Machine learning estimates for G20 subnational urban GHG emissions from 2000–2020

Abstract

Reliable, comparable greenhouse gas (GHG) emissions data at the subnational level remain scarce, despite growing expectations for cities and regions to lead on climate action. Inconsistent reporting, methodological variation, and limited coverage of self-reported inventories hinder efforts to track progress and guide mitigation opportunities. To address these challenges, we develop a machine learning (ML) framework to estimate annual Scope 1 and 2 CO₂-equivalent emissions for subnational jurisdictions in G20 countries from 2000 to 2020. Our approach integrates publicly available geospatial, socioeconomic, and environmental data with self-reported inventories where available, and aligns predictions with subnational administrative boundaries. Compared to traditional downscaling or proxy-based approaches, our model improves spatial relevance and predictive performance while capturing locally specific emission drivers. This globally consistent, administratively-aligned dataset can serve as a baseline for assessing climate progress, especially in data-poor or inconsistent reporting contexts, and supports more targeted, data-informed policy decisions for urban and regional decarbonization.

Background & Summary

More than 14,000 city and regional governments have pledged climate actions under various voluntary initiatives¹. These local actors often submit climate action plans, which include emission reduction pledges and related commitments to more than 30 transnational subnational government-focused initiatives, as tracked by the UN Framework Convention on Climate Change (UNFCCC) Non-State Actor Zone for Climate Action (NAZCA). Nearly 300 large cities with over 500,000 in population² and 570 cities and regions in the G-20 have even pledged net-zero targets³. Yet despite their growing visibility in global climate governance, fewer than 10 percent have reported greenhouse gas emissions (GHG) inventories to track their progress over time³. When emissions are reported, they are frequently incomparable due to methodological differences, self-selection of emission sources included, treatment of consumption-based or “out-of-boundary” emissions, among others^4–6. This gap, between signalled intent and accountability, underscores a critical challenge: while cities and regions are increasingly seen as key players in advancing both local and global climate goals, the absence of consistent, high-quality emissions data limits our ability to evaluate their impact and guide future investments in needed decarbonization efforts.

To address challenges related to data gaps, accounting inconsistencies, and comparability across scales, researchers have adopted a range of methods. These include statistical downscaling of national emissions data to the subnational level^7–12, as well as activity-based spatial allocation approaches that use sector- and activity-specific proxies, such as the locations of power plants, industrial facilities, or patterns of residential and agricultural emissions. These proxy-driven methods often rely on geospatial datasets such as the Emissions Database for Global Atmospheric Research (EDGAR) or the Open-source Data Inventory for Anthropogenic CO₂ (ODIAC) datasets, which utilize satellite remote sensing and other geospatial data as proxies, to allocate emissions within countries at finer spatial resolutions. However, these existing approaches suffer from several shortcomings. Statistically-downscaled city-level datasets fail to adequately capture the impact of subnational climate efforts since these methodologies assume that cities follow national-level trajectories of coarse GHG emission proxies like GDP and population, drowning out any potential signal of mitigation progress. Large gridded, geospatial datasets have the advantage of providing consistent assessment methodology, complete spatial and temporal resolution, and comparability. However, they are not aligned with city administrative and decision-making boundaries, limiting their utility and application to understanding subnational climate action progress. These approaches tend to have overall lower prediction accuracy and modeling limitations, particularly for complex urban scenarios and broader generalizability is also questionable¹³.

The rapid advancement of artificial intelligence (AI), particularly machine learning (ML) techniques, has opened up new opportunities for addressing persistent data and analytical challenges in climate science. One of the most promising applications of ML is its ability to help fill critical data gaps, since ML methods excel at integrating large, heterogeneous datasets and identifying complex, non-linear relationships between key drivers of emissions, such as energy consumption, industrial activity, land use, and socio-economic indicators^14–16. These approaches have the potential to significantly improve over traditional subnational GHG estimation approaches by leveraging a more flexible approach where algorithms “learn” complex patterns in the data without relying on predefined assumptions¹³. Particularly for large, heterogeneous datasets, ML approaches could enable a more comprehensive and dynamic understanding of urban environments and emissions¹³. For example, Hsu et al.¹⁷ utilized a gradient-boosting “tree model” ML framework (XGBoost)¹⁸ and underlying satellite remote-sensing derived geospatial predictors to estimate likely annual emissions and mitigation performance of all local administrative areas in Europe between 2001 and 2018. Neural networks and deep learning frameworks are also being introduced for prediction¹⁹.

Generalizing a global ML-based approach to cities across the world has its challenges. Globally, cities are diverse across a range of variables. They differ greatly in their physical morphology, such as urban form, land use patterns, and infrastructure, which directly influence the spatial distribution and magnitude of emissions²⁰. Variations in economic development levels further complicate modeling, as cities in high-income regions often have distinct energy systems, transportation networks, and building codes compared to those in low- and middle-income countries. Moreover, cities span a wide range of population sizes, densities, and geographic extents, each of which affects emission sources and intensities in unique ways. The mix of emission sources, ranging from heavy industry and power generation to informal settlements and biomass use, also varies significantly, requiring models to account for locally specific drivers and activities. As a result, ML models trained on data from one set of cities may not generalize well to others unless they are carefully calibrated, incorporate locally relevant variables, and leverage multi-source data to capture these differences.

