Exploring the Spatial Coupling Characteristics and Influence Mechanisms of Built Environment and Green Space Pattern: The Case of Shanghai

Abstract

Urban expansion will squeeze the green space system and cause ecological fragmentation. The question of how to expand cities more scientifically and build eco-cities has become an important topic of sustainable urban construction. This paper takes Shanghai as a research case. A deep neural network combined with an attention mechanism model measures the comprehensive level of the built environment and green space pattern of urbanization and quantitatively analyzes the coordinated relationship between the two using the coupled degree of coordination model. Subsequently, the K-Means clustering model was used for spatial clustering to determine the governance and construction directions for different spatial areas and was, finally, combined with the LightGBM model plus SHAP to analyze the importance and threshold effect of the indicators on the degree of coupled coordination. The results of the study show that (1) the core area of the city shows a high state of coordination, indicating that Shanghai has a better green space construction in the central city, but the periphery shows different imbalances; (2) three different kinds of areas are identified, and different governance measures as well as the direction of urbanization are proposed according to the characteristics of the different areas; and (3) this study finds that the structural indicators of the built environment, such as Average Compactness, Weighted Average Height, and Land Use Diversity, have a significant influence on the coupling coordination degree and have different response thresholds. The results of the study provide theoretical support for regional governance and suggestions for the direction of urban expansion for sustainable urbanization.

1. Introduction

With the rapid advancement of China’s economy, the national urbanization process has accelerated dramatically [1]. Large-scale urban expansion has profoundly reshaped land use patterns [2], driving a high degree of built environment agglomeration. However, past urban development has often failed to reconcile the tension between spatial expansion and ecological sustainability [3,4], resulting in ecological degradation and environmental pollution [5,6]. Unregulated and rapid urban sprawl frequently leads to the fragmentation, marginalization, and functional weakening of urban green spaces [7,8,9]. Haase [10] demonstrated that urban green space exhibits pronounced ecological vulnerability in the face of expansion, particularly with substantial ecosystem service losses in peripheral zones. While urbanization serves as a key driver of economic and social progress, enhancing infrastructure and public welfare, it concurrently intensifies resource depletion and environmental stress [11]. Consequently, a growing body of research has emphasized the ecological impacts of urban spatial growth and the imperative to identify synergetic pathways between urban development and ecological conservation [12,13]. From an urban ecological perspective, green space systems play a strategic role in enhancing urban resilience and mitigating climate-related risks [14]. Therefore, identifying the coupled spatial patterns and interactive mechanisms between the built environment and green space is not only a critical issue in urban spatial governance but also a key indicator for guiding sustainable urban expansion.

The relationship between urbanization and the ecological environment remains a central concern in the discourse on sustainable urban development. A deeper exploration of the interactive mechanisms between the built environment and green space pattern serves as a critical entry point for understanding the dynamics of urban spatial expansion and ecosystem evolution [15]. In recent years, growing scholarly attention has focused on the impacts of urbanization on ecosystem patterns and service functions [16,17], with extensive research addressing issues such as land cover transformation induced by urban expansion [18], landscape fragmentation [19,20], and changes in ecosystem services [21,22]. Within these studies, the built environment is frequently employed as a key indicator to characterize urban spatial morphology and its externalities [23]. Meanwhile, green space configuration is widely regarded as a visible manifestation of urban ecosystem health and sustainability [24], as its spatial distribution, structural composition, and ecological connectivity directly influence ecosystem services [25] and residents’ environmental well-being [26,27].

The spatial coupling between built-up areas and green space essentially reflects the dynamic balance—or potential conflict—between urban development and ecological preservation [28]. Unregulated expansion of urban spaces and the over-concentration of the built environment inevitably encroach upon green spaces and disrupt their spatial integrity [29]. In contrast, scientifically planned urban expansion and well-maintained green spatial patterns can enhance urban resilience and support the broader goals of sustainable urban development [30]. While existing studies have explored the coupling coordination between the built environment and green space from perspectives such as spatial matching [31], supply–demand coordination of ecosystem services [32], and land use response relationships [33], several critical gaps remain. These include (1) limited attention to the influence mechanisms and threshold effects of coupling factors; (2) insufficient consideration of spatial heterogeneity across different urban units; and (3) a lack of fine-scale, micro-level analysis. Therefore, investigating the coupling characteristics and driving mechanisms between the built environment and green space patterns—along with the identification of their spatial differentiation patterns—holds significant theoretical and practical value. It provides essential insights for advancing sustainable urban development and optimizing the synergy between ecological and built-up spaces [34].

