Multidimensional Urbanization and Its Links to Energy Consumption and CO₂ Emissions: Evidence from Chinese Cities

Abstract

This study develops an integrated analytical framework to examine the interplay of urbanization, energy consumption, and CO₂ emissions at the city level in China. Utilizing the Entropy-TOPSIS method for multidimensional urbanization measurement, the GM_Combo model for spatial spillover analysis, and Random Forest for identifying emission drivers, we analyze data from 282 Chinese cities from 2006 to 2020. Results reveal significant hierarchical differences in urbanization, with K-means clustering identifying high, medium, and low urbanization groups reflecting diverse regional development pathways. Energy consumption increasingly drives emissions, while urbanization’s influence declines, indicating partial decoupling. Strong spatial spillovers highlight the need for regional coordination. Ecological assets provide moderate mitigation effects. These findings contribute to the literature by introducing a multidimensional urbanization index, uncovering nonlinear energy–emissions dynamics, and quantifying intercity spillovers, offering empirical support for tailored low-carbon policies and sustainable urban governance.

1. Introduction

Urbanization is one of the most profound global trends in the 21st century, reshaping cities and their surrounding environments at an unprecedented pace [1]. With the expansion of urban populations and infrastructure, regional energy use and carbon emissions patterns have undergone significant changes [2,3,4]. In recent years, the scale and speed of urban transformation have been unprecedented, resulting in rapid economic gains and escalating energy consumption and CO₂ emissions [5]. Particularly in major urban agglomerations, the growing demands for infrastructure, mobility, and industry have intensified environmental pressures, highlighting the urgent need for effective and context-sensitive low-carbon strategies [6,7,8,9].

With the acceleration of global urbanization, revealing the dynamic relationship between urban growth and environmental outcomes is not only critical for national planning, but also significant for achieving international climate goals (such as the Paris Agreement) [10]. Among these, China presents a particularly striking case owning to its scale, speed, and uneven urban transformation model. In order to develop effective and targeted mitigation strategies, it is necessary to gain a more detailed understanding of urbanization processes at the city level and to explore the spatial heterogeneity and dynamic interactions between urban development, energy consumption, and environmental outcomes across different cities [11].

Urbanization is a complex and multidimensional process involving demographic expansion, economic growth, and spatial transformation [12]. Prior studies have demonstrated a long-term coupled relationship between urbanization, energy use, and carbon emissions, generally showing their mutual reinforcement [13,14,15]. Temporally, the mechanisms through which urbanization and energy consumption drive emissions vary across different development stages [15,16,17]. Spatially, urbanization’s impact on energy consumption and CO₂ emissions exhibits significant heterogeneity, reflecting disparities and inequalities among cities [18,19,20]. Moreover, the relationships among urbanization, energy consumption, and CO₂ emissions are inherently complex and nonlinear [21,22].

However, there are currently three key limitations. First, most studies conceptualize urbanization solely through proxy variables such as economic indicators, population size, or land use [23], often omitting its ecological and spatial vitality dimensions [24,25]. This narrow framing restricts the ability to capture the multifaceted nature of urban growth and its interaction with the environment, preventing policymakers from identifying city-specific carbon dynamics [26]. Moreover, previous studies have primarily focused on macro-regional scales, such as national or provincial levels, with limited detailed analyses at the city level. This constraint hampers the ability to capture the complex nonlinear dynamics of urban carbon metabolism, which refers to the processes of carbon flow, transformation, and emission within urban systems, involving interactions among human activities, the built environment, and ecological factors [27,28]. In addition, reliance on conventional panel regression models often neglects spatial interdependence and nonlinear effects, failing to reflect how urbanization, energy use, and CO₂ emissions interact across space [29,30]. The role of urban ecological factors in shaping carbon emissions is also frequently overlooked [31,32,33].

These gaps are not only methodological limitations but also have practical impacts. Simplified urbanization indicators may mislead carbon reduction policies. One-size-fits-all governance ignores the differences among cities, and the lack of spatial analysis underestimates the emission externalities and spillover effects among regions. To address these issues, this paper constructs a comprehensive analysis framework at the urban scale, which includes three innovative aspects: First, it incorporates urban vitality and ecology into the measurement of urbanization, breaking through the traditional socioeconomic perspective. Second, it introduces spatial spillover models and interpretable machine learning to reveal the complex mechanisms among urbanization, energy consumption, and carbon emissions from the perspectives of space and interaction. Third, it assesses the ecological compensation potential of urban green assets and quantifies their role in reducing emissions at different stages of urban development. By identifying spatial patterns, analyzing driving mechanisms and incorporating ecological mitigation factors, this paper provides new empirical evidence for differentiated, phased and ecologically oriented low-carbon strategies in the context of rapid urbanization.

