CDIAC FFCO2 emissions estimates are based on fuel statistic data published as the United Nation Energy Statistics Database (Boden et al., 2017). Emissions estimates are calculated on a global, national, and regional basis and by fuel type in the method described in Marland and Rotty (1984). CDIAC also provides their own gridded emissions data products that indicate annual and monthly FFCO2 fields at a 1 × 1∘ resolution (Andres et al., 1996, 2011). ODIAC2016 is primarily based on the year 2016 version of the CDIAC national estimates (Boden et al., 2016), which were the most up-to-date CDIAC emissions estimates at the time of the data development (currently Boden et al., 2017, is the latest). We first aggregated the CDIAC national (and regional) emissions estimates to 65 countries and 6 geographical regions (North America, South and Central America, Europe and Eurasia, the Middle East, Africa, and Asia Pacific) defined in Oda and Maksyutov (2011) (see the country/region definitions are shown in Table 1 in Oda and Maksyutov, 2011). In addition to the national and geographical categories, we decided to include Antarctic fishery emissions, which are from fishery activities over the Antarctic Ocean (< 60∘ S, 1–4 kTC yr-1 over 1987–2007 by Boden et al., 2016), as an individual emissions region and distributed in the same way as Andres et al. (1996). Emissions from international bunkers and aviation are not included in national emissions by international convention. Thus, CDIAC gridded emissions data products do not include the emissions from international bunkers and aviation although the CDIAC/ORNL does have records of those emissions on a national and regional basis. ODIAC2016 includes those emissions to achieve comprehensive global FFCO2 gridded emissions fields.

In CDIAC emissions estimates, the global total emissions and national total emissions are obtained using different calculation methods (global fuel production vs. apparent national fuel consumption; see Andres et al., 2012) and the CDIAC national totals do not sum to the CDIAC global total due to the difference in calculation method and inconsistencies in the underlying statistical data (e.g., import–export totals) (e.g., Andres et al., 2012). We thus calculate the difference between the global total and the sum of national totals and scaled up national totals to account for the difference. Andres et al. (2014) reported that global total emissions estimates calculated with production data (as opposed to apparent consumption data) have the smallest uncertainty (approximately 8 %; 2σ). It is thus used as the reference for global carbon budget analyses (e.g., Le Quéré et al., 2016). Inversion analysis is an extended version of the global carbon budget analysis using atmospheric models. We thus believe that imposing transport models and/or inversion models in a consistent way with the global carbon budget analysis, as Le Quéré et al. (2016) have done, has significance, although we sacrifice the accuracy of the national and regional emissions estimates. Due to the global scaling, national totals in ODIAC2016 differ from the estimates originally reported by the CDIAC/ORNL. The difference between the CDIAC global total and the sum of national emissions is often a few percent and thus the magnitude of the scaling is often within the uncertainty range of national emissions (e.g., 4.0 to 20.2 %; Andres et al., 2014). The global scaling factors derived and used in this study are presented in Table A2.

The 2016 version of the CDIAC emissions estimates only covers years to 2013 (Boden et al., 2016). We thus extrapolated the 2013 CDIAC emissions to years 2014 and 2015 using the 2016 version of the BP global fuel statistical data (BP, 2017). Our emissions extrapolation approach is the same as Myhre et al. (2009) and Le Quéré et al. (2016). Emissions from cement production and gas flaring (approximately 5.7 and 0.6 % of the 2013 global total; Boden et al., 2016) were assumed to be the same as those in 2013. International bunker emissions were scaled using changes in national total emissions.

CDIAC national emissions estimates (prepared by fuel type) were recategorized into our own ODIAC emissions categories (point source, nonpoint source, cement production, gas flare, and international aviation and marine bunkers). Following Oda and Maksyutov (2011), the sum of emissions from liquid, gas, and solid fuels was further divided into point source emissions and nonpoint source emissions. The total emissions from point sources were estimated using national total power plant emissions calculated using Carbon Monitoring for Action (CARMA; Wheeler and Ummel, 2008) (Oda and Maksyutov, 2011). As mentioned earlier, CDIAC gridded emissions data products only indicate national emissions and do not include international bunker emissions (Andres et al., 1996, 2011). In contrast, EDGAR provides bunker emissions in their gridded data product (JRC, 2017). Peylin et al. (2013) show some models include international bunker emissions and some do not, although the difference due to the inclusion–exclusion of the international bunker emissions in the prescribed emissions could be corrected afterwards (Peylin et al., 2013). In ODIAC2016, we carry CDIAC international bunker emissions reported on a country basis to achieve the complete picture of the global fossil fuel emissions. Country total bunker emissions (aviation plus marine bunkers) were distributed using spatial proxy data adopted from other emissions inventories described later (see Sect. 4.3). Although the CDIAC/ORNL does not report emissions from international aviation and marine bunkers separately, we loosely estimated those two emissions using UN statistics. We estimated the fraction of aircraft emissions using jet fuel and aviation gasoline consumption and then the international bunker emissions were divided into aircraft and marine bunker emissions.

