Coupling Coordination Between Transport Development Level and Carbon Emission Intensity in China: Spatiotemporal Patterns and Regional Heterogeneity

Abstract

Under the strategic context of building a transportation powerhouse in China, the transportation sector faces the dual challenge of reducing emissions while improving efficiency. This study constructs a two-dimensional regional classification framework based on the “economic-carbon” dimension and systematically investigates the coordinated evolution of the development level (TD) and carbon emission intensity (TCEI) of the transportation systems in 31 provinces of China from 2014 to 2023, using methods such as entropy weight TOPSIS, the coupling coordination degree (CCD) model, kernel density estimation (KDE), spatial autocorrelation analysis, and the XGBoost-SHAP explainable machine learning framework based on transfer learning. The study finds that (1) TD shows a fluctuating upward trend, while TCEI continues to decline, with regional imbalances; (2) in terms of time, CCD shows a general upward trend with an N-shaped evolution; spatially, CCD presents a pattern of stronger coordination in the east and weaker in the west, with sustained regional heterogeneity, forming a development pattern of “Region I leading, Region II breaking through, Region III maintaining, Region IV catching up”; and (3) regarding the driving factors, freight volume, transport industry output value, and passenger turnover are the core driving factors of CCD, with significant regional heterogeneity in their mechanisms. This study provides a systematic analytical framework and differentiated policy tools for promoting coordinated regional development of green transportation.

1. Introduction

Under the intensifying global climate change and the constraints of carbon neutrality targets, the transportation sector, as a major source of carbon emissions in China, faces the dual challenges of improving efficiency and reducing emissions. Currently, carbon emissions from the transportation sector account for about 10% of the country’s total emissions [1], and with the acceleration of motorization, emissions are showing a rigid growth trend. After China proposed its “dual carbon” goals in 2020, the pressure on the transportation sector to reduce emissions has further increased. The Chinese government has explicitly proposed the “Transportation Powerhouse” strategy and advocates the “high-quality” development concept, aiming to promote the transformation of the transportation system from scale expansion to quality and efficiency enhancement, and from high-carbon dependence to a green, low-carbon model [2,3,4,5,6].

The core challenge of this transformation lies in how to coordinate the two tasks of “development” and “carbon reduction” in transportation [6], seeking a path for their mutual optimization. On the one hand, the continued development of the economy and society requires the the continuous improvement of the transportation development level (TD), ensuring smooth passenger and freight flow, reducing logistics costs, and meeting the mobility needs of the people. On the other hand, the severe carbon reduction pressure requires effective control and reduction in the carbon emission intensity (TCEI) per unit of transportation output. However, the relationship between TD and TCEI is complex; improving TD and reducing TCEI are not independent goals. Therefore, an in-depth exploration of the intrinsic interactive mechanisms and coordinated state between TD and TCEI is crucial for formulating scientific and precise transportation policies.

The research framework of this paper is as follows: Section 2 reviews the relevant literature and identifies research gaps; Section 3 outlines the regional division, indicator system construction, data sources, and research methods; Section 4 presents the spatiotemporal evolution characteristics of TD, TCEI, and CCD, and identifies the CCD driving factors and discusses research limitations; Section 5 summarizes the research conclusions, proposes policy recommendations.

2. Literature Review

2.1. Research on TD

Transportation development level (TD) is the core indicator for measuring the comprehensive effectiveness of a transportation system in supporting economic and social development. Existing research mainly focuses on three aspects: defining the concept of TD, evaluation methods, and influencing factors.

In terms of concept definition, early studies focused on evaluating technical efficiency in a single dimension, such as transportation turnover output efficiency [7] and energy utilization efficiency [8]. With the deepening of the sustainable development concept, scholars have gradually expanded TD to encompass economic efficiency, social benefits, and environmental performance as a multi-dimensional concept [4,5]. Recent studies using the DPSIR model in the fields of rail transit [9] and low-carbon passenger transport [10] show that TD evaluation should systematically integrate the coordination between demand response, facility supply, and resource allocation. Regarding evaluation methods, data envelopment analysis (DEA) [8,11] and its extended models, such as Super-SBM, EBM-DEA, etc., are widely used for measuring efficiency in transportation. These methods can handle multiple inputs and outputs and identify inefficiencies. Other studies have compared random frontier analysis (SFA) methods [12], which distinguish between random errors and technical inefficiencies, providing a basis for econometric analysis of panel data. Additionally, the entropy weight-TOPSIS method and its variations, due to their objective weighting and comprehensive evaluation advantages, have been widely applied in multi-dimensional assessments. For example, scholars have used the entropy weight-TOPSIS method to calculate the contribution of each indicator to urban rail transit development [13], and the AHP-TOPSIS method to assess sustainable urban transportation strategies from economic, social, and environmental perspectives [14]. The fuzzy TOPSIS method has also been used to assess sustainable transportation systems [15]. Regarding influencing factors, research focuses on infrastructure investment [16,17], technological progress [18,19], and institutional environment [20,21]. The built environment of infrastructure has the most significant impact on travel behavior. Optimizing transportation infrastructure and improving public transportation services can promote the reduction in private vehicle use and guide low-carbon travel modes [17]. Technological innovations, such as intelligent transportation systems and new energy vehicles, improve transportation efficiency and energy utilization, thus enhancing TD [18].

2.2. Research on TCEI

Transportation carbon emission intensity (TCEI) is a key indicator for measuring the quality of low-carbon transportation. Many scholars have conducted extensive research on carbon emissions in the transportation sector [22,23,24,25,26], with a few studies addressing transportation carbon emission intensity and its regional disparities. Relevant studies mainly focus on measurement methods, spatiotemporal evolution characteristics, and influencing factors.

In terms of measurement methods, TCEI measurement includes two levels: total carbon emission accounting and intensity index calculation. In the total emission accounting level, direct measurement methods based on energy consumption statistics and emission factors are widely used, with the IPCC guidelines and China’s provincial greenhouse gas inventory compilation methods providing standardized technical norms [27,28]. The input-output method tracks the implicit carbon emissions in upstream sectors such as transportation infrastructure construction and vehicle manufacturing by constructing multi-regional input-output tables, making it suitable for life cycle assessments and inter-regional carbon transfer studies [29]. In terms of intensity index calculation, empirical research has found that the carbon emission per unit GDP is more suitable as a long-term sustainability evaluation unit for China [30,31]. In recent years, the application of remote sensing inversion and big data mining technologies has provided new methods for monitoring carbon emissions at high spatial and temporal resolutions [32]. In terms of spatiotemporal evolution characteristics, TCEI in China shows an overall downward trend but with significant regional differences, presenting a spatial pattern of “low in the southeast, high in the northwest,” with the largest differences in the eastern regions and the smallest in the central regions [33]. Regarding influencing factors, research has focused on economic growth, energy structure, industrial structure, population size, and urbanization [32,33,34,35], exhibiting significant nonlinearity and regional heterogeneity. The impact of economic growth on TCEI follows an inverted U-shape [34], while energy intensity varies regionally [34]. The driving factors significantly differ in different quantile ranges, reflecting threshold effects and conditional dependence [34,35]. Studies based on XGBoost-SHAP machine learning techniques have revealed the role and marginal contribution of variables in carbon emission intensity [36], which supports the framework of this study.