This study aims to (1) explore urbanization clusters in Chinese cities and understand urban development trends; (2) understand the spatial spillover effects and internal mechanisms of urbanization, energy consumption, and CO₂ emissions from the perspectives of space and interaction; and (3) evaluate the offsetting effect of ecological factors on carbon emission pressure. These findings are expected to provide references for formulating differentiated low-carbon strategies suitable for different urbanization stages.

The rest of the paper is organized as follows. Section 2 presents the data and methodology, followed by Section 3 which gives the main empirical results. Section 4 presents discussions, followed by conclusions in Section 5.

2. Data and Method

2.1. Data Sources

This study analyzed data from 282 prefecture-level cities across China for the period 2006–2020. The data used in this study were obtained from multiple sources to ensure a comprehensive and multidimensional analysis of urbanization, energy consumption, and CO₂ emissions.

Night-time Light (NTL) and Net Primary Productivity (NPP) Data: NTL data is a type of remote sensing data and is from the Harvard Dataverse platform [34]. In this study, this data is obtained from self-coded cross-sensor-based corrections. This data can be used as a proxy for urbanization and economic vitality, providing valuable information for understanding the spatial distribution and development of cities [35]. NPP represents the portion of organic carbon fixed by vegetation after deducting its own respiration consumption, and the data in this study are derived from the NASA MOD17A3HGF dataset version 6.1. The resulting NPP data were then adjusted by applying a scale factor of 0.0001. Due to satellite orbit issues, the dataset for Hainan Province was missing NPP data for 2020; therefore, data from 2021 were used to fill this gap.

Regional Gross Domestic Product (RGDP) and Population Data (POP): RGDP was obtained from national and provincial statistical yearbooks (www.stats.gov.cn), and this indicator reflects the economic development of each urban area. Population data comes from World Pop [36], a database that provides raster data on world population from 2000 to 2020.

Green Space Area (GSA) and Built-Up Urban Area (BUA) data: Information on urban green space area and built-up area was obtained from the China Statistical Yearbook and Land Use Dataset (www.stats.gov.cn). These indicators are useful for measuring the process of spatial urbanization and its relevance. GSA data represents the total area of green space in a city, which usually includes parks, green belts, gardens, and other types of public green space; GSA is closely related to the ecological quality of a city.

Energy Consumption (EC) and CO₂ Emission Data: The total energy consumption data for each city were obtained from the municipal statistical yearbooks, provincial statistical yearbooks, and statistical bulletins (www.stats.gov.cn). It includes electricity, gas, natural gas, and liquefied petroleum gas (LPG). CO₂ emission data are from the website of the China Urban Greenhouse Gas Working Group (http://www.cityghg.com/toCauses?id=4, accessed on 1 August 2025), including CO₂ emission parties in the categories of agriculture, service, industry, living, transportation, and energy, of which the total CO₂ emissions are selected for this paper.

2.2. Variable Descriptions

To provide a comprehensive overview of the data used in this study,

Table 1. Summary of variables and descriptive statistics.

Variable	Unit	Mean	Median	Min	Max	Std Dev
EC	ton	2,207,944.893	1,009,307.901	12,508.1486	39,215,809.99	3,735,554.63
NTL	index	25,175.1454	11,949.90935	292.565223	379,598.3199	39,201.91029
NPP	gC/m2/year	27,316.13032	18,103.29439	65.402098	345,246.9614	31,825.00712
POP	person	3,921,269.384	3,146,469.682	212,617.7649	27,699,601.14	3,057,244.859
RGDP	RMB	16,897,270.95	8,911,765.899	451,615.4057	303,337,741.9	26,734,527.2
BUA	km2	13,340.24823	7500	600	156,500	18,457.92243
GSA	ha	6918.39406	2994.5	24	164,611	15,044.44634
CO2	ton	3443.064716	2376.5	87	33,487	3564.492098
UI	index	0.092012513	0.071553573	0.008868741	0.703911781	0.081579017

summarizes the key variables, their measurement units, and descriptive statistics for the dataset covering 282 Chinese cities from 2006 to 2020.