To address these limitations, the dataset we introduce in this paper utilizes a machine learning framework to predict urban greenhouse gas emissions for subnational governments in G20 countries from 2000–2020. Our approach leverages a wide array of publicly available geospatial, socioeconomic, and environmental data, integrating these with self-reported emissions inventories from cities wherever available. A key distinguishing feature of our methodology is the use of the Global Administrative Areas Database (GADM) v.4.1²¹, which provides a globally consistent hierarchy of administrative boundaries. By aligning emission estimates with officially recognized city and municipal governance units, this approach enhances the relevance and usability of the data for local decision-makers. Unlike traditional statistical downscaling methods or other proxy-based geospatial modeling approaches, which often assume that urban emissions follow national-level trends and thus fail to capture local mitigation actions or unique city characteristics, our ML model is designed to extract additional, non-linear insights from diverse data sources, improving both the completeness and accuracy of city-level emissions estimates. While this dataset is not intended to replace cities’ own greenhouse gas inventories, we aim for it to serve as a comparable, globally consistent baseline that can support cities in tracking progress, identifying gaps, and informing climate action planning, particularly in contexts where self-reported data are unavailable, inconsistent, or outdated and when other gridded products underestimate emissions in smaller subnational entities.

Method

Workflow

The dataset workflow of this study is illustrated in Fig. 1. It starts with the cleaning and standardization of self-reported emissions data, followed by the spatial alignment of subnational actors. Then we extract environmental predictor variables from 2000 to 2020 for all subnational administrative units in the scope of this study. We conduct data validation and quality checks before model training and implementation. Finally, we compare predicted emissions with external datasets. Each of these steps is described in the following sections.

Fig. 1.

Workflow diagram of data processing, model training, and validation.

Study area

Our study focuses on the G20 (the European Union, Argentina, Australia, Brazil, Canada, China, France, Germany, India, Indonesia, Italy, Japan, South Korea, Mexico, Russia, Saudi Arabia, South Africa, Turkey, the United Kingdom, and the United States), which collectively comprise nearly 80% of global greenhouse gas emissions²². We include subnational entities across all administrative levels with available data that meet our quality controls, excluding only the highest administrative units (i.e., Regions) unless we determine more than 50% of their land area was urban. This urban-rural classification was based on a comprehensive assessment of human settlement characteristics, including population density, built-up area, socioeconomic development, and long-term urbanization trends¹². We applied this filter because our study focuses on urban GHG emissions that inform city climate actions, and removing rural-dominant regions helps the model to better represent our area of study, reducing statistical noise and improving performance during training, testing, and prediction. The final dataset includes 5,972 cities and 116 regions in the G20 and a total of 9,664 self-reported emissions entries for several years, with 2,273 entities reporting emissions more than once. Figure 2 shows the geographical distribution and type of subnational entities included in our final training dataset.

Fig. 2.

Map of cities (n = 5,972) and regions (n = 116) with self-reported emissions used to train and test our machine learning model.

While not representative of all subnational jurisdictions globally, the entities included in our dataset offer a revealing snapshot of those that have voluntarily self-reported their greenhouse gas emissions, which tend to be located in more resource-rich areas such as Europe and North America. The majority of reporting entities in our training dataset are from Europe, followed by North America and East Asia and the Pacific (see Table 1 for summary statistics). Among these, reporting entities in Europe tend to have the smallest administrative territories and the lowest populations, representing the lowest average self-reported emissions. Notably, this region has the second-highest average GDP per capita. Reporting entities in North America have the highest average GDP per capita and the third-highest average emissions. Those from East Asia and the Pacific report the second highest emission values amongst all regions, but have the highest average population and administrative area. In Latin America and the Caribbean, our dataset includes reporting entities from Argentina, Brazil, and Mexico. They have administrative territories similar in size to those in North America, but, on average, have twice the population and lower emissions. In Eastern Europe and Central Asia, reporting entities are limited to just two countries: 46 from Croatia and 16 from Turkey. Croatian entities resemble their western European counterparts, with smaller populations, lower self-reported emissions, and smaller administrative areas. Turkish entities, by contrast, are far larger in population, averaging 100 times more than those from Croatia, and report more than 5 million tons of CO₂ annually. In Southest Asia, our dataset only includes reporting cities and regions from Indonesia, which have relatively low GDP per capita but substantial average emissions, suggesting high-emitting local activities amid slower economic development. Finally, due to limited self-reported data, Sub-Saharan Africa is only represented by a small number of entities, all of which are major cities: Johannesburg, Ekurhuleni, Tshwane, KwaDukuza, Cape Town, and eThekwini.

Table 1.

Summary of reporting entities in training data.

Region	Entities (n)	Self-reported Emission Data Points	Average Population	Average GDP Per capita	Average Emissions (tons CO2e)	Administrative Area (km2)
Europe	5,314	7,975	47,466 (±450,297)	39,905.09 (±12,209.67)	338,611.2 (±4,505,677.97)	217.01 (±2,367.97)
North America	324	824	991,624 (±3,279,134)	61,826.16 (±22,305.76)	12,454,455.87 (±37,069,396.64)	12,626.87 (±53,209.97)
East Asia and the Pacific	284	575	5,368,728 (±16,193,127)	33,549.78 (±15,506.36)	39,778,130.81 (±121,421,800.8)	21,642.65 (±110,260.03)
Latin America and Caribbean	86	151	2,078,074 (±563,8822)	18,602.95 (±10,217.11)	9,825,536.59 (±23,618,735.85)	12,139 (±38,168.77)
Eastern Europe and Central Asia	62	97	553,345 (±1,946,072)	20,419.64 (±8,514.42)	2,431,632.42 (±7,398,410.31)	1,984.12 (±4,576.56)
South Asia	12	32	2,909,023 (±5,339,328)	10,495.21 (±4,295.22)	9,089,828.51 (±19,602,242.65)	214.57 (±378.64)
Sub-Saharan Africa	6	10	3,654,199 (±1,828,698)	15,741.19 (±4,358.23)	48,841,156.5 (±28,994,336.29)	2,919.85 (±2,136.73)

Self-reported CO₂ Emissions

We use a broad definition of subnational actors to include any subnational entity with jurisdiction over an administrative area, regardless of their denomination, or level of autonomy. To standardize geographic units across countries, we rely on a dataset built from the Global Administrative Areas Database (GADM) version 4.1²¹ enhanced with more recent official boundaries for 17 countries with outdated boundary data, which organizes subnational areas according to each country’s internal administrative hierarchy. In this system, administrative level 1 (ADM_1) represents the highest subnational division, and ADM_5 represents the lowest available level for any given country. Because national administrative structures vary, a unit such as a “Municipality” may appear at ADM_2 in one country and ADM_3 in another, depending on how that country’s governance system is organized.

