Calculation of CO₂ Emissions from China at Regional Scales Using Remote Sensing Data

Abstract

Since industrialization, global carbon dioxide (CO₂) emissions have been rising substantially, playing an increasingly important role in global warming and climate change. As the largest CO₂ emitter, China has proposed an ambitious reduction plan of peaking before 2030 and achieving carbon neutrality by 2060. Calculation of CO₂ emissions inventories at regional scales (e.g., city and county) has great significance in terms of China’s regional carbon policies as well as in achieving the national targets. However, most of the existing emissions data were calculated based on fossil fuel consumptions and were thus limited to the provinces in China, making it challenging to compare and analyze the CO₂ emissions of different cities and counties within a province. Machine learning methods provided a promising alternative but were still suffering from the lack of availability of training samples at city or county scales. Accordingly, this study proposed to use the energy consumption per unit GDP (ECpGDP) and GDP to calculate the effective CO₂ emissions, which are the CO₂ emissions if all consumed energy was generated by standard coal. Random forest models were then trained to establish relationships between the remote sensing night-light data and effective CO₂ emissions. A total of eight predictor variables were used, including the night-light data, the urbanization ratio, the population density, the type of sensors and administrative divisions, latitude, longitude, and the area of each city or county. Meanwhile, the mean value of the five-fold cross-validation model was used as the estimated effective CO₂ emissions in order to avoid overfitting. The evaluation showed a root mean square error (RMSE) of 10.972 million tons and an overall Pearson’s correlation coefficient (R) of 0.952, with satisfactory spatial and temporal consistency. The effective CO₂ emissions of 349 cities and 2843 counties in China during 1992–2021 were obtained, providing a promising dataset for CO₂-emission-related applications.

1. Introduction

Carbon dioxide (CO₂) emissions have played an increasingly important role in contributing to global warming and climate change over the past decades. The World Meteorological Organization suggested that the average global temperature for the period 2013–2022 increased by 1.14 °C when compared to the pre-industrial period of 1850–1900 [1], and emissions reduction is critical to limit global warming to 1.5 °C [2]. China is the largest CO₂ emitter, with an ambitious reduction plan of peaking before 2030 and achieving carbon neutrality by 2060 [3]. To meet this goal, China has taken an active role in energy conservation and emission-reduction activities and formulated a series of emission-reduction policies, such as optimizing the industrial structure and developing and utilizing new energy sources [4]. Calculation of CO₂ emissions inventories at regional scales (e.g., city and county) has great significance to China’s regional carbon policies as well as to achieving the national targets [5].

At present, many studies have conducted estimations of China’s CO₂ emissions at the national and provincial scales, with two main categories. The energy consumption data are the basis of most existing methods to estimate CO₂ emissions. Many studies [6,7] have adopted the regional carbon emissions factor accounting method of the Intergovernmental Panel on Climate Change (IPCC) to calculate CO₂ emissions. Some studies [8,9] used this method to estimate the CO₂ emissions of China and its 30 provinces during 1997–2017, with the provincial carbon emissions inventory, including energy-related emissions (17 fossil fuels in 47 industries) and process-related emissions (cement production). Similarly, using the definitions provided by the IPCC regional emissions method, several studies [10,11] have developed a method for constructing production-based inventories of CO₂ emissions from Chinese cities, covering 47 socio-economic sectors, 20 energy types, and 9 primary-industry products. This method was further combined with the input–output model to calculate the consumption-based CO₂ emissions inventories of 13 Chinese cities, being the basis of regional CO₂ reduction policies [12]. Despite the success of these studies, the required energy consumption data are only available for provinces and a small number of central cities in China, covering a few selected years. Estimating CO₂ emissions at city and county levels is still challenging.