2.3. Mechanism Analysis of TD and TCEI Interaction

Improving TD provides the foundational support for optimizing TCEI, guiding the transportation system towards low-carbon and efficient development. At the same time, the improvement of TCEI constrains and regulates the path of TD improvement, forcing the upgrade of the transportation development model.

2.3.1. TD’s Impact on TCEI

The improvement in TD reduces TCEI through multiple pathways. First, from a resource allocation perspective, TD improvement can enhance transportation efficiency by optimizing transportation organization structures. The coordinated development of freight and passenger transport systems, the promotion of multimodal transportation, and the application of intelligent transportation systems [18] can significantly reduce carbon emissions and increase transportation output per unit of carbon emission. Second, from a technological innovation perspective, TD improvement promotes the research and application of low-carbon technologies. Specifically, lithium-ion battery energy storage technology and electric propulsion and hybrid powertrain technology have become the dominant clean energy approaches. The large-scale deployment of new energy vehicles powered by these technologies, together with the increased penetration rate of clean energy in the transportation sector, has significantly improved energy use efficiency and reduced dependence on fossil fuels [19,37,38]. Finally, from a spatial layout perspective, improvements in transportation infrastructure reduce TCEI by optimizing spatial layouts. The densification of network density shortens the average transportation distance, and optimizing public transportation services helps shift travel structures towards low-carbon modes [17].

2.3.2. TCEI’s Impact on TD

The reduction in TCEI has both a constraining and supporting effect on TD. From the constraining path, the implementation of policies such as carbon emission trading, fuel taxes, and carbon taxes [39,40] increases the operational costs of high-carbon transportation methods, forcing companies to optimize their transportation organization methods and improve energy management processes. Hard constraints, such as carbon neutrality goals and transportation sector emission reduction standards, require companies to continuously optimize energy management systems, thus promoting the systematic improvement of TD. From the supporting path, the reduction in TCEI creates favorable conditions for the upgrading of TD. Under low-carbon transformation policies, the promotion of new energy vehicles, the optimization of transportation structure, and the construction of green transportation infrastructure significantly improve transportation equipment performance and service quality while reducing TCEI. Incentive mechanisms, such as green loans and subsidies for new energy vehicles provided by the government and financial institutions [41], lower the financing costs for clean energy transportation projects. Effective environmental regulation can enhance innovation and competitiveness [42,43], pushing the transportation industry to reduce pollution and improve efficiency.

2.4. Research Gaps and Contributions

Currently, the academic research on TD and TCEI has yielded abundant results, but there are still several limitations. First, most existing research on regional analysis continues to use traditional administrative divisions, which may obscure the true development contradictions and policy needs within regions. Second, existing research mostly focuses on the independent measurement and evaluation of TD, TCEI, and low-carbon transportation systems in a single dimension, lacking studies that treat TD and TCEI as an interactive, coupled dynamic system. Third, the spatiotemporal evolution of their coupling coordination degree (CCD) and the complex driving mechanisms behind it have not been sufficiently explored.

To fill these gaps, this study first constructs a regional classification framework based on the “economic-carbon” dual dimension and a comprehensive evaluation index system for TD and TCEI. Then, the entropy weight TOPSIS method is used to measure the TD and TCEI levels of 31 provinces. The CCD model and kernel density estimation are applied to study the spatiotemporal evolution characteristics of TD-TCEI coordination. Finally, the transfer learning-based XGBoost-SHAP algorithm is used to identify the key driving factors and their nonlinear mechanisms for regional CCD. This study aims to answer the following three questions: (1) What are the spatiotemporal evolution characteristics of TD-TCEI’s CCD under regional differentiation in China? (2) What factors influence CCD, and how do their mechanisms and regional heterogeneity work? (3) Under the dual carbon goals, what differentiated measures can promote the coordinated development of CCD in different regions?

3. Materials and Methods

3.1. Research Area

Given that single-dimensional classifications are insufficient to fully capture regional characteristics, this study adopts a two-dimensional panel framework for regional classification. This approach avoids the overgeneralization issues associated with traditional administrative or economic regional divisions, allowing for a more nuanced classification of the research regions [28,44]. The two-dimensional panel framework is constructed based on the average values of per capita GDP and carbon emission intensity of each province during the study period [45]. As shown in Figure 1, the 31 provinces of China (excluding Hong Kong, Macau, and Taiwan) are divided into four regions.

3.2. Evaluation Indicators and Data Sources

3.2.1. Evaluation Indicators

(1)

Indicators for TD

This study constructs the TD analysis indicator system based on the PSR (Pressure-State-Response) model. The PSR model, due to its ability to systematically reflect the dynamic interaction between human activities and environmental systems, has been widely applied in the evaluation of green transportation systems. This model initiates the identification and analysis of transportation demand pressures, which reflect the service demands placed on the transportation system by socio-economic development. These pressures subsequently influence the system’s supply status, including infrastructure configuration and the level of transport capacity, which in turn affects regional economic efficiency and residents’ travel experiences. Ultimately, these changes in status drive government and transportation authorities to implement response measures through fiscal investments and human resource allocation, aiming to alleviate pressure, improve supply conditions, and promote the sustainable development of the transportation system, respectively (

Table 1. Transportation Development Level Evaluation Indicator System.

Primary Classification	Secondary Classification	Indicators and References	Weight
Utilization Dimension	Passenger Transport Intensity	Passenger Traffic Volume (PTV) [11,28]	0.1448
		Passenger Traffic Volume-Kilometer (PTVK) [23]	0.1358
	Freight Transport Intensity	Freight Traffic Volume (FTV) [11]	0.1147
		Freight Traffic Volume-Kilometer (FTVK) [23,28]	0.1629
Supply Dimension	Infrastructure	Road Density Index (RDI) = Total Urban Road Length/Built-up Area [5,9]	0.0652
		Per Capita Road Area (PRA) [5,9]	0.0385
	Vehicles	Vehicle Ownership Amount (VOA) = Civil Vehicles + Operational Highway Vehicles (10,000 units) [4]	0.1263
		Public Bus per 10,000 People (PBV) [4,5]	0.0303
Support Dimension	Financial Support	Transportation Financial Expenditure (TFE) [23]	0.0744
	Human Resource Support	Transportation Employment Personnel (TEP) [7,11]	0.1072

).

(2)

Indicators for TCEI

TCEI refers to the carbon emission intensity per unit of transportation output. Since total carbon emissions are easily influenced by factors such as economic scale, TCEI provides a more accurate reflection of the quality and efficiency of a region’s or country’s transportation decarbonization [23,24,25,26,27,28,29,30,31,32,33,34]. This study discusses TCEI, and the specific calculation formula is as follows:

$$TCE{I}_{ij}={\displaystyle \frac{T{E}_{ij}}{T{G}_{ij}}}$$

(1)

where i denotes the region, j denotes the year, TE represents total transport carbon emissions, and TG represents the transport industry output value.