We define the sum of the emissions from solid, liquid, and gas fuels as land emissions (see Fig. 1). Land emissions are further divided into two emissions categories (point source emissions and nonpoint source emissions) and then distributed at a 1 × 1 km resolution in the ways described in Oda and Maksyutov (2011): point source emissions are mapped using power plant profiles (emissions intensity and geographical location) taken from the CARMA database (Wheeler and Ummel, 2008) and nonpoint source emissions are distributed using nighttime light data collected by Defense Meteorological Satellite Program (DMSP) satellites (e.g., Elvidge et al., 1999). To avoid difficulty in emissions disaggregation, especially over bright regions, in nighttime light data (e.g., cities), Oda and Maksyutov (2011) employed a product that does not have an instrument saturation issue rather than a regular nighttime light product. ODIAC2016 employs the latest version of the special nighttime light product (Ziskin et al., 2010). The improved nighttime light data have mitigated the underestimation of emissions over dimmer areas seen in ODIAC v1.7 (Oda et al., 2010). Nighttime light data are currently available for multiple years (1996–1997, 1999, 2000, 2002–2003, 2004, 2005–2006, and 2010). In ODIAC2016, due to the lack of information, the emissions from cement production were spatially distributed as a part of nonpoint source emissions, although those emissions should have been distributed as point sources. This needs to be fixed in future versions of ODIAC emissions data.

In ODIAC v1.7, emissions from gas flaring were not considered (Oda and Maksyutov, 2011). Nighttime light pixels corresponding to gas flares often appear very bright and would result in strong point sources in emissions data (Oda and Maksyutov, 2011). We thus identified and excluded those bright gas flare pixels before distributing land emissions using another global nighttime light data product that was specifically developed for gas flares by NOAA, National Centers for Environmental Information (NCEI, formerly National Geophysical Data Center, NGDC) (Oda and Maksyutov, 2011). In ODIAC2016 we separately distributed CDIAC gas flare emissions using the 1 × 1 km nighttime light-based gas flare maps developed for 65 individual countries (Elvidge et al., 2009). Other than the 65 countries, the gas flare emissions were distributed as a part of land emissions.

Emissions from international aviation and marine bunkers were distributed using aircraft and ship fleet tracks. International aviation emissions were distributed using the AERO2k inventory (Eyers et al., 2005). The AERO2k inventory was developed by a team at the Manchester Metropolitan University and indicates the fuel use and NOx, CO2, CO, hydrocarbon, and particulate emissions for 2002 and 2025 (projected) with injection height at a 1 × 1∘ spatial resolution on a monthly basis. We used their column total CO2 emissions to distribute emissions to a single layer. International marine bunker emissions were distributed at a 0.1 × 0.1∘ resolution using an international marine bunker emissions map from EDGAR v4.1 (JRC, 2017). We decided not to adopt an international and domestic shipping (1A3d) map from EDGAR v4.2 as it includes domestic shipping emissions that we do not distinguish.

In Fig. 2, global emissions time series from different emissions data were compared to give an idea of agreement among them. We calculated the global total for each year from four gridded emissions data for the period of 2000–2016: CDIAC global total + projection (taken from ODIAC2016), CDIAC gridded data (hence, no international bunker emissions), and two versions of EDGAR gridded data (v4.2 and FastTrack). The uncertainty range (shaded in tan) is 8 % (2σ), estimated for CDIAC global by Andres et al. (2014). Those gridded emissions data are often used in global atmospheric CO2 inversion analysis (e.g., Peylin et al., 2013). To account for the difference in emissions reporting categories (e.g., fuel basis in CDIAC vs. emissions sector basis in EDGAR), the EDGAR totals were calculated as the total short cycle C (with the file name “CO2_excl_short-cycle_org_C”) emissions minus the sum of emissions from agriculture (IPCC code: 4C and 4D), land use change and forestry (5A, C, D, F, and 4E), and waste (6C) (see more details on emissions sectors documented in JRC, 2017). International aviation (1A3a) and navigation (1A3b) were thus included in values for EDGAR time series. The authors acknowledge the JRC has updated EDGAR emissions time series for 1970–2012 in November 2014 (JRC, 2017). This study, however, uses gridded emissions data, which are not fully based on the updated emissions estimates, in order to characterize differences from gridded emissions data, especially for potential data users in the modeling community.