Remote sensing provides a promising alternative in CO₂ emissions estimation on regional scales as it can capture the spatial details and dynamics of the Earth’s surface at a low cost. Conventional remote sensing requires a clear physical relationship between the observed optical or microwave signal and the research target, and it is challenging to directly measure CO₂ emissions inventories [13]. However, the night light sensor, originally intended for measuring cloud reflectance and nocturnal weather dynamics, was found to be highly correlated to human activities and economic indicators, e.g., GDP, population density, industrial activities, and the urbanization degree of a region [14]. It has also been used to indirectly predict the energy use and CO₂ emissions level of a region using data-driven methods. Han et al. [15] developed a night-light-based method to simulate the urban CO₂ emissions of 11 selected cities in the Yangtze River Delta of China from 2003 to 2013. Meng et al. [16] proposed a top-down carbon emissions estimation method using night-light data and energy statistics to estimate CO₂ emissions in urban areas of China from 1995 to 2010, assuming that the estimation method is scale invariant. Similarly, Chen et al. [17] first adopted the method of machine learning (PSO-BP algorithm) to establish the relationship between provincial CO₂ emissions and remote sensing night-light data. The trained relationship then was used to estimate the CO₂ emissions of 2735 counties in China from 1997 to 2017. Apart from the main input variable of night-light data, the latitude, longitude and area of a county were also used for improved estimation. Yang et al. [18] trained an ensemble-structure-based neural network model to estimate the carbon emissions at the municipal level in three northeastern provinces from 1998 to 2013, based on the weighted coefficient strategy of stable nighttime-light data. These studies confirmed the feasibility of using remote sensing night light to estimate CO₂ emissions and provided valuable CO₂ emissions data from city to county scales. However, the true values of CO₂ emissions used for training were all from energy consumption data, only covering the 34 provinces and a few central cities in China. The models trained for the provinces were commonly directly applied to estimate the values at regional scales, with the uncertainty caused by the scale-effect being unclear.

Different from the limited data of energy consumption, many cities and counties in China reported their energy consumption per unit GDP (ECpGDP) irregularly, providing an alternative data source for energy consumption. In this study, all the publicly available records of ECpGDP were collected and the corresponding effective energy consumption, in terms of standard coal and CO₂ emissions, were then calculated, providing a few ground truths at the city and county scales. They were then used to train a random forest model for CO₂ emissions, taking the night-light data as the main input. Different from the existing estimation methods [17,18], two extra explainable parameters of population density and urbanization ratio were included in the model inputs to provide a better estimation of CO₂ emissions. The trained model was then used to estimate the effective CO₂ emissions for cities and counties over a long period of 1992–2021. Since the effective CO₂ emissions are based on society’s energy consumption, the estimated effective CO₂ emissions are expected to deviate substantially from the fossil-fuel-related emissions. However, it depicts a worst-case scenario for society’s CO₂ emissions, with a clear urgency for energy transition. At the same time, this can provide policymakers with more direct, regionally accurate data at city and county scales, avoiding the consideration of the CO₂ emissions footprint.

2. Data and Preprocessing

Administrative zone map of China. China has a three-tier system of administrative divisions, being provinces, cities, and counties. A city consists of a few counties, with the number of counties varying substantially from east to west. The administrative divisions of cities and counties in China have changed slightly in the past four decades and the estimation was all based on the current administrative division system, including 370 cities and 2878 counties. Six focus areas were highlighted for further analysis in this study in view of the economy and urbanization, with a summary and the locations of these areas being listed in

Table 1. Six focus areas of China in view of economy and urbanization.

ID	Focus Area	Description
A	The three northeastern provinces	Traditional old industrial base including Heilongjiang, Jilin, and Liaoning provinces
B	The Beijing–Tianjin–Hebei region	The capital economic circle including Beijing, Tianjin, and Hebei provinces
C	The Chengyu city cluster	The major economic cities in western China including Chongqing and Sichuan provinces
D	The central Yangtze River cluster	The most dynamic cities in China including Hubei, Hunan, and Jiangxi provinces
E	The Yangtze delta	The leading cities of the Yangtze River economic area including Shanghai, Jiangsu, Zhejiang, and Anhui provinces
F	The Pearl delta	Early economically developed central cities including Guangzhou, Shenzhen, and other nine cities

