates the emissions to the point sources better than EDGAR as it filtered the abnormal and likely irrational emitting points in CARMA.

3.2.2. Seasonal Variation of Emissions in Inventories

We investigated the differences in seasonal cycles by the monthly mean in the year 2012 and 2019 for ODIAC and EDGAR in China and in the USA, which are shown in Figure 6. The trends of seasonal variations are consistent between inventory and predictions either in ODIAC or EDGAR. The seasonal inventory emissions in ODIAC are generally larger (6.9% per month) than the predictions from April/May to December in China. The seasonal emissions in EDGAR are generally larger (4.4% per month) than predictions from August to February in the following year while they are less than the predictions in the rest of the months in China. The seasonal inventory emissions in both ODIAC and EDGAR showed a smaller value in the USA (3.4% per month in ODIAC; 5.1% per month in EDGAR) than the predications. However, EDGAR showed the largest bias in July 2019, which was 9.1% less than prediction. The seasonal variations, moreover, presented significantly larger differences between ODIAC and EDGAR in China compared with the USA where EDGAR clearly showed a seasonal cycle with the peak value in December and the trough in July whereas ODIAC only showed an increasing trend up to a maximum in December without a cyclic variation. The seasonal cycles in the USA presented both in EDGAR and ODIAC showed two peaks in January and August, and two troughs in May and October.

The monthly time series of inventory emissions in ODIAC were created from 2011 to 2020 using the climatological seasonality based on the 2013 version of the CDIAC monthly gridded emissions which only covered from 2000 to 2010. The monthly CDIAC data from 2000 to 2010 were constructed using a proportional proxy methodology of monthly available statistical data from the 21 largest fossil fuel CO₂ (FFCO₂)-emitting countries to extend the utility of those data to the globe and to a longer time series [54]. Therefore, the monthly emissions time series from ODIAC, which mostly depend on the statistical CO₂ emission data used in CDIAC, have probably large uncertainties because of the statistical uncertainty of the data in CDIAC which likely did not account for the changes in CO₂ emissions in the recent years [2]. The monthly variations in ODIAC for the Chinese region in Figure 6 demonstrated this issue as well.

The seasonality in EDGAR is the aggregated temporal profiles of the sectors including power generation, agricultural emissions, shipping and aviation emissions for each greenhouse gas in the Northern hemisphere. It is constructed based on the modulations of monthly distributions in emissions from the sectors data which are from the previous studies and the Long Term Ozone Simulation (LOTOS) database of Builtjes [3]. Figure 6 shows that the seasonal variations in EDGAR tend to be more reasonable than ODIAC according to the seasonal cycle of human CO₂ emission activities in China, such as the increase in FFCO₂ for heating in winter and the decrease in summer in Northern China which account for the largest emissions in China. These results indicate that the monthly allocated approaches of emissions in ODIAC need to be rectified in China.

We, moreover, collected the areas with large biases especially in 2019 as the regions of interest (ROIs) in China (Figure 7), to evaluate the uncertainty of seasonal variations of emissions in the ROIs. Figure 8 presents the time series of mean emissions from 2009 to 2019 in ODIAC and EDGAR for each ROI, and the monthly variation of predictions in 2012 and 2019, respectively.

As shown in Figure 8, the emissions in the three cities that are much higher than the other areas are in the order of Shanghai, Beijing, and Shenyang in ODIAC. However, the emissions in EDGAR are in the order of Shenyang, Shanghai, and Wuhan, and all of these cities also show the largest biases. The differences of emissions between ODIAC and EDGAR in the cities are likely due to spatial allocating issues of emissions using the CARMA data as described above. The biases are larger in EDGAR than in ODIAC. The largest biases in EDGAR are −3.96 Mt CO₂ and −5.92 Mt CO₂ in Shenyang in October in 2012 and 2019, respectively. The largest biases in ODIAC, however, are −2.22 Mt CO₂ and −2.38 Mt CO₂ in Shanghai in December in 2012 and 2019, respectively.

The annual growth rate of emissions from 2009 to 2019 in EDGAR, 0.048 Mt CO₂ year⁻¹ (R² = 0.85), is greater than that of ODIAC, 0.012 Mt CO₂ year⁻¹ (R² = 0.48) in the three cities with the largest emissions. The annual growth rate in ODIAC is probably more reasonable than EDGAR as the growth rate of anthropogenic CO₂ emissions has been generally slowing down by the activities of reducing CO₂ emissions through development of clean energy, switching from coal to gas, and moving large air-polluting factories out of megacities such as Beijing and Shanghai after 2010, the starting year of the new 5-year economic development plan in China [55].