Examples of subnational entities in our dataset span a range of types, including provinces, states, counties, regions, districts, municipalities, and villages, with names varying by country and language. At the ADM_1 level, the top-tier subnational unit, common designations include Provinces (e.g., Argentina, Canada, China, Turkey, South Africa), States (e.g., Australia, Brazil, Germany, United States), and Regions (e.g., Belgium, Finland, France, Greece). ADM_1 also includes Counties (e.g., Estonia and Sweden) and Federal Districts (e.g., Argentina and Brazil). At the ADM_2 level, commonly used terms include Districts (e.g., Austria, Germany, Turkey), Municipalities (e.g., Brazil, Finland, Mexico), Cities (e.g., China, Indonesia, Japan), and Counties (e.g., United States). At the ADM_3 level, Municipalities remain the most dominant unit in countries such as Austria, Italy, and Poland, alongside similar structures like Communes (Luxembourg and Romania). Administrative units at ADM_4 units are less common but still present, for example Communes in Belgium, Cantons in France, and Municipalities in Spain. Supplementary Figure S1 provides a summary of the GADM levels by country for the reporting and predicted entities in our dataset.

From this broad definition, we collected open-sourced information for subnational actors participating in national or international climate action networks and reporting platforms between 2019 and 2024, including: the Joint Research Centre (JRC), Global Covenant of Mayors and Energy (GCOM), EU Covenant of Mayors, Carbon Disclosure Project (CDP), C40 Cities for Climate Leadership Group, Japanese Ministry of the Environment, Local Governments for Sustainability (ICLEI) Carbonn® Climate Registry, Net Zero Tracker, Under2 Coalition, China Carbon Neutrality Tracker (CCNT), US Climate Mayors, US Climate Alliance, We Are Still In and the Global Climate Action Portal (GCAP or now renamed the Non-State Actor Zone for Climate Action or NAZCA) (see Online-only Table 2 for complete data source information and references). The collection process was performed either through direct access to the source’s database, or indirect collection using HTML parsing tools such as BeautifulSoup v4.12.3. Online-only Table 2 lists the data sources used to compile the self-reported emissions data.

Once collected the data was transformed through two complementary processing steps: spatial alignment, to match each subnational reporting entity to its corresponding physical or administrative boundary, and standardization, to ensure consistency of actors’ data reported across different platforms. The first step included extensive use and improvements to the ClimActor R package²³, including improvements to the matching dictionary, the matching functions for automatic and machine-aided harmonization, and updates to the naming convention for subnational entities. The updated package²⁴ allows us to consider not only nominal (i.e., name) matching but also our database of administrative boundaries. This process allowed us to attribute the self-reported climate information to policy relevant entities (i.e., local governments) and to leverage spatially-explicit data (e.g., EDGAR) to evaluate emissions data at multiple levels of governance. For the purposes of standardization, we classified the highest administrative level (ADM_1) as “Regions,” and the remaining levels (ADM_2, ADM_3, ADM_4 and ADM_5) as “Cities.”

The second stage involved a comprehensive cleaning and standardization of climate-related records for each entity, aimed at producing a consistent, structured database for each data source. The specific steps in this process varied depending on the origin of the data, as climate information was reported through diverse formats, including questionnaire responses, online dashboards, and downloadable databases. Each required tailored processing methods depending on which variables were collected. The process also included internal-consistency checks and ad-hoc fixes to the values to correct for identified errors commonly found in self-reported climate data, such as emissions units differences or misplaced data records. To enable systematic comparison and robust analysis across actors and platforms, when multiple GHG gases were individually reported, we standardized all emission values and targets into carbon dioxide equivalent (CO₂eq) using 100-year Global Warming Potential (GWP-100) values, consistent with those employed in the Intergovernmental Panel on Climate Change (IPCC) assessment reports²⁵. Once each individual data source is harmonized we integrate all spatialized climate records into a single database consisting of a total of 24,084 records.

Several steps of data quality checks and data filtering methods were implemented. First, we filtered out duplicate emissions data, identified as multiple emission values from different initiatives collected for the same entities in the same reporting year. For instance, European cities participating in the Global Covenant of Mayors for Climate and Energy - EU Secretariat (EUCoM) directly report baseline and monitoring emissions inventories to the EU Joint Research Commission, which serves as the EUCoM Secretariat. The EU JRC regularly evaluates and validates the self-reported data and inventories, which involves assessing the data’s completeness (according to their reporting criteria), coherence, and treatment of any outliers²⁶. In 2021, the EU JRC published the first harmonized dataset for 6,200 EUCoM participants spanning 10 years of collected data. Since we regularly collected self-reported data points from the EUCoM website starting from 2018²⁷, we prioritized data for reporting entities from Kona et al.²⁶ since they had undergone an intensive evaluation and validation exercise. In other cases of duplicated emissions data, we kept only the highest total emissions value to avoid underreporting. Second, we calculated per capita emissions and excluded entries with implausible values outside the range of 0.02 to 80 tons per capita. The upper bound of 80 tons CO₂ per capita is informed by historical data from Qatar, which has the highest per capita emissions globally. While it peaked at 93.8 tons per capita between 1990 and 2023, most annual values are below 80 tons CO₂ per capita²⁸. The lower bound of 0.02 tons CO₂ per capita is set just below the lowest value in this dataset, which is observed in Federated States of Micronesia and the Marshall Islands (0.04 to 0.05 tons CO₂ per capita). This threshold sets a reasonable buffer to include plausible emission per capita values while excluding any errors due to misreporting. Third, for entities reporting emissions in multiple years, we visually examined time series plots and removed outliers showing abrupt spikes or drops that suggested reporting error. Finally, after training our preliminary model and generating predicted emissions, we reviewed actors with predicted percentage error greater than one standard deviation from the mean by manually comparing and cross-checking reported emissions with other data sources to correct or remove data with potential reporting errors.

