City-level process-related CO2 emissions in China 2000–2021

Statistical overview of the dataset

Between 2000 and 2021, process-related CO₂ emissions from the 11 industrial products showed a notable increase. Total emissions rose from 294 Mt in 2000 to 1,110 Mt in 2021, with an average annual growth rate of 6.5% and an annual increase of 39 Mt. Among the three industries, the mineral industry contributed the largest absolute increase in emissions, the metal industry exhibited the highest relative growth rate, while the chemical industry experienced comparatively slower growth.

The mineral industry was the largest contributor to process-related CO₂ emissions. Emissions increased from 168 Mt in 2000 to 632 Mt in 2021, with an average annual increase of 22 Mt and a growth rate of 6.5%. Figure 2a shows that between 2014 and 2021, emission hotspots in this industry gradually shifted from northern and northeastern China to southern and southwestern regions. This spatial shift is largely driven by rapid economic growth and rising construction demand in southern provinces after 2014^35,36. The fastest-growing cities were Meizhou in Guangdong province (0.4 Mt/yr), Sanming in Fujian (0.36 Mt/yr), and Chengde in Sichuan (0.32 Mt/yr). In contrast, Shijiazhuang in Hebei (−0.7 Mt/yr), Xuzhou in Jiangsu (−0.59 Mt/yr), and Changchun in Jilin (−0.39 Mt/yr) experienced substantial declines. Figure 2d presents the emission composition of the mineral industry in 2021, showing that cement accounted for 99%, while the contribution from plate glass was relatively minor.

Spatial distribution and structure of process-related CO₂ emissions. (a–c) Average annual growth of process-related CO₂ emissions in the mineral, chemical, and metal industries from 2014 to 2021. Warmer colors (yellow to red) indicate emission increases, while cooler colors (light blue to blue) represent emission reductions. d-f Process-related CO₂ emissions composition of the mineral, chemical, and metal industries in 2021.

The chemical industry showed slower growth, with total emissions increasing from 93 Mt in 2000 to 185 Mt in 2021. The average annual growth rate was 3.3%, with an annual increase of 4 Mt. As illustrated in Fig. 2b, emission growth was concentrated in northern cities, including Ordos in Inner Mongolia (0.32 Mt/yr), Dalian in Liaoning (0.31 Mt/yr), and Yulin in Shaanxi (0.3 Mt/yr). Meanwhile, emissions declined in Shijiazhuang in Hebei (−0.26 Mt/yr), Liuzhou in Guangxi (−0.13 Mt/yr), and Hengshui in Hebei (−0.12 Mt/yr). Figure 2e shows the 2021 emission composition: ammonia contributed 65%, ethylene 25%, while calcium carbide and soda ash accounted for smaller shares.

The metal industry experienced the largest relative growth, with emissions rising from 32 Mt in 2000 to 293 Mt in 2021. The average annual growth rate reached 11.1%, with an annual increase of 12 Mt. As shown in Fig. 2c, between 2014 and 2021, emission growth was concentrated in resource-intensive cities. The fastest-growing cities were Tangshan in Hebei (1.39 Mt/yr), Binzhou in Shandong (0.86 Mt/yr), and Baotou in Inner Mongolia (0.48 Mt/yr). In contrast, emissions declined in Shanghai (−0.16 Mt/yr), Beijing (−0.1 Mt/yr), and Tianjin (−0.09 Mt/yr). Figure 2f presents the 2021 emission composition for the metal industry: crude steel accounted for 82%, followed by aluminum (15%), while lead, zinc, and other metals contributed relatively minor shares.

Uncertainty analysis

Uncertainty analysis is a key tool for evaluating the quality of CO₂ emission inventories^37,38. According to the IPCC, uncertainties primarily originate from activity data and emission factors⁵. Two commonly used methods for uncertainty assessment in previous studies are error propagation and Monte Carlo simulation. Although error propagation is relatively easy to implement, it relies on the assumptions of linearity and normally distributed input variables. These assumptions make it unsuitable when input parameters have large uncertainties or skewed distributions²⁷. In contrast, Monte Carlo simulation does not depend on such assumptions and can flexibly accommodate probability density functions of any shape and range, providing a more accurate representation of input uncertainties and their propagation in emission estimates³⁹. Therefore, this study adopts the Monte Carlo method to assess uncertainties in the estimation of process-related CO₂ emissions.

In our analysis, we assigned a coefficient of variation (CV) of 5% to activity data obtained from official statistical yearbooks and bulltins^27,40. For data imputed using missForest models, the CV was set at 15%^5,41. Emission factor CVs follow the values recommended by the IPCC (see Table 2)⁵. Emissions from ferroalloy production were excluded from the uncertainty analysis due to the complexity of its carbon sources (e.g., petroleum coke, graphite, biomass), which makes the quantification of its emission factor uncertainty particularly challenging.

Probability distributions were derived based on a comprehensive literature review^42,43. When the uncertainty range is less than ±60%, it is typically assumed to follow a normal distribution^44,45. Accordingly, we assumed normal distributions for both activity data and emission factors. A total of 20,000 random samples were generated to estimate the uncertainty range of process-related CO₂ emissions at the 95% confidence level. The results show that the total inventory uncertainty ranges from −3.87% to +3.91%. At the city level, Zhoushan (Zhejiang province) exhibited the highest uncertainty (−18.74% to +18.78%), while Ürümqi (Xinjiang Uygur Autonomous Region) had the lowest (−1.90% to +1.93%).