4. Discussion

The overestimations of emissions in China and underestimations in the USA, shown in Figure 4 and Figure A4 as well as Table A1 in Appendix A, are most likely caused by the uncertainty of inventory emissions as we can assume that the uncertainties of satellite-observed data are consistent spatially at a global scale due to their unified instrumental characteristics and retrieval algorithms. We also evaluated the performances of ODIAC and EDGAR by comparing them with CHRED. The emissions in CHRED are believed to be the most detailed and reasonable in representing the spatial pattern in China as it applied more detailed survey data of anthropogenic CO₂ emissions from point sources for the spatial allocation of the national total. Accordingly, using CHRED as the criterion, the spatial contrasts of inventory emissions between CHRED with PKU, ODIAC, and EDGAR are plotted for all of the grids which are shown in Figure 9, and for the means calculated based on XCO₂ segments (segmenting-XCO₂-based) and based on the provincial regions (province-based), respectively, which are shown in Figure 10.

The spatial distributions of emissions in PKU, ODIAC, and EDGAR are not highly correlated with CHRED in the grid unit where all R² are approximately 0.5, and the standard deviations (Std) are from 15 Mt CO₂ to 17 Mt CO₂ (indicated in Figure 9). It can be seen from Figure 10 that all inventories show better consistency with CHRED in the segmenting-XCO₂-based emissions (where R² is 0.88–0.96 and the Std are 1.5–3.1 Mt CO₂) than those in the province-based emissions (where R² is 0.53–0.84 and the Std are from 9 Mt CO₂ to 130 Mt CO₂). PKU emissions have the minimum deviation to CHRED while EDGAR shows the largest deviation (130 Mt CO₂) in the province-based emissions that is much larger than PKU (9 Mt CO₂), and ODIAC (39 Mt CO₂). The better agreement of PKU, ODIAC, and EDGAR with CHRED in the segmenting-XCO₂ emissions implies that spatial allocations of emissions in these inventories could be constrained by the XCO₂ as the clustering analysis of XCO₂ integrates the strong emitting sources and weak sources regions. The large inconsistency in the province-based emissions, especially in EDGAR, indicates its high uncertainty in the spatial disaggregation by using the proxy data. The predicted emissions for ODIAC and EDGAR, moreover, only demonstrate minor variations to the original emissions as shown in Figure 9 and Figure 10.

5. Conclusions

In this study, we applied two data-driven approaches that use satellite observations from GOSAT and OCO-2 to evaluate the performance of emission inventories from the widely used CHRED, PKU, ODIAC, and EDGAR. The evaluation was carried out by quantifying the correlations of the four inventories with XCO₂ anomalies derived from the satellite observations, and comparing emission inventories with emissions predicted by a machine learning-based model GRNN trained using satellite-observed XCO₂. In addition, we assessed the uncertainties of ODIAC and EDGAR for their spatial distribution and seasonal variations in 2012 and 2019, respectively, by GRNN.

Our results show that the spatial characteristics of four inventories have a good correlation with the spatially clustering XCO₂. However, EDGAR has the lowest correlation with the XCO₂ enhancements and at the same time has the largest deviation from CHRED based on the province-based emissions. This is due to the inexact attribution of the large emitting sources spatially. The inventories from both ODIAC and EDGAR show larger uncertainties in Asia, especially in megacities, than in Europe and the USA. The prediction biases in either ODIAC or EDGAR show systemic errors which are linearly dependent on the magnitude of emissions in China, whereas they are weakly dependent on the emissions in the USA. Generally, the biases in EDGAR are spatially larger than those in ODIAC whereas the seasonal variations in EDGAR tend to be more reasonable than ODIAC. The long-term growing trend in EDGAR, however, is likely overestimated comparing with ODIAC in 2019.

The emissions’ inventories have large uncertainty in their spatial and temporal disaggregation of the national emissions. This is due to their intense use of proxy datasets to represent anthropogenic emissions, such as the fossil fuel consumption from industries which are the primary point emission sources, however, satellite-observed night lights, population density, and many other proxy datasets are not accurate enough to characterize the anthropogenic fossil fuel emissions in China [56]. The results in this study indicate that the satellite observations can reveal the spatial and temporal uncertainty of inventory emissions. With more and more satellite data becoming available, the errors from using satellite-observed XCO₂ can be continuously reduced [57]. Therefore, satellite observations have great potential to improve the spatial and temporal disaggregation of emissions in inventories. The application of satellite observations as a top-down data-driven approach, as demonstrated in this study, would be expected to assist in verifying the emission inventories, as suggested in the IPCC report.

Author Contributions

Conceptualization, L.L. (Liping Lei) and S.Z.; Data curation, L.L. (Liping Lei), Z.Z. and C.M.; Formal analysis, M.S., K.G., L.L. (Luman Li) and L.L. (Liping Lei); Methodology, L.L. (Liping Lei), L.L. (Liangyun Liu) and S.Z.; Software, S.Z., M.S. and H.S. All authors have read and agreed to the published version of the manuscript.