2.3. Methodology

As shown in Figure 1, this study develops an integrated analytical framework that systematically investigates the relationships among urbanization, energy consumption, and CO₂ emissions through a sequential workflow. First, a multidimensional urbanization index is constructed using the entropy weight-technique for order preference by the similarity to ideal solution (Entropy-TOPSIS) method, and cities are classified based on their development patterns. Next, spatial analysis and interpretable machine learning models are employed to explore spatial associations and underlying mechanisms. Finally, regression modeling is used to evaluate the ecological mitigation potential of green assets. Together, this framework forms a comprehensive analytical chain that links pattern identification, mechanism interpretation, and emission reduction assessment.

2.3.1. Entropy-TOPSIS-Based Urbanization Index Construction

To comprehensively measure urbanization across Chinese cities, this study constructed a multidimensional urbanization index using the Entropy-TOPSIS method. The selection of indicators was grounded in five fundamental dimensions of urbanization: economic development, population concentration, spatial expansion, urban activity, and ecological condition. Specifically, RGDP was used to capture the level of economic development; POP represented demographic concentration; BUA reflected the extent of spatial expansion; NTL indicated urban activity intensity; and GSA was used to characterize the ecological greenness of urban environments. In addition, NPP is incorporated as a functional ecological indicator. The integration of multi-source data—ranging from statistical yearbooks to remote sensing products—not only enhances the representativeness of each dimension but also contributes to the completeness and reliability of the urbanization information [37].

The entropy method was employed to assign objective weights to each indicator based on the degree of information variability, minimizing subjective bias. These weights were then integrated into the TOPSIS framework to rank cities by their relative urbanization level, by calculating their proximity to an ideal urbanization scenario and distance from the least favorable one [38]. The entropy method calculates the objective weight $w_j$ for each indicator j as follows:

Normalize the indicator matrix $\mathrm{X}=\left[x_{ij}\right]$ to obtain $p_{ij}$:

$$p_{ij}={\displaystyle \frac{x_{ij}}{{\sum }_{i=1}^n{x}_{ij}}}$$

(1)

Calculate the entropy value $e_j$ for each indicator:

$$e_j=-k\sum _{i=1}^n{p}_{ij}\mathit{ln}{p}_{ij}$$

(2)

$$k={\displaystyle \frac{1}{\mathit{ln}n}}$$

(3)

Determine the weight $w_j$ of each indicator:

$$w_j={\displaystyle \frac{1-e_j}{{\sum }_{j=1}^m(1-e_j)}}$$

(4)

Compute the TOPSIS closeness coefficient $C_i$ for each city, representing its comprehensive urbanization level:

$$C_i={\displaystyle \frac{D_i^{-}}{D_i^{+}+D_i^{-}}}$$

(5)

$$D_i^{\pm }=\sqrt{\sum _{j=1}^m{w}_j(x_{ij}-x_j^{\pm })}$$

(6)

where $x_j^{+}$ and $x_j^{-}$ represent the ideal and anti-ideal values for indicator j, respectively.

It is important to clarify that the final urbanization index is a composite measure derived from the joint integration of all selected indicators through this workflow. Although the above formulas are presented in general form, each city’s final score $C_i$ is calculated by aggregating all weighted indicators simultaneously. Thus, the resulting value reflects a holistic urbanization profile that accounts for economic strength, population density, spatial growth, activity intensity, and ecological quality.

Inter-indicator correlations were calculated to assess the consistency of the index structure over time. The results, provided in Figure S1 of the Supplementary Materials, confirm stable relationships among the variables, indicating the robustness of the composite index.

2.3.2. Clustering Analysis

To identify city groups with similar urban characteristics, the derived multidimensional urbanization index is subjected to unsupervised clustering. The K-means algorithm is employed as the primary method due to its efficiency in handling large datasets and minimizing within-cluster variance [39]. As a robustness check, the K-medoids algorithm is also applied, which is more resilient to outliers since it minimizes the sum of dissimilarities rather than squared Euclidean distances [40].

K-means partitions the dataset into K clusters by iteratively assigning observations to the nearest cluster centroid and updating centroids based on the mean of cluster members. In contrast, K-medoids selects actual observations as cluster centers (medoids), offering better robustness when data contain noise or outliers [41].

