CO2 emissions from C40 cities: citywide emission inventories and comparisons with global gridded emission datasets

Abstract Under the leadership of the C40 Cities Climate Leadership Group (C40), approximately 1100 global cities have signed to reach net-zero emissions by 2050. Accurate greenhouse gas emission calculations at the city-scale have become critical. This study forms a bridge between the two emission calculation methods: (a) the city-scale accounting used by C40 cities—the Global Protocol for Community-Scale Greenhouse Gas Emission Inventories (GPC) and (b) the global-scale gridded datasets used by the research community—the Emission Database for Global Atmospheric Research (EDGAR) and Open‐Source Data Inventory for Anthropogenic CO2 (ODIAC). For the emission magnitudes of 78 C40 cities, we find good correlations between the GPC and EDGAR (R 2 = 0.80) and the GPC and ODIAC (R 2 = 0.72). Regionally, African cities show the largest variability in the three emission estimates. For the emission trends, the standard deviation of the differences is ±4.7% yr−1 for EDGAR vs. GPC and is ±3.9% yr−1 for ODIAC vs. GPC: a factor of ∼2 larger than the trends that many C40 cities pledged (net-zero by 2050 from 2010, or −2.5% yr−1). To examine the source of discrepancies in the emission datasets, we assess the impact of spatial resolutions of EDGAR (0.1°) and ODIAC (1 km) on estimating varying-sized cities’ emissions. Our analysis shows that the coarser resolution of EDGAR can artificially decrease emissions by 13% for cities smaller than 1000 km2. We find that data quality of emission factors (EFs) used in GPC inventories vary regionally: the highest quality for European and North American and the lowest for African and Latin American cities. Our study indicates that the following items should be prioritized to reduce the discrepancies between the two emission calculation methods: (a) implementing local-specific/up-to-date EFs in GPC inventories, (b) keeping the global power plant database current, and (c) incorporating satellite-derived CO2 datasets (i.e. NASA OCO-3).


Introduction
Cities currently account for about 70% of anthropogenic greenhouse gas (GHG) emissions (UN-Habitat, 2019). Urban emissions could increase as the global population percentage living in cities is expected to increase from 55% in 2018 to 68% in 2050 (U.N. 2019). Given the significance of urban GHG emissions, many cities have signed pledges to reduce GHG emissions as a contribution to mitigating the worst effects of climate change. The C40 Cities Climate Leadership Group (hereafter, C40) is a network of mayors collaborating to reduce urban GHG emissions and influence the global climate agenda. As of 2021, C40 comprises 97 cities, which account for a population of 582 million and 20% of the global Gross Domestic Product (C40 2021). Under the leadership of C40 and six international organizations, 1122 cities have signed onto the Race to Zero campaign, which aims to halve GHG emissions by 2030 and reach net zero emissions by 2050 (UNFCCC 2022).
The Global Protocol for Community-Scale Greenhouse Gas Emission Inventories (GPC) is a guideline tool for cities to estimate annual GHG emissions (C40, ICLEI, and WRI 2014). The GPC provides the principles and rules for calculating emissions consistent with IPCC Guidelines, but it does not require specific methodologies. Cities collect activity data and emission factors (EFs) based on availability and consistency with their country's national reported emissions. Since 2014, C40 cities have reported GPC-based GHG emission inventories to the Carbon Disclosure Project (CDP). The consistency and transparency of GPC inventories enable cross-city emissions comparisons, emission mitigation policy assessments, and consumption-based carbon footprint estimations (Lopes Toledo and Lèbre La Rovere 2018, Nangini et al 2019, Salvia et al 2021, Wiedmann et al 2021, Kongboon et al 2022.
The emissions research community has developed gridded CO 2 emissions datasets at various scales, including urban (Meng et (Oda et al 2018) are well-established emissions datasets with frequent updates covering up to recent years globally. The EDGAR and ODIAC's global coverage and frequent updates enable estimating CO 2 emissions from global cities, such as the C40 network. EDGAR and ODIAC have been used to estimate urban emissions for serving as the first guess for atmospheric inversion studies or directly for policy-relevant purposes (Oda et al 2017, Crippa et al 2019, 2021. Correspondingly, assessing the accuracy of global gridded emissions at sub-national scales has become critical (Hogue et al 2016, Gately and Hutyra 2017, Gurney et al 2019b, Chen et al 2020. Comparisons of the gridded emission datasets and local inventories have revealed uncertainties and biases associated with individual activity data, EFs, and downscaling methods (Gately and Hutyra 2017, Hutchins et al 2017, Wang and Cai 2017  Chen et al (2020) found large variability between the ODIAC and emission inventory estimates, ranging from −62% (ODIAC < Inventory) for Manhattan, USA, to +148% for Sao Paulo, Brazil.
We assess the level of agreement between city-level emission accounting (GPC inventories) and global gridded emission datasets (EDGAR and ODIAC) in estimating citywide CO 2 emissions and their trends. We choose the two gridded datasets-EDGAR and ODIAC-given their global coverage over a long-term period (1970-2018for EDGARv6.0 and 2000for ODIAC 2020b. We expand on the previous literature by Chen et al (2020) by (a) including more cities for a total of 78 globally, (b) updating the most recent emission inventory reported (2014-2019), (c) using GPC inventories as opposed to using emission estimates from the literature, (d) including both EDGAR and ODIAC's emission estimates, and (e) comparing emission trends. Further, we investigate the source of discrepancies between GPC inventories and gridded emission datasets. The results presented in this study can inform city staff in charge of local emission inventory and developers of gridded emission datasets about the current level of agreement for urban CO 2 emission estimates on the global scale and provide potential opportunities for improvements in the emission estimating methods.