and Figure 1, respectively. These areas were as follows: (A) the three northeastern provinces, being the traditional old industrial base but undergoing a recession in the past two decades; (B) the Beijing–Tianjin–Hebei region; (C) the Chengyu city cluster that is located in western China; (D) the central Yangtze River cluster, being the most dynamic city cluster in China; (E) the Yangtze delta, the largest city cluster in China, and (F) the Pearl delta, being the early economically development central cities. These focus areas are well-developed areas with much higher CO₂ emissions than the rest of China. Additionally, the longitudes and latitudes of the geometric center of each city or county and the corresponding areas were calculated and used as input variables in the model.

Night-light data. The data collected by the visible and infrared (OLS) sensors aboard the Defense Meteorological Satellite Program (DMSP) and the Visible Infrared Imaging Radiometer Suite (VIIRS) aboard the Suomi National Polar-Orbiting Partnership were used in this study. The spatial resolution of OLS is 3000 m, with the pixel size of the nighttime light (NTL) products being produced to 1000 m. The OLS NTL data were collected by six DMSP satellites (F10, F12, F14–F16, and F18), operating from 1992 to 2013. The VIIRS sensors are the successors of OLS starting from 2011. VIIRS is able to acquire new NTL remote sensing images (day/night band, DNB) with an enhanced spatial resolution of 750 m. The pixel size of the VIIRS NTL products was preprocessed to 500 m.

Since the two kinds of remote sensing data are obtained by different satellite sensors, there are differences in sensor parameters, data quality, time span, spatial–temporal and radiometric resolution, etc., which makes the two sets of data incompatible and limits the long-term study of night-light data. Fortunately, the two kinds of satellites had an overlap of 2012–2013, making it possible to achieve a consistent dataset. At present, a large amount of research is undertaken to harmonize the two kinds of remote sensing data. For example, Li et al. [19] used the monthly averaged VIIRS and monthly OLS data to generate DMSP-OLS-like data using a power function model. Wu et al. [20] proposed a geographically weighted regression model to generate DMSP-like data, and a consistent NTL time series for 1996–2017 was obtained. Recently, the National Center for Earth System Science released a “DMSP-OLS-like” 1 km NTL remote sensing dataset (1992–2022) of China (available at http://geodata.nnu.edu.cn, accessed on 30 January 2023.). This dataset was produced by a pseudo-invariant pixel method [20]. This “DMSP-OLS-like” dataset was found to hold strong linear correlations with various economic and development statistics, with average R² values of 0.931 and 0.654 at the national and provincial scales, respectively. In this study, this “DMSP-OLS-like” 1 km NTL remote sensing dataset was directly used, considering its relatively higher accuracy. The annual mean of the digital number (DN) values (i.e., pixel values) of each city and county of China were calculated for 1992–2021.

Energy consumption per unit of GDP (ECpGDP) data. ECpGDP is an indicator of energy intensity, reflecting the energy inefficiency of an economy. In China, ECpGDP was defined as the ton of coal equivalent (TCE) per CNY 10,000. This indicator was calculated using the energy consumption data that are not publicly available. The data covered a long period of 1997–2019, including 603 records of 101 cities and 905 records of 235 counties. All the ECpGDP records of cities and counties available on China’s economic and social big data research platform (https://data.cnki.net/home, accessed on 30 January 2023.) were acquired in this study. The corresponding records of GDP were released by the National Bureau of Statistics annually, being also obtained from the same data platform. Figure 1a,b show the spatial distributions of the available records of energy consumption per unit GDP at the city and county scales, respectively, while the number of records for each year is provided in Figure 2a. Most records were collected from 2005 to 2017 as they were considered to be a critical indicator of social and economy development since the 11th five-year plan of China (2006–2011); however, there was no clear spatial pattern of the number of records as it was not compulsory to be released to the public, and thus, none of the cities or counties reported it continuously.

