Carbon emission data were derived from the CO₂ global gridded dataset in the EDGAR_2024_GHG greenhouse gas database, jointly released by the IEA and EDGAR (Johansson et al., 2017). First, the original 1 ° × 1 ° resolution data were upscaled to 1 km × 1 km using R, and then clipped to the territorial extent of China. Subsequently, the dataset was resampled to a 500 m × 500 m resolution in ArcGIS 10.8 to enhance spatial precision. Vector boundaries and attribute information of national HTZs were extracted via the AMap (Gaode) API using Python, and spatially overlaid with the resampled raster data. Annual total carbon emissions were then extracted for each zone, constructing a continuous time series covering 2008–2023. To ensure high boundary-matching accuracy, a dual “grid–administrative boundary” verification procedure was employed, complemented by alignment with township and subdistrict boundaries.

The Super Slack-Based Measure (Super-SBM) model (Tone, 2002) is an important extension of Data Envelopment Analysis (DEA). It overcomes the limitations of traditional radial models, which neglect slack variables and are unable to differentiate among efficient decision-making units with efficiency scores greater than 1. By directly optimizing input excesses and output shortfalls, the Super-SBM model improves the precision of efficiency measurement. It also incorporates undesirable outputs—such as carbon emissions—into the analysis, quantifying their excess levels and providing a more comprehensive reflection of environmental efficiency (Amirteimoori et al., 2024).

This study assesses the carbon emission efficiency of national HTZs using a non-oriented Super-SBM model with undesired outputs under the assumption of NIRS. The non-oriented setting simultaneously optimizes inputs and outputs, which aligns with the multi-objective characteristics of real production activities. The NIRS constraint prevents misinterpreting efficiency improvements as scale expansion, making it particularly suitable for contexts in which carbon emission efficiency improvements are driven primarily by technological progress rather than scale effects. This model is well-suited to assessing environmental efficiency within a multi-input–multi-output framework and provides a more accurate measurement of resource utilization and green performance. The model is expressed as follows (Equation 1): ρ * = min 1 + 1 m ∑ i = 1 m s i − / x i k 1 − 1 s 1 + s 2 ∑ r = 1 q 1 s r + / y r k + ∑ t = 1 q 2 s t b − / b r k s . t . ∑ j = 1 , j ≠ k n x i j λ j − s i − = x i k ∑ j = 1 , j ≠ k n y r j λ j + s i + = y r k ∑ j = 1 , j ≠ k n b t j λ j + s t b − = b t k λ j ≥ 0 , ∑ j = 1 , j ≠ k n λ j ≤ 1 s − , s + , s b − ≥ 0 i = 1 , 2 , … , m ; r = 1 , 2 , … , q ; j = 1 , 2 , … , n j ≠ k (1)

In a system with n Decision Making Units D M U S , the efficiency value is given by ρ ρ > 0 . Define the k -th D M U as D M U k k = 1 , 2 , … , n . Every D M U employs m inputs x i k i = 1 , 2 … , m to generate q 1 desirable outputs y r k r = 1 , 2 … , q 1 , and q 2 unwanted outputs b r k r = 1 , 2 … , q 2 .

Use s i − , s i + , and s t b − to express slack variables for inputs, desirable outputs, and unpleasant outputs. The linear combination coefficients λ j are given and restricted to be λ j ≥ 0 .