CO 2 emissions data: GPC inventory, EDGAR, and ODIAC
The C40 GHG Dashboard compiles the GPC GHG emission inventories reported by 82 C40 cities (46 countries) to CDP. The emission accounting years for these inventories vary by city, ranging from 1990 to 2020. The C40 GHG Dashboard data are obtained from the C40 knowledge hub website (www.c40knowledgehub.org/s/article/C40-citiesgreenhouse-gas-emissions-interactive-dashboard) (C40 2022). We isolate fossil fuel CO 2 (FFCO 2 ) emissions by applying a scaling factor of 0.982 to the total fossil-fuel GHG emissions from the Dashboard data (figure S1). See text S1 for detailed methods for isolating FFCO 2 emissions from each CO 2 emissions dataset.
ODIAC downscales national fossil CO 2 emission estimates from the Carbon Dioxide Information Analysis Center (CDIAC, (Marland and Rotty 1984)) into 1 km grids globally Maksyutov 2011, Oda et al 2018). The CDIAC estimates national fossil CO 2 emissions using energy statistics published by the United Nations. For spatial disaggregation, ODIAC uses the Defense Meteorological Satellite Program (DMSP, (Elvidge et al 1999)) nightlight imagery as a spatial proxy for non-point sources and the carbon monitoring action (CARMA, (Wheeler and Ummel 2008)) data to locate power plant emissions. ODIAC version 2020b (ODIAC2020b), available for 2000-2019, is obtained from the Center for Global Environmental Research, National Institute for Environmental Studies (NIES) website (https://db.cger.nies. go.jp/dataset/ODIAC/).

Aligning CO 2 emissions from GPC inventory, EDGAR, and ODIAC for comparison
We align the three datasets-GPC inventory, EDGAR, and ODIAC-by sector, time, and spatial area to enable comparisons between them. For the sectoral alignment, we include scope 1 FFCO 2 emissions from the five sectors-residential, commercial, industrial, energy, and transportation-from the GPC inventory, EDGAR, and ODIAC, respectively. Four C40 cities-Bangkok, Dakar, Hong Kong, and Phoenix-are excluded due to missing emissions data for one or more subsectors in the GPC inventories. See text S1 for detailed methods for isolating FFCO 2 emissions from each CO 2 emissions dataset.
For the temporal alignment of annual emission estimates, we choose the most recent accounting year with estimates from all three emissions datasets available for each city. Of the 82 C40 cities with selfreported inventories, 74 cities have inventories for years between 2014 and 2018, which align with the years available from EDGAR version 6.0 and ODIAC version 2020b. Four cities-Rotterdam, Bengaluru, Mumbai, and Washington, DC-have their first available GPC inventory for 2019, and we use EDGAR emissions for 2018 for the comparison. For the temporal alignment of emission trend estimates, we use the earliest and latest accounting years available before 2019 for each city. Of the 82 C40 cities, 39 cities have submitted at least two complete inventories between 2000 and 2018.
The types of geographical boundaries used to construct the GPC inventories are the administrative boundary of a local municipality (N = 55), broader metropolitan area (N = 14), province/district/state/county (N = 6), and others (N = 3, i.e. Comprehensive Land Use Plan). For the spatial alignments, we aggregate CO 2 emissions from EDGAR and ODIAC by sampling grids within the same geographical boundary used to construct the GPC inventory of each city (figure 1). For EDGAR grid cells (0.1 • ) overlapping the city boundary line, we scale emissions by the ratio of the grid area within the city boundary to the grid area of the 0.1 • grid cell. For ODIAC grid cells (1 km), we aggregate emissions of the grid cells whose center coordinates are within the boundary and do not attempt to scale emissions.