In recent years, machine learning models have been increasingly applied to studies of urban spatial structure, remote sensing monitoring, and urban ecosystems, owing to their powerful nonlinear fitting capabilities and automatic feature extraction advantages [35,36,37]. In the context of weighting coefficient computation, Deep Neural Networks (DNNs) exhibit strong capacity for multi-layered abstraction and can effectively capture nonlinear interactions among complex variables [38]. By incorporating an attention mechanism, the model can autonomously learn and assign weights to each feature, thereby enhancing the representation power of key indicators.

For spatial classification, the K-Means clustering algorithm, as a classical unsupervised machine learning technique [39], is widely used to identify spatial units with similar characteristics at a micro-urban scale [40]. This allows for the segmentation of urban space into functionally distinct zones [41,42], providing strategic support for ecological zoning and planning governance [43]. In addition, the Light Gradient Boosting Machine (LightGBM), a state-of-the-art gradient boosting algorithm, is renowned for its high computational efficiency, scalability to high-dimensional data, and robustness to noise [44]. It is increasingly employed in the identification of urban system driving factors and predictive modeling [45]. When integrated with SHAP (SHapley Additive exPlanations), LightGBM facilitates interpretable analyses of model outputs by quantifying the marginal effects [46], threshold responses [47], and interaction dynamics [48] of different influencing factors [49], thereby providing a more explainable decision-making foundation for the coordinated optimization of built and ecological environments.

Shanghai serves as one of the most representative megacities in China [50]. Since the onset of reform and opening-up, the city has experienced accelerated urbanization, with rapid expansion of built-up areas [51], continuous transformation of land use structures [52], and increasing compression and fragmentation of green space systems [53]. These trends have produced a complex pattern of spatial coupling between urban development and ecological systems [54,55], making Shanghai an ideal case for examining the interactions between built environments and green space patterns under conditions of intensive urbanization. Under the national strategies of “ecological civilization” and “resilient cities,” Shanghai has actively advanced its green transition, urban resilience renewal, and ecological city-building initiatives [56,57]. The city has accumulated substantial experience in ecological corridor reconstruction and urban boundary control, demonstrating strong policy-driven and governance-responsive characteristics [58]. Consequently, Shanghai not only provides theoretical insight and empirical depth for studies on urban construction and green space patterns, but also offers valuable lessons for other large-scale metropolitan regions.

To summarize, this paper takes Shanghai as a research case and focuses on the following three scientific questions: (1) How can the spatial characteristics of the built environment and green space patterns in the urbanization process, as well as the coupling and coordination relationship between the two, be identified? (2) Based on the spatial relationship between the built environment and green space patterns, how can urban spatial development policies be formulated in a manner appropriate to local conditions? (3) What are the differences in the importance of different built environment and green space pattern indicators for regional coupling coordination, and what is the threshold effect? In order to solve the above problems, this paper uses a deep neural network (DNN) combined with the Attention Mechanism’s weight calculation method to realize the quantification of the comprehensive index of urban construction level and green space pattern, and uses the coupling coordination degree model to identify different coordination degree relationship regions. Meanwhile, the K-Means clustering method is used to classify urban spatial units. As well as combining the LightGBM model and the SHAP interpretation algorithm, the importance and threshold effects of different influencing factors are systematically quantified to reveal the key regulatory mechanisms of different indicators. Finally, differentiated urban space optimization strategies and spatial expansion directions for urbanization development are proposed to guide the construction of sustainable cities.

2. Materials and Methods

2.1. Study Area

Shanghai is located at the estuary of the Yangtze River in eastern China (30°40′–31°53′ N, 120°52′–122°12′ E) and serves as a major national hub and an international center for finance and commerce (Figure 1). Situated within the northern subtropical monsoon climate zone, Shanghai experienced an average annual temperature of approximately 18.5 °C and a total annual precipitation of about 1414.2 mm in 2023 [59]. These favorable climatic conditions provide essential ecological support for maintaining biodiversity and ecosystem services within the city’s green space system.