The restriction of non-increasing returns to scale (NIRS) prevents irrational expansion by assuming that inputs do not enhance outputs proportionally. This condition is Equation 2: ∑ j = 1 , j ≠ k n λ j ≤ 1 (2)

Geospatial data’s spatial autocorrelation is tested using Moran’s I, which has values between −1 and 1 (Moran, 1950). Positive spatial autocorrelation is shown when Moran’s I > 0, which means that nearby regions have comparable characteristics and create “high–high” or “low–low” clusters. Negative spatial autocorrelation is shown when Moran’s I is less than zero, suggesting notable regional differences with “high–low” or “low–high” distributed patterns. A random spatial distribution with negligible spatial correlation is suggested by Moran’s I ≈ 0. The global Moran’s I statistic is calculated as follows (Equation 3): M oran ′ s   I = ∑ i = 1 n ∑ j = 1 n ω i j x i − x ¯ x j − x ¯ S 2 ∑ i = 1 n ∑ j = 1 n ω i j (3)

Let n denotes the denote the count of HTZs; x i , x j are efficiency observations for zones i and j ; x ¯ is the sample mean; S 2 is the overall variance; and ω i j is the spatial weight matrix element.

Assign n to the number of HTZs, x i , x j to efficiency observations for zones i and j , x ¯ to the sample mean, S 2 to the overall variance, and ω i j to the spatial weight matrix element.

The weights are based on economic–geographic distance (Equation 4). ω i j = G I O i − G I O j × 1 d i j 2 (4)

Let G I O i and G I O j represent gross industrial outputs for zones i and j , and d i j represent their geographic distance.

Analysis of localized spatial autocorrelation uses Local Moran’s I (Equation 5): M oran ′ s   I i = x i − x ¯ S 2 ∑ j = 1 n ω i j x j − x ¯ (5) I i measures zone i’s local spatial relationship with its neighbors.

The Geographical Detector is a statistical method used to identify spatial stratified heterogeneity and detect driving factors. Its core idea is that if an independent variable has a significant impact on a dependent variable, the two should exhibit similar spatial distribution patterns (Wang et al., 2010). In this study, the explanatory power of each driving factor on the carbon emission efficiency of national HTZs is evaluated using the determinant power (q-value) of the Geographical Detector. The q-value ranges from 0 to 1, with larger values indicating stronger explanatory power. A q-value of 0 suggests no spatial association between the two distributions, while a q-value of 1 indicates complete spatial consistency. The calculation is given as follows (Equation 6): q = 1 − 1 N δ 2 ∑ h = 1 L N h · δ h 2 (6) L counts subregions, N h the sample size in region h ; δ h 2 he dependent variable’s within-region variance, N the total sample size, and δ 2 the study area’s overall variance. The q value quantifies how much each driving component affects carbon emission efficiency spatially.

The Multi-Scale Geographically and Temporally Weighted Regression (MGTWR) model is an extension of the GTWR model that incorporates a multi-scale weighting mechanism across both spatial and temporal dimensions. This allows for a more precise characterization of spatiotemporal dependence and non-stationarity (Wu et al., 2019). While GTWR simultaneously accounts for spatial and temporal effects, it applies a single spatiotemporal bandwidth to all variables (Fotheringham et al., 2015), which limits its ability to capture the multi-scale differences among various driving factors. To address this limitation, MGTWR assigns independent bandwidth parameters to each variable in both the spatial and temporal dimensions, thereby enhancing the model’s capacity to describe and explain spatiotemporal heterogeneity. The model can be expressed as follows (Equation 7): Y i = β 0 u i , v i , t i + ∑ k = 1 p β k u i , v i , t i ; h k s , h k t · x i k + ε i (7) Y i represents carbon emission efficiency at location i and time t i , β 0 u i , v i , t i is the local intercept, β k u i , v i , t i ; h k s , h k t is the spatiotemporal coefficient of the k -th driver, x i k is the observed value of the k -th variable; and ε i is the random error.

Based on the assumption of NIRS, this study statistically assesses the carbon emission efficiency of 178 national HTZs between 2008 and 2023 using a non-oriented Super-SBM model with unwanted outputs. Their low-carbon transition features are methodically examined from three angles: regional disparities, park-level efficiency ratings, and industrial combinations.

