Diverse built environment pathways for bridging global ambitions with local initiatives in sustainable transportation transitions

Abstract

While global transportation accounts for 23% of energy-related CO₂ emissions, current decarbonization strategies often neglect the transformative potential of local built environment design. We investigate how micro-level built environment improvements contribute to macro-level transitions in China’s transportation sector by integrating the Global Change Analysis Model with high-resolution (1 km) spatial analysis. We demonstrate that localized actions targeting density, diversity, design, destination accessibility, and distance to transit can reduce the structural transportation service demand potential for four-wheeled light-duty vehicles (TSD-LDV4W) by 21% nationally by 2060, approaching the ambitious SSP1 sustainability scenario. Interprovincial and urban-rural analyses reveal significant regional disparities in transition pathways driven by variations in economic conditions, technological capabilities, and geographical contexts, underscoring the pivotal role of meso-level institutional and structural characteristics in steering sustainability transitions. Machine learning quantitatively characterizes stark differences between historic pedestrian-centric cores and newly developed car-oriented districts, crucial for China’s transition from rapid urban expansion to urban renewal. Our findings demonstrate that neighborhood-scale interventions, including targeted infrastructure changes and local policy implementations, can substantially advance macro-level climate goals. These results suggest that context-sensitive strategies tailored to diverse development trajectories are essential for achieving low-carbon transportation transformation.

Introduction

The United Nations Sustainable Development Goals (SDGs) highlight the importance of resilient infrastructure (Goal 9) and sustainable transport systems (Goal 11) in aligning low-carbon transport transitions with regional development. While transportation accounts for 23% of energy-related CO₂ emissions globally¹, decarbonization strategies often fail to effectively incorporate the transformative potential of local built environment design. By 2050, passenger demand will increase by 79% due to population growth and urbanization², which underscores the urgent need to tackle emissions across the entire spectrum of vehicle-related activities, from daily operation to production-related environmental impacts^3,4. As the automotive landscape undergoes a profound transformation, with vehicles increasingly electrified and power generation decarbonizing, a critical shift in the emissions profile is emerging. While fuel cycle emissions are on a downward trajectory, the production of materials and manufacturing processes is set to account for an increasingly significant proportion of overall emissions³. What is especially noteworthy is that the consumption rebound triggered by the improvement in production efficiency has led to a substantial increase in the consumption of automotive products^5,6. This dynamic highlights the intricate web of challenges and opportunities in pursuit of sustainable transportation, characterized by the multi-level perspective (MLP) of transitions, which is deeply embedded within the complex interplay between technological, social, and institutional factors driving systemic change^7,8,9.

Existing studies employ Integrated Assessment Models (IAMs) to forecast carbon emission trends resulting from changes in transportation technologies, primarily taking into account the influences of population and the level of economic development^10,11,12. However, a problematic assumption in most of these studies is that developing countries will follow the well-worn “automobile society” path that developed countries have taken. They anticipate that economic growth in the Global South will inevitably trigger a surge in automotive transportation demand¹³. This assumption contradicts the very essence of a “sustainable transition”, which is fundamentally about overhauling the meso-level institutional and structural framework to forge entirely new macro-level landscapes. Given the pressing need to reduce carbon emissions in the transport sector, a striking disparity emerges.

Even countries with comparable per capita GDPs, like Denmark and the United States, exhibit significant differences in car ownership rates. This divergence can be attributed to their distinct transition trajectories¹⁴. This pattern extends beyond the Denmark-U.S. comparison. European cities such as Amsterdam and Copenhagen have successfully reduced car dependency through comprehensive built environment interventions^15,16, achieving modal shares for cycling and public transport that exceed 50% of all trips, despite having per capita GDPs comparable to car-dependent cities in North America and Australia¹⁷. This demonstrates how compact urban form and transit-oriented planning can maintain low levels of private car dependency. Similarly, emerging economies in the Global South demonstrate diverse trajectories: while some cities like Bogotá and Curitiba have prioritized bus rapid transit and cycling infrastructure despite lower per capita incomes^18,19, illustrating how transit-led restructuring can transform travel behavior at scale²⁰. Others have followed car-oriented development patterns similar to those of developed countries in earlier decades. These international examples underscore that transportation demand patterns are not predetermined by economic development levels alone, but are fundamentally shaped by policy choices and built environment characteristics—a principle that is particularly relevant for China’s ongoing urbanization and urban renewal processes. The design of the built environment plays a decisive role in shaping travel patterns across different socioeconomic contexts. It highlights the varied strategies countries adopted when developing transportation infrastructure to fulfill service requirements. It is a clear indication that not all transport demands need to be satisfied by private cars. In fact, the structure of travel choices is heavily shaped by the built environment^21,22.

The relationship between built environment and transportation systems, especially Vehicle Miles Traveled (VMT), has been extensively studied at the city level^23,24. By enhancing the quality and characteristics of the built environment, cities can achieve a reduction in urban transportation demand^23,24,25,26. This, in turn, leads to a decline in private car ownership and usage, ultimately resulting in lower vehicle kilometers traveled (VKT) and a corresponding decrease in carbon emissions^27,28. The “5D” framework—comprising Density, Diversity, Design, Destination Accessibility, and Distance to Transit—introduced by Cervero, Kockelman, and Ewing in 2001^29,30, has gained prominence as a vital analytical tool for understanding how built environments influence transport patterns. However, while the significance of micro-level built environment is widely recognized, macro-level scenario analyses generally fall short in providing specific implementation pathways. Some studies adopt an overly narrow focus on the micro-scale, failing to establish a clear link between localized urban actions and their broader contributions to macro-level climate and sustainability goals³¹. In essence, the process of “translating” the global vision into actionable measures at the local scale remains unclear³².

Taking China as an example, this paper aims to integrate the impact of the micro-level built environment characterized by the “5D” dimensions at the scale of one kilometer grid, with IAMs for macro-level travel demand forecasting, drawing upon the theoretical framework of the Multi-Level Perspective (MLP) of transitions. The research seeks to address two pivotal questions:

Firstly, to what degree can micro-level actions propel the realization of macro-scale sustainability visions within the transportation domain? This question aims to probe the potency of localized initiatives in shaping and advancing broader sustainability goals, and to ascertain the extent of their influence on large-scale systemic changes.

Secondly, what is the magnitude of variation that may emerge in the actualization of these sustainability visions, given the regional disparities in meso-level regimes and policy institutions? This question recognizes the significant role that urban and regional contexts play in the transition process and endeavors to quantify the potential deviations in outcomes that could arise due to differences in institutional frameworks, policy approaches, and socio-economic conditions at the meso-level.

Our study makes several methodological and empirical contributions to the literature on transportation sustainability transitions. First, we develop an innovative analytical framework that integrates macro-level Integrated Assessment Models with high-resolution (1 km) spatial analysis of built environments—bridging a persistent gap between global climate scenario modeling and local urban planning practice. Second, we apply this framework at an unprecedented national scale, analyzing over 9.5 million grid cells across China’s diverse geographic and socio-economic contexts, moving beyond the city-level case studies that dominate existing research. Third, we provide the first quantitative assessment of how micro-level built environment improvements can contribute to achieving ambitious macro-level climate goals, demonstrating that local actions can yield a significant reduction in private vehicle travel demand. Finally, we employ machine learning to project long-term transitions between pedestrian-friendly and car-centric urban forms, identifying the specific built environment characteristics that either support or hinder sustainable mobility patterns—insights directly applicable to China’s ongoing urban renewal initiatives and relevant to other rapidly urbanizing nations.

Methods

Analytical framework

Our analytical framework integrates macro-level Integrated Assessment Models with micro-level spatial analysis to examine how built environment improvements contribute to transportation demand reduction. Figure 1 presents a comprehensive workflow diagram illustrating our methodological approach, which consists of four main components: (1) Step 1: Macro-level Scenario Analysis. We employ the Global Change Analysis Model (GCAM v7.1) to simulate national-level TSD-LDV4W under five Shared Socioeconomic Pathway (SSP) scenarios from 2015 to 2060, incorporating ambitious electric vehicle deployment targets. We then use GCAM-China (v6.0) to disaggregate these projections to 31 provincial-level administrative regions, capturing regional heterogeneity in transportation transitions. (2) Step 2: High-resolution Built Environment Characterization. We construct a comprehensive 1 km × 1 km gridded dataset covering mainland China, characterizing five dimensions of the built environment (5D framework): Density (population), Diversity (land use mix), Design (road network), Destination Accessibility (distance to city centers), and Distance to Transit (transit station density). Historical data (2014–2020) and projections (2025–2060) are compiled from multiple authoritative sources (detailed in Table 1). (3) Step 3: Spatial Allocation and Temporal Prediction of TSD-LDV4W. We develop the Built Environment Advantage Index (BEAI) using established elasticity coefficients from meta-analytical research to spatially allocate the 2015 baseline TSD-LDV4W across grid cells. Machine learning models incorporating 5D indicators and spatial proximity effects then predict future TSD-LDV4W for each grid cell, accounting for temporal improvements in built environment characteristics. (4) Step 4: Urban Typology Analysis. We classify grid cells into four area types based on historical urban development patterns: pedestrian-friendly districts (pre-1995 urban cores), car-centric districts (post-2005 developments), transition zones, and other non-urban areas. A Random Forest classification model trained on 2015 data—using 5D indicators and economic variables as predictors—projects future transitions between area types, revealing how urban renewal and expansion patterns influence transportation sustainability. Table 1 summarizes all datasets employed in this study.

Fig. 1: Analytical framework of this study.

This figure illustrates the four steps of our analysis framework. Firstly, a macro-level scenario analysis is conducted using the Global Change Analysis Model (GCAM). Then, a high spatial resolution characterization process of the built environment is carried out through GIS analysis. Subsequently, spatial allocation and temporal prediction of Transportation Service Demand for Four-wheeled Light-duty Vehicles in Passenger Transportation Sector (TSD-LDV4W) are performed. Finally, an analysis of urban typology is carried out.