To explore the coupling relationship and spatial heterogeneity between the built environment and green space pattern at a finer scale, we partitioned the entire municipal area of Shanghai into 81,619 grid units, each measuring 300 m × 300 m, using ArcGIS spatial analysis tools.

2.2. Data Sources and Preprocessing

To accurately reflect its spatial heterogeneity, this study endeavors to obtain high-precision data. The boundaries of Shanghai are sourced from the National Geographic Information Public Service Platform (https://www.tianditu.gov.cn/), while the building vector data is obtained from the Figshare website. This data was measured by Zhang et al. between 2022 and 2024 and represents building vector data with a resolution of 0.3–1 m (0.5 m in the core area) [60]. The population data is sourced from the Landscan website for the year 2023, road data is sourced from the 2023 road vector data on the OSM website, POI data and bus stop data are sourced from the 2023 Gaode Maps, and land use data is sourced from the Zenodo website, provided by Li et al. as 1 m land use data for China [61]. Green space pattern data is based on 1 m urban land use data, calculated using the landscapemetrics package in R. The obtained or calculated indicator data is scaled and converted to 300 m by 300 m grid cells.

2.3. Construction of Research Indicators

To comprehensively assess the spatial coupling relationship between the built environment and green space patterns in cities, this study developed the Built Environment Index (BEI) and the Green Space Pattern Index (GSPI) based on an indicator system. The BEI reflects built environment characteristics such as building density, road traffic, and human activity within urban areas, measuring the intensity of urban development and the concentration of human activity. We developed a comprehensive set of BEI indicators based on the 5D framework proposed by Ewing and Cervero [62]. The GSPI is used to measure the spatial structure, scale, and connectivity of urban green spaces, reflecting the integrity and ecological functions of the green space system. This index is based on the objectives of spatial heterogeneity analysis [63]. All candidate indicators selected to reflect the GSPI have undergone multiple collinearity screening, and variables with a variance inflation factor (VIF) exceeding 5 have been excluded.

The final indicator system is summarized in

Table 1. Indicator system for the built environment and green space pattern, along with corresponding VIF values.

Indicators	No.	Variable	VIF
Built Environment Indicators (BEI)	1	Building Density	3.418191
	2	Weighted Average Height	1.430039
	3	Road Density	1.342832
	4	POI Density	1.471579
	5	Green Space Ratio	1.043532
	6	Population Density	2.111864
	7	Bus Stop Density	1.189187
	8	Average Compactness	2.710779
	9	Land Use Diversity	1.488116
	10	Water Body Ratio	1.869330
Green Space Pattern Indicators(GSPI)	11	Green Coverage Ratio	2.965984
	12	Largest Patch Index	1.584769
	13	Number of Patches	2.914436
	14	Shape Index	1.237455
	15	Aggregation Index	3.330684

. To further ensure the independence of variables used in clustering and modeling, we visualized the inter-variable correlation structure using a Pearson correlation heatmap (Figure 2). The results indicate that most variables fall within a moderate to low correlation range (between −0.5 and 0.5), suggesting minimal linear redundancy and acceptable discriminative capacity. Although some variables (e.g., building density and average compactness) exhibit moderate correlation, they remain within reasonable limits and do not pose significant multicollinearity concerns.

Among these, building density is calculated as the ratio of building floor area to the area of the unit grid, road density is calculated as the ratio of road length to the area of the unit grid, POI density is calculated as the number of POIs divided by the area of the unit grid, population density is converted from raw data to the unit grid, bus stop density is calculated as the number of bus stops divided by the area of the unit grid, and water body ratio is calculated as the ratio of water body area to the area of the unit grid. Land use diversity refers to the number of land use types within a unit grid. Weighted average height refers to the average building height within a unit grid, calculated using building area weighting to prevent extreme outliers from being exaggerated. Average compactness refers to the complexity of building shapes, with the specific calculation formula as follows:

$$C_{avg}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left({\displaystyle \frac{4\pi A_i}{P_i^2}}\right)$$

(1)

Theorem 1.

In the formula, $C_{avg}$ is the average compactness, $A$ is the building area, $n$ is the number of buildings within the unit grid, $A_i$ is the area of the $i$-th building, $P_i$ is the boundary length of the $i$-th building, and ${\displaystyle \frac{4\pi A_i}{P_i^2}}$ is the compactness of the $i$-th building (with a circle as the ideal type).