GPC GHG emissions inventory
The GPC guideline requires cities to report total GHG emissions attributable to activities within the city's geographic boundary, termed the BASIC level. The BASIC level covers scope 1 (i.e. emissions from sources within the city) and scope 2 (i.e. emissions occurring due to the use of grid-supplied energy) emissions from stationary energy and transportation sectors, and scope 1 and scope 3 (i.e. emissions occurring outside the city due to activities within the city) emissions from the waste sector (Fong et al 2014). See figure 2 of Fong et al (2014) for the complete list of subsectors covered by BASIC level. It is important to note that the BASIC level covers emissions from gridsupplied energy consumption (scope 2) while excluding the emissions from energy generation supplied to the grid (scope 1).
On average, the percentage of FFCO 2 in total GHG emissions ('BASIC') is 57% for Africa, 77% for Central East Asia, 51% for South and West Asia, 84% for Europe, 76% for Latin America, 70% for North America, and 49% for South and West Asia cities, with the global mean of 69%. For the nine cities-Abidjan, Amsterdam, Barcelona, Buenos Aires, Copenhagen, Rio de Janeiro, Rotterdam, Venice, and Yokohamathe BASIC emissions are smaller than the FFCO 2 emissions as these cities are net exporters of electricity toward surrounding areas.
Socioeconomic factors such as population, area, and population density show scaling relationships with citywide CO 2 emissions (Ribeiro et al 2019). According to GPC inventories of 78 global cities, citywide FFCO 2 emissions show large variability, ranging from 0.4 MtCO 2 for the city of Sydney, Australia, to 59 MtCO 2 for Chengdu, China. Such variability is mainly due to the difference in population and land area within each city boundary. Population (and land area) is 14.7 million (14 312 km 2 ) for Chengdu and 0.25 million (26 km 2 ) for the city of Sydney (Note: the Sydney metropolis' covers 5.2 million people in 12 368 km 2 ). Figure 3 shows that the per-area FFCO 2 emission for Sydney is 13.4 MtCO 2 km −2 (citywide 0.3 MtCO 2 ), greater than Tokyo's per-area emissions of 10.8 MtCO 2 km −2 (citywide: 23.6 MtCO 2 ), highlighting the importance of using the matching geographical boundary for the emission comparison study. Cities with power plants within their geographical boundary tend to have larger scope 1 FFCO 2 emissions (i.e. Rotterdam). In this study, we do not consider emissions occurring due to the import/export of electricity. Figure 4 shows FFCO 2 emissions in GPC inventories for the 78 C40 cities compared to emission estimates from the two global gridded emission datasets: EDGAR and ODIAC. Accounting years of the GPC inventories chosen for comparison range from 2014 to 2019 (median year: 2018). EDGAR's estimates of FFCO 2 emissions show a good correlation with GPC inventory estimates, with an R 2 value of 0.80 and the slope of ordinary   least squares (OLS) fit as 1.09. Regionally, cities in North America show the highest R 2 value of 0.95, followed by cities in Europe (R 2 = 0.94). Across 78 cities, the mean and standard deviation of the relative difference between EDGAR and GPC inventory estimates are −25 ± 53% (Relative difference = (EDGAR − GPC)/mean(EDGAR, GPC) × 100). African cities show the largest relative difference (−70 ± 84%). For Latin American cities, EDGAR emissions tend to be consistently smaller than GPC inventory emissions: 11 of 12 Latin American cities show EDGAR emissions lower than GPC inventory emissions. Only Salvador, Brazil, shows EDGAR emissions greater than GPC emissions by 13%.