3.2.2. Data Sources and Treatment

This study uses data from 31 provinces in China (excluding Hong Kong, Macau, and Taiwan) as the research sample, with the study period spanning from 2014 to 2023. The TD and TG indicator data primarily comes from the China Statistical Yearbook published by the National Bureau of Statistics (https://data.stats.gov.cn/, accessed on June 2025). Data on CO₂ emissions (TE) are derived from the EDGAR Global Atmospheric Research Emission Database released by the European Commission Joint Research Centre (https://edgar.jrc.ec.europa.eu/, accessed on June 2025). The EDGAR database provides provincial annual emission tables by sector and high-resolution gridded data with a spatial resolution of 0.1° × 0.1°. In this study, industry-wide CO₂ emission data are extracted for regional division of the study area, while CO₂ emissions from the transportation sector are used to measure TE, with emissions expressed in kt CO₂ per year. A small share of missing values in the transport industry output value (TG) is supplemented by linear interpolation, accounting for only 1% of the total observations. To eliminate the influence of dimensional discrepancies on the analytical results, all indicators are standardized using the range normalization method.

3.3. Research Method

3.3.1. Entropy Weight TOPSIS Method

The entropy weight TOPSIS method is a comprehensive evaluation approach that combines objective weighting and multi-attribute decision-making. It is widely used to measure the development level of systems, with specific formulas referenced in [13]. This method avoids subjective bias in weighting through the entropy weighting technique, while the TOPSIS method reflects the relative distance between the evaluation object and the optimal or worst possible solutions, providing reliable foundational data for subsequent coupling coordination analysis. In this study, MATLAB R2024a was employed to normalize the indicator data associated with TD and TCEI. Subsequently, the Entropy-Weighted TOPSIS method was applied to get the corresponding comprehensive indices.

3.3.2. Coupling Coordination Degree Model

Coupling coordination refers to the synergistic interaction and balanced development between subsystems, whereby mutual promotion leads to an optimized overall state. The Coupling Coordination Degree (CCD) model often use for quantifying the level of coordinated development between multiple systems; by comprehensively integrating the interaction strength among systems with their overall development levels, the model depict the coordinated evolution characteristics within complex socio-economic-environmental systems [46,47,48,49,50]. The calculation steps for the coupling coordination degree between TD and TCEI are as follows:

(1)

Comprehensive Index Calculation

The integrated development index T is constructed to reflect the overall development level of TD and TCEI. The formula is defined as:

$$T=\alpha TD+\beta TCEI$$

(2)

where T serves as an integrated index, $\alpha$ and $\beta$ denote the corresponding weight. Considering the equal importance of TD and TCEI, the weight set $\alpha =\beta =0.5$ [48].

(2)

Coupling Degree Calculation

The coupling degree C is calculated to measure the interactive strength and synergy level between the two subsystems. The formula is defined as:

$$C=2\sqrt{{\displaystyle \frac{TD\times TCEI}{{(TD+TCEI)}^2}}}$$

(3)

where C is the coupling degree, with a value range of [0, 1]. A higher C value indicates a stronger interactive relationship and higher synergy between TD and TCEI.

(3)

Coupling Coordination Degree

The final coordination level D is computed to represent the overall development state:

$$D=\sqrt{C\times T}$$

(4)

where D is the coupling coordination degree, with a value range of [0, 1]. A higher D value indicates a higher level of coordinated development between the transport economic subsystem and the transport carbon emission intensity subsystem. Referring to the distribution of D values in the sample, the grading threshold of D is determined as shown in

Table 2. CCD level grading standard.

CCD Level	CCD Range	CCD Level	CCD Range
Extreme Disorder	[0, 0.1]	Barely Coordination	(0.5, 0.6]
Severe Disorder	(0.1, 0.2]	Basic Coordination	(0.6, 0.7]
Moderate Disorder	(0.2, 0.3]	Intermediate Coordination	(0.7, 0.8]
Mild Disorder	(0.3, 0.4]	Good Coordination	(0.8, 0.9]
Near Disorder	(0.4, 0.5]	Excellent Coordination	(0.9, 1]

, and the classification ensures the rationality of hierarchical division.

3.3.3. Kernel Density Estimation (KDE)

Kernel Density Estimation (KDE) is a non-parametric method used to estimate the probability density function of random variables, thereby enabling the visualization of data distribution. This study employs KDE to analyze the dynamic evolution of CCD distribution. By plotting density curves for different years, changes in distribution characteristics can be observed, such as location, shape, and dispersion. The formula for KDE is:

$$f\left(x\right)={\displaystyle \frac{1}{nh}}{{\displaystyle \sum }}_{i=1}^{n}K\left({\displaystyle \frac{x-x_i}{h}}\right)$$

(5)

where $n$ is the number of samples, $h$ is the bandwidth that controls the smoothness of the curve, $K\left(u\right)$ is the kernel function $K\left(u\right)={\displaystyle \frac{1}{\sqrt{2\pi }}}{e}^{-{\displaystyle \frac{u^2}{2}}}$, and $x_i$ represents the observed data points.

3.3.4. Spatial Autocorrelation Analysis

Spatial autocorrelation analysis is used to examine whether the CCD values of provinces exhibit significant spatial dependence. This study employs both Global Moran’s I and Local Indicators of Spatial Association (LISA) to characterize the spatial clustering patterns of CCD at the global and local levels, respectively.

(1)

Global Moran’s I

Global Moran’s I is used to measure the overall degree of spatial autocorrelation across all provinces. The formula is:

$$I={\displaystyle \frac{n}{{{\displaystyle \sum }}_i{{\displaystyle \sum }}_j{w}_{ij}}}\cdot {\displaystyle \frac{{{\displaystyle \sum }}_i{{\displaystyle \sum }}_j{w}_{ij}\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{{{\displaystyle \sum }}_i{\left(x_i-\overline{x}\right)}^2}}$$

(6)

where $n$ is the number of provinces, $x_i$ and $x_j$ are the CCD values of provinces $i$ and $j$, $\overline{x}$ is the mean CCD value, and $w_{ij}$ is the element of the row-standardized spatial weight matrix $W$. The value of $I$ ranges from −1 to 1, where a positive value indicates positive spatial autocorrelation (spatial clustering of similar values), a negative value indicates negative spatial autocorrelation (spatial dispersion), and a value near zero indicates spatial randomness. Statistical significance is assessed using a permutation test with 999 random permutations.

(2)

Local Indicators of Spatial Association (LISA)

While Global Moran’s I capture the overall spatial pattern, it cannot identify local spatial heterogeneity. LISA decomposes the global statistic into individual provincial contributions, enabling the identification of local spatial clusters and outliers. The local Moran’s I for province $i$ is calculated as:

$$I_i={\displaystyle \frac{\left(x_i-\overline{x}\right)}{m_2}}{\displaystyle {\sum }_j}{w}_{ij}\left(x_j-\overline{x}\right)$$

(7)

where $m_2={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_i{(x_i-\overline{x})}^2}$ is the variance of the observations. Based on the signs of $x_i-\overline{x}$ and ${\displaystyle {\sum }_j{w}_{ij}{(x_j-\overline{x})}^2}$, each province is classified into four types: High–High hotspots (high CCD surrounded by high CCD neighbors), Low–Low coldspots (low CCD surrounded by low CCD neighbors), High–Low outliers (high CCD surrounded by low CCD neighbors), and Low–High outliers (low CCD surrounded by high CCD neighbors). The statistical significance of each local statistic is determined through a conditional permutation test with 999 permutations at the 5% significance level.

(3)

Spatial Weight Matrix

The spatial weight matrix is constructed based on a distance threshold derived from the geographic coordinates of provincial centroids. Specifically, the Euclidean distance matrix between all provincial centroids is first computed, and the average distance to the six nearest neighbors is used as the adjacency threshold. For each pair of provinces $i$ and $j$, $w_{ij}=1$ if the distance between them is less than the threshold, and $w_{ij}=0$ otherwise. The weight matrix is then row-standardized so that the weights for each province sum to unity.