2.3.1. Attention Mechanism and Deep Neural Network (DNN) for Weight Calculation

To determine the weights of each indicator and construct the comprehensive indices (BEI and GLPI), this study integrates the attention mechanism into a deep neural network framework (DNN-Attention), following a data-driven and interpretable modeling approach [64]. Unlike traditional methods such as the Entropy Weight Method (EWM) or Principal Component Analysis (PCA), which rely on the distributional characteristics of variables, the DNN-Attention model captures the nonlinear relationships among high-dimensional indicators and adaptively learns their contribution to the overall performance.

The Coupling Coordination Model (CCM) is an effective method for characterizing the interaction dynamics among different subsystems within urban systems [65]. It has been widely applied to assess the co-evolution and synergy among natural, social, and economic systems [66].

In this study, the CCM is employed to quantitatively evaluate the degree of coordination between the Built Environment Index (BEI) and the Green Space Pattern Index (GSPI), with the Coupling Coordination Degree (CCD) being calculated to reveal the interaction state between urban development (built environment) and ecological space (green space pattern) [67]. The CCD is a composite metric that reflects both the intensity of interaction (coupling) and the level of harmonious development (coordination) between two subsystems. A higher CCD value indicates better synergy and balance between urban construction and green space, while a lower value implies spatial mismatch or lack of integration.

The model consists of three main computational steps:

Coupling Degree (C)

The coupling degree function is first established to quantify the intensity of interaction between systems. It is defined as:

$$C={⌈{\displaystyle \frac{U_1\times U_2}{{\left(U_1+U_2\right)}^2/4}}⌉}^{0.5}$$

(2)

Theorem 2.

Here, U₁ and U₂ represent the normalized composite indices of the built environment and green space patterns in Shanghai, respectively. Both range within [0, 1]; A higher value of C indicates a stronger degree of interaction between the two subsystems.

Comprehensive Benefit Index (T)

$$T=a\times U_1+b\times U_2$$

(3)

Theorem 3.

In this study, both the built environment and green space patterns are assumed to hold equal importance in the urban system. Therefore, the weights are set as a = b = 0.5, reflecting a balanced development perspective.

Coupling Coordination Degree (D)

The final coordination level D is determined by integrating the coupling degree C and the comprehensive benefit index T:

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

(4)

Theorem 4.

The value of DDD, also known as the coupling coordination degree (CCD), lies within [0, 1]. A value closer to 1 indicates a higher level of coordinated development between the built environment and the green space pattern. CCD provides a quantitative basis for understanding the synergetic evolution of urban spatial and ecological subsystems.

2.3.2. K-Means Clustering Analysis

The K-Means algorithm is an unsupervised learning method based on distance metrics, capable of classifying spatial units based on their similarity in high-dimensional feature space. It iteratively minimizes intra-cluster variance while maximizing inter-cluster variance, ultimately assigning each spatial unit to the most similar cluster. In this study, we do not reduce the dimensionality of the original indicators. Instead, we directly input the full set of built environment and green space pattern indicators into the clustering model, preserving as much original spatial coupling information as possible. This high-dimensional approach enables finer-grained detection of coupling relationships. Then, we went through the elbow coefficient analysis, and the results showed that the K value decreases sequentially in the interval from 2 to 10 and the SSE decreases significantly for K = 3–4. Combined with the results of silhouette coefficient analysis, 2–3 values are high-value intervals, so K = 3 is a statistically reasonable and spatially interpretable choice.

Three coupling types were identified:

• Dispersed Natural Zone: Low in both urban and ecological development; includes scattered natural areas like wetlands and grasslands with weak spatial integration.
• Urban Dominated Coordination Zone: High built environment intensity with moderate green land support; coordination is maintained through compact development and vertical greening.
• Green-Oriented Transition Zone: Rich in green space but underdeveloped in urban infrastructure; low coordination reflects untapped synergy potential.

2.3.3. Nonlinear Modeling and Interpretation

LightGBM, a gradient boosting framework developed by Ke et al. [68], is known for its high efficiency and strong robustness to outliers and noise. It is especially suitable for modeling complex urban systems with many variables and high-dimensional features. However, as machine learning models are often treated as “black boxes,” this study adopts SHAP (SHapley Additive Explanations) to interpret the marginal contribution of each input variable. Derived from cooperative game theory, the SHAP method attributes feature importance based on Shapley values, allowing for a consistent and transparent explanation of model predictions. Through SHAP value ranking and distribution analysis, the key drivers, variable interactions, influence strengths, and sensitivity intervals can be clearly identified.