Urbanization ratio of cities and counties. The European Space Agency (ESA) Climate Change Initiative (CCI) and Copernicus Climate Change Service (C3S) landcover products were used to calculate the urbanization ratio of each city and county (Figure 3). The ESA CCI landcover products were updated annually from 1992 to 2015 using the SPOT4 satellite images [21], while the C3S landcover products were produced for 2016–2020 using Sentinel-2 data [22]. Though SPOT4 and Sentinel-2 had different spatial–temporal and radiometric resolutions, consistent global landcover maps were produced for the whole period of 1992–2020 at a spatial resolution of 300 m. The Earth’s surface was classified into 22 landcover types, including the urban target type with a fixed pixel value of 190. The products were released in the format of netCDF. In this study, the number of urban pixels of each city and county were calculated first. The value was then divided by the number of total pixels of a city or county to obtain the urbanization ratio. Since the CCI and C3S landcover product only covered 1992–2020, the urbanization ratio of 2021 was interpolated using a linear interpolation method.

Population density. The National Bureau of Statistics of China has produced seven national censuses, in 1953, 1964, 1982, 1990, 2000, 2010, and 2020, with the population of each city and county being publicly available. These datasets were also acquired from the China’s economic and social big data research platform. In this study, the population of non-census years was interpolated from the seven censuses using a spline function of the first order. The annual population of each city or county was then divided by the corresponding area (excluding Hong Kong, Macao, and Taiwan), resulting in the population density per km². The calculated population density ranged from 0.08 to 48,382/km², with the small values being mainly located in western cities and counties (Figure 4).

Fossil-fuel-based CO2 emissions data. Since the estimated CO₂ emissions in this study are effective values, the CO₂ emissions values estimated by the IPCC method were used for inter-comparison. Specifically, the carbon emissions of provinces available in the carbon accounting database (CEADs) were used in this study. These carbon emissions data were calculated according to the carbon emissions nucleic acid method of IPCC, covering 47 industries and 17 fossil fuels related to energy emissions [8,9,10,23]. The data cover the period of 1997–2019 and 30 provinces excluding Taiwan, Hong Kong, Macao, and the Tibet Autonomous Region.

3. Methodology

3.1. Calculation of the Effective CO₂ Emissions

The annual ECpGDP records were used to calculate the annual effective CO₂ emissions, being treated as the truth of the proposed estimation model. As previously mentioned, the ECpGDP was defined as TCE/CNY 10,000, and the energy consumption of a city or county i (EC_i) in TCE can thus be calculated by:

$${EC}_i={\mathrm{E}\mathrm{C}\mathrm{p}\mathrm{G}\mathrm{D}\mathrm{P}}_{\mathrm{i}}\times {\mathrm{G}\mathrm{D}\mathrm{P}}_i,$$

(1)

The effective CO₂ emissions (e_i) of a city or county i with a unit of million tons can then be calculated using:

$$e_i=\tau {\gamma }_0{\alpha }_0{\beta }_0{EC}_i\times {10}^{-6},$$

(2)

where τ is 44/12 (unitless), representing the mass ratio of carbon (C) to carbon dioxide (CO₂), and γ₀ is a unitless parameter of 0.9, indicating the carbon oxidation coefficient of standard coal. The values of α₀ and β₀ (unitless) are 29.27 and 24.74, respectively. The values of α₀ and β₀ represent the thermal conversion coefficient of standard coal and the carbon content of fuel [24], also known as the average carbon content of standard coal and the carbon emissions factor, respectively. The EC_i is an equivalent energy consumption of various energy sources, including fossil fuels and various renewable energy sources. Since the data source and the method to calculate the ECpGDP are not publicly available, the contribution of fossil fuels cannot be determined. The effective CO₂ emissions defined in Equation (2) can thus deviate significantly from the real CO₂ emissions, especially for urban areas with little consumption of fossil fuels but substantial energy consumption. However, the estimated effective CO₂ emissions are more straightforward in terms of the regional CO₂ reduction policies and domestic carbon emissions trading. Figure 2b shows the calculated effective CO₂ emissions of the 1508 samples, with the values of counties being mainly <20 million tons.