Acknowledgments

We are grateful for the ACOS-GOSAT v9r data and OCO-2 v10r data, which were provided by the ACOS/OCO-2 project at the Jet Propulsion Laboratory, California Institute of Technology and obtained from the ACOS/OCO-2 data archive maintained at the NASA Goddard Earth Science Data and Information Services Center. The datasets of ODIAC, EDGAR, PKU, and CEADs are freely available from http://db.cger.nies.go.jp/dataset/ODIAC/ (accessed on 20 January 2021), https://meta.icos-cp.eu/collections/unv31HYRKgullLjJ99O5YCsG (accessed on 20 January 2021) and http://inventory.pku.edu.cn/download/download.html (accessed on 22 January 2021), respectively. CHRED is available from the data developers upon request.

Conflicts of Interest

The authors declare no conflict of interest.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Figures and Table

Figure 1. Spatial distribution of the total CO2 emissions in the year 2012 (0.1 degree grid) for four inventories in China: (a) CHRED, (b) PKU, (c) ODIAC, and (d) EDGAR.

Figure 2. Correlation between emissions from bottom-up inventories and XCO2 from satellites in 2012. (a) Averaged emissions corresponding to different XCO2 in each segment of XCO2; (b) Correlation between bottom-up inventory emissions and satellite-observed XCO2 enhancement.

Figure 3. The correlation between bottom-up inventory emissions from ODIAC and EDGAR and satellite-observed XCO2 enhancement for ten years from 2010 to 2019. Different colors represent different inventory emissions data, where red represents EDGAR and blue represents ODIAC.

Figure 4. Averages of monthly inventories and predicted emissions based on GRNN using satellite observations in 2012 for (a) original ODIAC, (b) predicted emissions for ODIAC, (c) original EDGAR, (d) predicted emissions for EDGAR, (e) bias for ODIAC, and (f) bias for EDGAR.

Figure 5. Scatter plots of the bias corresponding to the emissions in inventory in 2012: (a) ODIAC in China, (b) EDGAR in China where red color is the result excluding two grids with abnormal emissions in the cities, Shanghai and Shenyang, (c) ODIAC in the USA, (d) EDGAR in the USA.

Figure 6. Seasonal cycles of emissions and predictions in ODIAC and EDGAR illustrated by the monthly average in 2012 for (a) China and (b) the USA, and in 2019 for (c) China and (d) the USA.

Figure 7. Locations of ROIs in China overlapped with the prediction biases for EDGAR in 2019 where the red circles tagged by C1–C7 are around the cities, and S1–S4 are around towns and villages.

Figure 8. Time series of averaged inventory emissions in the ROIs from 2009 to 2019 (left column) and the monthly prediction biases in the ROIs for 2012 (middle column) and 2019 (right column) for (a) ODIAC and (b) EDGAR.

Figure 8. Time series of averaged inventory emissions in the ROIs from 2009 to 2019 (left column) and the monthly prediction biases in the ROIs for 2012 (middle column) and 2019 (right column) for (a) ODIAC and (b) EDGAR.

Figure 9. Scatter plot of gridded emissions in 2012 in China from CHRED emissions (x-axis) in contrast to the PKU, ODIAC, and EDGAR emissions (y-axis).

Figure 10. CHRED (x-axis) in contrast to PKU, ODIAC, and EDGAR (y-axis) based on regional mean emissions in 2012 in China, in which (a) is the segementing-XCO2-based emissions which are from the clustering analysis of XCO2, and (b) is the province-based emissions which are averaged values for each province in China.

Table 1

Specifications of four selected anthropogenic CO₂ emission inventories.

Specification	CHRED	PKU	ODIAC	EDGAR
Covering area	China	Global	Global	Global
Period used	2012	2012	2009–2019	2009–2019
Resolution(Time/space)	Monthly/10 km	Monthly/0.1 deg	Monthly/1 km	Monthly/0.1 deg
Emission factor for raw coal(tC per ton coal)	0.518	0.518	0.746	0.713
Uncertainty	8%	19%	17.5%	15%
Energy statistic source	CESY	CESY	IEA	IEA
Point source	FCPSC	CARMA 2.0	CARMA 2.0	CARMA 3.0
Line source	national road, railway, navigation network, and traffic flows	Not available	Not available	OpenStreetMap and OpenRailway Map, Int.aviation and bunker
Area source	Population density, land use, human activity	Night-time light, Population density, Vegetation	Night-time light	Night-time light, Population density
Version	CHRED	PKU-CO2-v2	ODIAC2020	EDGARv4.3.2_bp
Released source	Data developer	http://inventory.pku.edu.cn/download/download.html (accessed on 10 January 2021)	http://db.cger.nies.go.jp/dataset/ODIAC/ (accessed on 10 January 2021)	https://meta.icos-cp.eu/collections/unv31HYRKgullLjJ99O5YCsG (accessed on 10 January 2021)
References	Cai et al. (2018) [4]	Wang et al. (2013) [5]	Oda et al. (2018) [2]	Janssens-Maenhout (2017) [3]

Note: FCPSC: the First China Pollution Source Census; CESY: China Energy Statistical Yearbook.