Before selecting the final prediction model, we compared linear regression, Random Forest, LightGBM, and XGBoost using the study data. As shown in

Table 2. Performance comparison of four machine learning models.

	Linear Regression	Random Forest	LightGBM	XGBoost
MAE	0.005321	0.002167	0.002086	0.002123
MSE	0.000088	0.000027	0.000024	0.000026
RMSE	0.009384	0.005216	0.004881	0.005097
R2	0.973189	0.991717	0.992746	0.992090
Residual plot

, LightGBM demonstrated the highest R² value and the lowest errors in other evaluation metrics, confirming its superior performance. Residual plots for LightGBM and XGBoost also showed stable, symmetric distributions centered around zero across the fitting range. Based on a comprehensive evaluation of all performance indicators, LightGBM was chosen as the modeling tool for further analysis.

2.4. Research Technical Framework Roadmap

This paper first constructs the Built Environment Index (BEI) and the Green Space Pattern Index (GSPI), then uses the DNN + Attention model to calculate their composite index. Based on their spatial coupling logic relationship (Figure 3), we use the coupling coordination degree model to calculate their coordination degree and employ the K-Means clustering model to identify three distinct regions. By combining the composite index and coordination degree, we analyze the characteristics of different clustering regions. Finally, we use the lightGBM + SHAP model to analyze the influence mechanisms of each indicator on coordination. Based on all research results, we propose spatial governance measures and urban development directions for different cluster regions (Figure 4).

3. Results

3.1. Spatial Distribution Characteristics of the Comprehensive Indices for Built Environment and Green Space Pattern

After developing the computational code for the DNN-Attention model, we utilized Python 3.12.7 to calculate the composite indices for both the built environment and green space pattern. The weight calculation results are presented in

Table 3. Weight coefficients of built environment indicators.

Variable	Attention_Weight
Building Density	0.09701068
Weighted Average Height	0.111377545
Road Density	0.082808815
POI Density	0.099431455
Green Space Ratio	0.10096676
Population Density	0.096694775
Bus Stop Density	0.11039864
Average Compactness	0.104573995
Land Use Diversity	0.08181346
Water Body Ratio	0.1149288

and

Table 4. Weight coefficients of green space pattern indicators.

Variable	Attention_Weight
Green Coverage Ratio	0.212185
Largest Patch Index	0.269961
Number of Patches	0.121384
Shape Index	0.185129
Aggregation Index	0.211341

. Among the built environment indicators, the top five contributors are the proportion of water bodies, weighted average building height, density of bus stops, average compactness, and proportion of green space, each with a weight coefficient of around 0.1. For the green space pattern indicators, the Largest Patch Index (LPI) holds the highest weight, followed by the Green Coverage Ratio (GCR), and then the Aggregation Index (AI), with all three indicators assigned weights exceeding 0.2.

In the weight analysis of building environment indicators, the water body ratio exhibits the highest attention weight (approximately 0.115), indicating that water bodies have a significant influence on the comprehensive assessment of urban built environment quality. This phenomenon is primarily due to water bodies being a key component of urban ecosystems, not only enhancing the ecological service functions of cities but also significantly influencing the surrounding building forms and spatial layout. The presence of water bodies is often associated with higher environmental quality and livability, hence their higher weighting in deep neural network models.

We integrated a composite index calculation module into the code, and after normalizing the resulting composite indices, imported them into ArcGIS 10.8.1 software for visualization using the natural breaks classification method, divided into eight levels. The results are shown in Figure 5 and Figure 6. The spatial variation in the composite Built Environment Index is pronounced, with high-value areas largely coinciding with the built-up urban zones [69]. The highest index values are concentrated in the northeastern city center, exhibiting a distinct clustering pattern consistent with the direction of urban development [70]. In contrast, the green space pattern in Shanghai predominantly exhibits moderate index values across most regions. High-value areas are distributed mainly in the peripheral suburban zones in the southwest and northeast. The high-value zones in the southwest are fragmented, while those in the northeast display a more linear spatial pattern. The central urban areas show a mixture of moderate and lower index values, with moderate values predominating.