3.3.5. XGBoost-SHAP Explainable Machine Learning Framework Based on Transfer Learning

Considering that some regions have small sample sizes, which may lead to overfitting of the model, this study adopts a multi-source domain transfer learning framework based on parameter transfer [51]. This framework transfers model knowledge from data-rich regions (source domain) to sample-scarce regions (target domain), effectively mitigating the small sample problem and improving the model’s generalization ability. It has been widely applied in small-sample research. Specifically, the region itself is treated as the target domain, and the remaining regions are used as the source domain for pretraining. This study employs Python 3.9.20 for data visualization and graphical presentation. The specific steps are as follows:

First, XGBoost is pre-trained on the source domain data using a common feature set [52]. Then, fine-tuning is performed on the small number of samples in the target domain, followed by feature importance calculation based on SHAP.

(1)

Fundamental Principles of the XGBoost Model

XGBoost (Extreme Gradient Boosting) employs an additive model to integrate multiple regression trees:

$${\hat{y}}_i={\displaystyle \sum }_{k=1}^K{f}_k\left(x_i\right),f_k\in \mathcal{F}$$

(8)

where K is the number of trees and F is the function space of regression trees. The model is optimized by minimizing a regularized objective function:

$$\mathcal{L}={\displaystyle \sum }_{i=1}^{n}l\left(y_i,{\hat{y}}_i\right)+{\displaystyle \sum }_{k=1}^K\Omega \left(f_k\right)$$

(9)

Here, $l\left(y_i,{\widehat{y}}_i\right)={(y_i-{\widehat{y}}_i)}^2$ is the squared loss function, and the regularization term $\Omega \left(f_k\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda {\parallel w\parallel }^2$ is used to control the model’s complexity, where T is the number of leaf nodes and w is the vector of leaf node weights.

(2)

Transfer Learning Strategy

To address the limited sample size in the target region, this study devises a two-stage parameter transfer scheme. Stage 1: Pre-training on the Source Domain. A base model $M_s=\left\{f_1,f_2,\dots ,f_K\right\}$ is trained on the source dataset $D_s$ to capture common feature relationships across regions. Stage 2: Fine-tuning on the Target Domain. Initializing with the pre-trained source model, training continues on the target dataset $D_t$. The model’s prediction is a composite of knowledge from the source domain and an increment learned from the target domain:

$${\hat{y}}_i^{\mathrm{target}}={\displaystyle \sum }_{k=1}^{K_x}{f}_k^{\mathrm{source}}\left(x_i\right)+{\displaystyle \sum }_{k=K_x+1}^{K_x+K_t}{f}_k^{\mathrm{target}}\left(x_i\right)$$

(10)

An Early Stopping mechanism is applied in both stages to prevent overfitting. To ensure knowledge transferability and non-redundancy, there is no overlapping data between the source and target domains.

(3)

Principles of SHAP Value Calculation

The SHAP (SHapley Additive exPlanations) method is used to quantitatively analyze the marginal contribution of each feature to the model’s output. SHAP values, based on the Shapley value from game theory, satisfy the axioms of local accuracy, consistency, and missingness:

$$f_j\left(x\right)={\displaystyle \sum }_{S\subseteq F\setminus \left\{j\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[f_{S\cup \left\{j\right\}}\left(x_{S\cup \left\{j\right\}}\right)-f_S\left(x_S\right)\right]$$

(11)

where $f_j$ is the SHAP value for feature $j$, F is the full set of features, S is a subset of features, and $f_S$ is the model output using only the feature subset S. A global feature importance ranking is constructed by calculating the mean of the absolute SHAP values for each feature.

4. Results and Discussion

4.1. Spatiotemporal Evolution of TD and TCEI

4.1.1. Spatiotemporal Evolution of TD

From 2014 to 2023, China’s TD exhibited an N-shaped fluctuating trajectory, as shown in Figure 2, decreasing from 0.31 in 2014 to 0.28 in 2022, and then rising again to 0.34 in 2023, with an average value of 0.31. From 2014 to 2018, with the significant increase in population mobility across China, TD showed a clear upward trend in all regions. However, after the outbreak of COVID-19, TD in all regions was impacted to varying degrees. Regional differences were significant: in 2014, the TD values by region, from high to low, were Region I (0.369), Region II (0.366), Region III (0.310), and Region IV (0.243). By 2023, these values had shifted to Region II (0.406), Region I (0.394), Region III (0.329), and Region IV (0.276).

From the perspective of usage intensity, TD in Region I began to decline continuously after 2017, with a more significant drop during the pandemic, primarily due to a decrease in passenger transport intensity. However, by 2023, TD rebounded rapidly, benefiting from the optimization of the port-rail intermodal transport system and improvements in multimodal transport efficiency, which helped reduce the negative impacts of the decline in passenger transport and supported a quick recovery post-pandemic. TD in Region II peaked at 0.42 in 2018 but then dropped significantly. On one hand, the pandemic severely restricted passenger and freight flow, causing a sharp decline in transport usage intensity. On the other hand, environmental policies slowed the growth of high-energy-consumption freight demand, leading to an overall decrease in transport usage intensity. TD in Region III showed weak overall growth, with passenger transport intensity continuing to be low due to the pandemic, and insufficient momentum for the growth of freight intensity. TD in Region IV remained the lowest overall, mainly due to passenger and freight transport intensities being low for an extended period, reflecting a dual dilemma of insufficient transport demand and underdeveloped transport markets in this region.

From the supply capacity perspective, TD in Region III showed a relatively steady growth trajectory, closely linked to the rapid improvement of infrastructure indicators and the continuous growth of transport capacity indicators. Transportation and logistics development significantly improved RDI and PRA, and the notable increase in VOA enhanced transport supply capacity. TD in Region IV remained low with small fluctuations, although infrastructure levels had significantly improved, they were still far below those of other regions, reflecting the limitations of weak infrastructure and insufficient transport capacity in the region. TD in Region I and Region II remained high due to supply support, but the pace of infrastructure development had slowed, signaling the regions had entered a phase of optimizing supply quality.

From the support capacity perspective, TD in Region I and Region II showed continuous high-level fluctuations, driven by fiscal and talent support which promoted the construction of intelligent transportation systems and the improvement of multimodal transport infrastructure. However, after 2018, the marginal effect of investment diminished, and the fiscal pressures during the pandemic further weakened the impact on TD improvement. TD in Region III and Region IV continued to fluctuate at low levels, with these regions generally having low infrastructure levels. On one hand, fiscal support had not reached the levels of Region I and Region II, and on the other hand, they faced the dual challenges of insufficient workforce and lack of high-level transportation management.

From the spatial distribution pattern shown in Figure 3, China’s TD exhibits a significant gradient characteristic of “high in the east, low in the west, with coastal regions outperforming inland.” Specifically: Region I, centered around Jiangsu and Zhejiang, forms a TD growth hub, relying on the integration strategy of the Yangtze River Delta and the development of intelligent transportation, with most areas falling within the high-value range. These regions are economically developed and generally located along the coast, with increasingly strengthened functions as comprehensive transportation hubs, offering transportation advantages. Region II, centered around Shandong, has TD concentrated in the medium to high-value range. This is attributed to the rich resources in these areas, high freight intensity, and the notable effectiveness of port-rail-road intermodal transport system construction. Region III, centered around Henan, Sichuan, and Anhui, displays significant internal variation in TD. The Yangtze River middle and lower reaches, along with the southwestern economic center of Sichuan, exhibit TD levels that are relatively medium to high, while border provinces show relatively low TD. Region IV, centered around Hebei and Liaoning, relies on the coordinated development of the Beijing-Tianjin-Hebei region and port resource integration to promote improvements in TD. However, the overall level still requires enhancement.