3.2. CO₂ Emissions Model

The flowchart of building the CO₂ emissions model and the application of the model to estimate the CO₂ emissions at city and county scales are provided in Figure 5. The main inputs of the estimation model (

Table 2. The variables contained in a sample (x) of the random forest model and the maximum and minimum values of each variable used in data normalization.

Type (Source)	Variable Name	Minimum	Maximum	Variables
DMSP/OLS NPP/VIIRS	Type of sensors	0	1	1
Administrative zone map of China	Type of administrative divisions	0	1	1
Administrative zone map of China	X[°]	75.94	127.96	1
	Y[°]	21.84	49.66	1
	A[km2]	31.5	471,000	1
CCI and C3S landcover products	Urbanization ratio	0.00	0.84	1
National censuses	Population density[/km2]	2.81	12,596.65	1
DMSP/OLS NPP/VIIRS	Mean of DN values	0.03	37.36	1

) included the mean of the night-light data (mean DN value), the population density, and the urbanization ratio of a city or county. Moreover, the geographical coordinates (x, y) and the area (A) of each city and county were used as auxiliary input parameters to reflect the regional differences in CO₂ emissions, with x and y being the values of the geographical center [17,18]. The commonly used normalization method of z-score standardization was applied to the whole dataset, so that the data of different orders of magnitude were uniformly converted into the same order of magnitude to ensure comparability of the data. Notably, the mean and standard deviation required in the z-score standardization were those of the whole dataset. The effective CO₂ emissions were treated as the truth of the estimation model.

Random forest regression models were selected to estimate the effective CO₂ emissions, considering the complex non-linearity between the effective CO₂ emissions and the aforementioned explainable parameters. Random forest models have been widely used in landcover classification and surface parameter retrieval from remote sensing data [25,26,27]. In the training process of the random forest model, multiple small decision trees were built with different subsets and feature attributes, and they were then ensembled into a more powerful model [28]. Each subset is built with randomly selected samples and randomly selected feature attributes, and this randomization reduces the sensitivity of the decision tree to the training data, thereby preventing overfitting [29].

A total of 1508 samples were achieved, including 603 samples of 101 cities and 905 samples of 235 counties. This is a relatively large sample set compared to that of other studies focusing on the provincial scale [16,17,18,19], but is still small for a machine learning method [30]. Different from the common practice of splitting the sample set into training, validation, and testing sets, 5-fold cross-validation was adopted for the whole sample set, where the sample set was randomly split into 5 subsets. This resulted in 5 independent training and validation sessions, and 5 random forest models were trained. In each session, 4 of the 5 subsets were used for training with the out-of-bag subset (10% of the whole sample set) being used for validation. Since a few hyperparameters were required in the random forest model, an exhaustive grid search was carried out for the two most important parameters [31], i.e., the number of trees (10 to 70 with an interval of 10) and the maximum features for splitting a leaf node (‘auto’ and ‘sqrt’). It is somewhat considered as “cheating” in this step, as the hyperparameters were optimized for the validation set and the 5-fold validation accuracy represents the best results that can be achieved. In the estimating stage, all 5 models were used to estimate the CO₂ emissions, with the average of the five values being the final output. Moreover, the maximum values, minimal values, and standard deviation values of the 5 outputs were calculated, partly reflecting the potential estimation uncertainty

3.3. Accuracy Metrics

Four commonly used accuracy metrics were used to reflect the accuracy of the trained models, namely, bias, Pearson’s correlation coefficient (R), root mean square error (RMSE) in million tons, and mean absolute percentage error (MAPE). The RMSE and MAPE provide the absolute and relative errors. Since the CO₂ emissions estimated in this study were effective values, with the absolute values being less critical, MAPE was treated as the main indicator.