Appendix A

Table A1

Statistics of inventory emissions results in different regions.

Units (Mt CO2)			China	USA	India	Europe
2012	ODIAC	Inventory	9696	4763	1827	3453
		Prediction	9396	4904	1751	3423
		Total bias	−300	141	−76	−30
		Mean bias	−0.27	0.15	−0.28	−0.05
		Maximum bias	0.32	2.24	0.11	1.54
		Minimum bias	−11.42	−0.19	−6.30	−3.16
		Standard deviation	0.78	0.27	0.59	0.32
	EDGAR	Inventory	11,546	4844	2843	4119
		Prediction	11,334	4997	2883	4049
		Total bias	−212	153	41	−70
		Mean bias	−0.19	0.16	0.15	−0.11
		Maximum bias	1.22	3.70	6.79	1.57
		Minimum bias	−13.68	−0.81	−1.51	−7.75
		Standard deviation	0.74	0.41	0.53	0.54
2019	ODIAC	Inventory	10,372	4707	2455	3068
		Prediction	9947	4754	2363	3252
		Total bias	−425	47	−92	184
		Mean bias	−0.38	0.05	−0.34	0.28
		Maximum bias	0.22	2.71	0.01	8.64
		Minimum bias	−12.37	−0.45	−5.36	−2.15
		Standard deviation	1.04	0.16	0.62	0.73
	EDGAR	Inventory	12,996	4752	3504	3777
		Prediction	12,641	4773	3224	3889
		Total bias	−355	21	−280	113
		Mean bias	−0.32	0.02	−1.03	0.17
		Maximum bias	3.12	3.21	1.20	8.15
		Minimum bias	−73.43	−4.87	−31.06	−1.35
		Standard deviation	3.00	0.48	2.32	0.66
Number of sampling grids			1105	964	273	664

Figure A1. Workflow for training the GRNN neural network model and evaluating emission inventory using the predicted emissions based on GRNN trained using the satellite XCO2 observations.

Figure A2. The gridded mean of (a) XCO2 and (b) SIF in the Northern Hemisphere from April 2009 to December 2020.

Figure A3. Temporal variations of (a) averaged emissions of ODIAC, EDGAR, and (b) XCO2 and SIF in the Northern Hemisphere.

Figure A4. Averages of monthly inventories and predicted emissions based on the GRNN using satellite observations in 2019 for (a) original ODIAC, (b) predicted emissions for ODIAC, (c) original EDGAR, (d) predicted emissions for EDGAR, (e) bias for ODIAC, and (f) bias for EDGAR.

Figure A5. Scatter plots of the bias corresponding to the inventory emissions in 2019 for (a) ODIAC in China; (b) EDGAR in China where red color is the result of excluding two grids with abnormal emissions in the cities of Shanghai and Shenyang; (c) ODIAC in the USA, and (d) EDGAR in the USA.

Figure A5. Scatter plots of the bias corresponding to the inventory emissions in 2019 for (a) ODIAC in China; (b) EDGAR in China where red color is the result of excluding two grids with abnormal emissions in the cities of Shanghai and Shenyang; (c) ODIAC in the USA, and (d) EDGAR in the USA.

Figure A6. The distributions of emitting point sources from (a) CARMA and (b) heat source detected by satellite observations in China.

Figure A6. The distributions of emitting point sources from (a) CARMA and (b) heat source detected by satellite observations in China.

References

1. Oda, T.; Maksyutov, S. A very high-resolution (1 km × 1 km) global fossil fuel CO₂ emission inventory derived using a point source database and satellite observations of nighttime lights. Atmos. Chem. Phys.; 2011; 11, pp. 543-556. [DOI: https://dx.doi.org/10.5194/acp-11-543-2011]

2. Oda, T.; Maksyutov, S.; Andres, R.J. The Open-source Data Inventory for Anthropogenic Carbon dioxide (CO₂), version 2016 (ODIAC2016): A global, monthly fossil-fuel CO₂ gridded emission data product for tracer transport simulations and surface flux inversions. Earth Syst. Sci. Data; 2018; 10, pp. 87-107. [DOI: https://dx.doi.org/10.5194/essd-10-87-2018] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31662803]