4.1.2. Spatiotemporal Evolution of TCEI

Between 2014 and 2023, China’s TCEI showed a continuous downward trend, as shown in Figure 4, decreasing from 3.74 tons per 10,000 yuan to 2.65 tons per 10,000 yuan, with a cumulative reduction of 29.3%, and an average value of 3.32 tons per 10,000 yuan during the period. Regional differences were significant: in 2014, the average TCEI for each region, from low to high, was Region I (1.72), Region II (2.77), Region III (3.90), and Region IV (6.03). By 2023, these values had decreased to 1.38, 2.18, 2.53, and 4.36 tons per 10,000 yuan, respectively. Specifically: Region I benefited from the early promotion of transportation electrification and new energy applications in economically developed areas, along with an improved transportation network and high penetration of new energy vehicles, which kept TCEI at the lowest level. Region II achieved a steady decline in TCEI through the green transformation of the coastal port cluster in Shandong and the promotion of clean energy heavy trucks. However, energy-rich provinces such as Inner Mongolia, with substantial coal export demand and a high proportion of traditional transportation methods, limited the reduction in TCEI. Region III benefited from the application of clean energy in transportation, leading to a relatively rapid decrease in TCEI. Region IV, although showing a significant reduction, still had a high TCEI due to its generally vast and sparsely populated areas, long transportation distances, and abundant resources, which resulted in high freight intensity. This reflects the structural constraints faced by resource-rich and border regions in the low-carbon transformation of transportation.

From the spatial distribution patterns shown in Figure 5, China’s TCEI exhibits a significant gradient characteristic of “low in the east, high in the west, and low in the south, high in the north.” Specifically: Region I, centered around Zhejiang, Shanghai, and Jiangsu, generally has TCEI values in the low range. Region II, centered around Shandong, has TCEI primarily concentrated in the medium range, due to the high transportation demand in the region, resulting in a relatively higher level. Region III, centered around Yunnan, Guizhou, and Sichuan, shows significant internal differences in TCEI. The middle and lower reaches of the Yangtze River have relatively low TCEI, while the southwestern region, with a high proportion of traditional transport methods, still exhibits a higher TCEI. Region IV, centered around Tibet, Qinghai, and Xinjiang, saw the largest reduction in TCEI, but its absolute value remains high, indicating that the low-carbon transformation of transportation in these regions still faces significant challenges. These spatial patterns are closely related to the progress of transportation greening and transport structure in each region.

4.2. The Spatiotemporal Evolution of CCD

4.2.1. The Temporal Evolution of CCD

During the study period, the overall CCD remained in the moderate coordination stages shown in Figure 6, with an average annual value of 0.7, exhibiting a slow N-shaped evolution. The temporal evolution can be divided into three stages: Steady improvement stage (2014–2016): CCD increased from 0.6880 to 0.6924, with an average annual growth rate of 0.64%. High-level fluctuation stage (2017–2019): CCD decreased from 0.7028 to 0.6974. Despite a slight decline, it remained at the moderate coordination level. Impact recovery stage (2020–2023): CCD decreased from 0.6936 to 0.6923 due to the impact of the pandemic in 2022, and then rebounded to 0.7198 in 2023, with an annual growth rate of 3.97%. This marked a shift towards a high-level coordination state. Overall, before 2017, CCD from high to low was Region I, Region II, Region III, and Region IV. After 2017, Region II surpassed Region I, with CCD from high to low being Region II, Region I, Region III, and Region IV. Among these, Region IV experienced the most significant CCD growth. This evolutionary trend may be related to the implementation of the Transportation Powerhouse strategy, the deepening of the dual carbon policies, and the combined impact of the pandemic, resulting in heterogeneous development in different regions in terms of TD improvement and TCEI control.

The CCD evolution characteristics of the four regions exhibit significant spatial differentiation and path differences. Region I’s CCD increased from 0.757 to 0.768, consistently maintaining a high coordination level. Its coordination advantage stems from a high-quality transportation infrastructure system: the enhanced efficiency of port hubs and the green transformation of rail transit networks have facilitated highly efficient and low-carbon coordinated development. As the national leader in green transportation development, the deep integration of multimodal transport and the green upgrading of ports have become the core driving forces for maintaining high-level coordination. Region II’s CCD increased from 0.739 to 0.772, rising to first place among the four regions. However, its coordination improvement process was accompanied by “high performance-high intensity” structural contradictions. On one hand, the improved transportation infrastructure network and efficient transportation organization system promoted rapid TD growth; on the other hand, the demand for large-scale bulk material export, driven by resource-based industries, and the high proportion of traditional road transport led to slower reductions in TCEI. This asymmetrical development model carries risks of high-carbon path dependence. Region III’s CCD increased from 0.703 to 0.716, showing a slow improvement trend. The main driving force behind the coordination improvement came from progressive improvements in the supply side. The continuous investment in transportation infrastructure due to the “Central Rise” and “Western Development” strategies laid a hardware foundation for TD improvement, while abundant clean energy resources such as hydropower and wind power promoted the green transformation of the energy structure. However, the relatively slow promotion of carbon reduction technologies and transportation structure adjustments hindered the rapid decline of TCEI, resulting in a slower CCD improvement rate compared to Region II and Region IV. Region IV’s CCD surged from 0.578 to 0.658, showing the fastest CCD growth among the regions. On one hand, the large-scale development and utilization of clean energy such as wind and hydropower significantly reduced TCEI. On the other hand, the low starting point of CCD and the substantial room for improvement led to a noticeable growth effect.

The dynamic evolution of the CCD kernel density estimation (KDE) for the four regions during the study period is shown in Figure 7. As a non-parametric descriptive tool, KDE results are inherently sensitive to bandwidth selection and the relatively small sample size at the regional level. Using a combination of Silverman’s rule-of-thumb and cross-validation for bandwidth determination, with robustness checks conducted within a ±20% bandwidth range, the analysis reveals the following patterns. From the number of peaks, the kernel density curves for Region I and Region II exhibit a bimodal feature, while those for Region III and Region IV show a unimodal feature, indicating limited regional differentiation within the latter two groups. From the distribution position, the main peak of all regions displays a fluctuating rightward shift, most notably in Region IV, suggesting an overall upward trend in CCD levels. From the distribution shape, the height of the main peak increases and the width expands across regions, reflecting a modest growth in absolute differences in CCD and continued regional heterogeneity. From the distribution extension, Region I shows slight right-tail extension, whereas Regions III and Region IV exhibit noticeable left-tail extension, implying that a subset of provinces within these regions maintains CCD values below the regional average. These patterns collectively indicate a tendency toward alleviation of polarization in CCD distribution, particularly the reduction in extreme low-value clustering in Region IV.