GPC inventory, EDGAR, and ODIAC
ODIAC shows a good correlation with GPC inventory estimates for the 78 C40 cities, with an R 2 of 0.72 and the slope of OLS fit as 1.08. Regionally, European cities show the highest R 2 of 0.91. Across all 78 cities, the mean and standard deviation of the relative difference between ODIAC and GPC inventory estimates are 12 ± 62%. For six cities in South and West Asia, ODIAC emission estimates are consistently larger than the GPC inventory emission estimates (RD = 82 ± 36%). African cities show a relative difference of −11 ± 111%, the largest variability across all regions. Figure 5 shows the percentage change rates of FFCO 2 emissions between the earliest and latest accounting years for which GPC emissions inventories are available. For the 46 C40 cities with more than one accounting year, the temporal gaps between the two accounting years range from 1 to 18 years (median: 5 years). Of the 46 C40 cities, 25 cities show declines in FFCO 2 emissions over time. New Orleans, USA, shows the fastest decline rate of −13.1% yr −1 between 2005 and 2018. The rapid decline of CO 2 emissions in New Orleans results from the 2016 closure of the Michoud power plant (EPA 2022a). The median emissions decline rate is −2.0% yr −1 for the 25 cities. The other 21 cities have shown increased emissions since their earliest accounting years, ranging from +0.1% yr −1 to +7.3% yr −1 (median: +1.5% yr −1 ). The transportation sector emissions have declined for 26 out of 46 cities since the first accounting years. For the residential/commercial/industrial sector, 26 out Figure 5. Percentage changes in FFCO2 emissions between the earliest and the latest accounting years of the GPC inventories for 46 C40 cities (Change = (CO2,Latest − CO 2,Earliest )/CO 2,Earliest /∆years × 100, unit: % yr −1 ). Two accounting years are shown next to city names. The colored bars indicate the sectoral emission changes, and the diamond symbol indicates the net emission change. Thirty-two cities in figure 3 are excluded as they only have one year of available data or missing data in the time series. Figure S4 shows the total percentage changes in emissions between the two accounting years.

Trends in scope 1 FFCO 2 emissions for C40 cities 3.2.1. GPC emissions inventory
of 46 cities show declines in emissions. Energy sector emissions have declined in 19 cities and increased in another 19, while the remaining 8 cities have no incity energy sector CO 2 emissions. Figure 6 compares the trends in FFCO 2 emissions for 39 C40 cities, estimated from the GPC inventory, EDGAR, and ODIAC. For the comparison analysis, we calculate the trends in emissions using the emissions for two accounting years as followings: (Trend = (CO 2,Latest year before 2019 − CO 2,Earliest year )/ CO 2,Earliest /∆years × 100, unit: % yr −1 ). The comparison analysis excludes seven cities listed in figure 5-Dubai, Houston, Istanbul, Lisbon, Los Angeles, Moscow, and Quito-as they have less than two GPC inventories before 2019. Out of 39 cities in the comparison, the number of cities that show the same directional changes in emissions is 23 for GPC vs. EDGAR and 24 for GPC vs. ODIAC (Quantiles I and III in figures 6(a) and (b)). The mean and standard deviation of differences in trend estimates (Trend Gridded Dataset − Trend GPC ) is 0.5 ± 4.7% yr −1 for EDGAR vs. GPC and −0.3 ± 3.9% yr −1 for ODIAC vs. GPC.