3. Janssens-Maenhout, G.; Crippa, M.; Guizzardi, D.; Muntean, M.; Petrescu, A. Edgar v4.3.2 global atlas of the three major greenhouse gas emissions for the period 1970–2012. Earth Syst. Sci. Data Discuss.; 2017; 11, pp. 959-1002. [DOI: https://dx.doi.org/10.5194/essd-11-959-2019]

4. Cai, B.; Liang, S.; Zhou, J.; Wang, J.; Cao, L.; Qu, S.; Xu, M.; Yang, Z. China high resolution emission database (CHRED) with point emission sources, gridded emission data, and supplementary socioeconomic data. Resour. Conserv. Recycl.; 2018; 129, pp. 232-239. [DOI: https://dx.doi.org/10.1016/j.resconrec.2017.10.036]

5. Wang, R.; Tao, S.; Ciais, P.; Shen, H.Z.; Huang, Y.; Chen, H.; Shen, G.F.; Wang, B.; Li, W.; Zhang, Y.Y. et al. High-resolution mapping of combustion processes and implications for CO₂ emissions. Atmos. Chem. Phys.; 2013; 13, pp. 5189-5203. [DOI: https://dx.doi.org/10.5194/acp-13-5189-2013]

6. Andres, R.J.; Boden, T.A.; Higdon, D. A new evaluation of the uncertainty associated with CDIAC estimates of fossil fuel carbon dioxide emission. Tellus B Chem. Phys. Meteorol.; 2014; 66, 23616. [DOI: https://dx.doi.org/10.3402/tellusb.v66.23616]

7. Han, P.; Zeng, N.; Oda, T.; Lin, X.; Crippa, M.; Guan, D.; Janssens-Maenhout, G.; Ma, X.; Liu, Z.; Shan, Y. et al. Evaluating China’s fossil-fuel CO₂ emissions from a comprehensive dataset of nine inventories. Atmos. Chem. Phys.; 2020; 20, pp. 11371-11385. [DOI: https://dx.doi.org/10.5194/acp-20-11371-2020]

8. Andres, R.J.; Boden, T.A.; Higdon, D.M. Gridded uncertainty in fossil fuel carbon dioxide emission maps, a CDIAC example. Atmos. Chem. Phys.; 2016; 16, pp. 14979-14995. [DOI: https://dx.doi.org/10.5194/acp-16-14979-2016]

9. Gurney, K.; Razlivanov, I.; Song, Y.; Zhou, Y.; Benes, B.; Abdul-Massih, M. Quantification of fossil fuel CO₂ emission on the building/street scale for a large US city. Env. Sci Technol.; 2012; 46, pp. 12194-12202. [DOI: https://dx.doi.org/10.1021/es3011282]

10. Oda, T.; Bun, R.; Kinakh, V.; Topylko, P.; Halushchak, M.; Marland, G.; Lauvaux, T.; Jonas, M.; Maksyutov, S.; Nahorski, Z. et al. Errors and uncertainties in gridded carbon dioxide emissions inventory. Mitig. Adapt. Strateg. Glob. Change; 2019; 24, pp. 1007-1050. [DOI: https://dx.doi.org/10.1007/s11027-019-09877-2]

11. Lauvaux, T.; Miles, N.L.; Deng, A.; Richardson, S.J.; Cambaliza, M.O.; Davis, K.J.; Gaudet, B.; Gurney, K.R.; Huang, J.; O’Keefe, D. et al. High-resolution atmospheric inversion of urban CO₂ emissions during the dormant season of the Indianapolis flux experiment (INFLUX). J. Geophys. Res. Atmos.; 2016; 121, pp. 5213-5236. [DOI: https://dx.doi.org/10.1002/2015JD024473]

12. Chevallier, F.; Broquet, G.; Zheng, B.; Ciais, P.; Eldering, A. Large CO₂ emitters as seen from satellite: Comparison to a gridded global emission inventory. Geophys. Res. Lett.; 2020; 49, e2021GL097540.