Comparison of self-reported inventories to globally-gridded emission products

To illustrate some of the challenges that arise when utilizing globally-gridded datasets to understand subnational climate action, Fig. 3 shows a comparison of the records of self-reported subnational greenhouse gas inventories data against the widely-used EDGAR gridded dataset for a selection of the most commonly- reported sectors by cities (e.g., buildings, energy, industry, waste and ground transportation). As observed, gridded territorial emissions show a high correlation with self-reported inventories at larger administrative areas such as states and regions (ADM_1: R² = 0.90). However, this correlation weakens at finer scales, where emissions are often systematically underestimated. These differences stem from several sources, including the limited spatial resolution of globally gridded datasets like EDGAR, differences in how emissions are allocated at the local level, and the challenge of capturing locally-specific activity data. For instance, gridded datasets often assign emissions to the location where they physically occur, such as a power plant or landfill, rather than to the location where electricity is consumed or waste is generated. In contrast, self-reported emissions inventories more often account for these indirect or distributed emission sources, leading to higher reported inventory values. These attribution and granularity issues underscore the importance of aligning methodological assumptions when integrating global and local emissions data.

Fig. 3.

Comparison between self-reported subnational greenhouse gas emissions inventory data vs EDGAR territorial emissions for urban relevant sectors (e.g., buildings, energy, industry, waste and ground transportation).

Feature selection

A critical step in our modeling framework is the identification of key predictors of city-level self-reported CO₂ emissions. Building on the machine learning framework developed by Hsu et al.¹⁷ for predicting emissions and evaluating abatement performance in European cities, we draw from the existing literature on urban emission sources and drivers to construct a globally applicable set of predictors^5,13,14,16.

To capture historical emission trends and their alignment with self-reported inventories, we incorporate annual fossil fuel CO₂ emissions from the Open-source Data Inventory for Anthropogenic CO₂ (ODIAC), which provides globally gridded data (2000–2023, 1 km × 1 km) on emissions from fossil fuel combustion, cement production, and natural gas flaring²⁹. Additionally, we utilize the Emissions Database for Global Atmospheric Research (EDGAR v8.0 GHG), which provides sector-specific, gridded emissions of CO₂, CH₄, and N₂O from 1970 to 2022 at a 0.1° resolution. We focus on fossil-based sources, including combustion, industrial processes (e.g., metal and mineral production), solvent use, and agricultural activities⁷. These sectoral emission data are consolidated into six main IPCC-based categories: transport, industry, agriculture, energy, buildings, and waste. CH₄ and N₂O emissions are processed similarly.

Recognizing that urban CO₂ emissions are not solely driven by stationary fossil fuel sources, we incorporate additional environmental indicators as proxies. Building on our previous study¹⁷, we included temperature-driven energy demand indicators such as heating and cooling degree days (HDD and CDD) calculated with the monthly NASA MERRA-2 temperature product from 2000 to 2020 (Bosilovich et al.³⁰). The HDD and CDD are measured as the total temperature deviation from the reference temperature (T_base) (see Eqs. 1–2 below):

$${\mathit{HDD}}_{\mathit{it}}={\mathrm{\Sigma }}_{m=1}^{12}\max (0,T_{\mathit{base}}-T_{\mathit{itm}})D_{\mathit{tm}}$$ 1

$${\mathit{CDD}}_{\mathit{it}}={\mathrm{\Sigma }}_{m=1}^{12}\max (T_{\mathit{itm}}-T_{\mathit{base}},0)D_{\mathit{tm}}$$ 2

where ${\mathit{HDD}}_{\mathit{it}}$ and ${\mathit{CDD}}_{\mathit{it}}$ represent heating and cooling degree days for region i in year t, capturing how temperatures deviate from the reference (base) temperature $T_{\mathit{base}}$, is set to 15.5 °C for HDD and 22 °C for CDD³¹. $T_{\mathit{itm}}$ denotes the mean temperature for region i in year t and month m, and $D_{\mathit{tm}}$ is the number of days in month m of year t. We assume that temperature deviations are uniform across all days within a month, and the monthly contributions are summed across the year to obtain the annual HDD and CDD for each region. To better capture variation across regions and address the issue of zero CDD values observed in many European areas, we created a new variable, Temperature Difference, defined as the sum of CDD and HDD. This variable reflects the total number of degree days during which temperatures deviate from the baseline, whether due to heat or cold.

Air pollution, particularly nitrogen dioxide (NO₂) and fine particulate matter (PM_2.5) are often co-emitted with greenhouse gases, especially from fossil fuel combustion processes and transportation sources³². We included global satellite-derived PM_2.5 data from 2000 to 2020 at a 0.01° × 0.01° resolution developed by the Atmospheric Composition Analysis Group^33,34 and the annual ground-level NO₂ concentrations from 2005 to 2019 at approximately 1 km resolution from the same research group³⁵. We further applied linear extrapolation to the NO₂ data at the administrative unit level to fill gaps between the original dataset coverage and the scope of our study for years from 2000 to 2005 and 2020. Dust surface mass concentration (DUSMASS), black carbon surface mass concentration (BCSMASS), sulfur dioxide surface mass concentration (SO2SMASS), and sulfate surface mass concentration (SO4SMASS) are also included in our analysis from the MERRA-2 Monthly Mean Aerosol Diagnostics, Version 5.12.4 from 2000 to 2020³⁶. Black carbon and sulfur oxides (SOx) are closely associated with energy use intensity and fossil fuel combustion³⁷. While dust can originate from human activities and land cover and land use changes, these may also contribute to increased CO₂ emissions³⁸.