Table 1 Summary of datasets used in this study

Calculation of TSD-LDV4W at national and provincial scales

TSD-LDV4W was calculated using the modules within the IAMs. IAMs are important tools for quantitatively analyzing key processes and interactions in human socio-economic activities and global resource-environment systems¹⁰. They have been used in IPCC reports to provide policy insights for mitigating global climate change and promoting sustainable development³³. A large number of studies use IAMs to analyze energy consumption and greenhouse gas emissions in the transportation sector^11,12. We used one of the widely used IAMs, GCAM version 7.1, and the newly released GCAM-China version 6.0 to calculate the TSD-LDV4W at the national level and across the 31 provincial-level administrative regions of mainland China. Hong Kong, Macau, and Taiwan are excluded from this analysis due to data availability constraints and their distinct administrative and statistical systems. GCAM covers 32 regions globally, with a modeling timeframe extending from 2015 to 2100, using a 5-year time step. GCAM is a partial equilibrium, dynamic recursive model that connects the global economy, energy system, water, agriculture, and land-use systems with a reduced-form climate model. Assumptions for the model’s components are outlined in the GCAM documentation.

This study focuses specifically on passenger transportation demand for four-wheeled light-duty vehicles (TSD-LDV4W), deliberately excluding freight transport, for several reasons. First, passenger travel behavior is directly influenced by the 5D built environment characteristics that affect individuals’ daily activity patterns, mode choices, and trip generation—the mechanisms central to our analytical framework. Second, private passenger vehicles represent the fastest-growing and most policy-relevant transport segment in China’s ongoing urbanization and motorization process, and are the primary target of demand management strategies aimed at reducing car dependency. Third, freight transport operates through different spatial logics that respond to policy levers (industrial zoning, logistics hub placement, freight rail investment) distinct from the 5D urban form factors we examine.

The passenger sector consists of five hierarchical levels, corresponding to different modes (e.g., road, rail), sub-modes (e.g., bus, light-duty vehicle), size classes, and drivetrain technologies. Their market share in future periods largely depends on income, prices, elasticities, and the time value of transportation. In GCAM, the transportation service demand (TSD) (e.g., passenger-km for passenger service, and tonne-km for freight service) in region $R$ and modeling period $T$ is estimated by the following Eq. (1):

$${{TSD}}_{R,T}=\,{{TSD}}_{R,\,T-1}{\left(\frac{{Y}_{R,\,T}}{{Y}_{R,T-1}}\right)}^{\alpha }{\left(\frac{{P}_{R,\,T}}{{P}_{R,T-1}}\right)}^{\beta }\left(\frac{{N}_{R,\,T}}{{N}_{R,T-1}}\right)$$

(1)

Where $Y$ represents per-capita GDP in region $R$ and modeling period $T$, $P$ presents the total service price aggregated across all transportation modes (e.g., road, rail). $N$ is the population, and $\alpha$ and $\beta$ are hyperparameters representing the income and price elasticities, respectively. Population enters with an implicit elasticity of 1.0, meaning transportation demand scales proportionally with population when other factors are held constant—a standard assumption in transportation demand modeling reflecting that more people generate proportionally more travel, absent changes in per-capita behavior. The methods for calculating transportation subsector competition and transportation technology costs can be found at http://jgcri.github.io/gcam-doc/demand_energy.html#inputs-to-the-module.

This study uses 2015 as the baseline year because it is GCAM (version 7.1)‘s final calibration year, for which complete, validated historical data are available across all economic sectors, energy systems, and regions globally. All GCAM simulations initialize from this comprehensively calibrated 2015 baseline, ensuring consistency between our analysis and the broader integrated assessment modeling literature. Projections then extend from 2015 to 2060 in 5-year time steps, covering the period relevant to China’s carbon neutrality commitment (2060) and allowing sufficient time for urban form transitions. Given the extensive promotion and deployment of electric vehicles (EVs) since 2015, we imposed restrictions on the share of new electric vehicle sales while keeping all other model assumptions unchanged. In GCAM, a “policy-portfolio-standard” can include three types of policies (tax, subsidy, RES), as specified by the “policyType” tag. Following the work of Ou et al.³⁴, this study also adopts the RES policy type, where RES specifies a renewable energy standard, that is, specifying a share of a sectoral output. Ou et al.³⁴ explain the functioning of the RES policy type. The method for solving minimum percentage constraints and intensity targets in GCAM is to credit technologies that meet or exceed the target, while not crediting or penalizing those that fall short. The RES code setting follows the example of the Energy Intensity Standard in the GCAM documentation (http://jgcri.github.io/gcam-doc/policies_examples.html#res).

Using real-world data from the China Automobile Dealers Association’s China Passenger Cars Association on the penetration rate of new energy vehicles (including EVs, hybrid vehicles, etc.) for 2023³⁵, we approximately estimate the share of EVs in China’s new car sales market in 2025. We have also set an ambitious growth trajectory for EVs, targeting 70% EV sales by 2060. This trajectory follows a pattern of rapid growth followed by slower growth (the specific setting of the EV sale target is provided in Supplementary Table S1). This ambitious EV growth scenario serves as the benchmark for comparing with the local effects of micro-level built environment improvements. Based on the transportation modes and vehicle type classifications in GCAM, we made assumptions about annual travel per vehicle and load factors for three vehicle types: Car, Large Car and Truck, and Mini Car. The specific values assumed are provided in Supplementary Tables S2, S3.

The global GCAM treats China as a single regional unit, limiting the ability to conduct subnational analysis. Therefore, we used the newly released GCAM-China to analyze provincial-level heterogeneity. GCAM-China disaggregates the energy-economic system of China into 31 provincial-level sub-regions and six electricity grid regions, which are also integrated into the global GCAM. In GCAM-China, electricity generation and end-use energy demand (for buildings, transportation, and industry) are modeled at the provincial level. To control for other factors and variations, and to facilitate comparison between the national and provincial levels, we applied the same EV sale targets across all provinces as those used at the national level.

Calculation of high-resolution 5D indicators

Given data availability and temporal completeness, we selected representative publicly available 1 km resolution raster data from published sources.

Population density represents “Density”. Historical population density data were sourced from WorldPop, specifically the “Unconstrained individual countries 2000–2020 UN adjusted (1 km resolution)” dataset, as it meets our 1 km resolution requirement and includes data for the baseline year 2015. WorldPop’s top-down modeling methods use a global database of administrative unit-based census and projection counts from 2000 to 2020, employing detailed geospatial datasets to disaggregate these counts to grid cell-based data. The assumption of unconstrained datasets is that no settlement dataset is accurate enough to identify all residential settlements/buildings globally and therefore be used as a mask over the 2000–2020 period to map uninhabited areas. To address this, we used raster data with residential area markers as a mask for extraction in the subsequent steps. We checked the data for variability and removed a cell from the original dataset in Siming District, Xiamen, which contained an extreme population outlier. We then adjusted this value to match neighboring areas, thereby increasing the reliability of the data.

For future population data, we must also use 1 km resolution grid data, maintaining continuity and correlation with historical data, and aligned with the future prediction years of GCAM. Considering these factors, we selected the global 1 km grid-scale multi-scenario population distribution dataset for 2020–2100 published by Wang et al.³⁶. The advantage of this spatially explicit population dataset is that it is based on the WorldPop dataset, combined with a random forest algorithm for model construction and data prediction, and validated by comparison with the WorldPop dataset at both the sub-national and grid levels. The spatial resolution of this gridded population dataset is 30 arc-seconds (~1 km), with a 5-year time interval, and it includes five shared socioeconomic pathway (SSP) scenarios. To facilitate comparison with the SSP2 scenario—which serves as the reference pathway for multiple socio-economic indicators in GCAM and is consistent with the GDP projections adopted in this study—we selected the SSP2 population trajectory from among the five available datasets for subsequent calculations.

Land Use Mix as an Indicator of “Diversity”. This study focuses on the small-scale “functional mix” related to people’s work, daily activities, and travel demand. Thus, we aim to comprehensively categorize the land use types of residential and employment areas. To achieve this, two separate data sources were used: one for urban areas and another for non-urban areas.

Data on land use types for urban areas are relatively scarce. Currently, we have identified only one nationwide dataset for Urban Land Use Categories in China, the EULUC-China dataset³⁷. This dataset, developed by Tsinghua University and over 30 other research institutions, includes five major land use categories—Residential, Commercial, Industrial, Transportation, and Public Management and Service—under the “Essential Urban Land Use Categories (EULUC)”. This dataset uses 10-meter satellite images, OpenStreetMap, nighttime lights, Points of Interest (POI), and Tencent social big data from 2018 as input features. This is a vector polygon dataset, which we converted into a 30 m resolution raster dataset to match with the land use data for non-urban areas. For non-urban areas, we selected the 2020 30-meter resolution raster layer from China’s Multi-Period Land Use/Cover Change Remote Sensing Monitoring Dataset (CNLUCC), developed collaboratively by five specialized geographic institutions under the Chinese Academy of Sciences. The database uses U.S. Landsat satellite imagery as its primary information source, and it is constructed through manual visual interpretation to create China’s national-scale multi-period land use/land cover (Land-Use and Cover-Change Project) database. The land use is categorized into primary and secondary classifications, with seven primary types: Cropland, Woodland, Grassland, Water, Built-up Land, Ocean, and Unused Land. The secondary classifications further divide these into 26 types.

To better align with our research goal of characterizing the land use mix related to human habitation and daily work, we selected six primary land use types from this classification system. However, within Built-up Land, we retained the secondary classifications for non-urban areas, including urban land, rural settlements, and other built-up areas (e.g., factories, large industrial zones, quarries, etc.). As a result, the final land use classification for non-urban areas consists of 9 types (6 primary + 3 secondary). We used the EULUC-China raster range for urban areas and the remaining area from the CNLUCC dataset, excluding the EULUC-China range, for non-urban areas. By performing raster mosaicking, we obtained a complete 30 m resolution land use type image that includes both urban and non-urban areas, which is divided into 20 land use types (see Supplementary Table S4).