4. Results

4.1. Estimation Accuracy of the Proposed Model

The five-fold cross validation resulted in five models, and the predictions over the five validation subsets were compared with the truth, covering the whole dataset (Figure 6a). Satisfactory validation accuracy metrics were achieved with an overall R and RMSE of 0.952 and 10.972 million tons, respectively. A small overall bias of −0.086 mill tons was observed; however, the large values were substantially underestimated. This is generally consistent with the applications of machine learning methods in other studies, demonstrating that predicted values always regress towards the mean [32]. A further investigation of the range-specific bias and MAPE was carried out in Figure 6b, including four ranges of 0–70, 70–140, 140–210, and >210 million tons. A small positive bias of 1.07 million tons was observed, which became negative and substantially decreased as the true value increased, being −75.86 million tons for the range of >210 million tons. The validation accuracy at the county scale was thus slightly better than that of the city scale because the CO₂ emissions of counties were relatively small (Figure 2b). However, the difference in MAPE was relatively small, with the largest MAPE of 28.87% being observed in the range of >210 million tons. The MAPE values of the cities or counties within the range of 70–140 million tons were relatively small, being 11.69%.

The MAPE values of each city and county were also calculated to show the spatial variations in the model performance (Figure 7). While the MAPEs of a few cities or counties have a large value of >50%, e.g., those in the northeast of China, 75 of the 100 cities and 183 of the 235 counties had a small MAPE of <10%, suggesting a satisfactory spatial consistency. Seven cities and sixteen counties were observed to have a large MAPE of >70%. They were mainly cities or counties with low effective CO₂ emissions or 1–2 samples, and a small estimation bias can result in a large MAPE.

Figure 8 shows the MAPE values of each calendar year. In general, the MAPE values showed a fluctuating upward trend, suggesting that the performance of the estimation models gradually decreased. Relatively large fluctuations were observed before 2005, being mainly caused by the limited number of samples in this period (Figure 2a). The MAPEs of 2005 to 2016 were stable in the range of 12% to 22%, being partly related to the large number of samples in this period. A large MAPE of 43% was observed in 2017. This was caused by two samples having low effective CO₂ emissions of <1 million tons. After removing these samples, the MAPE was ~25%, being consistent with other years around this time period.

4.2. Effective CO₂ Emissions of China: Results at the City and County Scales

The effective CO₂ emissions of each city from 1992 to 2021 were estimated using the trained models. As was previously mentioned, five models were built using five-fold cross-validation and the averaged estimations were treated as the effective CO₂ emissions; the values of 370 cities of four selected years are presented in Figure 9. In general, the CO₂ emissions of all cities have been on the rise over the past four decades. The largest city-level emissions increased from 159.19 million tons in 1992 to 297.21 million tons in 2021, with an average annual increase of 86.7%. Meanwhile, the smallest city-level emissions increased from 4.83 to 11.84 million tons.

The geographical pattern of China’s CO₂ emissions has changed substantially in the past three decades. In 1992, most of the cities had relatively low effective CO₂ emissions of <15 million tons and the eastern coastal areas had higher emissions compared to the whole country, especially in the two megacities of Shanghai and Beijing. The hotspots were mainly located in four focus areas of the Beijing–Tianjin–Hebei region (B), the Yangtze delta (E), the Chenyu city cluster (C), and the Pearl delta (F). This is mainly due to the relatively early economic development of these areas, and at the same time, it also drives the development of surrounding cities to a certain extent, so that the high-carbon-emissions areas show a trend of spatial expansion. While the effective CO₂ emissions of all cites increased from 1992 to 2000, a similar spatial pattern was observed in 2000, and the hotspots within each focus area were enhanced or extended in space. Most cities in northwest China still had a low amount of effective CO₂ emissions of <15 million tons, except for the capital city of Xinjiang (Urumqi). However, this changed substantially in 2010, and only the cities of the Qinghai–Tibet Plateau maintained relatively low effective CO₂ emissions. A few hotspots were observed outside the focus areas, generally being the central cities of each region, e.g., the capital cities of the three northeastern provinces (Figure 9a). From 2010 to 2021, the emissions growth in most cities began to accelerate, with 95% of cities having CO₂ emissions of >15 million tons in 2021. The imbalance between the eastern costal area, central China, and the northwest area was further reduced. However, the three northeastern provinces and the eastern part of the middle Yangtze cities had much smaller increases compared to other areas, being mainly caused by the economic stagflation of these areas in the past decade.