Comparison between GPC inventory and gridded datasets
Two African cities-Johannesburg and Tshwane, South Africa-show the largest discrepancies in the trend estimates, having opposite directional changes. For Latin American cities, the trend estimates showed overall good agreement with a mean difference of 0.3% yr −1 for EDGAR vs. GPC and −1.0% yr −1 for ODIAC vs. GPC. However, the trend estimates for Latin American cities showed significant variability: the standard deviation of the difference is ±6.0% yr −1 for EDGAR vs. GPC and ±4.5% yr −1 for ODIAC vs. GPC, both values being the second largest following African cities (figures 6(b) and (d)). European cities showed the best agreement in the trend estimates, having a trend difference of 1.7 ± 2.0% yr −1 for EDGAR vs. GPC and 0.4 ± 1.2% yr −1 for ODIAC vs. GPC, followed by North American cities and E. SE Asian & Oceanic cities. The city of New Orleans, USA, shows a noticeably larger difference than other North American cities (the GPC inventory: −13.1% yr −1 , EDGAR: −1.9% yr −1 , ODIAC: −1.6% yr −1 ). Figure 6. The upper panels show FFCO2 emissions trends estimated from the GPC inventory, EDGAR (a), and ODIAC (c). Different colors and shapes of symbols indicate C40 regions that cities belong to, as labeled in (b). The lower panels show differences in the trend estimates between EDGAR and GPC (b) and ODIAC and GPC (d) (∆Trend = Trend Gridded Dataset − TrendGPC). The box and whisker plot indicates the minimum, lower quartile, median, upper quartile, and maximum differences. The black circle symbol indicates the mean differences. See figure S5 for the direct comparison between EDGAR and ODIAC.

Spatial resolutions of ODIAC (1 km) and EDGAR (0.1 • )
We investigate the impact of spatial resolutions of a gridded product-1 km vs. 0.1 • -on estimating citywide CO 2 emissions as follows. First, ODIAC's 1 km grid cells are spatially aggregated into 0.1 • grid cells. Then, citywide FFCO 2 emissions for 78 C40 cities are estimated from 0.1 • gridded ODIAC using the same area scaling approach used for EDGAR (see section 2.2). Figure 7(a) shows the percentage differences of citywide CO 2 emissions estimated from 0.1 • gridded and the original 1 km ODIAC. We find that gridding of 1 km pixels into 0.1 • pixels decreased emission estimates for 67 out of 78 C40 cities, while 11 cities showed an increase in emission estimates. The magnitude of decrease or increase in emission estimates tends to be smaller for larger area cities: the mean percentage differences in emission estimates between the 0.1 • and 1 km ODIAC is −13% for the cities smaller than 1000 km 2 (N = 47), −2.6% for cities between 1000 km 2 and 3000 km 2 (N = 21), and −0.7% for cities larger than 3000 km 2 (N = 10). Boston, USA (figures 7(b) and (c)) and Tel Aviv, Israel (figures 7(d) and (e)) show the largest (+183%) and smallest (−63%) percentage differences, respectively.
We discuss the contributing factors to the varying impacts of spatial resolutions and their policy implications in section 4.2.

Spatial disaggregation errors in EDGAR and ODIAC
As a measure of variability among three citywide FFCO 2 emissions estimates (i.e. EDGAR, ODIAC, GPC inventory), the coefficient of variation (CV) is calculated for each of the 78 C40 cities (CV = σ/µ × 100, where µ and σ and the mean and standard deviation of the three emission estimates). Figure 8 shows that values of the CV tend to be smaller for cities with larger geographical areas (i.e. the larger number of EDGAR/ODIAC grid cells aggregated). A simple regression model CV = e/ √ City Area is applied to data points in each region, assuming that the grid cell-level uncertainty reduces by the square root of N when aggregated to citywide sum emissions. Cities in North America showed the best goodnessof-fit statistic with the chi-squared (χ 2 ) value of 222, followed by cities in Europe (χ 2 = 325). Cities in Africa showed the lowest goodness-of-fit statistic (χ 2 = 3081). The estimated coefficient e, the CV at a single cell level (1 km 2 ), was lowest for European cities (383%) and highest for African cities (1373%).  (e)). To estimate the citywide emissions from 0.1 • cells, we use the spatial area scaling approach as described in section 2.2. The top panel (a) shows the percentage differences in emission estimates from 0.1 • and 1 km ODIAC for 78 C40 cities. The panel indicates the mean percentage differences for three area ranges-smaller than 1000 km 2 , between 1000 km 2 and 3000 km 2 , and greater than 3000 km 2 . The number of cities with negative and positive differences are shown (N negative , N positive ). The lower panels show the maps of two cities with the highest (Boston, USA (b) and (c)) and the lowest percentage differences (Tel Aviv, Israel (d) and (e)). See section 4.2 for the discussion on the impact of spatial resolutions of a gridded emission product on estimating citywide emission estimates.