4.2.2. Spatial Distribution Characteristics of the CCD

To analyze the spatial heterogeneity of CCD, four time snapshots from 2014, 2017, 2020, and 2023 were selected, and spatial visualization was conducted using ArcGIS 10.8, as shown in Figure 8. Region I consistently maintains a high CCD level. In 2014, six provinces, including Jiangsu, Zhejiang, and Anhui, had already reached a good coordination level. By 2017, Guangdong and Hubei had upgraded to good coordination, making all nine provinces in the region achieve at least moderate coordination. However, in 2020, the impact of the pandemic caused provinces like Hubei and Shanghai to regress, reducing the number of provinces with good coordination to three. By 2023, Hubei had returned to good coordination, and the region returned to a high and stable state. Region II shows strong CCD stability. Shandong and Inner Mongolia consistently maintained at least moderate coordination throughout the study period, with Shandong remaining in good coordination throughout, becoming the only region with continuous stability at or above moderate coordination. This stability is attributed to the sustained high performance of TD and steady decline in TCEI, although the structural issues of high carbon emissions still constrain the region’s ability to achieve higher coordination levels. Region III shows steady CCD improvement. The number of provinces with at least moderate coordination increased from 9 in 2014 to 11 in 2023, with a net increase of 2 provinces, the largest improvement among the four regions. Although the number dropped to 8 in 2020 due to the pandemic, it quickly recovered in 2023. Sichuan upgraded from moderate to good coordination, and provinces like Yunnan and Guizhou also saw improvements in coordination. This steady improvement is mainly driven by the dual forces of infrastructure development supporting TD growth and the advantages of clean energy driving the rapid decline in TCEI. Region IV achieved a significant low-level CCD leap. In 2014, only 3 provinces reached moderate coordination or above, with Xinjiang and Qinghai at barely coordinated or even near-disordered levels. By 2017, Xinjiang had risen to barely coordinated, and Guangxi and Shanxi had reached moderate coordination. During the pandemic in 2020, Xinjiang, Qinghai, and Shanxi saw an upward trend despite the adverse circumstances. By 2023, Xinjiang and Gansu had risen to primary coordination, and Hebei remained stable at good coordination. The number of provinces with moderate coordination or above increased from 3 to 4. Although Qinghai and Tibet experienced fluctuations and regression, the region as a whole made the leap from low coordination to moderate coordination, mainly driven by steady TD improvement and significant reductions in TCEI.

Overall, the spatial evolution across the four snapshots reveals a distribution pattern of higher CCD in the east and lower CCD in the west, with regional differences gradually narrowing. The number of provinces achieving good coordination or higher increased from 8 in 2014 to 11 in 2023, while the number of provinces below barely coordinated decreased from 6 to 3. To further understand the differentiated pathways driving CCD changes in different regions, the next step will involve using machine learning methods to analyze the impact of various internal indicators of TD and TCEI.

4.2.3. Spatial Autocorrelation of the CCD

To further study the spatial heterogeneity of CDD, a global Moran type I and local spatial correlation index (LISA) analysis was carried out, and the results are shown in

Table 3. Global Moran’s I for CCD (2014–2023).

Year	Moran’s I	Z-Score	p-Value
2014	0.348632	3.2393	0.004
2015	0.295396	2.6438	0.018
2016	0.338438	3.1344	0.007
2017	0.372692	3.3177	0.001
2018	0.382329	3.4809	0.002
2019	0.321307	3.5135	0.003
2020	0.326249	2.8339	0.01
2021	0.424316	3.6917	0.001
2022	0.433423	4.0869	0.001
2023	0.340013	2.9805	0.007

. Study finds that the Global Moran’s I values for CCD remain positive and statistically significant across all years during the study period (p < 0.05). This indicates the existence of persistent positive spatial autocorrelation in CCD distribution, revealing that provinces with similar CCD levels exhibit significant spatial clustering rather than random distribution. Notably, the temporal trajectory of Global Moran’s I exhibits a fluctuating pattern that closely mirrors the N-shaped evolution of CCD: the values rose from 0.349 in 2014 to 0.382 in 2018, declined to 0.326 in 2020, surged to 0.433 in 2022, and moderated to 0.340 in 2023. This synchronous co-movement suggests that the spatial clustering intensity of CCD is closely coupled with its overall level, implying that regional spatial spillover effects constitute a key mechanism underlying the coordinated evolution of TD and TCEI.

The LISA analysis further reveals the local spatial association structure underlying the regional heterogeneity of CCD, as shown in Figure 9. Across the four time snapshots, significant spatial clusters are predominantly concentrated in Region I and Region IV, while Regions II and III exhibit no statistically significant local clustering, reflecting the spatial transitional nature of these two regions.

In Region I, Anhui, Jiangsu, Jiangxi, and Hubei form a persistent and statistically significant High–High (HH) cluster across all four years, indicating that these provinces and their neighbors consistently maintain high CCD levels. This stable HH cluster, centered around the middle-lower Yangtze River region, provides statistical evidence for the spatial spillover effects generated by the Yangtze River Delta integration strategy and the coordinated development of regional transportation networks. Notably, Shanghai shifted from a non-significant status in 2014 and 2017 to a significant Low–High (LH) outlier in 2020 and 2023, indicating that this municipality’s CCD fell below its high-performing neighbors during and after the pandemic period, forming a localized depression within the otherwise high-coordination cluster. Regions II and III show no significant LISA clusters across all four years, indicating that the CCD levels of provinces within these regions do not form statistically significant spatial agglomerations with their neighbors. For Region II, this reflects its spatial transitional position between the high-coordination eastern cluster and the low-coordination western periphery. For Region III, the absence of significant clustering is consistent with the substantial internal variation in CCD levels identified earlier, where provinces such as Sichuan exhibit relatively high CCD while neighboring provinces remain at lower levels, resulting in offsetting spatial effects. In Region IV, Ningxia and Gansu appear as significant Low–Low (LL) coldspots in 2014 and 2017, confirming the existence of a low-coordination spatial trap in the northwestern frontier during the early study period. However, both provinces lost their LL significance by 2020 and 2023, and Qinghai only briefly appeared as a significant LL coldspot in 2020 before returning to non-significance in 2023. This progressive dissolution of the LL cluster suggests that the low-coordination spatial trap in Region IV is gradually loosening, consistent with the “Region IV catching up” development pattern identified in the temporal analysis.

Overall, the LISA analysis identifies two dominant spatial structures: a persistent HH cluster in Region I and a gradually dissolving LL coldspot in Region IV. The stability of the HH cluster confirms that high-level coordination in the eastern region is reinforced by spatial spillover effects, while the dissolution of the LL cluster provides evidence that the spatial lock-in effect constraining western regions is being progressively weakened. The absence of significant clustering in Regions II and III underscores their transitional roles in the national CCD spatial pattern. These findings, combined with the Global Moran’s I results, provide a comprehensive spatial statistical foundation for the subsequent analysis of differentiated driving mechanisms across regions.

4.3. Driving Factors Analysis

4.3.1. Model Selection and Validation

The performance of the baseline XGBoost model, trained solely on region-specific data, was compared with that of the transfer-learning XGBoost model following pretraining on data from other regions, as shown in

Table 4. Model Performance Comparison: Baseline vs. Transfer Learning.