All four global total values obtained from four gridded emissions data agree well within 8 % uncertainty. The difference between ODIAC and CDIAC gridded data (3.3–5.7 %) was largely attributable to the international bunker emissions and global correction. ODIAC (where the total was scaled by CDIAC global total) and the two versions of EDGAR showed minor differences in magnitude (0.3–2.7 %) and trend, which are largely attributable to the differences in the underlying statistical data (e.g., UN Statistics Division vs. the US Energy Information Administration from different inventory years) and the emissions calculation method (fuel basis vs. sector basis). Global total estimates at 5-year increments are shown in Table 1. For the years 2014 and 2015, we estimated the global total emissions at 9.836 and 9.844 PgC. Boden et al. (2017) reported the latest estimate for 2014 global total emissions as 9.855 PgC. Our projected 2014 emissions estimate was lower than the latest estimate by approximately 0.02 PgC (0.2 %).

Figure 3 shows the same type of comparison as Fig. 2, but for the top 10 emitting countries (China, US, India, Russian Federation, Japan, Germany, Islamic Republic of Iran, Republic of Korea (South Korea), Saudi Arabia, and Brazil, according to the 2013 ranking reported by CDIAC). We aggregated all four gridded emissions fields to a common 1 × 1∘ field and sampled using the 1 × 1∘ country mask used in CDIAC emissions data development. The annual uncertainty estimates for national total emissions (2σ) are made following the method described by Andres et al. (2014) and values are shown in Table 2. In the analysis presented in Fig. 3, emissions from international aviation (1A3a) and navigation (1A3b) are excluded. All four national total values sampled from four gridded emissions data at a 1 × 1∘ resolution often agree within the uncertainty estimated by Andres et al. (2014). Systematic differences of ODIAC from CDIAC gridded data can be largely explained by (1) global correction (the total was scaled using CDIAC global total) and (2) the differences in emissions disaggregation methods. Although ODIAC is expected to indicate slightly higher values than CDIAC gridded data (often a few percent) because of the global correction (note global correction can be negative, despite the depiction in Fig. 1), ODIAC sometimes indicates values lower that CDIAC gridded data by more than a few percent (see Japan in Fig. 3 as an example). This is due to a sampling error using the 1 × 1∘ country map in the analysis. The aggregated 1 × 1∘ ODIAC field is slightly larger than that of CDIAC, especially because the coastal areas depicted a high resolution in the original 1 × 1 km emissions field. This type of sampling error was discussed in Zhang et al. (2014). ODIAC employs a 1 × 1 km coastline and a 5 × 5 km country mask as described in Oda and Maksyutov (2011). Thus, the use of a 1 × 1∘ CDIAC country map results in missing some land mass (hence, CO2 emissions). Similar sampling errors can happen for countries that are physically small and island countries, depending on the resolution of analysis. Despite the sampling errors, the authors used the CDIAC 1 × 1∘ country map to perform this comparison analysis with CDIAC gridded data as a reference. The lower emissions indicated by ODIAC or EDGAR in this analysis do not always mean the national total emissions are lower. The emissions estimates at a national level often agree well even among different emissions inventories (e.g., Andres et al., 2012).

The global total emissions fields of CDIAC gridded emissions data and ODIAC2016 for the year 2013 (the most recent year CDIAC indicates) are shown in Fig. 4. Emissions fields are shown at a common 1 × 1∘ resolution. The major difference seen between two fields is primarily due to inclusion–exclusion of emissions from international bunker emissions that largely account for the differences indicated in Table 1. A breakdown of ODIAC 2013 emissions fields are presented by emissions category in Fig. 5. The emissions fields for point sources, nonpoint sources, cement production, and gas flaring were produced at a 1 × 1 km resolution in the ODIAC 3.0 model, but as mentioned earlier, we focus on the 1 × 1∘ version of ODIAC2016 in this paper. In CDIAC gridded emissions data, the emissions over land are distributed by population data without fuel type distinction. In the ODIAC 3.0 model, we have added additional layers of consideration in the emissions modeling from the conventional CDIAC model and add the possibility of future improvement with improved emissions proxy data.