To determine the optimal number of clusters, the elbow method is applied. This approach evaluates the total within-cluster sum of squared errors (SSEs) for different values of K (ranging from 1 to 10 in this study). The SSE typically decreases as K increases, but the marginal gain drops beyond a certain point—the “elbow”. This inflection point reflects a balance between model simplicity and cluster compactness. The elbow point was identified using a geometric detection algorithm [42], and the results indicated that three clusters provided the optimal partitioning (Figure 2).

2.3.3. Spatial Autocorrelation Analysis

Spatial dependence is evaluated using Global Moran’s I and Local Indicators of Spatial Association (LISA). Global Moran’s I is defined as follows:

$$I={\displaystyle \frac{N\sum _{i=1}^N\sum _{j=1}^N{w}_{i,j}(M_i-\overline{M})(M_j-\overline{M})}{\sum _{i=1}^N{(M_i-\overline{M})}^2\left(\sum _{i=1}^N\sum _{j=1}^N{w}_{i,j}\right)}}$$

(7)

where $M_i$ and $M_j$ are the observed values of the target attribute features on the study objects i and j, respectively, and ${\mathrm{w}}_{\mathrm{i},\mathrm{j}}$ is the spatial weight. LISA decomposes global statistics into local components to detect spatial clusters and outliers.

2.3.4. Random Forest Modeling

A Random Forest model is employed with CO₂ emissions as the dependent variable and urbanization and energy consumption as predictors [43]. The model outputs include variable importance and prediction accuracy [44]. This ensemble method captures nonlinear relationships and avoids multicollinearity.

2.3.5. GM_Combo Model

To capture spatial dependence in CO₂ emissions, this study employs the GM_Combo model [45,46], which integrates both spatial lag and spatial error components. The model is specified as follows:

$$y=\rho Wy+X\beta +\mu$$

(8)

$$\mu =\lambda W\mu +\epsilon$$

(9)

where y is the standardized CO₂ emissions vector, X contains the standardized variables EC and UI, and W is the spatial weight matrix constructed using K-nearest neighbors (KNN), which reflects the adjacency and strength of spatial connections between cities. This matrix is key to capturing spatial autocorrelation, describing how geographically close cities influence each other. While KNN was selected for its simplicity and effectiveness in our context, we also acknowledge other commonly used spatial weight constructions such as contiguity-based and distance-based matrices. ρ is the spatial lag coefficient, representing the influence of neighboring cities’ CO₂ emissions on the emissions of a given city, λ is the spatial error coefficient, capturing the spatial correlation in the error terms, and ε is the white noise error term.

In this study, the GM_Combo model serves four main purposes. First, by incorporating the spatial lag term, the model identifies spatial spillover paths and intercity linkages in CO₂ emissions, enabling the assessment of whether a city’s emissions are significantly influenced by its neighboring cities. A significantly positive ρ indicates the presence of a “pollution diffusion” effect. Second, the inclusion of the spatial error term allows the model to control for spatial autocorrelation arising from omitted variables or regional heterogeneity (e.g., geographic conditions, policy environments), thereby enhancing the robustness of the estimations. Third, compared to single spatial lag or error models, the GM_Combo model provides a more comprehensive representation of spatial interactions and heterogeneity among variables in the urbanization process, substantially improving model fit and explanatory power. Finally, by quantifying spatial spillover effects, the model offers important policy insights for formulating regionally coordinated emission reduction strategies, helping to avoid the “island effect” of isolated urban mitigation efforts and promoting more integrated carbon governance.

2.3.6. Urban Ecological Offset Model

To evaluate the capacity of ecological systems to mitigate urban carbon pressure, we construct an urban ecological offset model using CO₂ emissions as the pressure variable. Urbanization-related factors (NTL, POP, RGDP, BUA, EC) are treated as urban pressure drivers, while NPP and GSA represent ecological mitigation variables [47].

First, we build a baseline model with only urban pressure factors:

$$\widehat{y_1}={\beta }_0+\sum _{i=1}^k{\beta }_i{X}_i$$

(10)

Then, we add ecological factors to form a full model:

$$\widehat{y_2}={\gamma }_0+\sum _{i=1}^k{\gamma }_i{X}_i+\sum _{j=1}^k{\delta }_j{Z}_j$$

(11)

where $X_i$ are urban pressure variables, $Z_j$ are ecological mitigation variables (NPP and GSA), and $\widehat{y_1}$ and $\widehat{y_2}$ represent the fitted CO₂ emissions from the baseline and full models, respectively.