Model	Region	R2	RMSE	MAE
Baseline R2	Region I	0.8083	0.0371	0.0185
	Region II	−1.2994	0.1042	0.0795
	Region III	0.9185	0.0283	0.0199
	Region IV	0.7826	0.0737	0.0311
Transfer Learning R2	Region I	0.9777	0.0127	0.0102
	Region II	0.9714	0.0116	0.0091
	Region III	0.9678	0.0178	0.0133
	Region IV	0.765	0.0767	0.0407

. To ensure robust validation and prevent data leakage, a random train/test split allocating 70% of the observations to training and 30% to testing was adopted, with the test set strictly reserved for final evaluation. In the transfer-learning variant, all observations from the target region were explicitly excluded during pretraining on the source dataset, followed by fine-tuning on the target-region training samples. Hyperparameters were carefully tuned based on preliminary grid search in the baseline model (using 3-fold cross-validation) and fixed at optimized values with early stopping to enhance generalization. These procedures yielded the performance metrics reported in Table 4. The results indicate that the performance of the model improved significantly after applying transfer learning. Specifically, the R² for Region II increased from a negative value to 0.97, with the R² for all regions, except Region IV, exceeding 0.96. Additionally, the RMSE and MAE for all regions were both below 0.08, validating the effectiveness of this model framework in small sample scenarios and providing a reliable basis for the subsequent analysis of feature importance quantification.

4.3.2. Identification of Key Factors

Based on the trained XGBoost model, the SHAP interpretability framework was used to assess the relative contribution and mechanism of each original indicator in the system to CCD. The importance of each factor was ranked using the average SHAP value of the samples, quantifying the marginal contribution of each factor to the improvement of CCD. The SHAP analysis revealed the heterogeneity of regional driving factors, as shown in Figure 10.

In Region I, the dominant driving factors, ranked from high to low, are FTV, TG, and PTVK. The SHAP values for FTV are mainly distributed between 0 and 0.05, with the majority concentrated on the right side, indicating that the increase in freight volume consistently exerts a positive driving effect on CCD in this region. The SHAP values for TG are mainly distributed between 0.02 and 0.04, suggesting that the growth of transport industry output value contributes positively to CCD improvement by enhancing the economic efficiency of the transportation system. The SHAP values for PTVK are distributed on both sides, with a slightly higher concentration on the left, indicating that passenger turnover intensity exhibits a mixed and marginally negative influence on CCD in this region. In Region II, the dominant driving factors, ranked from high to low, are FTV, TEP, and TG. The SHAP values for FTV are entirely distributed on the right side, mainly concentrated between 0.04 and 0.05, showing that the freight volume in this region significantly drives CCD positively. The SHAP values for TEP are distributed on both sides, reflecting the bidirectional impact of transportation employment personnel on CCD in this region. The SHAP values for TG are entirely distributed on the right side between 0.03 and 0.05, indicating that the transport industry output value plays a stable and positive role in promoting CCD. In Region III, the dominant driving factors, ranked from high to low, are TG, FTVK, and PTV. The SHAP values for TG are broadly distributed between 0.01 and 0.04, with points dispersed on both sides, reflecting the highly heterogeneous and complex impact of transport industry output value on CCD in this region. The SHAP values for FTVK are mainly distributed around 0.03 on both sides, with a higher concentration on the left, indicating that freight turnover intensity predominantly exerts a negative driving effect on CCD in this region. The SHAP values for PTV are also distributed around 0.03 on both sides, with a higher concentration on the left and several notable positive outliers, showing that passenger volume exhibits a mixed but occasionally strong positive influence on CCD improvement. In Region IV, the dominant driving factors, ranked from high to low, are FTV, FTVK, and PTVK. The SHAP values for FTV are distributed on both sides, with a wider spread toward the negative direction, reflecting significant differentiation in the impact of freight volume on CCD in this region. The SHAP values for FTVK are mainly distributed between 0.01 and 0.03, with points relatively evenly scattered on both sides, indicating the complex and bidirectional influence of freight turnover intensity on CCD. The SHAP values for PTVK are similarly distributed on both sides, suggesting that passenger turnover intensity also exhibits substantial variability in its effect on CCD in this region.

Overall, the four regions exhibit differentiated driving mechanisms. Regions I and II maintain high and stable CCD levels through FTV, but the driving direction in Region II is more consistent, while Region I shows more driving differentiation. Region III focuses on TG as the core factor, yet its effect direction is highly heterogeneous. Region IV is dominated by FTV, FTVK, and PTVK, with highly differentiated effects from various factors, reflecting the complexity and instability of the driving mechanisms. This heterogeneity reveals the essential differences in the CCD improvement paths across regions, confirming the necessity of promoting TD–TCEI coordinated development according to local conditions. Due to the complex driving mechanisms of different indicator factors on CCD, the subsequent analysis will use dependency plots to further explore the specific roles of each factor in influencing CCD.

4.3.3. Single-Factor Importance Impact Analysis

To further analyze the specific impact of different factors, dependency plots were used to conduct an in-depth analysis of the top three influential indicator factors across all regions, as shown in Figure 11. The plots display the dependency relationship between the dominant factors for CCD and SHAP values in different regions. The data were fitted using the Generalized Additive Model (GAM), where the red curve represents the GAM fit, the pink area indicates the 90% confidence interval, and the bottom frequency histogram shows data density. The dependency plot analysis identifies three typical driving patterns as follows:

(1)

Threshold-Crossing Type: The impact of TG on CCD in Region III shows a typical “suppression-then-promotion-followed-by-stabilization” characteristic. The TG-SHAP curve shows that the SHAP values cross zero and turn positive when TG reaches approximately 0.2, indicating that a low level of transport industry output value initially has a significant negative impact on CCD. Once TG exceeds 0.3, the SHAP values stabilize at a positive level of 0.04, reflecting the emerging scale effect of transportation economic output. In Region IV, the FTV-SHAP curve exhibits the sharpest threshold transition, starting at approximately −0.125, crossing zero at around 0.15, and stabilizing at 0.02 when FTV exceeds 0.3, reflecting the critical importance of a minimum freight volume for CCD in underdeveloped transportation regions. Similar threshold effects were also observed in various indicators across different regions. For instance, the FTV-SHAP curve in Region I shows that the SHAP values turn positive when FTV reaches approximately 0.2, and stabilize at around 0.075 when FTV exceeds 0.5. In Region II, the SHAP value for TEP turns positive at approximately 0.35, and stabilizes at around 0.03 when TEP exceeds 0.55. The FTVK-SHAP curves in Regions III and IV, and the PTVK-SHAP curve in Region I exhibit analogous patterns with varying thresholds. In these scenarios, the driving effect on CCD follows a pattern of initial suppression, followed by threshold-crossing, and then diminishing marginal effects.

(2)

Continuous Promotion Type: In Region I, the impact of TG on CCD remains consistently positive, with SHAP values rising steadily from 0.02 to around 0.04, indicating that the transport industry output value in this region has reached a higher level, and its sustained growth continues to enhance CCD. In Region II, FTV shows a positive upward trend, with SHAP values rising from 0.015 to between 0.05 and 0.06, indicating that freight volume has reached a scale economy effect and continues to drive CCD upward. The TG-SHAP curve in Region II exhibits a similar consistently positive pattern. This pattern reflects how high-coordination regions maintain a stable CCD promotion mechanism by leveraging existing advantages.