We then created a 1 km × 1 km grid for the entire country, consisting of 25,397,415 cells, of which 9,501,225 are valid pixels (i.e., non-empty pixels covering areas within China’s administrative boundaries), accounting for 37.41% of the total. We intersected the grid with the land use data layer and recalculated the geometry to determine the area of each land use type within each 1 km resolution cell. To quantitatively calculate the diversity of land use types within each 1 km grid, we used the Shannon Diversity Index (SHDI) to compute the “diversity” indicator^24,38,39. The calculation formula for the SHDI is as follows:

$${SHDI}=-\mathop{\sum }\limits_{i=1}^{m}{P}_{i}ln{P}_{i}$$

(2)

$${P}_{i}=\frac{{A}_{i}}{\mathop{\sum }\limits_{i}{A}_{i}}$$

(3)

Where ${SHDI}$ represents the Shannon Diversity Index, ${P}_{i}$ represents the area proportion of the land use type $i$, and ${A}_{i}$ is the area of the land use type $i$. When the ${SHDI}$ value equals 0, it indicates that the cell contains only one land use type. A higher ${SHDI}$ value indicates a greater variety of land use types within the cell. Since algebraic derivations and estimates regarding land use mix diversity will be conducted later, we calculated the ${SHDI}$ using the original values instead of the standardized ${SHWI}$. After calculating the ${SHDI}$, we linked the results to the attribute table of the grid.

Road Network as the Indicator of “Design”. We acquired road data for all types of roads in China spanning from 2014 to 2024 via OpenStreetMap (OSM). OSM is open-source and widely accessible, providing a long-term temporal coverage of historical data. Researchers have employed indicators such as Distance between Entrances⁴⁰, Road Network Density^41,42, Road Connectivity⁴³, and Intersection Density^44,45 to measure road design. Given data availability and computational convenience, we calculated road network density for each 1 km grid cell, which is defined as the road length per square kilometer, as an indicator of the “Design” quality for that cell. We calculated the road network length using the spatial connectivity of the national-scale grid based on “computational geometry”. We acknowledge that road network density is a simplified topological proxy for design quality that does not capture important qualitative aspects such as road hierarchy, pedestrian infrastructure quality, street width, sidewalk conditions, or the functional importance of different road segments. However, for national-scale analysis covering over 9.5 million grid cells, road network density offers several advantages: (1) it is consistently measurable across all regions using openly available data; (2) previous research has established significant correlations between network density and travel behavior at large spatial scales; and (3) it captures the fundamental design principle that denser, more connected street networks generally support shorter trips and non-motorized transport.

Distance to City Center as the Indicator for “Destination Accessibility”. We used city centers to represent “destinations”, as they generally provide greater access to services and amenities. This approach is also commonly adopted in many studies⁴². Since our study covers the entire country at a 1 km resolution, using straight-line Euclidean distance would introduce significant errors, as it does not reflect the actual driving distance along roads. To approximate “driving distance” more accurately, we employed the “Cost Distance” method in GIS analysis. We assigned time costs (in minutes) based on road width to roads of different grades and types, as well as non-road areas. The time cost for each cell is the reciprocal of the corresponding driving speed, with the time cost representing resistance to mobility. Higher-grade roads allow higher driving speeds and thus exhibit less resistance. Time costs were calculated following operations like converting polylines to raster data and performing raster mosaicking. The driving speeds and road widths for different road grades were based on China’s road design standards⁴⁶. For road design types not directly classified in OSM, interpolation was performed based on practical experience, as detailed in Supplementary Table S6. We recognize that using city centers as a single destination type is a simplification of the complex, polycentric nature of accessibility. In reality, accessibility depends on proximity to multiple types of destinations (employment centers, commercial areas, educational facilities, healthcare services, etc.) with varying frequencies of need and trip generation potential. However, systematic data on the full hierarchy of destinations and transit service levels are not consistently available across all Chinese cities at the required spatial resolution. Our approach follows established practice in large-scale built environment research, where city centers serve as reasonable proxies for areas of high accessibility, as they typically concentrate employment, services, and multimodal transit connections.

The total number of railway stations, metro stations, and bus stops represents “Distance to Transit” indicator. The site data for transit stops is sourced from OSM, covering the period from 2014 to 2024. We calculated the total number of stops for each 1-km grid cell, which serves as an indicator of the “Distance to Transit” quality.

The future data for the three indicators—Diversity, Design, and Distance to Transit—are from our hypothetical simulation. For diversity, the main challenge was the lack of existing predicted data on urban land use types from 2020 to 2060. This challenge was also highlighted in Fu et al.²⁸, so we followed their approach and applied a multiplication factor directly to the original diversity values. The land use mix will improve, but the extent and growth rate of improvement will differ between urban built-up areas and other regions (e.g., suburbs and rural areas), with urban areas showing slightly greater improvement than other regions. Given the future expansion of urban built-up areas, we used predictive data from Huang et al.⁴⁷ as the basic data to differentiate between future built-up and non-built-up areas. This dataset predicts the future expansion patterns of Chinese cities at a 1 km resolution under the SSP scenarios and delineates urban growth boundaries (UGBs) for China up to 2100. Due to data limitations, we made a simple assumption about the growth multiplier factor for diversity, as detailed in the Supplementary Table S5. Since our basic diversity data are from 2018 and 2020, we assumed that the diversity in 2020 is 1, and then retroactively adjusted the diversity value in 2015. Starting in 2025, we assumed there would be improvements with a growth rate that initially accelerates and then decelerates.

For the indicators of “Design” and “Distance to Transit”, we use linear fitting to extrapolating each cell’s data from 2014–2024 to 2060 for prediction. Considering the stability of China’s administrative divisions and the locations of city centers, the same dataset is used to predict “Destination Accessibility”. All raster images for the entire country are standardized with a uniform mask and projection coordinates to facilitate processing in GIS software.

Calculation and prediction of TSD-LDV4W for each year

To construct the national-scale TSD-LDV4W data at a 1-km resolution, we first calculated the TSD-LDV4W for each cell in 2015 based on the relative weight of each cell. We used the elasticity coefficients between the 5D indicators and transportation service demand derived from Fu et al.²⁸, who conducted a comprehensive meta-analysis of 26 empirical studies examining the relationships between built environment characteristics and vehicle travel demand. Their synthesis provides robust, generalizable estimates regarding how each 5D dimension affects travel behavior in various urban contexts. The elasticity coefficients obtained from this meta-analysis represent the percentage change in vehicle travel that is associated with a 1% change in each built environment indicator. It is not feasible to directly calculate TSD-LDV4W using matrix computation based on the absolute values because of the different dimensions and measurement methods of the 5D indicators. Therefore, we employed the relative relationships among the elasticity coefficients to develop the Built Environment Advantage Index (BEAI), which serves as spatial weights for allocating the built-environment-driven TSD-LDV4W potential across grid cells. In this context, grid-level TSD-LDV4W is operationally defined as an intrinsic travel demand propensity attributable to the built environment. This approach assumes that areas with more favorable built environment characteristics, such as higher density, greater diversity, better design, greater accessibility, and better transit access, will generate proportionally lower travel demand, consistent with the extensive empirical literature synthesized by Fu, et al.²⁸. Given the absence of a direct physical measure for the localized travel demand attributable solely to urban structure, we operationalize the grid-level TSD-LDV4W as a proxy for built-environment-driven demand potential.

$$\begin{array}{l}{BEAI}=\left(\frac{\left|{\beta }_{1}\right|}{{\sum }_{i=1}^{5}\left|{\beta }_{i}\right|}\right){Densit}{y}_{{std}}+\left(\frac{\left|{\beta }_{2}\right|}{{\sum }_{i=1}^{5}\left|{\beta }_{i}\right|}\right){Diversit}{y}_{{std}}+\left(\frac{\left|{\beta }_{3}\right|}{{\sum }_{i=1}^{5}\left|{\beta }_{i}\right|}\right){Desig}{n}_{{std}}\\ +\left(\frac{\left|{\beta }_{4}\right|}{{\sum }_{i=1}^{5}\left|{\beta }_{i}\right|}\right){DestinationAccessibilit}{y}_{{std}}\\ +\left(\frac{\left|{\beta }_{5}\right|}{{\sum }_{i=1}^{5}\left|{\beta }_{i}\right|}\right){DistancetoTransi}{t}_{{std}}\end{array}$$

(4)

Where ${\beta }_{1}\,$= −0.08, ${\beta }_{2}\,$= −0.13, ${\beta }_{3}\,$= −0.03, ${\beta }_{4}\,$= −0.06, ${\beta }_{5}$ = −0.04²⁸. Due to the differences in units and the evaluation criteria of the influencing factors^48,49, we standardized the 5D indicator data. ${Densit}{y}_{{std}}$, ${Diversit}{y}_{{std}}$, ${Desig}{n}_{{std}}$, ${DestinationAccessibilit}{y}_{{std}}$, and ${DistancetoTransi}{t}_{{std}}$ represent the standardized values of the 2015 5D indicators, calculated using the range normalization method. These normalized weights reflect the relative importance of each 5D dimension in explaining vehicle travel behavior based on extensive empirical evidence.

Since we did not directly calculate the regression coefficients between the 2015 TSD-LDV4W and the 2015 5D indicators, but instead used the relative relationships of elasticity coefficients, we used machine learning to calculate the future TSD-LDV4W, which simplified large-scale grid computations. In geospatial research, spatial autocorrelation is commonly observed, indicating that geographic entities are not independent but spatially interdependent⁵⁰. To account for spatial proximity effects beyond the 5D characteristics of each grid cell, we included geographic coordinates (X and Y in the Krasovsky_1940_Albers projection), as additional features in the machine learning model. Additionally, we incorporated the mean and standard deviation of all variable values within a 5 km × 5 km square neighborhood into the model. Five-fold cross-validation revealed that the model incorporating spatial proximity outperformed the baseline model. Specifically, R² increased from 0.9795 to 0.9813, and RMSE decreased from 0.0429 to 0.0339. After being trained on the independent and dependent variables in 2015, the machine learning model calculated the future TSD-LDV4W based on the 5D indicator data for subsequent years. Given the large number of grid cells in this study and the potential “ecological fallacy” resulting from cumulative regression errors for each grid, machine learning offers a simple and efficient approach.