In Fig. 6, we compared the four global gridded products over land and also calculated differences from ODIAC2016 (shown in Fig. 7; histograms are presented in Fig. A1). It is often very challenging to evaluate the accuracy and uncertainty of gridded emissions data because of the lack of direct physical measurements on grid scales (Andres et al., 2016). Recent studies have attempted to evaluate the uncertainty of gridded emissions data by comparing emissions data to each other (e.g., Oda et al., 2015; Hutchins et al., 2016). The differences among emissions were used as a proxy for uncertainty. However, it is important to note that such evaluation does not give us an objective measure of which one is closer to truth, beyond characterizing the differences in emissions spatial patterns and magnitudes from methodological viewpoints (e.g., emissions estimation and disaggregation). Some of the gridded emissions data are partially disaggregated using commercial information, and users are often not authorized to fully disclose the information used. This thus makes the comparison even less meaningful and/or significant. Oda et al. (2015) also discussed that emissions inter-comparison approaches often do not allow us to evaluate two distinct uncertainty sources (emissions and disaggregation) separately. In addition, because of the use of emissions proxy for emissions disaggregation (rather than mechanistic modeling), such comparison can be only implemented at an aggregated, coarse spatial resolution. These issues will be further discussed in Sect. 7.

Because of the limitation mentioned above, here we compared emissions data only to characterize the differences that can be explained by the differences in emissions disaggregation methods. We implemented this comparison exercise using the 2008 emissions field aggregated at a 1 × 1∘ resolution. The year 2008 is the most recent year for which all the four emissions fields are available. The major emissions spatial patterns (e.g., emitting regions such as North America, Europe, and East Asia) are overall very similar as the correlations were driven by national emissions estimates (which we already saw to be in good agreement earlier), but we do see differences due to emissions disaggregation at the subnational level. Because of the use of nighttime light, ODIAC did not indicate emissions over some of the areas (e.g., Africa and Eurasia) while others are indicated. In particular, EDGAR shows emissions over those areas that are largely explained by line source emissions such as transportation. Overall, ODIAC tends to put more emissions towards populated areas than suburbs. This is also explained by the lack of line sources. In EDGAR v4.2, domestic fishery emissions can be seen, but not in EDGAR FT. Even in these two EDGAR versions, we can confirm the subnational differences in the United States, Europe, and China.

Figure 8 shows time series of regional fossil fuel emissions aggregated over 11 land regions defined in the TransCom transport model intercomparison experiment (e.g., Gurney et al., 2002). The global seasonal variation and the associated uncertainty have been presented and discussed in Andres et al. (2011). Here monthly total emissions values were calculated for eleven TransCom land regions and presented with the associated uncertainty values (see Table 3). The monthly total values were calculated both excluding international bunker emissions (hence, land emissions only) and including the emissions. The uncertainty range was calculated with mass weighted uncertainty estimates of countries that fall into the TransCom regions. The uncertainty ranges shown in Fig. 8 show annual uncertainty plus the monthly profile uncertainty (12.8 %; reported by Andres et al., 2011). Monthly time series are presented for land-only emissions and land and international bunker emissions (here, largely aviation emissions). As described earlier, the emissions seasonality was adopted from Andres et al. (2011). The patterns in the emissions seasonality are often largely characterized by the large emitting countries within the regions (e.g., US for region 2 and China for region 8). Since Andres et al. (2011) used geographical closeness (also, type of economic systems) to define proxy countries, the countries in the same TransCom regions can have similar or the same seasonal patterns in their emissions.

As we can see in Fig. 4 (panel plot for aviation emissions), aviation emissions are intense over North America, Europe, and Asia. Global total aviation emissions was approximately 0.12 PgC yr-1 in 2013 and it often does not account for a large portion of the global total (1.2 % of the global total in 2013). However, considering the fact that those emissions are concentrated in particular areas such as North America, Europe, and East Asia, rather than evenly distributed in space, and are often imposed at the surface layer in transport model simulation, care must be taken to achieve an accurate atmospheric CO2 transport model simulation (Nassar et al., 2010). Aviation emissions were often around 0.5–5.1 % of the land total emissions over most regions, but also reached 12.7 % (North American Boreal).

In the production of ODIAC2016, we used several versions or editions of CDIAC estimates (e.g., global estimates, national estimates, and monthly gridded data). This could often happen in emissions data production, as some of the underlying data are not updated ro upgraded at the time of emissions data production (we often start updating emissions data after new fuel statistical data are released). We sometimes accept the inconsistency and try to use the most up-to-date information available. For example, we could use GCP's emissions estimates (e.g., Le Quéré et al., 2016) to constrain the global totals, if CDIAC global total emissions estimates are not available. The way we obtained emissions estimates for each version is often described in the NetCDF header information of the emissions data product. The use of the CARMA power plant estimates for estimating the magnitude of the point source portion of emissions is hard to eliminate, although ideally this is done using emissions estimates that are fully compatible with CDIAC estimates. We are currently examining UN statistical data (which CDIAC emissions estimates are based on) to assess the ability of explaining power plant emissions.