13. Jonas, M.; Marland, G.; Krey, V.; Wagner, F.; Nahorski, Z. Uncertainty in an emissions-constrained world. Clim. Change; 2014; 124, pp. 459-476. [DOI: https://dx.doi.org/10.1007/s10584-014-1103-6]

14. Schneising, O.; Heymann, J.; Buchwitz, M.; Reuter, M.; Bovensmann, H.; Burrows, J. Anthropogenic carbon dioxide source areas observed from space: Evaluatement of regional enhancements and trends. Atmos. Chem. Phys.; 2013; 13, pp. 2445-2454. [DOI: https://dx.doi.org/10.5194/acp-13-2445-2013]

15. Keppel-Aleks, G.; Wennberg, P.O.; O’Dell, C.W.; Wunch, D. Towards constraints on fossil fuel emissions from total column carbon dioxide. Atmos. Chem. Phys.; 2013; 13, pp. 4349-4357. [DOI: https://dx.doi.org/10.5194/acp-13-4349-2013]

16. Lei, L.P.; Zhong, H.; He, Z.H.; Cai, B.F.; Yang, S.Y.; Wu, C.J.; Zeng, Z.C.; Liu, L.Y.; Zhang, B. Assessment of atmospheric CO₂ concentration enhancement from anthropogenic emissions based on satellite observations. Chin. Sci. Bull.; 2017; 62, pp. 2941-2950. [DOI: https://dx.doi.org/10.1360/N972016-01316]

17. Sheng, M.; Lei, L.; Zeng, Z.-C.; Rao, W.; Zhang, S. Detecting the Responses of CO₂ Column Abundances to Anthropogenic Emissions from Satellite Observations of GOSAT and OCO-2. Remote Sens.; 2021; 13, 3524. [DOI: https://dx.doi.org/10.3390/rs13173524]

18. Bovensmann, H.; Buchwitz, M.; Burrows, J.P.; Reuter, M.; Krings, T.; Gerilowski, K.; Schneising, O.; Heymann, J.; Tretner, A.; Erzinger, J. A remote sensing technique for global monitoring of power plant CO₂ emissions from space and related applications. Atmos. Meas. Tech.; 2010; 3, pp. 781-811. [DOI: https://dx.doi.org/10.5194/amt-3-781-2010]

19. Liu, D.; Lei, L.P.; Guo, L.J.; Zeng, Z.C. A Cluster of CO₂ Change Characteristics with GOSAT Observations for Viewing the Spatial Pattern of CO₂ Emission and Absorption. Atmosphere; 2015; 6, pp. 1695-1713. [DOI: https://dx.doi.org/10.3390/atmos6111695]

20. Bie, N.; Lei, L.; He, Z.; Zeng, Z.; Cai, B. Specific patterns of XCO₂ observed by gosat during 2009–2016 and assessed with model simulations over China. Sci. China Earth Sci.; 2020; 63, pp. 384-394. [DOI: https://dx.doi.org/10.1007/s11430-018-9377-7]

21. Hakkarainen, J.; Ialongo, I.; Tamminen, J. Direct space-based observations of anthropogenic CO₂ emission areas from OCO-2. Geophys. Res. Lett.; 2016; 43, pp. 400-411. [DOI: https://dx.doi.org/10.1002/2016GL070885]

22. Yang, S.; Lei, L.; Zeng, Z.; He, Z.; Zhong, H. An Evaluatement of Anthropogenic CO₂ Emissions by Satellite-Based Observations in China. Sensors; 2019; 19, 1118. [DOI: https://dx.doi.org/10.3390/s19051118]

23. Mustafa, F.; Bu, L.; Wang, Q.; Yao, N.; Shahzaman, M.; Bilal, M.; Aslam, R.W.; Iqbal, R.; Mustafa, F.; Bu, L. et al. Neural Network Based Estimation of Regional Scale Anthropogenic CO₂ Emissions Using OCO-2 Dataset Over East and West Asia. Atmos. Meas. Tech.; 2021; 14, pp. 7277-7290. [DOI: https://dx.doi.org/10.5194/amt-14-7277-2021]

24. Pinty, B.; Janssens-Maenhout, G.; Dowell, M.; Zunker, H.; Brunhes, T.; Ciais, P.; Dee, D.; Denier van der Gon, H.; Dolman, H.; Drinkwater, M. et al. Toward an operational anthropogenic CO₂ emissions monitoring & verification support capacity—Baseline requirements, model components and functional architecture. Eur. Comm. Jt. Res. Cent.; 2020; 101, pp. 1439-1451.

25. IPCC. Summary for Policymakers. Global Warming of 1.5 °C; Masson-Delmotte, V.; Zhai, P.; Pörtner, H.-O.; Roberts, D.; Skea, J.; Shukla, P.R.; Pirani, A.; Moufouma-Okia, W.; Péan, C.; Pidcock, R. et al. World Meteorological Organization: Geneva, Switzerland, 2018; pp. 6-13.

26. Goddard Earth Science Data Information and Services Center (GES DISC) at National Aeronautics and Space Administration (NASA). Available online: https://oco2.gesdisc.eosdis.nasa.gov/data/ (accessed on 19 January 2021).