Finally, since transportation service demand is generated by actual residents, we cross-referenced the land use classification in the diversity layer to select cells in urban areas and rural settlements. Then, we removed uninhabited cells in non-urban areas to eliminate the statistical impact of large uninhabited areas on the sample. After extracting the key areas, each raster image contained a total of 19,606,916 cells, with 150,277 valid pixels. Among these valid pixels, 114,313 were located in urban areas, while 35,964 were in non-urban areas.

Delineation and projection of old and new urban areas

Based on historical data from the Global Urban Boundaries (GUB) dataset⁵¹, we selected the 1995 GUB as the old city center. Given their earlier development history relative to the period when most Chinese cities entered the automobile era (early 21st century), we assume these areas exhibit more compact urban design, smaller block scales, and narrower, denser street networks, which are hallmarks of typical pedestrian-friendly districts. Considering the general timeline of new town expansion in Chinese cities, we defined newly developed areas as the spatial difference between the 2018 and 2005 GUB boundaries. These areas were mostly developed after 2005, when car ownership in China was rapidly increasing, and thus are considered representative of car-centric districts. The area between pedestrian-friendly and car-centric districts is defined as the “transition zone.” Since it shares characteristics of both, it is excluded from classification as either a typical pedestrian- or car-oriented district. All other areas are categorized as non-urban.

Considering that the relationship between built environment characteristics and urban area types involves complex, non-linear interactions that machine learning algorithms can capture more flexibly than linear models, we employ a supervised machine learning classification approach rather than traditional regression or rule-based methods. We trained a supervised machine learning classification model using 2015 data to predict area types and then selected the best-performing model for future area type projections. Our use of single-year 2015 training data for projections to 2060 is based on the space-for-time substitution principle, which is widely used in spatial analysis. China’s vast spatial expanse and regional development heterogeneity imply that in 2015, different cities simultaneously exhibit development stages spanning multiple decades. By training on this spatially diverse dataset, we are effectively training on conditions that span the development spectrum, thereby enabling the model to project future conditions that fall within the observed spatial range. Moreover, our 5D indicator datasets have different temporal coverage (some starting 2014, others from 2018 to 2020), making consistent multi-year training infeasible. The 2015 training year represents the optimal balance among data completeness, temporal proximity to our baseline year, and the availability of validated area type classifications derived from GUB temporal analysis. Predictor variables include the 5D built environment indicators, Local Affluence Level (cell-level GDP), and Prefectural GDP (GDP of the corresponding prefecture-level city)⁵², reflecting spatial economic variation. After testing several algorithms (including K-Nearest Neighbors and Support Vector Machines), we chose the Random Forest (RF) classification model because it demonstrated the best performance [ROC AUC (Area Under the Receiver Operating Characteristic Curve) = 0.86]. We conducted 5-fold cross-validation within 2015 data to evaluate the model’s generalization across spatial subsets. The RF model is well-suited for this task because it is robust to non-linear relationships and interactions between predictors, and it enables us to quantify the relative importance of each feature in distinguishing these area types via SHAP values. The model’s hyperparameters were fine-tuned and optimized.

To examine the influence of individual predictor variables on transitions between area types from 2015 to 2016, we computed Z-scores to identify deviations associated with specific transition types:

$$z\,{score}=\frac{{chang}{e}_{{mean}}-{overal}{l}_{{mean}}}{{overal}{l}_{{std}}}$$

(5)

Here, ${overall\_mean}$ and ${overall\_std}$ represent the mean and standard deviation of each predictor variable (e.g., diversity) across all grid cells. ${change\_mean}$ refers to the mean value of each predictor variable for a specific type of transition (e.g. from old city center to newly developed area).

Results

IAM-predicted vehicle travel demand amid EV-driven automotive transition

We begin by establishing baseline transportation demand projections using GCAM and GCAM-China, which serve as comparative benchmarks for assessing the impact of built environment improvements. Given the potential of electric vehicles (EVs) to curb emissions in the transportation sector, governments worldwide have rolled out a range of incentives to accelerate the penetration of EV sales and increase their usage³⁴. To gain insights into future trends, various Integrated Assessment Models (IAMs) have been employed to forecast the growth trajectory of transportation service demand (TSD). TSD is measured in passenger-kilometers, representing the cumulative kilometers traveled by all residents in a specific region over a given year^11,34. Leveraging the Global Change Analysis Model (GCAM) and a provincial-level representation of the China energy system (GCAM-China), this research estimates the evolution of transportation service demand associated with four-wheeled light-duty vehicles (TSD-LDV4W) within the framework of fundamental transition occurring in China’s automotive system, which is being propelled by large-scale EV deployment in China. The results of this analysis will lay the groundwork for comparing the effectiveness of local measures in reducing TSD-LDV4W in subsequent sections. In light of China’s pledge to peak carbon emissions by 2030 and achieve carbon neutrality by 2060, we have selected 2030 and 2060 as pivotal observation years. Additionally, data from 2025, the year most proximate to the current situation, has been included for comparative purposes.

The global push to mitigate climate change is a driving force of varying degrees of strength, spurring the low-carbon transition in the transport sector. This divergence in impetus leads to notable disparities in both the overall volume and peak levels of TSD-LDV4W across different Shared Socioeconomic Pathway (SSP) scenarios, as illustrated in Fig. 2. In all SSP scenarios within GCAM’s global framework, China’s TSD-LDV4W initially shows an upward trend before eventually declining. However, the magnitude and timing of the peak vary. SSP1, which represents the most sustainable pathway with strong climate mitigation efforts and sustainable development patterns, reaches an early peak of 1.86 × 10¹² passenger-kilometers in 2020 and declines to 1.17 × 10¹² passenger-kilometers by 2060, a 37% reduction from the peak. In contrast, SSP2 (middle-of-the-road) and SSP3 (regional rivalry) reach their peaks in 2035, whereas SSP4 (inequality) and SSP5 (fossil-fueled development) peak later, in 2040. Compared to SSP1, SSP2-SSP5 decline more moderately to 2.12 × 10¹², 2.25 × 10¹², 2.56 × 10¹² and 2.47 × 10¹² passenger-kilometers by 2060, representing only 14%, 12%, 8% and 17% reductions from the peak. When considering the total calculated TSD-LDV4W, the SSP1 scenario registers the lowest value. Meanwhile, the peak TSD-LDV4W of the SSP5 scenario, which is the highest among all, is 1.6 times that of SSP1, underscoring a significant gap. The relatively small differences among SSP2-SSP5 scenarios compared to SSP1 reflects a fundamental insight: without transformative changes in development patterns and strong climate ambitions, transportation demand trajectories cluster around high-growth pathways driven by rising incomes and continued car-oriented development. Only SSP1, with its emphasis on sustainable consumption, compact urban development, and modal shift, achieves substantially lower demand. As the deployment of EV continues to expand, the proportion of TSD-LDV4W met by EVs rises steadily each year, aligning with the predetermined EV share targets. By 2060, this proportion is expected to reach ~70%. At the same time, the market share of gasoline-powered vehicles undergoes a rapid decline, ranging from 14% (SSP2) to 22% (SSP5) by 2060.

Fig. 2: Projection of Total Chinese Transportation Service Demand for Four-wheeled Light-duty Vehicles in Passenger Transportation Sector (TSD-LDV4W) under five Shared Socioeconomic Pathway (SSP) scenarios: 2015–2060 using Global Change Analysis Model (GCAM).

a–e Respectively show the changes in the total Chinese TSD-LDV4W under different SSP scenarios. Different colors are used to distinguish between various automotive powertrain technologies: NG stands for Natural Gas, FCEV for Fuel Cell Electric Vehicles, and BEV for Battery Electric Vehicles. Y-axis: TSD-LDV4W in million passenger-kilometers; X-axis: Year from 2015 to 2060.

Based on the results calculated by GCAM, we employed GCAM-China to decompose the national transportation system to the provincial level (Fig. 3). With large-scale deployment of electric vehicles, TSD-LDV4W in 2030 and 2060 is projected to decrease by 2% and 9%, respectively, compared to the model’s built-in reference scenario. This reduction in TSD-LDV4W resulting from increased EV deployment has been similarly validated in models from countries such as the United States³⁴. We must clarify an important conceptual distinction: the modest reductions in TSD-LDV4W projected by GCAM-China when EV deployment constraints are imposed do not represent direct effects of vehicle electrification on travel demand. Rather, these reductions emerge from GCAM’s system-wide economic modeling of the transport sector. As EV deployment increases, the model accounts for changes in service provision costs, modal competition, and consumer responses to altered price signals. According to the interlinkages between the transport sector and other systems in GCAM, rising per capita GDP directly increases demand for passenger transport services, while boosting wage levels and overall living standards. In line with sustainable development goals promoting electric vehicle adoption, this shift encourages consumers to choose faster and lower-emission modes of travel. However, it also raises the overall cost of service provision, which may ultimately reduce travel demand⁵³.

Fig. 3: Projections of Total Chinese Transportation Service Demand for Four-wheeled Light-duty Vehicles in Passenger Transportation Sector (TSD-LDV4W) from 2015 to 2060 using the GCAM-China provincial model highlight substantial disparities across provinces.

This figure presents the provincial TSD-LDV4W results from 2015 to 2060, based on the GCAM-China provincial model simulation. In the bar chart, different colors represent various vehicle powertrain types, and the shading of each province block reflects the initial TSD-LDV4W level in 2015, with darker gray indicating higher values. This analysis covers the 31 provincial-level administrative regions of mainland China; Hong Kong, Macau, and Taiwan are excluded due to their distinct administrative systems and data availability constraints in GCAM-China and built environment datasets. Y-axis: TSD-LDV4W in million passenger-kilometers; X-axis: Year from 2015 to 2060.