We also processed gridded electricity consumption, population, and gross domestic product (GDP) as key socio-economic drivers of urban emissions. These variables are derived from Chen et al.³⁹ (electricity consumption, 1992–2019, 1 km), Schiavina et al.⁴⁰ (population, 1975–2030, 100 m), and Kummu et al.⁴¹ (GDP, 1992–2022, 30 arc-seconds), respectively.

To generate the environmental variables used for model prediction, we applied zonal statistics to each administrative unit included in the study. This analysis was conducted across two platforms. Variables such as population, GDP, electricity consumption, and emissions from EDGAR were processed using the ReduceRegion function in Google Earth Engine (GEE). In parallel, other spatial datasets, including ODIAC emissions, weather data, and air pollution metrics (PM_2.5, NO₂, dust, and SO_x), were processed using the rasterstats Python package version 0.15.0⁴². Depending on the nature of each variable, either the mean or the sum within each administrative boundary was calculated (see Table S1 for more details).

Model specification

Since subnational CO₂ emissions are shaped by a complex non-linear, interdependent interactions of environmental, energy-related and socioeconomic factors, we employ AutoGluon, an automated machine learning (AutoML) framework that integrates various models, including deep neural networks, into a hierarchical ensemble⁴³. AutoGluon automatically manages cross-validation, out-of-fold prediction tracking, and data shuffling, collectively mitigating overfitting and improving generalization⁴⁴.

Given the heterogeneity of our input dataset, which includes variables collected at different spatial resolutions and temporal frequencies, AutoGluon is well suited for managing multi-source, scale-inconsistent features. For training and testing, we compile features from multiple sources, such as remote sensing satellites, ground-based monitoring, model-derived estimates, and statistical reports, including socioeconomic indicators, satellite-derived emission proxies, climate variables, and environmental pollution metrics. We allocate 80% of the full dataset to training, determined by a standard 80/20 random split, with the remaining 20% reserved for testing and validation.

Figure 4 below presents the feature importance plot (Fig. 4a) and model performance based on the test dataset (Fig. 4b). Overall, the predictive model demonstrates strong predictive capability, with a coefficient of determination (R²) of 0.77 and a Mean Absolute Percentage Error (MAPE) of 38.57%, both calculated using the original (non-log-transformed) emission values. To facilitate visualization, we display the predicted emissions versus true emissions on a logarithmic scale. The feature importance table reports the relative contribution of each predictor to the model’s overall predictive performance, among which population and GDP are the most important features, jointly accounting for approximately 50% of the model’s explained variance. Additional contributors include CO₂ emissions from the building sector, electricity consumption, and longitude, which together explain another 10% of the variance. We explored the effect of the predictors on model predictions using SHAP (SHapley Additive exPlanations)⁴⁵ which shows that population, GDP have the highest impact on the prediction and are positively correlated with the results, indicating that higher population or GDP strongly increases the predicted emissions up to (Figure S2). Other predictors follow the same pattern but their effect is less intense, such as the case of electricity consumption, CO₂ from building and CH4 from transport sector. Overall, the importance and direction of the predictors of our model are consistent with previous studies that predicted emissions for subnational entities in specific regions¹⁷.

Fig. 4.

(a) Feature importance describing which variables contribute to the overall machine learning model’s performance; and (b) model performance (R² and MAPE) comparing the self-reported emissions data and the final predicted CO₂eq emissions of subnational units. The error bar presents uncertainty bound due to feature shuffling. Performance metrics are calculated based on the original scale.

We benchmark AutoGluon against the widely used traditional machine learning model XGBoost, used in Hsu et al.¹⁷ to evaluate its relative performance in handling high-dimensional features. As shown in Supplementary Table S2, AutoGluon achieves more balanced performance between model accuracy and generalization capability⁴³, with a substantially lower MAPE compared to XGBoost, motivating our selection of AutoGluon as the primary predictive model. The final trained model is used to generate subnational-level CO₂eq emissions estimates from 2000 to 2020, producing a consistent emissions dataset.

Data limitations

Our dataset and model are limited by several constraints. First, most of the data are self-reported by subnational actors that participate in international climate initiatives, although some of the data are derived from national-level efforts (e.g., the China Carbon Neutrality Tracker, see Table S1). While there are most certainly subnational entities outside of these initiatives pledging climate actions and reporting emission inventories, we are unable to systematically capture and assess them due to the unwieldy nature and infeasibility of collecting every instance outside of common networks and reporting platforms. Second, due to the nature of our spatial standardization, we are unable to include entities that have no available spatial boundaries (e.g., municipal corporations, regional cooperative associations, or associations of municipalities). Although in some cases we were able to use OpenStreetMaps to identify an alternative spatial boundary, in some cases no authoritative data sources were available to validate this information. Finally, the availability of Global South data is still a major limitation for the scope of the study. Specifically, we could only find 628 self-reported emissions data points for cities and regions outside of the G20, and we could only identify geographic boundaries for 265, either due to unavailable geometries or by insufficient information on the data sources to unequivocally identify the entity’s geometry.