3.2. Spatial Clustering Patterns and Coupling Relationship Identification

K-Means, an unsupervised machine learning algorithm, was applied using Python to classify the spatial units into three distinct clusters: the Dispersed Natural Zone, Urban Dominant Coordination Zone, and Green-Oriented Transition Zone (Figure 7). Visualization conducted in ArcGIS revealed that the Dispersed Natural Zone is scattered primarily across the urban periphery, with some clusters observed in the eastern, western, and northern outskirts, while the remaining areas exhibit a dispersed distribution. The Green-Oriented Transition Zone is also located mainly in suburban areas, showing clustered patterns in the northern region and a tendency for aggregation in the south, but overall remains relatively dispersed. Comparison with the spatial distribution of the Green Space Pattern Index indicates a strong spatial correspondence between high-value green space areas and this zone. The Urban Dominant Coordination Zone largely overlaps with areas exhibiting high Built Environment Index values, predominantly within the built-up urban core, where clustering is more pronounced compared to the Built Environment Index alone. Furthermore, using the coupling coordination degree (CCD) model to analyze the coordination between the built environment and green space patterns, the CCD values were classified into three levels via the natural breaks method: low coordination (0–0.240560), medium coordination (0.240561–0.308737), and high coordination (0.308738–0.649624). The spatial distribution of these levels, shown in Figure 8, indicates that high coupling primarily concentrates within the Urban Dominant Coordination Zone, with a central cluster radiating outward. Medium coordination areas are highly fragmented and predominantly located adjacent to high coordination zones. Conversely, low coordination zones form extensive clusters mainly in the northern and southern parts of Shanghai, largely coinciding with what we define as the Green-Oriented Transition Zone. The low level of harmonization in this region is mainly due to the fact that the ecological resources are abundant and the built environment is not yet developed, reflecting an “unsynergized” situation that contains a huge potential for ecological development and urban expansion, and therefore requires a different governance strategy.

3.3. Mechanism Analysis of Influencing Factors

3.3.1. Overall Feature Importance Analysis

This study employs the LightGBM model combined with SHAP values to quantify and rank the contribution of each feature to the coupling coordination degree (CCD). As illustrated in Figure 9, the feature exerting the greatest influence on CCD is average compactness, with a SHAP value of 0.018338, followed by weighted average height (0.010148) and land use diversity (0.009895). The remaining features, in descending order of importance, include road density (0.009449), water body ratio (0.007212), bus stop density (0.006814), population density (0.0068), building density (0.004580), POI density (0.004053), Green Space Ratio (0.003491), GCR (0.001023), SI (0.000096), AI (0.000049), LPI (0.000020), and NP (0.000016). Notably, built environment indicators demonstrate substantially greater influence on CCD, with SHAP values consistently exceeding 0.001, whereas green space pattern indicators, except for GCR, present negligible effects with SHAP values mostly below 0.0001. The SHAP bee swarm plot reveals a positive correlation between higher values of built environment indicators and increased coupling coordination degree.

3.3.2. Threshold Effects

After analyzing the global significance of indicators related to the built environment and green space patterns, we further investigated the threshold effects of all variables. The results indicate that most indicators exhibit a linear correlation with coupling coordination and have a positive impact on coupling coordination once they exceed specific thresholds. Specifically, the thresholds that positively influence coupling coordination are as follows: building density > 0.09, point of interest density > 0, bus stop density > 0, green space ratio > 0.02, population density > 216.23, average compactness > 0.2, land use diversity > 6, water body ratio > 0.07, GCR < 0.53, and SI < 1.29. Meanwhile, the LPI, NP, and AI indicators fluctuate around the zero point (Y = 0), exhibiting a non-linear impact on coupling coordination. Additionally, when the weighted average height > 0.01 and road density > 0, the impact on SHAP decreases, and a dispersed point distribution phenomenon occurs (Figure 10).