EFs in GPC inventories
The GPC inventories submitted through the City Inventory Reporting and Information System include a spreadsheet documenting EFs. Figure 9 shows the overview of the EFs documented for 43 cities across six regions ( figure 9)  and 87% of those EFs are international scales. Subnational scale EFs account for 1% of the EFs for African cities' Inventory. East. Southeast Asian & Oceanian cities' EF profiles are close to those of North American and European cities, having an average reference year of 2014 and an international EF portion of 33%. Latin American cities' EF profile has an average reference year of 2010 and an international EF portion of 51%. Also, the GPC guideline recommends cities provide data quality assessment for EFs used for emission calculations, using a High-Medium-Low rating. European and North American cities reported the largest percentage of 'High' quality EFs (62%), while the African cities reported the largest percentage of 'Low' quality EFs (87%).

Discrepancies in citywide CO 2 emission estimates between the two emission calculation methods: citywide emission inventory (GPC) and global gridded emissions (EDGAR and ODIAC)
For 78 C40 cities, the relative difference in annual citywide FFCO 2 emission is −25 ± 53% between EDGAR and GPC inventory and 12 ± 62% between ODIAC and GPC inventory. Among the seven global regions, cities in Africa showed the largest variability in the relative difference (1σ = ±84% for EDGAR vs. GPC, 1σ = ±111% for ODIAC vs. GPC). This result is consistent with Chen et al (2020), who reported a large variability in the emission discrepancies between local inventories and ODIAC for 14 global cities, ranging from −62% (ODIAC < Inventory) for Manhattan, USA, to +148% for Cape Town, South Africa. Chen et al (2020) reported that the ODIAC CO 2 emission estimates are higher than the local inventories for Cape Town (+148%), Sao Paulo (+43%), and Beijing (+40%). Our analysis shows that ODIAC's FFCO 2 emissions are higher than the GPC inventories for Cape Town (+91%) and Sao Paulo (+31%). Meanwhile, EDGAR's estimates are lower than the GPC inventories for Cape Town (−25%) and Sao Paulo (−60%).
For 39 C40 cities, the difference of trend estimates is 0.5 ± 4.7% yr −1 for EDGAR vs. GPC and −0.3 ± 3.9% yr −1 for ODIAC vs. GPC. The 1 sigma values of ±4.7% yr −1 (EDGAR vs. GPC) and ±3.9% yr −1 (ODIAC vs. GPC) are greater than the rates of emission reduction that many C40 cities have pledged: i.e. net zero by 2050 from the baseline year of 2010 (−2.5% yr −1 ). This finding suggests that emission estimates from global gridded emission datasets must be significantly improved for city-level policy applications, such as emission mitigation progress tracking. The following sections discuss three primary sources of uncertainty-spatial disaggregation, point emission sources, and EFs.

Sources of emission discrepancies: spatial resolutions of EDGAR (0.1 • )
ODIAC has been frequently used in urban CO 2 emission studies given its 1 km spatial resolution (Janardanan et al 2016, Lauvaux et al 2016, Han et al 2020c. Meanwhile, EDGAR's 0.1 • resolution has been considered more suitable for regional/country scale analysis (Gately et al 2013, Han et al 2020a, Crippa et al 2022. Figure 7 shows that gridding 1 km pixel of ODIAC into 0.1 • grid cells results in decreases in citywide emissions for 67 out of 78 C40 cities, while 11 cities show increase in emissions. When 1 km grid cells are aggregated into 0.1 • grid cells, emission gradients smooth out throughout the 0.1 • pixel. Such a smoothing effect can lead to either overestimation or underestimation in the citywide emissions depending on the spatial gradient of emissions surrounding the city boundary. For most cities, emissions tend to increase as the pixel gets closer to the urban center. In such cases, gridding to a coarser resolution results in underestimations in citywide emissions, as shown for Tel Aviv, Israel (figures 7(d) and (e)). For Tel Aviv, Israel, the west of the city is mostly the Mediterranean sea, with only marine emissions, resulting in the largest underestimation among 78 cities (−63%). When the 1 km pixels outside the city have larger emissions than those within the city, gridding leads to overestimating citywide emissions. For Boston, USA, several power plants are located outside of the city boundary (i.e. Kendall Power Plant and MIT Central Utility Plant, both located in Cambridge, MA), contributing to the most significant overestimation in citywide emissions among 78 cities (+183%) (figures 7(b) and (c)). Figure 7 also shows that the emission smoothing effect tends to decrease with larger area cities, as the fraction of pixels that overlap with city boundaries decreases for larger cities. Our analysis suggests that the coarse spatial resolution of EDGAR has less of an impact on cities larger than 3000 km 2 (Mean RD = −0.7%). For the cities between 1000 km 2 and 3000 km 2 , relatively small underestimations are expected (−2.6%) when using EDGAR. For the cities smaller than 1000 km 2 , we recommend careful investigation of emission spatial gradients before using EDGAR, as significant bias can be introduced (−63% (Tel Aviv) ∼ +183% (Boston)).