Another aspect to consider is the differences in the way that subnational governments report their GHG emissions and the scopes, sectors and GHG they consider in their inventories. Supplementary Figure S3 shows the differences in GHG coverage in self-reported emissions data from G20 countries by regions (i.e., whether a subnational entity reports only CO₂ emissions or also additional GHG gases such as CH₄, N₂O and F-gases in CO₂eq). Subnational entities in Europe and Eastern Europe and Central Asia regions show over 60% of self-reported emissions covering CO₂ and other GHG gases, with fewer than 20% providing insufficient information to evaluate the GHG coverage. By contrast, other global regions show close to 50% of self-reported emissions data with insufficient information to unequivocally identify their GHG coverage even when the emissions units are CO₂eq. While we have not considered global regions or countries as predictors to reduce regional overfitting, these inconsistencies in the self-reported subnational emission data could still be a source of bias, particularly outside of Europe, where some of the predicted results could be underestimated due to the inherent characteristics of the training data for countries in those areas.

Data Records

We provide model-predicted CO₂eq emissions data for G20 subnationals (refer to Supplementary Figure S2 for which GADM levels are available by country) from 2000 to 2020. Each row in the CSV file represents a specific administrative unit in a given year. The variables included are:

• WB_region: One of World Bank geographic regions (North America, Latin America and the Caribbean, East Asia and the Pacific, Europe, Eastern Europe and Central Asia, Sub-Saharan Africa, and South Asia)

• ISO: The ISO 3166-1 alpha-3 country codes (e.g., USA for the United States of America)

• Country: Name of the country that includes the subnational unit

• Region: The highest administrative level (typically admin_1) containing the observation (e.g., “State of New York” is the state-level administrative name for the unit “County of Albany, NY”)

• Name_full: Human-readable name of the subnational unit, standardized for clarity and consistency (aligned with the ClimActor naming conventions, see Hsu et al., 2020) (e.g., “County of Albany, NY”)

• Name_short: Simplified Name_full (e.g., “Albany” is short for “County of Albany, NY”)

• Year: Calendar year of the emissions estimate.

• Pred_emissions: Predicted annual CO2eq emissions (in metric tons) for the administrative unit based on our model.

• lat: Latitude of the administrative unit geometry’s centroid.

• long: Longitude of the administrative unit geometry’s centroid.

The model-predicted CO₂eq emissions data are available for download from the Data-Driven EnviroLab UNC Dataverse page⁴⁶.

Data examples

Predicted changes in greenhouse gas emissions at the subnational level reflect the geographic, economic and developmental trajectories of regions within the world’s largest economies. Figure 5 shows the percent change in predicted emissions between 2000 and 2020 across administrative level 1 units (e.g., states or provinces) in G20 countries. Regions in parts of North America and Europe exhibit emission declines, notably several US states including the District of Columbia (38%) and Maryland (35%), and some areas in Europe such as Bulgaria’s Province of Vratsai (31%) or Greece’s Epirus and Western Macedonia (35%) show marked emission declines over the 20-year period. In contrast, rapid emission increases are evident across much of Turkey, India, China, Southeast Asia and South America such as Turkey’s Province of Denizli that saw an increase of over 300% or India’s State of Sikkim that increased emissions by 220% in the same period. At a national level, we also find consistency in emissions reduction trends, as in the case of the US, we observe a 16.3% reduction in aggregate emissions from 5,422.584 MTCO₂ in 2005 to 4,536.724 MTCO₂ in 2020–consistent with official U.S. Environmental Protection Agency reports⁴⁷. On the contrary, China has seen an increase of 91% for the whole period and 28% for the period between 2005 and 2010, which is consistent with the 31% increase reported in total CO₂eq emissions through their nationally determined contribution (NDC)⁴⁸.

Fig. 5.

Percentage change of predicted greenhouse gas emissions between 2000 and 2020 for administrative level 1 of G-20 countries.

Examining more closely into individual countries and finer-scale geographic units allows for insight into the spatial distribution of greenhouse gas emissions within countries. Figure 6 showcases three selected countries: France, Canada, and South Africa, and illustrates how greenhouse gas emissions vary across administrative levels within these countries, offering insights into which subnational areas may be driving greenhouse gas emissions. The maps of France and South Africa show that while total emissions at the administrative level 1 are relatively similar; when examining emissions at administrative level 2 and 3, it becomes clear that states and provinces containing major cities are primary emitters. In Canada, the high emissions are mainly observed in major cities located in Ontario and Quebec, showing consistency with the country’s population and GDP centers. Although not densely populated, several administrative level 3 units in the Alberta region also contribute significant amounts of greenhouse gas emissions due to local oil and gas industries.

Fig. 6.

Predicted emissions in 2020 at the administrative levels 1–3 for selected countries. Note: The black dots represent the representative cities of the selected counties, namely Paris, Ottawa, and Johannesburg from left to right.

Examining temporal trends in predicted emissions for selected urban and regional areas reveals further heterogeneity in emission trajectories and underlying drivers. As shown in Fig. 7, many cities exhibit diverging emission paths over time. Tokyo Prefecture (Japan) and Los Angeles (United States) show relatively stable emissions except for the year or so leading up to 2020, driven consistently by population (panel C), GDP (panel D) and electricity consumption growth (panel E), despite estimated declines in territorial CO₂ emissions (panel B). These patterns reflect the differences in socioeconomic development and energy use that drive varying trends in predicted emissions, highlighting their relevance as key predictors. While we see consistent temporal patterns between predicted emissions and underlying predictors, we also observe higher values of predicted CO₂eq emissions compared to territorial CO₂eq emissions. This difference, also noted in Fig. 3, highlights the difficulties of accurately estimating emissions using gridded datasets, particularly in smaller geometries when high-emitting facilities are located outside or near boundaries of administrative entities.

Fig. 7.