The emissions seasonality in ODIAC2016 is based on Andres et al. (2011) and it can be further extended to an hourly scale using the TIMES scaling parameter. We note that the emissions seasonality was based on the top 10 emitting countries' fuel statistics and Monte Carlo simulation (Andres et al., 2011). The emissions seasonality for countries other than the top 10 could be less robust. Also, because of the use of Monte Carlo, the seasonality is different over different editions of monthly emissions data. It is also important to note that the repeated use of climatological (mean) seasonality for recent years (described in Sect. 5) could be a source of uncertainty and bias. Andres et al. (2011) estimated the monthly uncertainty as 12.8 % (2σ) in addition to the annual emissions uncertainty. As we often impose fossil fuel emissions, care must be taken when applied to inversions. Ultimately, as carried out by Vogel et al. (2013), we might be able to evaluate temporal profiles from statistical data and improve them (but only to limited small locations).

As mentioned earlier, the evaluation of gridded emissions data is often very challenging and most of the emissions data studies share this difficulty. Although the emissions estimates are made on global and national scales with small uncertainties (e.g., 8 % for the global scale by Andres et al., 2014), considerable errors seem to be introduced when the emissions are disaggregated (e.g., Hogue et al., 2016; Andres et al., 2016). Andres et al. (2016), for example, estimated the uncertainty associated with CDIAC gridded emissions data on a per grid cell basis with an average of 120 % and a range of 4.0 to 190 % (2σ). Hogue et al. (2016) looked closely at CDIAC gridded emissions data over the US domain and estimated the uncertainty associated with the 1 × 1∘ emissions grids as ±150 %. Those errors seem to be unique to the disaggregation method (Andres et al., 2016). Future funding may allow us to pursue a full uncertainty analysis of the ODIAC emissions data and model, akin to the Andres et al. (2016) approach but accounting for the greater-than-one carbon distribution mechanisms utilized in the ODIAC emissions modeling framework. All of the spatially distributed gridded emissions data mentioned in this paper suffer from the same basic defect: they use proxies to spatially distribute emissions rather than actual measurements. In addition, evaluating emissions distributions based on a nighttime light proxy can be challenging as the connection between CO2 emissions and proxy is less direct compared to population (e.g., per capita emissions). A combined use of emissions proxy and geolocation data (e.g., power plant location) would also add additional difficulties to finding a comprehensive measure of the uncertainty because of different types of error and uncertainty sources (e.g., Woodard et al., 2015). As finer spatial scales are approached, the defect of the proxy approach becomes more apparent: proxies only estimate emissions fields. The ODIAC data product has been used not only for global simulations at an aggregated spatial resolution, but also at very high spatial resolution (e.g., Ganshin et al., 2012; Oda et al., 2012, 2017; Lauvaux et al., 2016). Thus, an emissions evaluation at a high resolution has become an important task. One approach we could take for evaluating high-resolution emissions fields is comparing to a local finely grained emissions data product such as Gurney et al. (2012), acknowledging the limitations of the approach discussed earlier. Another approach would be evaluating emissions data in concentration space rather than emissions space. As reported in Vogel et al. (2013) and Lauvaux et al. (2016), with radiocarbon measurements and/or good, spatially dense CO2 measurements, a high-resolution transport model simulation can provide an objective measure for emissions data evaluations (e.g., model–observation mismatch and emissions inverse estimate).

While the quality (i.e., bias and uncertainty) of the gridded emissions estimates remains unquantified for most of the emissions data mentioned in this paper, the emissions data are still used because sufficient measurements in space and time are not presently available to offer a better alternative. At the very least, we presented the uncertainty estimates over the aggregated TransCom land regions. We believe that the regional uncertainty estimates are highly useful for atmospheric CO2 inversion modelers, more than uncertainty estimates at a grid level, which still do not seem to be ready for use. Inversion studies often aggregate flux estimates over the TransCom land regions to interpret regional carbon budgets, while flux estimations in their models are performed at much higher spatial resolutions (e.g., Feng et al., 2009; Chevallier et al., 2010; Basu et al., 2013). Taking advantage of the ODIAC emissions dataset being based on the CDIAC estimates, we adopted the updated uncertainty estimates reported by Andres et al. (2016) and obtained the regional uncertainty estimates. Those estimates are new and readily usable for the inversion studies, especially when interpreting the regional estimates.