Figure 10 shows the effective CO₂ emissions of each county in 1992, 2000, 2010, and 2021. Similarly, the emissions of most counties kept increasing from 1992 to 2021. While the Beijing–Tianjin–Hebei region, the Yangtze delta, the Chenyu city cluster, and the Pearl delta had relatively higher CO₂ emissions, the spatial patterns at the county scale were substantially different from those observed at the city scale (Figure 9). Little difference was observed between the eastern and western counties in 1992 and 2000. This was mainly caused by the large difference in the administrative areas, with the eastern counties commonly having a much smaller area. For example, the areas of the largest and the smallest county are 220,000 km² and <100 km², respectively. The hotspots in 2010 and 2021 generally matched those of the city scale. A few counties of the three northeastern provinces had reduced CO₂ emissions from 2010 to 2021.

The standard deviation of the five models was estimated as an indicator of the model uncertainty. For simplicity, only the standard deviations of 1992 and 2021 are presented in Figure 11, being the calendar years with the smallest and largest uncertainties, respectively. The spatial patterns of the standard deviation and the mean values were generally consistent. The areas and/or periods with low effective CO₂ emissions commonly had a small degree of uncertainty. In other words, the uncertainties were greater over the areas and/or periods with higher CO₂ emissions. For example, 99% of counties have an uncertainty of <2 million tons in 1992, while Beijing had an uncertainty of >25 million tons during the same year. Moreover, the western areas had greater uncertainties than other areas.

4.3. Inter-Comparison with the CO₂ Emissions from the IPCC Accounting Method

An inter-comparison between the estimated effective CO₂ emissions and the value calculated by the IPCC accounting method was carried out for the whole country. Since the accounting method was based on the consumption data of fossil fuels, emissions have been denoted as fuel-based emissions. Briefly, the fuel-based CO₂ emissions of 30 provinces from 1997 to 2019 were acquired [8,9,11,23], and the national CO₂ emissions were calculated excluding Hong Kong, Hainan, Tibet, and Taiwan. The temporal evolution of the national fuel-based CO₂ emissions was compared with that of the effective CO₂ emissions in Figure 12.

In general, both the fuel-based CO2 emissions and the estimated effective emissions values presented an upward trend from 1997 to 2019 and can be roughly divided into three stages. In the first stage of 1997–2001, the fuel-based CO₂ emissions were at a relatively low value of ~3200 million tons and changed little over time, while the estimated effective values increased slightly from 6031 to 6801 million tons. As expected, the effective CO₂ emissions were much higher than the fuel-based values. This was mainly due to the different estimation methods. The effective CO₂ emissions values used in this study were calculated based on GDP and energy consumption per unit of GDP, including all kinds of energy consumptions. Notably, only nine samples were available before 2005 (Figure 2a); thus, the trained models can have large uncertainties in this period. This, together with the fact that models overestimated the small values (Figure 6b), may have resulted in an overestimation of effective CO₂ emissions in these years. From 2001 to 2013, both estimations of CO₂ emissions showed a steadily increasing trend, with an average increase of 636 million tons per year. This was mainly driven by the rapid industrialization process of the first decade of the 21st century, leading the growth of fuel-based energy demand and, thus, CO₂ emissions. While the effective CO₂ emissions were still much higher than the fuel-based values in this period, the ratio between the fuel-based and effective emissions increased from 48% to 79%, suggesting that the fuel energy was increasingly important in this stage. After 2013, the growth rate of the fuel-based emissions began to decline, while the growth rate of effective emissions kept on increasing, with the gap increasing from 2965 to 8683 million tons as a result of the transition from traditional fossil fuels such as coal and oil to renewable energy.