4. Discussion

4.1. Spatial Distribution of Built Environment and Green Space Patterns

From the perspective of spatial distribution characteristics, the composite index of the built environment exhibits a pronounced pattern of spatial concentration. High-value areas are predominantly clustered within the city center and core urban districts, displaying a clear agglomeration pattern radiating outward from the urban core. This spatial configuration aligns with the centripetal concentration of population, industry, and transportation resources during urbanization [71,72,73] and is particularly driven by the dual forces of historical urban layouts and land-intensive development policies [74,75,76,77]. The urban core manifests spatial characteristics of high building density, intensive land use, and concentrated functional structures.

In contrast, the composite index of green space patterns is characterized by a more dispersed spatial distribution, predominantly exhibiting medium values across the urban landscape. High-value green space zones are primarily situated at the urban periphery, notably in the southwestern suburbs and the northeastern waterfront areas. The fragmented high-value patches in the southwest may be attributed to topographic constraints and higher ecological sensitivity, whereas the linear high-value distribution in the northeast likely reflects policy-driven ecological control corridors along rivers and coastal zones. Furthermore, with respect to coupling degree, green space pattern indices at medium-to-high levels appear to represent an optimal balance for harmonizing urbanization and ecological environment—corroborating findings by Kamble et al. [78] regarding ideal urban green space accessibility densities.

This spatial heterogeneity reflects a land-use logic dominated by urban expansion accompanied by green space retreat during city development [79,80,81]. The pattern of clustered high built environment indices in the core area juxtaposed with peripheral green space highs essentially embodies the spatial contradictions driven by rapid urban expansion and land commodification [82,83]. On one hand, dense development in the central city prioritizes economic and residential demands, marginalizing green spaces [84]; on the other hand, ecological policies promote the formation of “ecological belts” or “green wedges” on the urban fringe [53]. Recent policy measures have partially alleviated tensions between green space patterns and built environments. For example, Shanghai’s “Master Plan for Urban Development (2017–2035)” emphasizes the creation of a green, open ecological network, ensuring that ecological land (including green plazas) constitutes no less than 60% of the municipal land area, establishing a multi-tiered, interconnected, and multifunctional ecological spatial system described as “double rings, nine corridors, and ten zones” to mitigate development pressure. Moreover, in peripheral zones, regulatory planning and ecological redline protections safeguard ecological spaces from encroachment, thereby maintaining the connectivity and integrity of green spaces [85].

4.2. Coupling Relationships and Spatial Clustering

A comparative analysis of spatial clustering zones and coupling coordination degree reveals a notable spatial congruence and logical linkage between the two. The clustering results identify three distinct regional typologies that collectively reflect the coordination status of Shanghai’s grey–green urban spaces.

Urban Dominant Coordination Zones are mainly concentrated in core built-up areas. These zones usually correspond to redevelopment areas or functional core areas and are characterized by concentrated planning resources and strong policy orientation [86]. These zones generally show high to moderately high values for the Built Environment Index, while the Green Space Pattern Index tends to be moderate to slightly below average. The relatively high degree of central coupling and coordination in this region indicates that strong green support has been achieved through multi-level green space, vertical greening, pocket parks, etc., under the conditions of dense urban development [87,88,89,90], but the degree of coordination in the periphery is lower than that in the center, and comparing the Built Environment Index with the Green Space Pattern Index, we know that the lack of green leads to a lower degree of coupling. Therefore, the region needs to build more green space and other strategies to improve its green support capacity [91,92].

Green-Oriented Transition Zones have lower built environment value but higher green space pattern indices, reflecting strong ecological reserve capacity. Their level of coupling and coordination is moderately low, and most are located in ecological buffer zones at the urban–rural transition zone or the edge of urban expansion. These areas have low urbanization levels and relatively intact and large-scale green space systems. However, due to lagging infrastructure and low development intensity, synergies between the built environment and green space systems have not yet been established. Whether these areas require urbanization development, the construction of ecological green belts, or measures to protect farmland, among other policies, requires comprehensive consideration. In London, the Metropolitan Green Belt policy has successfully contained urban sprawl and preserved ecological buffers. However, it also constrains housing supply and may redirect suburban growth, raising spatial equity and governance concerns. Debates persist over balancing green-belt protection with sustainable urban densification.