The ecological offset effect is calculated as the difference between the two model outputs:

$$∆y=\widehat{y_1}-\widehat{y_2}$$

(12)

A positive Δy suggests that ecological factors help offset CO₂ emissions associated with urbanization, reflecting the regulatory role of green infrastructure and vegetation. Two fitting methods, linear and nonlinear, are used to depict the relationship between urban stressors, ecological mitigation elements, and the resulting carbon offset effect.

3. Results

3.1. Urban Clustering and Urbanization Rate Results

The number of clusters was chosen using the elbow method based on the sum of squared errors. To categorize Chinese cities based on their urbanization extent, we applied the K-means algorithm to a multidimensional urbanization index, classifying 282 cities into three distinct groups: high, medium, and low urbanization levels (Figure 3). The high urbanization group includes six megacities—Beijing, Shanghai, Guangzhou, Shenzhen, Tianjin, and Chongqing—characterized by advanced economic development, dense populations, and extensive infrastructure, driven by national policies favoring coastal economic zones and global trade integration. The medium group comprises 36 regional centers, such as Nanjing, Wuhan, and Hangzhou, which serve as economic and administrative hubs within their regions, benefiting from targeted urban development policies and proximity to major transport networks. The remaining cities fall into the low urbanization group, often located in inland or less economically prioritized regions, where geographic constraints and limited industrial investment hinder rapid urban growth. The sporadic spatial distribution of high urbanization cities and the relatively discrete pattern of medium urbanization cities reflect regional disparities in policy support, economic opportunities, and resource availability.

The clustering results indicate that from 2006 to 2020, urban stratification across Chinese cities has remained continuous and stable, with clear hierarchical differences among the three groups (Figure 4). Highly urbanized cities—primarily major metropolitan areas such as Beijing, Shanghai, and Guangzhou—exhibit stable and mature development trajectories with relatively limited internal variation. Cities in the medium urbanization group have experienced rapid growth but increasing internal dispersion, reflecting divergent development paths among regional centers. In contrast, cities in the low urbanization group have shown slower development yet relatively little internal disparity, suggesting a more gradual and steady urbanization process.

This stratification is driven by causal factors such as capital accessibility, infrastructure development, and regional policy incentives, which tend to favor coastal and eastern cities over those located inland. Highly urbanized cities have experienced the fastest urbanization rates, forming core urban hubs that drive national economic growth. In contrast, cities in the low urbanization group have made limited progress due to constraints such as reduced investment and industrial activity. The medium urbanization group, meanwhile, has undergone intensified internal differentiation, shaped by local governance and patterns of economic specialization. These patterns highlight the significant regional imbalances in China’s urbanization process and their implications for sustainable development. Therefore, urban policy should take into account not only the current development status but also the trajectories of growth and long-term potential for promoting more balanced regional development.

As a robustness check, we applied K-medoids clustering. This method yielded more balanced group sizes but produced less interpretable patterns in the context of urban development hierarchies. For comparison, K-medoids results are provided in Figure S3 of the Supplementary Materials.

3.2. Exploration of Spatial Distribution Patterns and Assessment of Spatial Spillover Effects

3.2.1. Exploration of Spatial Distribution Patterns

This paper examines the spatial characteristics of urbanization, energy consumption, and CO₂ emissions using both global and local spatial autocorrelation methods.

Global Moran’s I analysis (Figure 5) reveals statistically significant and increasing spatial clustering for urbanization, energy consumption, and CO₂ emissions from 2006 to 2020, indicating that these variables exhibit non-random spatial dependence driven by regional factors. The urbanization index shows moderate spatial clustering (Moran’s I ranging from 0.102 to 0.114), reflecting economic agglomeration and policy-driven urban development in eastern coastal regions. CO₂ emissions exhibit the strongest clustering (Moran’s I rising from 0.138 to 0.236), driven by the concentration of industrial activities and energy-intensive infrastructure in urban centers. Energy consumption displays lower but increasing spatial dependence (Moran’s I from 0.034 to 0.081, p = 0.05 in 2006, p < 0.01 thereafter), likely due to the concentration of energy-intensive industries and shared energy infrastructure in urban clusters [19]. These patterns of spatial clustering highlight the role of geographic and economic factors in shaping the distribution of urbanization and its environmental impacts.