Overall, the carbon emission efficiency of national HTZs has evolved through four stages: quick improvement, volatile adjustment, steady development, and short-term fall (Figure 1). In 2009, efficiency reached a temporary peak under the combined influence of the global financial crisis and China’s domestic demand expansion policies. Between 2011 and 2014, the expansion of energy-intensive industries and increasing regional development disparities led to a significant decline in efficiency. In 2016, the initial implementation of carbon market pilot programs began to take effect, driving a gradual recovery in efficiency. From 2017 to 2020, guided by the “carbon peaking” target, continued investment in clean energy and industrial restructuring sustained efficiency growth. Although the COVID-19 pandemic posed challenges in 2020, the digital economy helped maintain efficiency at a relatively high level. Between 2021 and 2023, efficiency declined to approximately 0.84, primarily due to a rebound in industrial carbon intensity, rising carbon investment elasticity, and shortages of hydropower.

Building on this foundation, regional comparison results (Figure 2) reveal a clear “higher in the south, lower in the west” spatial pattern of carbon emission efficiency across the eight integrated economic regions. The Southern Coastal region exhibits the highest efficiency (0.878), followed by the Northeast (0.863) and the Middle Reaches of the Yangtze River (0.862), which show similar levels. The Eastern Coastal (0.856), Southwest (0.850), and Northern Coastal (0.849) regions fall into the intermediate range, while the Yellow River Midstream (0.800) and Northwest (0.772) regions display relatively low efficiency levels. To further uncover intra-regional characteristics, the following section examines carbon emission efficiency from the perspective of park-level ratings.

In order to describe national HTZs' carbon emission efficiency levels across different regions, a rating system was established based on average efficiency values: Zones achieving an average efficiency ≥1.0 are designated “Excellent”, those in the range of 0.9–1.0 and 0.7–0.9 are considered “Good” and “Average”, respectively, while those below 0.7 are regarded as “Poor.” Figure 3 illustrates the spatial distribution and rating results of national HTZs.

In the Southern Coastal Economic Region (Fujian, Guangdong, Hainan), which comprises 22 national HTZs, 50% are rated “Excellent” or “Good.” Putian (1.062), Sanming (1.033), and Zhanjiang (1.010) perform particularly well. Leveraging locational, policy, and technological advantages, this region has achieved synergy between ecological and economic benefits. Parks have pursued diversified green transition pathways through industrial decarbonization, green finance, and technological innovation.

In the Northeast Economic Region (Liaoning, Jilin, Heilongjiang), 31.3% of the 16 national HTZs are rated “Excellent” or “Good.” Changchun (1.083), Yanji (1.065), and Daqing (1.050) lead the region. Efficiency improvements have been achieved through smart energy management, CCUS applications, and cleaner production, with technological upgrading and resource recycling emerging as key drivers.

In the Middle Reaches of the Yangtze River Economic Region (Hubei, Hunan, Jiangxi, Anhui), 37.6% of the 38 national HTZs fall into the “Excellent” or “Good” categories. Yichun promotes green production through its full lithium battery industry chain, while Xiaogan leverages 5G industrial internet to increase the share of green electricity. However, parks such as Jiujiang and Xianning are constrained by a single energy structure and low industrial energy efficiency.

In the Southwest Economic Region (Yunnan, Guizhou, Sichuan, Guangxi, Chongqing), 33.3% of the 22 national HTZs fall into the “Excellent” or “Good” categories. Yuxi has built a zero-carbon park, and Guiyang has implemented green access standards, both significantly reducing emissions. However, Panzhihua’s vanadium–titanium industry faces technological constraints limiting its mitigation potential, and Beihai lacks a well-developed circular system. Overall, the region’s green development foundation remains weak, requiring accelerated technological upgrading and structural optimization.

In the Northern Coastal Economic Region (Beijing–Tianjin–Hebei, Shandong), 25% of the 20 national HTZs achieve “Excellent” or “Good” ratings. Zhongguancun in Beijing has taken the lead through zero-carbon park initiatives and AI-based energy management. The Yellow River Delta has optimized its energy structure via carbon capture and integrated “source–grid–load–storage” systems. The regional promotion of “AI + energy” models and expansion of blue carbon pilot projects are contributing to coordinated emission reduction.