27. Crisp, D.; Frankenberg, C.; Messerschmidt, J.; Wennberg, P.O.; Wunch, D.; Yung, Y.L. The ACOS CO₂ retrieval algorithm—Part II: Global XCO₂ data characterization. Atmos. Meas. Tech.; 2012; 5, pp. 687-707. [DOI: https://dx.doi.org/10.5194/amt-5-687-2012]

28. O’Dell, C.W.; Eldering, A.; Wennberg, P.O.; Crisp, D.; Gunson, M.R.; Fisher, B.; Frankenberg, C.; Velazco, A. Improved retrievals of carbon dioxide from Orbiting Carbon observatory-2 with the version 8 ACOS algorithm. Atmos. Meas. Tech.; 2018; 11, pp. 6539-6576. [DOI: https://dx.doi.org/10.5194/amt-11-6539-2018]

29. Kiel, M.; O’Dell, C.W.; Fisher, B.; Eldering, A.; Nassar, R.; MacDonald, C.G.; Wennberg, P.O. How bias correction goes wrong: Measurement of XCO₂ affected by erroneous surface pressure estimates. Atmos. Meas. Tech.; 2019; 12, pp. 2241-2259. [DOI: https://dx.doi.org/10.5194/amt-12-2241-2019]

30. Zeng, Z.-C.; Lei, L.; Hou, S.; Ru, F.; Guan, X.; Zhang, B. A regional gap-filling method based on spatiotemporal variogram model of columns. IEEE Trans. Geosci. Remote Sens.; 2014; 52, pp. 3594-3603. [DOI: https://dx.doi.org/10.1109/TGRS.2013.2273807]

31. Zeng, Z.-C.; Lei, L.; Strong, K.; Jones, D.B.A.; Guo, L.; Liu, M.; Deng, F.; Deutscher, N.M.; Dubey, M.K.; Griffith, D.W.T. et al. Global land mapping of satellite-observed CO₂ total columns using spatio-temporal geostatistics. Int. J. Digit. Earth; 2017; 10, pp. 426-456. [DOI: https://dx.doi.org/10.1080/17538947.2016.1156777]

32. Guo, L.J.; Lei, L.P.; Zeng, Z.-C.; Zou, P.F.; Liu, D.; Zhang, B. Evaluation of spatio-temporal variogram models for Mapping XCO₂ using satellite observations: A Case Study in China. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2015; 8, pp. 376-385. [DOI: https://dx.doi.org/10.1109/JSTARS.2014.2363019]

33. He, Z.; Lei, L.; Zhang, Y.; Sheng, M.; Wu, C.; Li, L.; Zeng, Z.-C.; Welp, L.R. Spatio-temporal mapping of multi-satellite observed column atmospheric CO₂ using precision-weighted kriging method. Remote Sens.; 2020; 12, 576. [DOI: https://dx.doi.org/10.3390/rs12030576]

34. Xiao, J.; Li, X.; He, B.; Arain, M.A.; Beringer, J.; Desai, A.R.; Emmel, C.; Hollinger, D.Y.; Krasnova, A.; Mammarella, I. et al. Solar-induced chlorophyll fluorescence exhibits a universal relationship with gross primary productivity across a wide variety of biomes. Glob. Change Biol.; 2019; 25, pp. e4-e6. [DOI: https://dx.doi.org/10.1111/gcb.14565] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30614142]

35. Grace, J.; Nichol, C.; Disney, M.; Lewis, P.; Quaife, T.; Bowyer, P. Can we measure terrestrial photosynthesis from space directly, using spectral reflectance and fluorescence?. Glob. Change Biol.; 2007; 13, pp. 1484-1497. [DOI: https://dx.doi.org/10.1111/j.1365-2486.2007.01352.x]

36. Meroni, M.; Rossini, M.; Guanter, L.; Alonso, L.; Rascher, U.; Colombo, R.; Moreno, J. Remote sensing of solar-induced chlorophyll fluorescence: Review of methods and applications. Remote Sens. Environ.; 2009; 113, pp. 2037-2051. [DOI: https://dx.doi.org/10.1016/j.rse.2009.05.003]

37. Malmström, C.M.; Thompson, M.V.; Juday, G.P.; Los, S.O.; Randerson, J.T.; Field, C.B. Interannual variation in global-scale net primary production: Testing model estimates. Glob. Biogeochem. Cycles; 1997; 11, pp. 367-392. [DOI: https://dx.doi.org/10.1029/97GB01419]

38. Land Processes Distributed Active Archive Center (LP DAAC). Available online: https://ladsweb.modaps.eosdis.nasa.gov/ (accessed on 14 January 2021).

39. Copernicus Climate Data Store (CDS). Available online: https://cds.climate.copernicus.eu/ (accessed on 24 January 2021).

40. Stephen, M.; Stephanie, W.; Calvin, L. VIIRS day/night band (DNB) stray light characterization and correction. Proceedings of the SPIE; San Diego, CA, USA, 25–29 August 2013.