5. Discussion

Acceptable validation accuracy (RMSE: 10.972 million tons) was achieved in the cross-validation, but the models underestimated the low values and overestimated the high values. This resulted in overestimations in the periods and areas with low effective CO₂ emissions, e.g., the national emissions of 1992–2001, and a potential overestimation in the values after 2015. However, the spatial pattern and its temporal evolution were captured, providing valuable information about CO₂ emissions at the finer scale of the city and county levels. Despite the promising results, the data used in this study do not fully reflect the CO₂ emissions of individual regions. For example, some cities in western China had low degree of social-economic development; thus, their night light intensities and population densities were low, but the CO₂ emissions were actually at a high level. The proposed method can thus potentially underestimate the effective CO₂ emissions, resulting in greater uncertainties in these areas (Figure 11). Notably, the CO₂ emissions of a city do not precisely equal the total emissions of all its constituent counties, owing to the model’s inherent challenge in maintaining consistent accuracy across diverse scales. The proposed methods tended to have a lower accuracy for the recent years (Figure 8). This was also related to the underestimations at high values. In view of the main input data, night-light data may not fully reflect the social and economic activities, because the night light of a city can saturate earlier than the peaking of CO₂ emissions. Accordingly, more publicly available data sources related to CO₂ emissions should be included for improved accuracy. However, this is still challenging because most statistics cannot fully cover the whole period of 1992–2021. Moreover, the uncertainties of other data sources such as comprehensive industry reports, satellite imagery, and government environmental databases are generally unknown and the potential benefit is unclear.

The proposed method used all of the publicly available records of ECpGDP in the training and validation, but the samples (1508) are still insufficient to train an “optimal” model that fully captures the CO₂ emissions at the city and county scales. While some more samples can be included in the future, they may be still limited. Accordingly, data augmentation techniques may be more promising to improve the generalizability of the models [33].

The effective CO2 emissions metric was derived under the assumption that all energy originates from standard coal, making it not directly comparable to actual CO₂ emissions (Figure 12) and, consequently, less applicable to the climate change models [34]. However, the primary motivation of this study was to provide an alternative to an improved regional CO₂-emission-reduction policy. The estimated effective CO₂ emissions provide an analytical lens that serves to fill crucial gaps in the regional dataset. It is essential to note that this approach, while not perfectly aligning with real-world emissions, still exhibits a noteworthy correlation, as evidenced by the high R² values. This correlation suggests a parallelism between the effective CO₂ emissions model and actual emissions, highlighting the potential utility of the proposed method. The real CO₂ emissions caused by human activities within a city or county can be different from the fuel-based CO₂ emissions, especially for areas with cross-region energy transfer. For example, the real CO₂ emissions in eastern China may be much larger than the value estimated by the IPCC accounting method as a result of the China’s west–east power transmission. In contrast, the ECpGDP and effective CO₂ emissions consider all kinds of energy consumption, thus being more straightforward for policy makers, without requiring consideration of the footprint of CO₂ emissions. Furthermore, the concept of effective CO₂ provides a robust baseline for assessing progress in emission-reduction efforts. For a city or county, a lower effective CO₂ definitely means lower CO₂ emissions, with the reduction being attributed to operating in an energy-efficient way and/or using lower carbon energy sources. This indicator can be used together with the statistics of renewable energy, providing a complete picture of the national and regional efforts of emissions reduction. For example, the growing gap between effective CO₂ emissions and fuel-based emissions during the period from 2009 to 2019 (Figure 12) reflects, to a certain extent, the growing proportion of clean energy utilization and the considerable achievements of energy consumption transformation measures.

6. Conclusions

The effective CO₂ emissions of China at the city and county scales were estimated and evaluated in this study using random forest models and a five-fold cross-validation, with an overall R and RMSE of 0.952 and 10.972 million tons, respectively. The spatial and temporal patterns of effective CO₂ emissions were captured, being consistent with those of social and economic development. This study successfully provided the first carbon emissions estimations for cities and counties in China, which provides a key reference for the formulation of regional carbon policies in China.