Sources of emission discrepancies: spatial disaggregation error in the EDGAR and ODIAC
Andres et al (2016) reported an uncertainty of 120% (2σ) for CDIAC's FFCO 2 emitting grid cell over the annual timescale, and Oda et al (2019) reported that the emission uncertainty for a 1 km grid cell decreases by 80% when aggregated at 100 km. Our study shows that the CV for three FFCO 2 emission estimates (i.e. EDGAR, ODIAC, and GPC inventory) tend to decrease with the increasing geographical area of cities. When a regression model of CV = e/ √ City Area is applied, North American cities showed the lowest standard error (s = 14), and African cities showed the highest standard error (s = 55). The regional variation in the standard error is induced by regional biases in (a) spatial proxies used for disaggregating national emissions (i.e. population settlements for EDGAR and nighttime light intensity for ODIAC), (b) point emission source error, (c) EFs used in GPC inventory, and (d) uncertainty in national FFCO 2 emissions: Andres et al (2012) reported that uncertainty in national FFCO 2 emissions ranges from a few percent (i.e. 3%-5% for the U.S.) to ∼50% for countries with poorly maintained statistical infrastructures of energy data.
A relatively higher standard error for a given region implies that the spatial disaggregation error has a relatively small impact on the region, and the other sources of uncertainty-national emissions, point sources, and EFs-have large contributions to the variability in emission estimates. We discuss the impact of discrepancies in EFs and point sources in the following sections.

Sources of emission discrepancies: point emission sources
Two cities-New Orleans, USA, and Amman, Jordan-show the largest discrepancies in the emission trend estimates. The case study of Amman, Jordan, shows inconsistent accounting of power plants between GPC inventory, EDGAR, and ODIAC lead to large variability in the trend estimates (EDGAR: 12.3% yr −1 , ODIAC: 2.3% yr −1 , GPC inventory: 2.6% yr −1 ). The case study of New Orleans, USA, reveals that EDGAR and ODIAC have some point sources that are not keeping up with the up-to-date operation status, inducing significant error in the trend estimates (GPC inventory: −13.1% yr −1 , EDGAR: −1.9% yr −1 , ODIAC: −1.6% yr −1 ).
For the U.S. power sector, the U.S. EPA Clean Air Markets Program Data (CAMPD) provides operation status, emissions, and geolocations of individual power-generating facilities (EPA 2022b). The CAMPD dataset enables a thorough evaluation of uncertainty in the U.S. region's point source emission in gridded emission datasets and local inventories (Ahn et al 2020, Liu et al 2020). Currently, such a detailed power plant database is not available on a global scale. The Carbon Monitoring for Action (CARMA) is the only global database that gathers CO 2 emissions for ∼50 000 individual power plants but has no longer been active since 2012 (Wheeler and Ummel 2008). Continuous efforts to develop a global database of power plant emissions would be essential for accurately tracking urban GHG emission mitigations for global cities. The lack of local-specific and up-to-date EFs data has been considered a major challenge in developing the GHG emissions inventory (Satterthwaite 2007, Pitt and Randolph 2009, Ibarra-Espinosa and Ynoue 2017, Bai et al 2018, Nagendra et al 2018. Baltar de Souza Leão et al (2020) analyzed GHG inventories for 24 Brazilian cities and found that 17 inventories possess incomplete activity data for major sectors. Arioli et al (2020) reviewed 73 journal articles that report city-level GHG inventory and found that most cities lack local-specific transportation data. Li et al (2017) adopted the GPC to estimate GHG emissions in Beijing, China, and reported that the missing data was the major challenge in developing the GPC inventory. Driscoll et al (2015) conducted the gap analysis of GHG emission inventories for Trondheim, Norway: they reported that the most pressing need is local-specific data, such as local transportation activity data. Our findings suggest that promoting and supporting global south cities to develop and adopt local-specific and up-to-date EFs should be prioritized to reduce discrepancies in citywide emission estimates.