These results offer a lucid depiction of the influence mechanism through which macro-level landscape pressure operates on meso-level regimes within the framework of the Multi-Level Perspective theory^9,54. A close examination of the results reveals the tangible impact of macro-level visions on the private car transport sector. The pronounced disparities in TSD-LDV4W observed across different SSP scenarios serve as a compelling testament to how global climate visions, acting as a source of pressure of varying intensities, prompt national level policy makers to set specific emission reduction targets and formulate tailored policies to steer the transformation of the private car transport sector.

It is evident that distinct national policies can generate a wide range of macro-level effects⁵⁵. Moreover, the manifestation of this pressure is not uniform across regions. Instead, it varies according to the existing progress each region has made and its capacity to respond to future visions. Some regions, due to their advanced state of development or more proactive approach, may be better equipped to handle the pressure and implement effective policies, while others may face greater challenges in adapting to the changing landscape.

Micro-level built environment: a catalyst for macro-level TSD-LDV4W reduction

IIn order to demonstrate the influence of the local built environment on TSD-LDV4W, we employed the built-environment-driven structural travel demands within each one-kilometer grid to depict the structural spatial distribution of TSD-LDV4W throughout the country. Drawing on prior research concerning the elasticity coefficients between urban built environment indicators and transportation demand, we developed the “Built Environment Advantage Index (BEAI)” using 2015’s 5D built environment data as weights for the spatial allocation of travel demands. Secondly, the projections for the future 5D built environment indicators are grounded in historical urban development patterns. We specifically extrapolated future trends from the urban evolution that took place between 2014 and 2024. Historical records indicate a general improvement in cities’ 5D indicators, including denser road networks and a greater number of public transit stations in China. Through the adoption of this rational extrapolation method, we can integrate macro-level predictions with micro-level forecasts. Then, using the 5D forecast data for future years, we applied the elasticity coefficients to estimate changes in TSD-LDV4W over time, accounting for the spatial proximity characteristics of each grid cell in the computation. Finally, we visualized the spatial distribution of TSD-LDV4W as depicted in Fig. 4a. Each grid cell’s pixel value represents the relative spatial share of structurally allocated potential TSD-LDV4W burden attributable to the built environment characteristics at that location.

Fig. 4: Variations and spatial distribution of Chinese Transportation Service Demand for Four-wheeled Light-duty Vehicles in Passenger Transportation Sector (TSD-LDV4W) potential across 2015 to 2060, driven by micro-level built environment improvements.

a shows the spatial distribution of national built-environment-driven structural TSD-LDV4W propensity in 2060 at a 1 km x 1 km resolution. b, c, and d are local zoom-ins of (a), with (b) depicting the TSD-LDV4W distribution in the Beijing-Tianjin-Hebei region, c showing the TSD-LDV4W distribution in the Yangtze River Delta, and d showing the TSD-LDV4W distribution in the Pearl River Delta. e Shows the probability density distribution of grid-level TSD-LDV4W values nationally for four time points: 2015 (gray), 2025 (blue), 2030 (green), and 2060 (yellow). Y-axis: Probability density (density units); X-axis: TSD-LDV4W per grid cell (passenger-kilometers per km² per year). The leftward shift and shape evolution of the density curves over time—from right-skewed (2015) through bimodal saddle-shaped (2025, 2030) to left-skewed (2060)—illustrate the changing distribution of travel demand intensity across the national territory, with an increasing proportion of grid cells exhibiting lower TSD-LDV4W values as built environment improvements take effect.

Our model projects a notable decrease in the overall TSD-LDV4W level nationwide from 2015 to 2060, driven by the extrapolation of observed improvements in 5D built environment indicators from 2014 to 2024. Specifically, the total national TSD-LDV4W drops from 1.58 × 10¹² passenger-kilometers in 2015 to 1.24 × 10¹² passenger-kilometers in 2060, marking a 21% reduction. This result underscores the substantial potential for reducing private car ownership and the consumption of materials for vehicle production when the characteristics of built environments are taken into account. When juxtaposed with the five SSP scenarios in GCAM, the TSD-LDV4W reduction resulting from built environment improvements (which essentially reflects the impact of local actions) is close to that in the SSP1 scenario, and significantly outweighs the reductions in other SSP scenarios. This modeled outcome suggests that improving the local micro-level built environment could make the achievement of SSP1-level transportation demand reduction more attainable, even under SSP2 socio-economic trajectories.

Our analysis uncovers a significant correlation between the spatial distribution of TSD-LDV4W and regional characteristics. Notably, while economically prosperous regions such as the Beijing-Tianjin area, the Yangtze River Delta, and the Pearl River Delta exhibit elevated concentrations of TSD-LDV4W, a distinctive planar spread of these vehicles is observed across the North China Plain. This phenomenon can primarily be attributed to the region’s predominantly flat topography, which facilitates a dense and relatively uniform distribution of rural settlements throughout the plains. In contrast, travel demand in the southern coastal regions is more concentrated in urban centers and their immediate surroundings. This pattern arises because urban development in these areas is often constrained by mountainous landscapes and the limitations imposed by the marine environment.

The density distribution plot of TSD-LDV4W values (Fig. 4e) depicts a leftward shift in the probability density curve over time, indicating a decreasing trend in TSD-LDV4W values. The profiles of these curves demonstrate notable annual variations: the 2015 curve exhibits a right-skewed peak, which transitions into a saddle-shaped profile by 2025 and 2030, and eventually evolves into a left-skewed peak by 2060. The significant and substantial disparities in TSD-LDV4W values across different locations underscore that the local built environment characteristics captured by the 5D indicators (encompassing Density, Diversity, Design, Destination Accessibility, and Distance to Transit) are strongly associated with different travel behavior patterns in various regions. These spatial variations in built environment quality drive differential travel demand intensities, as residents in areas with favorable 5D characteristics (dense, mixed-use, well-connected neighborhoods with good transit access) face lower barriers to non-auto travel and shorter necessary trip distances.

Regional disparities in private car dependence

To investigate how disparities in regional economic development influence the diversity of low-carbon transition pathways in the transport sector, we conducted a comparative analysis by juxtaposing the outcomes of provincial-level model simulations within GCAM-China with those derived from province-level simulations based on population-weighted distributions in 2015 and 2060 (Fig. 5a).

Fig. 5: Regional differences in Transportation Service Demand for Four-wheeled Light-duty Vehicles in Passenger Transportation Sector (TSD-LDV4W) in 2015 and 2060.

a Compares the population-weighted allocation results of TSD-LDV4W in each province and the simulation outcomes generated by GCAM-China for the years 2015 and 2060. b and c illustrate the distribution of TSD-LDV4W values in urban and rural areas calculated from the spatial distribution of TSD-LDV4W in Fig. 4 for 2015, 2030, and 2060, respectively.

The comparison between the population-weighted distribution results of TSD-LDV4W in each province and the simulation outcomes generated by GCAM–China for the years 2015 and 2060 (see Fig. 5a) indicates that by 2060, the impact of inter-provincial economic development disparities in China on provincial TSD-LDV4W has significantly diminished. At a macro level, the total population has emerged as the primary determinant of regional TSD-LDV4W. The correlation coefficient between the population-weighted distribution results of TSD-LDV4W and the simulation results of GCAM-China increased from 0.63 in 2015 to 0.73 in 2060.

The projection shows a more equitable development future with narrowing regional disparities in socio-economic development levels, while the differences in the micro-level built environment, which is profoundly influenced by the process of urbanization, emerge as a key factor influencing TSD-LDV4W. According to the spatial distribution of TSD-LDV4W, in 2060, the share of TSD-LDV4W in urban areas accounts for 75.97% of the national total, reflecting population concentration in cities. However, the mean TSD-LDV4W per grid cell in rural areas is 2.62% higher than in urban areas in 2060. Figure 5b, c reveals the overall change in distribution of grid-level TSD-LDV4W in urban and rural areas from 2015 to 2060 calculated from the spatial distribution of TSD-LDV4W. Improvements in the local built environment have great potential to reduce TSD-LDV4W in both urban (15.59%) and rural areas (17.06%). With increasing population moving into cities—urban residency projected to increase by 6.28 percentage points between 2015 and 2060—changes in travel volume in urban areas are pivotal to alterations in the overall travel pattern. This reflects fundamental differences in settlement patterns, service accessibility, and available transportation alternatives between urban and rural contexts.

In urban areas, despite high population densities generating substantial aggregate travel demand, residents have access to diverse transportation alternatives—extensive public transit networks (buses, metros, light rail), walkable destinations within neighborhoods, cycling infrastructure, and short trip distances due to mixed land use. These alternatives substantially reduce per-capita reliance on private cars. Even in car-centric urban districts, the presence of transit options and proximate destinations mitigates TSD-LDV4W intensity compared to rural settings.

In rural areas, by contrast, settlement dispersal and limited transportation alternatives create car dependency. Rural residents face greater distances to services, minimal public transit, rural road infrastructure limitations, and dispersed settlement patterns. Therefore, when rural residents undertake trips, they must disproportionately rely on private vehicles for mobility, generating higher per-capita and per-area TSD-LDV4W despite lower population densities.

Urban renewal as transition opportunity

When we focus our analysis on urban built-up areas, we observed that in the majority of cities, the rapid urbanization and transportation motorization that China has experienced over the past two decades have had a profound impact on the urban built environment. The central old urban areas, boasting a long-standing development history and well-established built environments, display relatively low TSD-LDV4W levels. In contrast, the peripheral areas register the highest TSD-LDV4W, where new towns have emerged over the past two decades, developing in tandem with mass migration of the population to urban areas. To cater to the demands of automobile consumption, these new towns exhibit significant differences in characteristics compared to the traditional city centers of the pedestrian-oriented era.