(3)

U-Shaped Crossing Type: In Region III, the impact of PTV on CCD follows a U-shaped curve of “promotion-then-decline-and-tends-to-zero,” with SHAP values starting at approximately 0.15, crossing zero when PTV reaches 0.05, and gradually stabilizing near zero when PTV exceeds 0.5, reflecting that the initial growth of passenger volume, which outpaced infrastructure supply, led to a rapid attenuation of its positive contribution to CCD. In Region IV, the PTVK-SHAP curve exhibits a more complex U-shaped pattern of “promotion-then-decline-followed-by-recovery”: the SHAP values start at approximately 0.04, cross zero, and turn negative when PTVK reaches 0.05; then recover and cross zero again when PTVK reaches approximately 0.3; and continues to rise to around 0.03 when PTVK exceeds 0.45. This reflects that as passenger turnover intensity initially outpaced supply capacity, the positive effect temporarily reversed, but eventually recovered as supply–demand balance was restored. This pattern reveals the supply–demand imbalance phase that low-coordination regions may experience during their development.

Overall, the core driving factors for CCD, ranked from high to low, are FTV, TG, and PTVK. The dependency plot analysis also reveals three core patterns: FTV and TG exhibit a critical threshold effect across most regions, where the initial impact on CCD may be negative, but once the threshold is crossed, the effect turns positive and gradually stabilizes. This threshold varies by region, with Regions III and IV exhibiting sharper transitions from deeply negative initial values, while Region I shows a more gradual transition. PTVK exhibits U-shaped adjustment characteristics in low-coordination regions, reflecting the supply–demand imbalance phase during development. In high-coordination Regions I and II, the dominant factors have reached a steady state or scale economy range, and the marginal effect tends to stabilize. In low-coordination Regions III and IV, the dominant factors are still in the threshold-crossing or U-shaped adjustment phase, reflecting the complexity and uncertainty of their CCD improvement process. These nonlinear relationships provide scientific evidence for the formulation of differentiated regional policies. High-coordination regions should focus on optimizing existing advantages while seeking new drivers, whereas low-coordination regions need to accelerate crossing the threshold while balancing the pace of investment to avoid supply–demand imbalances.

4.4. Limitations

This study still has several limitations that require improvement in future research. First, regarding the indicator system, the current TD evaluation framework mainly focuses on internal factors of the transportation system. Future research should construct a comprehensive ecological impact index, integrating multi-dimensional indicators such as carbon emissions, air quality, and noise pollution, to achieve a coordinated assessment of transportation development and ecological protection. Second, given that this study heavily relies on various construction methods and derived indicators, future work should incorporate global sensitivity analysis [53] to examine the robustness of the reported results with respect to key modeling choices. Deep learning can be integrated with variance-based multi-method global sensitivity analysis (GSA) in transportation network evaluation [54]. Future, GSA can also be applied in conjunction with the XGBoost framework to enhance the reliability of the overall model.

Third, this study is based on provincial-level data and lacks a micro-level perspective, such as enterprise competition, residents’ travel equity, and welfare levels. Future CCD studies could incorporate micro-level data to enhance the social inclusiveness and sustainability of the dual carbon strategy. Finally, regarding the driving mechanism, although this study identifies key driving factors and their nonlinear effects, the interaction effects and transmission pathways between variables have not been fully explored. Future research could employ structural equation modeling or mediation effect analysis to reveal the complex impact mechanisms.

5. Conclusions and Recommendations

5.1. Research Conclusions

This study, based on the “economic-carbon” two-dimensional classification framework, used the entropy weight TOPSIS method, coupling coordination degree (CCD) model, kernel density estimation (KDE), spatial autocorrelation analysis, and XGBoost-SHAP machine learning framework to systematically analyze the coordinated evolution of transportation system performance (TD) and carbon emission intensity (TCEI) across 31 provinces in China from 2014 to 2023. The main findings are as follows:

First, the transportation low-carbon system exhibits non-synchronous evolutionary characteristics of “performance fluctuations, steady emission reductions.” In terms of time, TD shows an N-shaped fluctuation, significantly impacted by the pandemic and policy adjustments. TCEI continuously declined, demonstrating clear evidence of the success of the low-carbon transformation. In terms of space, TD shows a gradient differentiation of “high in the east, low in the west, and coastal regions outperforming inland.” TCEI shows a gradient characteristic of “low in the east, high in the west, and low in the south, high in the north”.

Second, CCD exhibits a three-dimensional evolutionary pattern of “temporal N-shaped evolution, spatial gradient convergence, and internal heterogeneity and differentiation.” In terms of time, during the study period, CCD overall moved from primary coordination to good coordination, experiencing three phases: steady improvement, high-level fluctuations, and impact recovery, following an N-shaped trajectory. Kernel density analysis revealed that the polarization phenomenon in CCD has alleviated, with low-coordination provinces gradually improving, but regional internal heterogeneity continues to be prominent. In terms of space, a multi-level coordinated development pattern was formed with “Region I leading, Region II breaking through, Region III maintaining, Region IV catching up”.

Third, the driving mechanism exhibits characteristics of “triple core, nonlinear threshold, and regional heterogeneity.” The core driving factors for CCD are FTV, TG, and PTVK. These factors exhibit significant threshold effects and regional heterogeneity. The nonlinear mechanisms reveal that different regions need to adopt tailored strategies to cross thresholds and achieve CCD improvement based on their specific stage of development.

In conclusion, this study provides a systematic analytical framework for regional transportation green transformation, revealing the spatiotemporal patterns and driving mechanisms of TD-TCEI coordinated development. It offers important insights for achieving the transportation powerhouse strategy under the dual carbon goals.

5.2. Recommendations

First, a multi-stakeholder collaborative governance system should be built, and cross-departmental coordination should be strengthened. Guided by the transportation powerhouse strategy and the dual carbon goals, differentiated policies should be formulated by region and stage, with a cross-departmental coordination mechanism established. Local governments should optimize planning and industrial layout based on local economic-carbon characteristics, while transportation enterprises should actively participate in carbon trading and improve transportation infrastructure networks.

In addition, regional collaborative linkages should be strengthened to narrow the CCD spatial gap. On one hand, leverage the demonstration effect of high CCD regions, where they can export intelligent transportation solutions and multimodal transport models, and share green transformation and “shift from road to rail” implementation experiences with regions that lack them. On the other hand, establish a cross-regional coordination mechanism by forming a transportation low-carbon alliance in key regions and setting up a green cooperation fund to support infrastructure interconnection. Additionally, promote transportation document standardization and facilitate customs clearance, explore cross-regional carbon emission trading, and optimize resource allocation through market-based methods.

Finally, policies should be implemented based on the key determining factors of CCD coordination. The threshold effects and U-shaped adjustment patterns revealed in the dependency plot analysis provide precise targets for policy design. Regions should adjust their development direction based on the temporal evolution of CCD and progress in crossing thresholds. High-economic low-carbon regions should maintain their existing scale economy advantages in freight volume and transport industry output value while optimizing passenger turnover configuration to sustain high-level coordination. High-economic high-carbon regions should focus on enhancing freight transportation efficiency through “shift from road to rail” and multimodal transport, gradually moving towards greener development. Low-economic low-carbon regions should accelerate the growth of transport industry output value to cross the critical threshold, particularly focusing on the development of transportation electrification. Low-economic high-carbon regions should prioritize building minimum freight volume capacity to cross the threshold as quickly as possible, while carefully managing the pace of passenger turnover expansion to avoid supply–demand imbalances.