Remote sensing of atmospheric CO 2
Emerging satellite observations of atmospheric CO 2 , such as NASA Orbiting Carbon Observatory 3 (OCO-3), can provide independent citywide emission estimates (Nassar et  . The major sources of uncertainties in the satellite-based approach are atmospheric transport (wind) and isolating urban enhancement signals from the background level, which is more challenging for the growing season due to active biospheric uptake (Wu et al 2022). With further improvements in methodologies, satellite-based estimates could bridge gaps between city inventories and gridded datasets. We suggest the following ten cities-Rotterdam, Yokohama, Buenos Aires, Kuala Lumpur, New York City, Amsterdam, Seoul, Paris, Boston, and Addis Ababa-as testbed sites for developing satellite-based emission estimate methods given (a) high per-area CO 2 emissions (>30 ktCO 2 km −2 , figure 3) and (b) periodic emissions reporting from the GPC method. Improvements in the three emission tracking approacheslocal emission inventories, gridded emission datasets, and atmospheric observations-would complement each method and accelerate progress toward accurate, consistent, and transparent emissions tracking.

Implications for stakeholders of urban climate action planning
The GPC provides detailed and consistent guidelines for cities to construct GHG emission inventories. According to GPC inventories of 78 global C40 cities, on average, scope 1 FFCO 2 emissions account for 69% of total GHG emissions.
Our study shows that ODIAC could be used to provide first-order estimates for citywide annual FFCO 2 emissions, especially for cities lacking reliable data-collecting systems. For 78 global cities' annual FFCO 2 emissions, ODIAC agrees to GPC inventory within ±62% (1σ). The level of agreement between ODIAC and GPC inventory varies by region: African cities show the largest variability between ODIAC and GPC. ODIAC (1 km resolution) and EDGAR (0.1 • resolution) tend to show better agreement toward GPC inventory for larger cities. We recommend using EDGAR for cities larger than 1000 km 2 . For emission trend estimates, our study shows that neither the current version of ODIAC2020b nor EDGARv60 are suitable for tracking citywide emission mitigation progress at a policy-relevant scale.
The data quality of EFs used in GPC inventories varies by region, with the highest quality for European and North American and the lowest for African and Latin American cities. Updating local-specific and up-to-date EFs, which would require investments in city-scale data-collecting systems, will narrow the discrepancies in emission estimates. We also acknowledge that emerging satellite remote sensing data (i.e. NASA OCO-3) could be utilized to bridge gaps between various emission estimation methods.

Data availability statement
No new data were created or analysed in this study.

Acknowledgments
The C40 GHG Dashboard data (versions 03.2021, 02.2022, and 05.2022) are obtained from the C40 knowledge hub website (www.c40knowledgehub. org/s/article/C40-cities-greenhouse-gas-emissionsinteractive-dashboard). The emission inventories in the City Inventory Reporting and Information System (CIRIS) format spreadsheet files are obtained via communication with C40 and CDP (www.cdp. net/en/data). We obtained geographical boundary shapefiles via communication with C40, and the shapefiles are available upon request. Annual sector-specific EDGARv6.0 for fossil CO 2 sources are obtained from the JRC EDGAR website (https:// edgar.jrc.ec.europa.eu/dataset_ghg60#sources). The version of the ODIAC data product (ODIAC2020b) is available from the Global Environmental Database website hosted by the Center for Global Environmental Research (CGER), National Institute for Environmental Studies (NIES), Japan (http://db.cger. nies.go.jp/dataset/ODIAC/). The authors appreciate the helpful comments provided by Tomohiro Oda at the Universities Space Research Association (USRA). This work was supported by a Grant from the Wellcome Trust (Award No. 216075/Z/19/Z).