In examining the evolution of light-duty vehicle use and urban development in China, we identified four spatial typologies, two of which were classified as representative pedestrian-friendly and car-centric districts. Prior to the 21st century, car ownership in China remained low, and daily mobility primarily relied on walking, cycling, and public buses. According to official Chinese statistical yearbooks, private passenger car ownership increased from 2.06 vehicles per 1000 people in 1995 to 192.92 in 2024. Consequently, urban development in the 20th century can be characterized as pedestrian-oriented, featuring compact urban forms and smaller block sizes. After 2005, China experienced a steady increase in car ownership, coinciding with widespread construction and expansion of new urban districts. As such, post-2005 urban development can be broadly categorized as a car-centric expansion model. Areas developed between 1995 and 2005, shaped under less consistent planning policies, are treated as transition zones, exhibiting neither strongly pedestrian-friendly nor distinctly car-centric characteristics. Areas falling outside these three categories are collectively referred to as “other non-urban areas”, comprising predominantly rural settlements, along with scattered small towns and remote township centers. Drawing on historical data from the Global Urban Boundaries (GUB) dataset⁵¹, we designated the 1995 GUB as typical pedestrian-friendly districts. Subsequently, we identified the disparity in GUB changes between 2018 and 2005 as typical car-centric districts. In 2015, the total TSD-LDV4W for pedestrian-friendly districts, car-centric districts, transition zones, and other non-urban areas amounted to 1.77 × 10¹¹, 4.69 × 10¹¹, 1.57 × 10¹¹, and 7.78 × 10¹¹ passenger-kilometers, respectively (Fig. 6). Although car-centric districts are increasing over time, TSD-LDV4W continues to decrease due to significant changes in the 5D indicators resulting from improvements and optimization of the urban built environment. There is substantial potential for reducing private car travel demand by constructing or transforming more districts into pedestrian-friendly areas.

Fig. 6: TSD-LDV4W differences between car-centric and pedestrian-friendly districts highlight the potential of urban renewal in transforming car-oriented travel.

a–d show the TSD-LDV4W distribution across the four area types in 2015 in 2060.

An important observation from Fig. 6 is that the mean values of TSD-LDV4W show similar declining trends across area types, but the magnitude of decline varies, and absolute TSD-LDV4W levels remain dramatically different across area types. In 2060, car-centric districts still generate higher TSD-LDV4W per unit area (4.02 million passenger-kilometers) than pedestrian-friendly districts (3.85 million passenger-kilometers), meaning that area type classification remains highly consequential for aggregate national demand. Methodologically, our projections apply 5D indicator improvements based on observed 2014–2024 trends uniformly across the national territory (with differentiation between urban built-up areas and non-urban areas as specified in Supplementary Table S5). This approach assumes that policy efforts to improve the built environment are determined by the historical development trends of the geographic location where each grid is situated, reflecting local urbanization and infrastructure development strategies. From a policy perspective, this finding suggests that built environment improvements can reduce car travel across all urban form types, not only in traditionally walkable neighborhoods. Even car-centric districts benefit substantially from strategic interventions that reduce (though do not eliminate) car dependency. This is crucial for China’s context, where preventing further lock-in of car-dependent lifestyles in expanding peripheral areas may be more feasible than fully transforming existing car-centric districts into pedestrian-friendly ones.

A critical insight from our analysis is that car-centric district area expansion and total TSD-LDV4W reduction are not contradictory, but reflect different spatial dynamics. This occurs for the following three reasons. First, intensive margin effects dominate extensive margin effects. While car-centric areas expand spatially (extensive growth), the TSD-LDV4W per unit area or per capita within all area types—including car-centric districts—declines substantially due to projected 5D improvements (intensive reduction). Our model projects that built environment quality continues improving even in newly developed areas based on observed 2014–2024 trends. Second, temporal quality improvements play a role. A “car-centric” district developed in 2060 has fundamentally better 5D characteristics than a car-centric district from 2015. Third, urbanization effects contribute to this pattern. As population increasingly concentrates in urban areas, residents move from rural areas (with high per-capita TSD-LDV4W due to dispersed settlement patterns) into urban areas where built environment improvements are most pronounced, further contributing to national-level demand reduction. These findings underscore that quality improvements within expanding car-centric districts—through better transit provision, road network connectivity, and land-use mixing—offer substantial potential for demand management, even as urban expansion continues.

China’s urbanization has entered a new phase of development from large-scale new city construction to a stage centered on optimizing existing urban stock through urban renewal. To delve deeper into the disparities in how various built environment factors influence pedestrian-friendly and car-centric districts, and to identify the key aspects of the built environment that urban planning and management should prioritize to curb private car travel demand, we employed a machine learning supervised classification model. A Random Forest (RF) classification model was trained on 2015 data to discern the relationship between seven predictor variables—including five local 5D built environment indicators (Density, Diversity, Design, Destination Accessibility, and Distance to Transit) and two spatial economic indicators (Local Affluence Level, defined as cell-level GDP, and Prefectural GDP, representing the GDP of the corresponding prefecture)—and the four area classifications. Following a 5-fold cross-validation process, the model achieved a ROC AUC (Area Under the Receiver Operating Characteristic Curve) of 0.86. Using the Shapley method, we computed the average absolute SHAP values for each area category (Fig. 7).

Fig. 7: A supervised classification machine learning model predicts future transitions between pedestrian-friendly and car-centric districts.

This figure shows the SHAP feature importance for the four area types, based on the mean absolute values, where larger values indicate greater contributions of the predictor variables to the predictions.

The results reveal distinct travel behavior patterns between pedestrian-friendly and car-centric districts. Nationally, average TSD-LDV4W is lower in pedestrian-friendly districts than in car-centric districts. SHAP value analysis shows that across all areas, local affluence level (GDP in one grid cell) emerges as the most influential factor. However, a comparison of average TSD-LDV4W in 2060 versus 2015 reveals a steeper decline in pedestrian-friendly districts than in car-centric districts, showing that as the regional economic development disparities decreasing, the influence of the local built environment on travel behavior patterns becomes increasingly decisive. Among built environment variables, density and destination accessibility contribute most to distinguishing pedestrian-friendly from car-centric districts, whereas distance to transit contributes the least. This may reflect China’s nationwide efforts to expand public transit infrastructure and promote service equalization between urban and rural areas.

However, the importance of built environment and regional economic characteristics varies among different area types. In pedestrian-friendly districts, Density and Regional GDP carry greater weight than in car-centric districts. Conversely, in car-centric districts, Destination Accessibility and Diversity factors are more pivotal. This finding implies that urban planning and management should tailor their efforts to enhance specific dimensions and elements of the built environment for different area types. To achieve this, it is essential to implement nuanced policy measures. These measures should aim to improve the overall urban transportation environment, reduce private car travel demand, and steer the sustainable development of urban transport.

The trained random forest classification model is then applied, with projected predictor variables to predict urban area classifications for future years. By examining the future predicted changes for the four area types, several notable observations can be made (Fig. 8). From 2015 to 2060, the area of pedestrian-friendly districts is expected to grow from 37,622 sq km in 2015 to 44,890 sq km in 2060, representing a growth of 19%. On the other hand, the area of car-centric districts is projected to increase by 72,384 sq km, marking a substantial growth of 73%. It is evident that urban growth, following the development trajectory of 2015, is predominantly characterized by the expansion of car-centric areas.

Fig. 8: While area types undergo mutual conversions, their transition patterns differ markedly.

This figure selects the years 2015, 2030, and 2060 to provide a more detailed view of the transformations between the four area types across these three years.

To effectively navigate the evolving landscape of urban spatial development strategy, it is crucial that we identify the built-environment and regional economic characteristics that predispose a specific area to transform into a pedestrian-friendly district during the urban renewal. By calculating the Z-scores of each characteristic across different types of change, we identified characteristics that are significantly higher or lower than the average for each type of change (as illustrated in Fig. 9).

Fig. 9: Built environment and economic factors exert distinct influences on the transformation of different area types.

This figure shows the Z-scores for different predictor variables by change type.

An analysis of area changes of different area types between 2015 and 2060 reveals that pedestrian-friendly districts in traditional urban centers are expanding, though modestly. While many transition zones have been converted into pedestrian-friendly districts, ~27% of those identified as pedestrian-friendly in 2015 have shifted toward car-centric forms. Many of these newly car-centric districts are found on the suburban fringes of small and medium-sized cities or within the cores of more remote urban centers. Population density and local affluence level emerge as the most critical factors in sustaining pedestrian-friendly districts, suggesting that urban renewal efforts should prioritize preserving the economic and demographic vitality of traditional city centers. Otherwise, particularly in the context of population decline, these areas may shift toward other typologies, thereby undermining their role in promoting sustainable mobility. Transition zones account for a relatively small share of total area and exhibit a more balanced pattern of transformation, although earlier-developed zones tend to shift toward car-centric districts.

Furthermore, under current urban development trajectories, car-centric districts are projected to remain the dominant spatial form moving forward. Reducing dependence on automobile-based travel will require targeted improvements in the design dimension of the built environment. Notably, ~11% of car-centric districts have transformed into pedestrian-friendly districts, primarily in the suburban areas of large and mega cities.

Additionally, a substantial share of other non-urban built areas—41% by 2030 and 20% by 2060—have transitioned into car-centric districts. These non-urban built areas primarily comprise discontinuous settlements and built-up land within peri-urban clusters, including rural villages and remote town centers. Future shifts in transport modes in these areas are closely tied to broader lifestyle transitions. Historically, these areas have existed as loosely connected suburban clusters—such as township and village-level centers—where self-sufficiency dominates and access to services and transport remains largely local. In future projections, many of these regions are expected to shift toward car-centric forms, driven by the convenience of motorized transport, which expands the spatial reach of both daily services and mobility. As a result, these areas become increasingly car-dependent and more likely to transition into car-oriented regions. Land use diversity plays a relatively influential role in this transition. To reduce automobile travel demand in these areas, urban planning should prioritize enhancing land use diversity, thereby addressing local service provision gaps that emerge alongside rising living standards.