Trends in predicted emissions and key emission drivers for representative subnational entities from 2000–2020. Note: Panel (A) reports the trend in machine learning-predicted CO2eq emissions; Panel (B) shows EDGAR CO2 emissions; Panels (C), (D), and (E) show trends in population, total GDP, and electricity consumption, respectively. The triangles (·) in panel A represent self-reported emissions inventories.

Looking at specific subnational entities, we further compare our predicted emissions with the self-reported emissions for the areas in column A of Fig. 7. These areas are representative regions or cities in their respective countries and have their absolute percentage error ranging from less than 1% for some observations in Istanbul (TUR), Tokyo (JPN), to a maximum value of 60% for one observation in Johannesburg (ZAF). The MAPE for each city, calculated from all available points in its respective time series, shows that most cities have an MAPE of less than 25% with values ranging from 1.2% for Istanbul (TUR), 6.3% for Delhi (IND), 13.1% for Berlin (DEU), 18.3% for Buenos Aires (ARG), and 24.5% for Paris (FRA), with only Johannesburg (ZAF) having a value of 44.5%. Overall, the predicted results for these individual cities is consistent with the accuracy observed in the overall MAPE in model training, while still highlighting modeling challenges in locations with limited or inconsistent emissions reporting (See Data Limitations)

Technical Validation

Due to the scarcity of publicly available, self-reported subnational greenhouse gas inventories, there is no comprehensive external validation dataset against which we can fully assess our model. In fact, our framework effectively represents, to our knowledge, the most comprehensive subnational emissions dataset for the G20 currently available, and it can be readily scaled to other countries and globally. However, to evaluate the plausibility and performance of our predictions, we instead compare our estimates to alternative city-level emissions datasets: statistically down-scaled emissions from the Global Covenant of Mayors for Climate and Energy (GCoM) Data Portal for Cities (abbreviated hereafter as DPfC) and the Global Gridded Daily CO₂ Emissions Dataset (GRACED) database.

Validation against other datasets

The DPfC dataset, developed by the World Resources Institute (WRI) and GCoM, provides city-scale greenhouse gas emissions estimates using the Common Reporting Framework (CRF). The dataset uses sector-specific and statistical downscaling methodologies to estimate emissions across various sectors, such as buildings and stationary energy, transportation and mobile energy, and waste. These estimates are derived through a combination of national and regional statistics, and the local contextual information. Emission factors are also adjusted to align with specific municipal boundaries. Emissions from Industrial Processes and Product Use (IPPU) and Agriculture, Forestry and Other Land Use (AFOLU) are not required by this framework but are reported by some countries⁴⁹.

As of June 2025, we downloaded available data for G20 countries, including Brazil, Canada, Denmark, Indonesia, India, Japan, Mexico, and the United States. Each of these countries contains at least one year of emission data from 2000 to 2020, allowing for comparative analysis with our model results. In most countries, IPPU emissions are not available, and emissions from AFOLU are excluded from our comparison analysis to make it consistent with the scope of sectors in our self-reported emission dataset.

Figure 8 shows the comparison results for each country between our model’s predicted emissions and the DPfC data for Brazil, Canada, Denmark, and Japan. Overall, our predictions are higher than the DPfC emissions estimates in representative countries, as demonstrated through Fig. 8, which shows that for most countries of interest, our estimated greenhouse gas emissions values are above the parity line. Denmark, Canada and Mexico are the only countries where we found a R² value over 0.5, indicating a relatively strong relationship between our prediction and the DPfC dataset. In Denmark and Mexico, the regression slopes are close to 1, indicating that the variation in the emissions is captured. The poor model fit (R² = 0) in the remaining countries does not support a reliable interpretation of the regression parameters. However, a pattern of slopes greater than 1 and with most predicted emissions above the parity line is observed in countries such as Brazil, Japan, Indonesia, and the United States. This trend suggests a systematic offset between the two datasets. One possible explanation for these discrepancies is that almost all actors from the DpfC do not include IPPU emissions: for example, fewer than 1% of Brazilian cities report IPPU data. Additionally, transportation sector emission data is not available for actors from Indonesia, India, and Japan, nor does their data documentation include the specifications for these two sectors. These sector differences likely result in a significant underestimation of actor annual emissions in the DPfC dataset.

Fig. 8.

Comparison of annual emissions for subnational actors: Model predicted emissions vs. Data Portal for Cities (abbreviated as DPfC).

In addition, we extracted CO₂ emission data for G20 countries from the GRACED dataset, which provides 0.1° × 0.1° gridded emissions at a daily resolution from cement production and fossil fuel combustion across major sectors (industry, power, residential, ground transportation, international aviation, domestic aviation, and international shipping)⁵⁰. We compared our model-predicted self-reported emissions with GRACED data across multiple administrative levels to evaluate our predictions (Fig. 9). The comparison shows strong overall consistency, with a coefficient of determination (R²) of 0.82 for the full sample and the highest consistency at administrative level 1 (R² = 0.87). We also find that consistency decreases at finer administrative levels, reflecting increasing discrepancies at smaller subnational scales. These differences may partly stem from the ~20% uncertainties reported in GRACED estimates and, more importantly, highlight the challenges of reconciling emission estimates at city or finer administrative levels. In addition, GRACED reports total emissions including international aviation and shipping, which may lead to larger results when downscaled the results to cities compared to our predicted emissions.

Fig. 9.

Comparison between model predicted emissions v.s GRACED emissions in 2019. Note: (a) all administrative levels; (b) administrative level 1; (c) administrative level 2; (d) administrative level 3; (e) administrative level 4.

Uncertainty analysis

To evaluate the robustness and reliability of our modeling framework, we assess uncertainty from variability in feature attributions, predictive uncertainty, and epistemic uncertainty captured by our ensembled model.