LISA analysis provides nuanced insights into the geographic patterns of spatial dependence (Figure 6a–c). Low–low urbanization clusters consistently appear in the Northwest, where geographic isolation and limited economic investment restrict urban growth. In contrast, high–high urbanization clusters in the Pearl River Delta, Yangtze River Delta, and Beijing–Tianjin regions are driven by concentrated industrial activity, favorable policies, and advanced infrastructure, fostering rapid urban development. The expansion of these clusters over time, with transitional clusters emerging in central and western cities like Chongqing since 2015, reflects the influence of regional development policies and infrastructure networks. For energy consumption (Figure 6b), high-value clusters in urbanized areas result from industrial concentration and high urban energy demand, while dispersed low-value clusters indicate uneven access to energy infrastructure. CO₂ emissions (Figure 6c) show persistent high-value clusters in industrialized regions like the Yangtze River Delta and Beijing–Tianjin–Tangshan, driven by heavy industry and transportation, with low-value clusters in Guangxi linked to lower industrial activity. Cities like Chongqing and Changsha, exhibiting low–high clusters for energy consumption and CO₂ emissions, highlight their role as emerging economic hubs within less urbanized regions, driven by targeted development policies.

Overall, the high-value clustering areas of the three elements are roughly the same, but the scale of clustering is CO₂ emission, EC, and UI, in descending order. However, the low-value clustering areas show a large difference among the three. This indicates that the spatial supply allocation of each factor is generally the same for highly urbanized cities, while the low-value allocation of each factor has its own similarities and differences. These local clustering patterns underscore regional disparities in urban and environmental dynamics, setting the stage for analyzing spatial spillover effects in the subsequent section to inform coordinated regional policy strategies.

3.2.2. Assessment of Spatial Spillover Effects

Understanding the significant spatial autocorrelation of UI, EC, and CO₂ emissions, to further investigate the spatial relationship between UI, EC, and CO₂ emissions. This study adopted the GM_Combo model to examine spatial spillover effects among UI, EC, and CO₂ emissions by capturing two key components of spatial dependence: the spatial lag (ρ), reflecting the influence of neighboring cities’ emissions on a city’s own emissions (spatial autoregression), and the spatial error (λ), representing correlation in unobserved factors such as regional policies (spatial error dependence). The model results are shown in

Table 2. Regression results of the GM-Combo model.

Year	EC Coefficient	UI Coefficient	ρ	λ	Pseudo R2	Spatial Pseudo R2
2006	0.4298 ***	0.3097 ***	0.0623 ***	0.0202 ***	0.5766	0.5363
2010	0.2761 ***	0.3846 ***	0.0767 ***	0.0187 ***	0.5254	0.4582
2015	0.3559 ***	0.3342 ***	0.0487 *	0.0540 ***	0.5312	0.4729
2020	0.7253 ***	−0.0673	0.0674 ***	0.0173	0.6056	0.5632

Note: * p < 0.05, *** p < 0.001.

.

EC’s effect on CO₂ emissions is significant and increases over time (0.4298 in 2006 to 0.7253 in 2020, p < 0.001), while UI’s effect weakens, becoming insignificant in 2020 (−0.0673, p = 0.3655), suggesting potential decoupling. The spatial lag coefficient (ρ) is significant across all years (p < 0.05), peaking in 2010 (0.0767, p < 0.001), reflecting strong spillovers driven by regional industrial cooperation. The spatial error coefficient (λ) is significant in 2006 (0.0202, p < 0.001), 2010 (0.0187, p < 0.001), and 2015 (0.0540, p < 0.001), but insignificant in 2020 (0.0173, p = 0.3655), indicating reduced spatial heterogeneity, possibly due to uniform policies. Model fit improves over time (Pseudo R²: 0.5766 in 2006 to 0.6056 in 2020). These findings highlight that CO₂ emissions are amplified by neighboring cities’ activities, necessitating regional strategies like carbon emission trading in clusters like the Yangtze River Delta or cross-city urban planning to mitigate transboundary pollution and promote low-carbon urbanization.

3.3. Random Forest Model

As shown in Figure 7, EC consistently exhibits the strongest explanatory power, increasing in significance from 0.61 in 2006 to 0.81 in 2020. This trend suggests that the role of energy use in driving emissions has become more prominent over time, likely due to intensifying industrial activity and increasing urban energy demand. In contrast, the effect of urbanization on CO₂ emissions declined from 0.39 to 0.19, suggesting that while urbanization initially had a significant effect on emissions, its relative impact has diminished. This may be due to structural change or the improvement and gradual decoupling of urban efficiency, and furthermore, the results are largely the same as those of the GM_Combo Model.