41. Integrated Carbon Observation System(ICOS). Available online: https://meta.icos-cp.eu/collections/unv31HYRKgullLjJ99O5YCsG/ (accessed on 18 January 2021).

42. Kuenen, J.J.P.; Visschedijk, A.J.H.; Jozwicka, M.; van der Gon, H.A.C.D. Tno-macc_ii emission inventory; a multi-year (2003–2009) consistent high-resolution european emission inventory for air quality modelling. Atmos. Chem. Phys.; 2014; 14, pp. 10963-10976. [DOI: https://dx.doi.org/10.5194/acp-14-10963-2014]

43. Steinbach, J.; Gerbig, C.; Denbeck, C.R.; Karstens, U.; Minejima, C.; Mukai, H. The CO₂ release and oxygen uptake from fossil fuel emission estimate (coffee) dataset: Effects from varying oxidative ratios. Atmos. Chem. Phys.; 2011; 11, pp. 6855-6870. [DOI: https://dx.doi.org/10.5194/acp-11-6855-2011]

44. Guan, D.; Liu, Z.; Geng, Y.; Lindner, S.; Hubacek, K. The gigatonne gap in China’s carbon dioxide inventories. Nat. Clim. Change; 2012; 2, pp. 672-675. [DOI: https://dx.doi.org/10.1038/nclimate1560]

45. Keppel-Aleks, G.; Wennberg, P.O.; Schneider, T. Sources of variations in total column carbon dioxide. Atmos. Chem. Phys.; 2011; 11, pp. 3581-3593. [DOI: https://dx.doi.org/10.5194/acp-11-3581-2011]

46. Hutchins, M.G.; Colby, J.D.; Marland, G.; Marland, E. A comparison of five high-resolution spatially-explicit, fossil-fuel, carbon 395 dioxide emission inventories for the United States. Mitig. Adapt. Strateg. Glob. Change; 2017; 22, 947. [DOI: https://dx.doi.org/10.1007/s11027-016-9709-9]

47. Veldt, C. Updating and Upgrading the PHOXA Emission Data Base to 1990; TNO report Netherlands Organisation for Applied Scientific Research: Apeldoorn, The Netherlands, 1992; pp. 92-118.

48. Friedrich, R.; Reis, S. Emissions of Air Pollutants—Measurements, Calculations and Uncertainties; Springer: Berlin/Heidelberg, Germany, 2004; 335.

49. Specht, D.F. A general regression neural network. IEEE Trans. Neural Netw.; 1991; 2, pp. 569-576. [DOI: https://dx.doi.org/10.1109/72.97934]

50. Cigizoglu, H.K.; Alp, M. Generalized regression neural network in modelling river sediment yield. Adv. Eng. Softw.; 2006; 37, pp. 63-68. [DOI: https://dx.doi.org/10.1016/j.advengsoft.2005.05.002]

51. Center for Global Development (CGD). Available online: https://www.cgdev.org/ (accessed on 24 January 2021).

52. Wang, C.; Corbett, J.J.; Firestone, J. Improving Spatial Representation of Global Ship Emissions Inventories. Environ. Sci. Technol.; 2008; 42, pp. 193-199. [DOI: https://dx.doi.org/10.1021/es0700799]

53. Builtjes, P.J.H. The LOTOS Long Term Ozone Simulation Project; Summary Report TNO Technical Report R92/245 TNO-MW TNO: Delft, The Netherlands, 1992.

54. Andres, R.J.; Gregg, J.S.; Losey, L.; Marland, G.; Boden, T.A. Monthly, global emissions of carbon dioxide from 355 fossil fuel consumption. Tellus B Chem. Phys. Meteorol.; 2011; 63, pp. 309-327.

55. The People’s Republic of China on the 13th Five-Year Plan for National Economic and Social Development Program [EB/OL]. Available online: http://www.gov.cn/xinwen/2016-03/17/content_5054992.html/ (accessed on 15 January 2021).

56. Ma, C.; Yang, J.; Xia, W.; Liu, J.; Zhang, Y.; Sui, X. A Model for Expressing Industrial Information Based on Object-Oriented Industrial Heat Sources Detected Using Multi-Source Thermal Anomaly Data in China. Remote Sens.; 2022; 14, 835. [DOI: https://dx.doi.org/10.3390/rs14040835]

57. Liu, L.Y.; Chen, L.F.; Liu, Y.; Yang, D.X.; Zhang, X.Y.; Lu, N.M.; Ju, W.M.; Jiang, F.; Yin, Z.S.; Liu, G.H. Satellite remote sensing for global stocktaking: Methods, progress and perspectives. Natl. Remote Sens. Bull.; 2022; 26, pp. 243-267.