Feature attribution uncertainty

We examine the variability of feature importance using the 99th-percentile bounds (p99_low and p99_high) derived from repeated permutations (Fig. 4). These bounds provide an empirical estimate of the stability associated with the relative importance of each feature. For most predictors, the resulting intervals are narrow, indicating that the model consistently attributes similar levels of importance to key features across permutations. This stability suggests that the model’s feature attribution is not significantly affected by the randomness in the permutation and sampling variability.

Predictive uncertainty

To quantify uncertainty in the predicted emissions themselves, we trained a quantile regression model using the same feature set as the main model. This approach yields a full predictive distribution for each observation, from which we extract the 2.5th and 97.5th percentiles to construct 95% prediction intervals (PIs).

On the held-out test set (N = 1,933), the model produces well-calibrated uncertainty intervals, with a PI coverage of 94.5% compared to our model outputs, which is the ensemble mean. Only 2.9% of values fall below and 2.6% above the PI, indicating that the intervals are neither systematically under- nor over-dispersed. As expected for heteroscedastic environmental data, PI uncertainty increases with emission magnitude, but our coverage remains stable (≈0.93–0.99) across emission deciles. These results indicate that the predictive distribution is well-behaved and reflects realistic levels of uncertainty in emissions estimation.

Epistemic uncertainty

We further quantify the Epistemic Uncertainty (EU), which evaluates uncertainty arising from model structure rather than data noise⁵¹. EU is computed as the median absolute deviation of the predictions produced by all individual models within the AutoGluon ensemble for a random sample of 20% of the administrative entities per country. The median EU is 2.14 × 10³ tCO₂e, while the median relative EU is 20.2% based on the median robust coefficient of variation (RCV) across all ensembles and timeseries. Due to the inherent heteroscedasticity of the emissions data, the average EU is higher (87.92 × 10³ tCO₂e) with an average relative EU of 23.7%. Importantly, the relatively low median EU indicates strong consensus among ensemble members, consistent with the overall model accuracy (MAPE) and the well-calibrated PIs described above.

Additionally, we examine the EU distribution by country to explore how uncertainty varies across global regions. The highest median relative EU values are observed in Saudi Arabia (55%) and South Africa (34%), the only G20 representatives from the Middle East & North Africa and Sub-Saharan Africa regions. These higher uncertainties reflect limited temporal reporting, sparse emissions data, and inconsistent sectoral coverage in these regions. In contrast, Europe exhibits the lowest EU (11%), consistent with its denser, more complete, and more consistent self-reported inventories. Other regions, including East Asia and the Pacific, Eastern Europe and Central Asia, South Asia, Latin America, and North America, show intermediate EU values ranging from 20% to 32%. Additional uncertainty intervals for the ten case-study cities in Fig. 7 are provided in Figure S4. These plots illustrate that, even at fine administrative levels, predictions remain broadly consistent across ensemble models. Elevated uncertainties in some Global South cities correspond to limited data availability and sectoral inconsistencies, as discussed in the Data Limitations section.

In summary, the combination of feature attribution analysis, quantile-based predictive uncertainty, and epistemic uncertainty demonstrates that the model produces stable, well-calibrated, and interpretable emissions estimates. The primary sources of uncertainty arise not from model instability but from incomplete reporting, heterogeneity in sectoral coverage, and limited temporal depth in certain regions. These insights will help guide future data collection and model refinement efforts.

Usage Notes

Predicted results should be considered the total yearly CO₂eq emissions for Scope 1 + 2 for Building, Energy, Industrial, Waste and Ground Transportation sectors for the corresponding year. Emissions are reported at multiple subnational administrative levels within G20 countries. Users are provided with both time series prediction results for all G20 administrative units as well as the trained machine learning model product, which can be used to generate emissions predictions for other areas of interest within G20 countries. Model inputs should be transformed using Hyperbolic arcsine transformation and output is generated under the same transformation, predicted values at tons of CO₂ equivalent (tCO₂eq) are obtained by using inverse hyperbolic arcsine.

It is important to emphasize that these predicted emissions estimates are not intended to replace official or locally produced greenhouse gas inventories by subnational governments themselves. Rather, they serve as a supplemental data product designed to provide consistent, spatially complete estimates that enable time series analyses, identification of emission hotspots, and integration into climate impact or mitigation scenario modeling. For cities or regions with limited technical or institutional capacity to produce high-quality inventories, these estimates may provide a useful baseline or diagnostic starting point.

As demonstrated in our data comparison with other city-level emission datasets, such as the GCoM Data Portal for Cities or GRACED, the dataset described in this paper offers the advantage of being fully consistent across space and time while still aligning to administrative boundaries relevant for policy and governance. In particular, compared to these other datasets, the data provided in this study is better able to accurately capture smaller administrative subnational units’ emissions. While gridded products are useful for global comparability, they often do not conform to subnational jurisdictions, limiting their relevance for local decision-making. Conversely, our model is specifically designed to work at the administrative level, making it more suitable for understanding subnational trends and informing city or regional climate planning.

Author contributions

Y.Y., X.W. and D.M. contributed equally to the study. Y.Y., X.W., and D.M. collected data and conducted the analysis. X.W., D.M., and A.H. cleaned and compiled the data. Y.Y. and X.W. contributed to the method. A.H. conceptualized and supervised the study. Y.Y., X.W., D.M., and A.H. wrote the manuscript. All coauthors reviewed, edited, and approved the manuscript.

Data availability

Machine learning and plotting were performed using Python and R. The final trained machine learning model, data, and materials are publicly available via the Data-Driven EnviroLab Dataverse [10.15139/S3/N5SVSP]⁴⁶.

Code availability

Machine learning and plotting were performed using Python and R. The final trained machine learning model, data, and materials are publicly available via the Data-Driven EnviroLab Dataverse [10.15139/S3/N5SVSP]⁴⁶.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41597-026-06691-9.