As shown in Figure 8, the model’s predictions are progressively closer to the actual CO₂ emissions, with the fit improving over time. These results indicate that the UI and EC together are increasingly capable of explaining spatial and temporal emission patterns, validating the robustness of the model in identifying major urban carbon drivers.

3.4. Evaluating the Offset Effect of Ecological Assets on Urban CO₂ Emissions

3.4.1. Preliminary Exploration of the Relationship Between Urbanization and Ecological Factors

The results (

Table 3. Directional effects of urban factors on CO₂ emissions.

Variable	Coefficient	Std. Error	p-Value
Constant	2302.858	153.281	0.000 ***
NTL	688.795	1917.427	0.071 *
EC	24,300.000	849.595	0.000 ***
NPP	−2177.562	894.879	0.010 **
POP	−1630.320	1250.642	0.193
RGDP	3094.086	2705.198	0.025
BUA	1155.344	1733.608	0.051
GSA	−3257.085	1788.221	0.069

Note: R² = 0.433, n = 1128. Significance: * p < 0.05, ** p < 0.01, *** p < 0.001.

) provide an overview of the linear relationships between urban characteristics and CO₂ emissions, yielding a moderate R² of 0.43. As shown in Figure 9, the estimated coefficients offer clear directional insights into the roles of different factors. Among them, EC stands out as the most significant and influential predictor (p < 0.001), confirming its dominant contribution to urban emissions.

In contrast, ecological indicators, particularly NPP (p = 0.015) and green space area (GSA) (p = 0.069), exhibit negative coefficients, suggesting their potential roles in emission mitigation. While GSA’s effect is weaker and marginally significant, the negative direction aligns with theoretical expectations and supports the inclusion of ecological variables in urban low-carbon policy frameworks.

Other socioeconomic factors—such as POP, RGDP, and BUA—did not show statistically significant impacts. However, the signs of their coefficients remain consistent with conceptual assumptions, implying possible nonlinear relationships or multicollinearity effects that merit further investigation. Overall, despite the model’s modest explanatory power, the results reinforce the importance of considering both anthropogenic and ecological dimensions when analyzing urban CO₂ emissions.

3.4.2. The Mitigation Effect of Ecological Factors on Urban CO₂ Emissions

To enhance the robustness of the analysis, we used a linear regression model and a Random Forest regression model, respectively, incorporating both urban stressors and ecological factors. As shown in Figure 10, there was a slight increase in the explanatory power of CO₂ emissions, indicating an increase in the contribution of ecological factors. The Random Forest model was favored over the linear regression, which had less explanatory power. Detailed linear regression results are shown in Supplementary Figure S3, which was not used in the final analysis due to its limited effectiveness.

The offsetting effect of ecological assets on urban CO₂ emissions was assessed using two Random Forest Models: Model 1 included only urban stressors and Model 2 included both urban stressors and ecological factors. Model 1 explained 92.61% of the variation in CO₂ emissions (R-squared = 0.9261) with a mean squared error (MSE) of 0.0008. Adding ecological factors to Model 2 increased the R-squared to 0.9307, but the MSE remained the same, suggesting that the impact on prediction accuracy was minimal.

The ecological mitigation effect was positive but small, with an average marginal effect of 0.0003 and a median of 0.0002. This indicates that, holding other factors constant, a unit increase in ecological indicators (such as NPP or GSA) is associated with a 0.03% reduction in CO₂ emissions. While the magnitude appears modest, the consistent positive sign across models suggests a statistically stable effect. These findings imply that ecological factors—particularly vegetation productivity and green space—play a measurable role in offsetting emissions, and should be considered in urban planning strategies aimed at low-carbon development.

4. Discussion

4.1. Urbanization Dynamics and Spatial Inequality

This study first conceptualizes urbanization as a multidimensional process. By constructing a comprehensive urbanization index that integrates demographic, economic, spatial, ecological, and regional vitality dimensions, it transcends the traditional measurement framework that relies solely on population, economy, and land use [12]. This more holistic analytical framework not only enriches the conceptual understanding of urbanization but also provides a replicable approach for evaluating environmental impacts in rapidly urbanizing regions.