In the Eastern Coastal Economic Region (Shanghai, Jiangsu, Zhejiang), 16.7% of the 24 national HTZs fall into the “Excellent” or “Good” categories. Zhangjiang has improved efficiency through waste heat recovery, photovoltaic power generation, and ground-source heat pumps, while Xiaoshan Linjiang has advanced a “materials–manufacturing–recycling” closed-loop transformation. These efforts have produced notable results in energy optimization and carbon reduction, with strong demonstration effects.

In the Yellow River Midstream Economic Region (Shaanxi, Shanxi, Henan, Inner Mongolia), none of the 21 national HTZs are rated “Excellent,” and only Ankang is rated “Good.” The region’s industrial structure is dominated by coal, chemicals, and metallurgy, leading to high carbon intensity. While Xi’an and Yulin have begun developing full hydrogen and photovoltaic industry chains, and some parks are introducing carbon capture technologies with demonstration potential, the region faces pronounced issues of industrial homogeneity and resource misallocation.

In the Northwest Economic Region (Gansu, Qinghai, Ningxia, Tibet, Xinjiang), all 11 national HTZs fall into the “Medium” category. Despite increases in renewable energy capacity, multiple constraints—including grid regulation limitations, the concentration of heavy industries, coal dependence, technological lag, and insufficient policy support—have hindered efficiency improvements.

To analyze the impact of dominant industries on carbon emission efficiency, this study identifies the two industries with the highest shares of total industrial output in each national HTZs, based on the 2018 revised Catalogue for the Review of Chinese Development Zones issued by six ministries and industry data from the official websites of each park. These dominant industries are then used to construct an industrial combination–carbon emission efficiency matrix (Figure 4). The results reveal substantial differences in efficiency across industrial combinations: high-efficiency combinations exhibit synergistic effects, whereas low-efficiency combinations face significant risks of efficiency loss.

Industrial combinations led by the new energy sector achieve the highest efficiency levels. The combination of new energy and energy-saving industries reaches an efficiency score of 1.018. Combinations of new energy with high-tech services (0.971) and new materials (0.895) also perform well. These combinations are typically supported by well-developed green infrastructure and highly integrated industrial chains, with extensive applications in photovoltaics, energy storage, and carbon nanomaterials, resulting in outstanding resource use efficiency.

Several “strong–weak” industrial combinations also show considerable improvement potential. The electronic information industry alone has an efficiency of 0.908, but its combination with biopharmaceuticals yields an efficiency of 0.882, reflecting the role of intelligent manufacturing technologies in improving energy efficiency within the pharmaceutical sector. The combination of advanced manufacturing and aerospace reaches an efficiency of 0.932, indicating that precision manufacturing offers energy-saving advantages in complex processing stages. Such combinations generally feature complete chains from R&D to application, demonstrating strong industrial linkage effects.

In contrast, low-efficiency combinations reflect issues of structural mismatch and insufficient integration. The combination of light industry and new materials achieves an efficiency of only 0.785, while the combination of high-tech services and electronic information is even lower at 0.739. These results expose weaknesses such as high energy consumption in traditional manufacturing, inadequate technological integration, and severe resource redundancy. When light-asset services and heavy-asset manufacturing fail to connect effectively, synergistic effects are difficult to achieve, leading instead to increased energy consumption and emissions. The efficiency of the combination of high-tech services and aerospace is only 0.769, indicating a pronounced mismatch between their production chains and demand structures; the low degree of industrial linkage makes it difficult to achieve effective synergy.

Overall, high-efficiency industries exert a positive spillover effect, but the magnitude of this improvement is bounded. Conversely, combinations led by low-efficiency industries significantly drag down overall efficiency, indicating that both the technological characteristics of dominant industries and their capacity for coordination with supporting industries jointly determine overall carbon efficiency in HTZs.