To evaluate the potential impact of excluding ferroalloy emissions, we conducted a sensitivity analysis assuming a conservative CV of 50%, which showed that their inclusion would alter the total uncertainty range by less than 0.1 percentage point.

Comparison with existing emission estimates

In addition to uncertainty analysis, we further evaluated the reliability of our activity data and emission estimates by comparing them with national statistics and results from previous studies. First, we compared the collected output data for 11 industrial products with national-level statistics (see Fig. 3). Since this study covers 289 cities and not the entire country, the activity data for all 11 products are generally lower than national statistics. This discrepancy is primarily due to the exclusion of cities not included in our study.

Relative differences in activity data between this study and national statistics. The relative difference in activity data is calculated as (this study’s data/national data) − 1. A negative value indicates that the activity data collected in this study are lower than the national statistics.

Among all products, the outputs of calcium carbide, aluminum, lead, and zinc in our dataset are more than 20% lower than the corresponding national totals. These products are likely produced in smaller or less-developed cities that are not part of our study. Additionally, our data are primarily sourced from statistical yearbooks and bulletins, which only include industrial enterprises above the designated size. As a result, outputs from small enterprises are often excluded, leading to further underreporting. In contrast, the reported outputs of cement, plate glass, ethylene, ammonia, soda ash, crude steel, and ferroalloy in this study are only slightly lower than the national totals. This consistent underestimation suggests that the dataset is reliable—it does not overreport industrial output, which is crucial for ensuring the accuracy of CO₂ emission estimates.

To further assess the accuracy of our emission estimates, we compared them with those reported by Yu et al.²⁷ and Hu et al.¹⁸, two of the most comprehensive national-scale studies on China’s process-related CO₂ emissions. Since no existing dataset provides city-level coverage of multiple industrial products, these national-level estimates serve as the most appropriate reference point for validating our results. As shown in Fig. 4, our relatively lower emission estimates reflect a narrower spatial coverage by design, rather than an underestimation. Unlike national-level studies, our study improves spatial resolution by providing city-level CO₂ emission estimates. Furthermore, the use of localized emission factors and more complete output data further improves the accuracy and completeness of the estimates.

Process-related CO₂ emissions from 11 industrial products compared with other studies. Please note that the ranges of the y-axis are different in each subplot.

In the mineral industry (Fig. 4a–b), our estimates for cement and plate glass are consistently lower than those of Yu et al. and Hu et al. during 2000–2020. For cement, this is mainly due to methodological variation—our study calculates emissions based on cement output, whereas both Yu et al. and Hu et al. use clinker-based estimates. On average, our cement emissions are 16% lower than those of Yu et al. and 22% lower than those of Hu et al. For plate glass, both our study and Yu et al. both apply a China-specific emission factor (0.07 t CO₂/t), while Hu et al. adopts the default value (0.1 t CO₂/t), resulting in estimates that are 39% higher than ours. The 13% lower estimate compared to Yu et al. is mainly attributable to cities not included in our sample.

In the chemical industry (Fig. 4c–f), our emission estimates for calcium carbide, ethylene, ammonia, and soda ash are consistently lower than those of Yu et al. and Hu et al. All three studies use the same emission factor for calcium carbide (1.15 t CO₂/t), so the 33% lower estimate in our study is due to unstudied cities. Similarly, we use the same emission factor for ethylene (2.25 t CO₂/t) as Hu et al., yet our results are approximately 10% lower. For ammonia, both our study and Yu et al. use a China-specific emission factor (2.97 t CO₂/t), while Hu et al. adopts the lower default emission factor (2.77 t CO₂/t). As a result, our estimates are 14% lower than Yu et al. and 8% lower than Hu et al. In the case of soda ash, although all three studies adopt the same emission factor (0.67 t CO₂/t), Hu et al. does not account for the fact that only 5% of China’s soda ash is produced via the natural soda process, which is the only one that generates CO₂ emissions. This omission leads to substantial overestimation—Hu et al.’s estimate is 96% higher than ours, while ours are 20% lower than Yu et al.’s.

In the metal industry (Fig. 4g–k), our emission estimates for ferroalloy, aluminum, lead, and zinc are consistently lower than those of Yu et al. and Hu et al., whereas our estimates for crude steel are slightly higher in certain years. The higher estimates for crude steel are due to data limitations: we lack city-level data on pig iron consumption and the technological breakdown of crude steel production via BF–BOF and EAF routes. Therefore, we use total crude steel output as a proxy and apply national-level process shares to estimate emissions. For ferroalloy, aluminum, and zinc (Fig. 4h,i,k), we use the same emission factors as Yu et al. (0.28, 1.5, and 3.12 t CO₂/t respectively). However, our results are 21%, 35%, and 29% lower, respectively, due to differences in spatial coverage. By contrast, Hu et al. adopts default emission factors (1.3, 1.6, and 1.72 t CO₂/t), which do not align well with China’s actual industrial practices. Consequently, our estimates are 83%, 29%, and 94% lower than those of Hu et al., respectively. For lead (Fig. 4j), we follow the same methodology as Yu et al., applying a weighted average based on different smelting processes. But due to the lack of city-level process shares, we rely on national-level proportions. Hu et al., on the other hand, uses a single default emission factor without distinguishing between smelting processes, resulting in 63% higher emissions compared to our study.