Unlike previous studies that focus primarily on national or provincial scales (e.g., Wang et al. and Heidari et al. [19,26]), this research adopts a city-level perspective, offering a more nuanced examination of intra-category differences among cities with similar urbanization levels. Based on the constructed urbanization index, 282 Chinese cities are consistently classified into high, medium, and low urbanization groups, revealing distinct regional stratification patterns. Compared to earlier studies on spatial inequality in urban infrastructure and policy support [33,48], our clustering-based approach provides a clearer and more concrete representation of urban hierarchies.

By identifying different types of urban clusters, this study proposes differentiated and stratified policy recommendations. It offers detailed empirical insights into how regional policy incentives and access to capital jointly shape urban environmental outcomes, laying both a theoretical and practical foundation for the development of targeted urban planning strategies.

4.2. The Dynamic Role of Urbanization and Energy Consumption in CO₂ Emissions and Their Spatial Linkages

Our analysis confirms a significant causal relationship between energy consumption and CO₂ emissions in Chinese cities, consistent with global trends observed in previous studies [49,50]. Unlike many earlier works that merely identify correlations, this study goes further by quantifying the relative importance of variables and their spatial dependencies, consistently identifying energy consumption as the primary driver of CO₂ emissions. From 2006 to 2020, its explanatory power has increased significantly, reflecting a deepening reliance on energy-intensive industries, delayed industrial upgrading, and growing residential energy demand.

Since the implementation of supply-side structural reforms in 2016, China’s energy structure has shifted toward electrification, reducing the use of coal gas and liquefied petroleum gas [51]. However, due to the continued dominance of coal-fired electricity—especially in highly urbanized areas such as Beijing and Shanghai—emissions have not been substantially reduced. Instead, through spillover effects and spatial dependence, the high-value clustering of CO₂ emissions has expanded far more than that of energy consumption. This spatial divergence underscores the necessity of region-specific energy policies to curb emissions more effectively.

Meanwhile, the influence of urbanization on CO₂ emissions has shown a gradual decline, suggesting that some cities are becoming more resource-efficient. Supported by compact urban forms, improved public transportation, and green building standards, urban expansion has become less environmentally costly and is beginning to exhibit signs of decoupling [52,53]. This trend implies that even as mature cities continue to grow, the marginal increase in CO₂ emissions per unit of urban development may be decreasing.

4.3. Policy Pathways for Urbanization, Emissions, and Ecological Mitigation

Building on the study’s findings, differentiated urban governance strategies are needed across the three stages of the urbanization–emission–mitigation pathway.

In low urbanization cities, where development is still accelerating, compact urban planning and green infrastructure integration should be prioritized to avoid future carbon lock-in. Given the relatively balanced intra-group dynamics, investments in efficient public transit, low-carbon industrial foundations, and ecological land use can shape a sustainable growth trajectory from the outset.

In medium urbanization cities, where emissions are intensifying and urban forms are diverging, coordinated land use planning and energy reform are crucial. The increasing explanatory power of energy consumption suggests the need for structural shifts—phasing out fossil fuel reliance, enhancing energy efficiency, and guiding urban expansion toward less energy-intensive patterns. Policy efforts must prevent uncontrolled sprawl and strengthen local institutional capacity to manage emission risks.

In high urbanization cities, ecological mitigation becomes increasingly important as traditional emission reduction avenues narrow. While urbanization’s impact on emissions is weakening, these cities still maintain high emission levels. Expanding green space, enhancing vegetation productivity, and embedding ecological functions into land use regulations can provide marginal but essential emission buffers. Given the observed weakening of spatial spillover effects, local interventions must be complemented by regional coordination to manage transboundary emission dynamics.

5. Conclusions

This study reveals significant regional disparities in China’s urban development from 2006 to 2020, as demonstrated by K-means clustering of cities into high, medium, and low urbanization groups. EC emerges as the primary driver of CO₂ emissions in Chinese cities, with the influence of the UI declining over time, particularly post-2015, due to improved urban efficiencies. Strong spatial spillover effects, captured by the GM_Combo model’s significant ρ, underscore the necessity of regional coordination in urban clusters like the Yangtze River Delta. Additionally, ecological assets such as GSA and NPP provide moderate but consistent mitigation effects on emissions. These findings contribute to sustainable urbanization research by introducing a multidimensional urbanization index that integrates ecological and vitality dimensions, uncovering nonlinear energy–emissions dynamics through Random Forest modeling, and quantifying intercity carbon spillovers to advocate for spatially coordinated governance.