Our comparative analysis demonstrates that car-centric districts are steadily expanding under current trends, while pedestrian-friendly districts are marginally increasing. The expansion of car-centric districts suggests that, although mixed-use planning is widely acknowledged as a key strategy for urban sustainability⁵⁶, it is often sidelined in the planning of new towns and satellite cities. These development modes often prioritize ‘regional functionality’—such as boosting economic output or creating growth hubs in peripheral areas—leading to large, homogeneous built forms that deprioritize livability and sustainability⁵⁷. As regional disparities in macroeconomic development continue to narrow, it remains critical to redesign and enhance the built environment at the micro scale to shift automobile-dependent travel behavior toward more sustainable mobility patterns. As the pace of rapid urbanization slows, China has to shift from large-scale, expansion-driven urban development to a human-centered approach that places greater emphasis on urban renewal under the “New-type Urbanization” strategy⁵⁸.

Discussion

This research introduces an innovative perspective that effectively bridges the gap between macro-level visions for mitigating climate change and micro-level built environment design, and aligning with local urban renewal initiatives. It demonstrates how micro-level local actions, such as improving the ‘5D’ aspects of the built environment, can significantly propel the realization of macro level goals, while simultaneously highlighting the spatial variations in the effects of these endeavors.

Reducing private car travel stands as a cornerstone of most transportation emission reduction goals^59,60. The Avoid-Shift-Improve (ASI) framework is widely adopted as a strategic approach to achieving sustainable transport transitions¹. The framework aims to minimize unnecessary trips and reduce overall travel demand—including urban planning interventions⁶¹—to mitigate the transport sector’s environmental and resource impacts.

Macro-level integrated assessment models simulate sustainable transport transitions using SSP-based scenarios¹¹. These models offer travelers choices across different technologies and modes, while dynamically balancing travel costs, energy transitions, and emission targets through interactions with sectors such as energy, land use, and natural resources⁵³. The assumptions embedded in different SSP scenarios outline alternative, plausible futures, incorporating diverse patterns of population growth, urbanization, technological advancement, and development pathways^62,63. While SSP scenarios delineate future aspirations and trajectories with a focus on macro-level techno-economic pathways, they fall short in providing concrete, actionable strategies to attain these goals, particularly by neglecting the influence of local built environments.

Our study proposed an actionable approach that elucidates both the rationale and methods by which localities can contribute to reducing private car travel demand. Notably, our population and GDP forecasts are based on the SSP2 scenario. The TSD-LDV4W outcome for 2060, amounting to 1.24 × 10¹² passenger-kilometers, can be achieved through improvements in the 5D built environment. This outcome closely aligns with the SSP1 target forecasted by GCAM (1.17 × 10¹² passenger-kilometers) and significantly undercuts the SSP2 projection (2.12 × 10¹² passenger-kilometers).

These findings suggest that even if macro-level population and economic indicators follow a business-as-usual trajectory, micro-level improvements in the 5D built environment can still reduce private car travel demand and steer the transport sector toward a more sustainable path—closely aligned with the SSP1 scenario⁶³. In short, achieving the SSP1 scenario will require not only managing macro-level development indicators, but also leveraging the transformative potential of micro-level built environment interventions. This approach needs coordinated action across multiple scales and policy domains^28,64. At the macro level, this entails managing population growth and distribution through strategic urbanization policies, steering economic development toward less material-intensive service sectors, and maintaining strong climate mitigation commitments that create consistent policy signals across all sectors^65,66.

At the meso-level concrete implementation pathways include reforming urban planning standards and regulations⁶⁷, integrating land use and transport planning^27,30, prioritizing urban renewal over greenfield expansion⁶⁸, shifting transportation infrastructure funding from highway and parking construction toward public transit expansion (especially metros and BRT in large cities, regular bus service in medium and small cities)⁶⁹, pedestrian infrastructure improvements, and cycling infrastructure. Our results show that Distance to Transit improvements have substantial demand reduction effects across all area types.

At the micro level, we need targeted interventions by area type and implementation involves neighborhood-scale design interventions²⁹. Our machine learning analysis (Figs. 7–9) reveals that different area types require emphasis on different 5D dimensions. Pedestrian-friendly districts benefit most from maintaining density and regional economic vitality; car-centric districts require design improvements and destination accessibility enhancement^70,71; non-urban areas need land-use diversity improvements to reduce travel distances for daily needs^72,73.

Critically, these interventions must be sustained over decades. Our projections assume that the 2014–2024 rate of 5D improvements continues—this requires consistent policy commitment across multiple political cycles⁶⁶, and governance structures that prevent backsliding toward car-oriented development during economic downturns²⁸.

As rebound effects from lower travel costs may counteract emission reductions from vehicle electrification and automation^74,75,76, macro-level transport transition models must account for the influence of micro-scale travel environments on behavior⁷⁷. Macro-level integrated assessment models typically estimate transport service demand based on total travel costs, including fuel and non-fuel expenditures, occupancy rates, wage levels, and travel speeds⁵³. However, substantial regional disparities in travel costs—marked by complex and dynamic fluctuations—undermine model generalizability. Moreover, outdated base-year cost calibrations introduce uncertainty into models’ ability to reflect recent socio-economic shifts and evolving mobility trends⁷⁸. Depending on the direction of actual changes, this could lead models to either overestimate or underestimate future travel demand. For example, if real-world travel costs have declined faster than calibrations assume (due to ride-sharing or autonomous vehicles)⁷⁹, models may underpredict demand. This bidirectional uncertainty underscores the value of our approach, which explicitly incorporates built environment changes rather than relying solely on cost-based behavioral responses^27,80.

These models also assume that consumers passively choose among competing technologies, without influencing the attributes of those technologies themselves⁵³. These assumptions overlook the bidirectional feedback between consumer behavior and mobility systems, and disregarding the interaction between technology uptake, industrial dynamics, and evolving consumption patterns. In the case of electrification, technological progress promotes EV adoption, triggering scale effects in vehicle manufacturing, lowering production costs, and reducing travel expenses—thus reinforcing adoption and intensifying rebound effects⁷⁴. This feedback loop contrasts with conventional model assumptions, which posit that rising income drives a preference for faster modes, thereby increasing service costs.

These model limitations motivate our integrated approach that explicitly accounts for built environment dynamics^81,82. Conventional IAMs assume that as vehicle costs decline (through electrification, automation, economies of scale), consumers simply increase their travel consumption, generating rebound effects that partially offset emission reductions from technological improvements⁸³. This assumption treats built environment characteristics as static background conditions rather than as dynamic policy variables⁸⁴. However, micro-scale improvements to the built environment alter this trajectory. Even amid declining future vehicle costs, enhancements in the 5D built environment can significantly reduce private car travel demand by addressing the fundamental attractiveness and feasibility of non-auto modes. When walking, cycling, and transit become more convenient through better urban design, such as shorter distances to destinations, the relative attractiveness of driving may decline even if the absolute cost of driving also declines. Integrating these improvements into model assumptions better captures the complex interplay between industrial development, technological innovation, consumer behavior, and urban form^84,85.

Cities play a pivotal role in mitigating the environmental impacts of transport systems⁸⁶, a role that extends beyond improvements to the 5D built environment^59,87. It also involves implementing soft local policies under the purview of urban authorities^88,89. For example, cities can integrate low-carbon and sustainability principles into new urban development plans and urban renewal process, and promoting the adoption of autonomous driving and shared mobility to reduce transport emissions⁹⁰. Enhancing transport sustainability has received growing attention in the context of urban renewal⁹¹, given its influence on carbon emissions and climate change through both mobility patterns and the built environment. Other studies suggest that, relative to business-as-usual scenarios, changes in urban form could reduce urban transport-related greenhouse gas emissions by up to 25% by mid-21st century^1,92,93. This magnitude of demand-side mitigation potential aligns with our research on urban form impacts. The convergence of our China-specific findings with international evidence strengthens confidence that built environment interventions represent a robust, generalizable strategy for transportation demand management across diverse national contexts. Despite its inherently local nature, the transition towards sustainability should be steered by a global outlook, which provides a contextual framework, enabling local stakeholders to grasp the wider ramifications of their decisions.

Our simulation findings provided compelling evidence that local initiatives are essential for cultivating sustainable practices and attaining overarching macro-level goals set by the United Nations^1,32,94. Regions at different stages of economic development show significant variation in implementing sustainable transport policies, infrastructure provision, travel patterns, and public attitudes towards sustainable transport modes. As national economic disparities diminish from 2015 to 2060, the influence of the local built environment on travel behavior patterns becomes increasingly decisive. Among the spectrum of local policies, especially in the process of urban renewal, neighborhood-scale urban design and the augmentation of supporting infrastructure emerge as paramount. This community-focused strategy not only grants greater adaptability but also curtails the trial-and-error costs linked to extensive overhauls, thereby mitigating the uncertainties stemming from advancements in intelligent automotive technologies and the rise of remote work models^95,96.

In our projections, despite the continued expansion of car-centric districts, both the total and average values of TSD-LDV4W decline significantly—driven by continued improvements in the micro-level built environment—highlighting the critical role of enhancing built environment quality, even within car-centric urban forms. In contrast, the modest expansion of pedestrian-friendly districts underscores the need for urban renewal that is more aligned with sustainability transitions. Developing and restoring pedestrian-friendly areas is closely tied to the modes of renewal implementation. Renewal efforts should prioritize preserving and enhancing historic urban fabrics, avoiding large-scale demolition that undermines pedestrian-friendly design principles. Embedding sustainability into all aspects of spatial design is essential for shaping future modes of living.

The projected reductions in TSD-LDV4W are not deterministic outcomes but extrapolations based on historical trends in 5D improvements over the past decade, and remain subject to substantial uncertainty. It means that the actual trajectory will depend heavily on the proactive role of urban planning. Different area types would require tailored adjustments and improvements across specific dimensions of the 5D built environment. This underlines why our study places particular emphasis on the critical role of urban planning in shaping sustainable transitions.

Our findings on the importance of the 5D built environment align with studies from many European cities, which often feature established pedestrian-centric cores that inherently suppress car demand compared to the more auto-centric urban sprawl common in the United States^16,17. However, China’s context is unique, involving a scale and speed of urbanization not seen in Europe^97,98. Furthermore, its challenges differ from many cities in the Global South, which may face different constraints such as informal settlements and rapid, unplanned motorization⁹⁹. Our study, therefore, adds a critical perspective on how a nation undergoing a massive, state-guided transition can leverage built environment design—a lesson that could be adapted for rapidly urbanizing regions in the Global South seeking to avoid car-dependent lock-in.

In GCAM’s framework, the model endogenously represents modal competition. Consumers choose among private vehicles (gasoline, hybrid, electric) and non-private modes like buses, rail, and non-motorized options. Their choices are based on generalized costs, including monetary (purchase and operating costs) and time costs (travel speed and schedule convenience)⁸¹.

As EVs are deployed under penetration constraints, lower operating costs reduce the generalized costs of private vehicle travel, creating a rebound effect that increases demand^100,101. But GCAM scenarios assume continued transit improvements, maintaining its competitiveness despite private vehicle cost changes. The net effect shows 2–9% TSD-LDV4W reductions when EVs are deployed compared to reference scenarios, indicating rebound effects are partially offset by broader system adjustments.

Our micro-level built environment analysis shows local urban form can offset rebound effects from declining vehicle costs²⁷. Five-dimensional built environment improvements can achieve 21% demand reductions, far exceeding those from EV deployment alone in GCAM projections¹⁰². Better built environments make non-auto modes more attractive regardless of private vehicle costs^103,104. High-density neighborhoods, diverse land-use, well-connected streets, high destination accessibility, and good transit provision increase non-auto mode attractiveness. These improvements affect fixed costs and qualitative experience of mode choice, not just marginal per-trip costs^105,106. So, built environment improvements can reduce private vehicle demand even when vehicle electrification and automation lower driving costs.

Urban-rural disparities offer insights into EV-transit-built environment interactions. In urban areas with extensive public transport, EV deployment has modest modal shift impacts as high-quality transit remains competitive despite lower private vehicle costs¹⁰⁵. Urban built environment characteristics support multiple mobility options, and policies can maintain transit competitiveness¹⁰⁷. In rural areas without viable public transport, EVs become the default motorized mode for long-distance trips^108,109. Poor rural built environment quality means cost reductions won’t enable walking or cycling for long-distance needs. This explains higher TSD-LDV4W intensity in rural areas despite lower population densities¹¹⁰. Rural mobility sustainability requires different strategies, such as improved rural transit, clustered rural settlement patterns, and enhanced digital services¹¹¹.

While our study employs established frameworks—including IAMs (GCAM), the SSP scenarios, and the 5D built environment framework—its contribution lies in the novel integration and application of these tools to address a critical gap in sustainability transition research. Previous studies using IAMs typically operate at national or regional scales, treating transportation demand as an aggregate function of population, GDP, and fuel prices, without spatial differentiation or explicit consideration of urban form^112,113. Conversely, built environment research has largely focused on individual cities or metropolitan areas, demonstrating local impacts on travel behavior but failing to connect these findings to global climate scenarios or national decarbonization pathways^84,114.

Our integrated framework addresses three key limitations in existing research. First, we have considered the spatial granularity in scenario analysis. By downscaling IAM projections to 1 km resolution and incorporating spatially explicit built environment characteristics, we enable policymakers to understand not just how much transportation demand might change nationally, but where these changes will occur and what local factors drive them. Second, we have made the quantification of micro-macro linkages. Rather than asserting that “built environment matters” based on city-level case studies, we quantify precisely how much micro-level interventions can contribute to macro-level goals, providing actionable metrics for policy evaluation. Third, we have used dynamic transition modeling. Our machine learning approach projects how different urban area types evolve over 45 years under realistic development trends, revealing that current trajectories favor car-centric expansion (73% growth) over pedestrian-friendly development (19% growth)—a finding that challenges assumptions underlying many sustainability plans¹¹⁵.

These contributions are particularly valuable for rapidly urbanizing nations facing decisions about urban expansion versus renewal, and for international climate policy discussions about the role of demand-side interventions in achieving Paris Agreement targets.

Our study has the following limitations. First, for the 1 km resolution grid data covering the entire country, we utilized multiple data sources. However, due to the temporal limitations of the data, we selected the most available data from existing databases. This selection includes using 2019 land use diversity data retroactively for 2015, as well as utilizing OpenStreetMap (OSM), an open-source, voluntarily provided geographic data platform (i.e., Volunteered Geographic Information), for bus station and road network data. Although OSM has become a widely used dataset for global geographic analysis¹¹⁶, its data quality is inevitably affected by issues such as inconsistent standards and delayed updates, as it relies on multi-source geographic data provided by the public. The OSM data for China also shows regional imbalances and variations across different spatial scales^116,117. We advocate for the public release and accessibility of more authoritative, standardized, and comparable data that can be obtained and compared at the same time.

Second, although this study develops BEAI to estimate TSD-LDV4W in each 1 km × 1 km grid cell, this is based on results derived from prior research rather than actual collected data. The machine learning approach used to estimate TSD-LDV4W in this study is particularly valuable for its integration of established variable importance relationships from previous research with the spatial proximity effects inherent in geospatial analysis at the level of individual grid cells. The advantage of the machine learning approach is that it can directly provide predictions without explicitly calculating regression coefficients, thus omitting this step. Moreover, this modeling framework is highly scalable, meaning that when more variables or more complex nonlinear relationships need to be considered in the future, the existing analytical framework can be directly applied. An ideal source of TSD data would be real-world travel data directly measured from actual mobility, which would replace the dependent variable derived from BEAI in this study and be input into the machine learning model to uncover nonlinear relationships. However, high-resolution transportation demand data is difficult to obtain¹¹⁸. Some scholars have used nationwide mobility survey data, measuring transportation demand by the number of trips made by residents on working days¹¹⁹, while others have employed traffic assignment models calibrated for case studies based on traffic counts. Additionally, these scholars used a population-weighted approach to allocate travel demand from traffic zones during the morning peak to higher-resolution grids^118,120. In the context of smart city policy implementation, such data should be accessible³¹, but we are currently unable to obtain it from government agencies or mapping companies. Therefore, we hope that, in the future, more scholars will validate our research using more detailed real-world data.

Third, we emphasize that, although our study focuses on the 5D of the built environment and aims to demonstrate how improvements in these indicators can reduce TSD-LDV4W, previous research shows that the elasticity coefficient between the built environment and TSD usually does not exceed 0.5^28,29. This indicates that the reduction in TSD-LDV4W due to urban and rural planning and its impact on climate change mitigation is relatively small compared to policies that directly target car use (e.g., limiting private car ownership or mandating a switch to electric vehicles) and demographic factors. Our spatial allocation algorithm focuses on the structural influence of urban form by defining TSD-LDV4W as an environment-driven propensity. While this isolates the 'carbon lock-in' risk by characterizing the intrinsic car-dependence of different land-use patterns, it may underestimate absolute traffic counts in some areas. However, we believe that micro-level actions can aggregate into a macro vision, which aligns with our research objective: we are not constructing a framework to replace IAMs, but rather exploring, based on the results of the IAMs, how much additional contribution the local actions can make to the low-carbon transition of transportation, beyond socio-demographic factors, and how much carbon reduction potential exists.

Fourth, our operationalization of the 5D built environment indicators involves simplifications necessitated by the national scale and data availability constraints. Most notably, our ‘Design’ indicator uses road network density as a proxy, which captures network connectivity but not qualitative design features such as pedestrian infrastructure, street aesthetics, traffic calming measures, or the functional hierarchy of road segments. Similarly, our ‘Distance to Transit’ indicator counts transit stations without weighting by service frequency, capacity, or multimodal connectivity. While these simplifications are standard practice in large-scale spatial analysis and have demonstrated validity in prior research, they inevitably miss nuances that detailed local studies could capture. Future research could enhance these indicators by incorporating additional data sources—such as street-level imagery for sidewalk quality assessment, transit schedule data for service frequency, and planning documents for functional road classifications—where such data become available at national scales.

Conclusions

This study fills a critical gap in transportation sustainability research by quantitatively demonstrating how micro-level improvements in built environment can contribute to achieving macro-level climate goals in China’s transportation sector. The key empirical findings are:

First, substantial demand reduction potential exists through built environment interventions. Our spatially explicit analysis projects that improvements in the 5D built environment characteristics (Density, Diversity, Design, Destination Accessibility, and Distance to Transit) can reduce national TSD-LDV4W by 21% from 2015 to 2060 (from 1.58 × 10¹² to 1.24 × 10¹² passenger-kilometers). This reduction is close to the ambitious SSP1 scenario outcome (1.17 × 10¹² passenger-kilometers), indicating that local urban planning actions can achieve reductions in travel demand by car comparable to the most sustainable development pathway.

Second, significant regional disparities characterize China’s transportation transition pathways. Analysis at provincial reveals that by 2060, inter-provincial economic development disparities will diminish as determinants of transportation demand, with the correlation between population-weighted and GCAM-China projections increases from 0.63 to 0.73. However, urban-rural differences in the quality of the built environment are emerging as increasingly decisive factors, with rural areas exhibiting a 2.62% higher mean TSD-LDV4W than urban areas, even though urban areas account for 75.97% of national total demand.

Third, current urban development trajectories favor car-centric expansion. Our machine learning projections indicate that if the trends from 2014 to 2024 continue, the area of car-centric districts will grow 73% (from 99,000 to 171,000 km²) by 2060, while pedestrian-friendly districts will expand by only 19% (from 37,622 to 44,890 km²). This trajectory is in conflict with sustainability goals and highlights the urgency of reorienting urban renewal policies toward compact, mixed-use development patterns.

Fourth, different area types require tailored built environment interventions. SHAP analysis reveals that pedestrian-friendly districts rely most critically on maintaining density and regional economic vitality, whereas car-centric districts benefit primarily from design improvements and enhanced destination accessibility. These differential sensitivities indicate that standardized planning approaches are inadequate; context-specific strategies must address the specific built environment deficits of different urban form types.

These findings demonstrate that bridging global climate ambitions with local planning initiatives requires the explicit integration of built environment considerations into transportation demand forecasting and policy design. The substantial demand reduction potential that we have quantified provides empirical support for giving priority to urban form interventions alongside vehicle technology transitions in comprehensive decarbonization strategies.