The Dispersed Natural Zone is the area with the smallest spatial extent and the most significant ecological dominant features among all types. Its distribution is highly fragmented and shows a pattern of fragmentation, mainly consisting of natural landscapes such as wetlands, grasslands, and inland water bodies. There are few human development activities in the area, and the green space pattern is structurally deficient, resulting in a low Built Environment Index (BEI) and Green Space Pattern Index (GSPI) and a correspondingly low coupling coordination degree (CCD). This area does not represent a typical dysfunctional urban area but is closer to an ecological conservation area with strong natural attributes and ecological buffer functions. Therefore, ecological protection, habitat connectivity maintenance, and environmental buffer zone construction should be the focus of planning to avoid incorporating it into conventional urban construction paths.

But, due to the single-year data limitation, the current spatial patterns reflect a static snapshot of coupling coordination zones. However, as cities evolve, especially under rapid urbanization and planning policy shifts, these zones may change significantly over time. The absence of temporal analysis limits the capacity to observe such transformation trajectories or long-term impacts of planning interventions.

4.3. Policy Recommendations

This study identifies the coupling coordination relationship and spatial heterogeneity between the urban built environment and green space patterns and delineates distinct cluster zones. These findings reveal the developmental shortcomings and potential foundations across different spatial units, offering valuable insights for ecological optimization, urban expansion management, the enhancement of grey–green space synergy, and sustainable urban development. Based on these results, we propose the following strategic recommendations:

4.3.1. Zonal Differentiation and Tailored Spatial Governance Mechanisms

• In the Urban Dominant Coordination Zone, due to its high coupling coordination, high-density built environment, and strong green space compensation characteristics, the future planning of the core area should prioritize enhancing the layered integration and multifunctional use of green infrastructure. In contrast, the surrounding areas with low coordination should focus on green space development. During the construction process, reference should be made to the threshold effects of indicator thresholds to set corresponding thresholds for each green space development, ensuring that construction achieves the optimal coordination range. Additionally, the promotion of vertical greening, rooftop gardens, and pocket parks is encouraged to facilitate micro-ecological renewal and quality improvement in densely populated urban core areas [93,94]. Simultaneously, protecting urban ecological red lines and achieving a sustainable development model that combines high-density built environments with green penetration is of critical importance. Taking Singapore as an example, the city implemented the “City in a Garden” strategy under extremely limited land resources and widely applied measures such as green roofs, three-dimensional green walls, and sky gardens. Since the launch of the “Sky Garden Incentive Programme” in 2009, Singapore has constructed over 100 hectares of green roofs and vertical greening as of 2023, effectively alleviating the urban heat island effect and improving air quality.
• The Green-Oriented Transition Zones possess substantial green space resources but relatively underdeveloped built environments, resulting in moderate coupling coordination. In this zone, emphasis should be placed on preserving green space patterns and integrating urban development with green space planning to avoid fragmented urban sprawl that undermines existing ecological configurations. Priority should be given to planning green space, protecting ecological land, and constructing ecological corridors to enhance its ecological reserve function and support a smooth transition towards green and low-carbon urban districts [95].
• The Dispersed Natural Zone is primarily composed of ecologically sensitive landscapes such as wetlands, grasslands, and water bodies, characterized by low built environment intensity and fragmented spatial distribution. Rather than reflecting development–ecology conflict, its low coupling coordination degree stems from minimal urban integration and functional disconnection. Therefore, planning in this zone should emphasize ecological protection over urban expansion. Strategic priorities include maintaining ecological integrity through strict land-use controls and redline enforcement, enhancing habitat quality via targeted ecological restoration, and improving spatial connectivity by integrating these dispersed patches into broader green space networks [96]. Where minimal development is unavoidable, eco-sensitive design and nature-based solutions should be adopted to ensure that ecological functions are not compromised. This approach ensures that the zone serves as a resilient ecological reserve, supporting regional biodiversity and landscape sustainability.

4.3.2. Urban Expansion Guidance and Institutional Support

We recommend establishing an integrated policy framework encompassing “coupling coordination–spatial classification–governance response,” incorporating coupling coordination degrees and spatial clustering outcomes into overarching urban renewal and spatial planning strategies. An indicator-based regulatory system can be implemented to establish zonal control benchmarks—such as maximum building density and minimum green space ratios—guided by the sensitivity and importance of key influencing indicators.

Notably, Urban Dominant Coordination Zones are predominantly located within central and adjacent high-density urban areas, where stable coordination between built environment and green space patterns has been achieved, reflecting mature urban functions and effective coupling mechanisms. Conversely, the Green-Oriented Transition Zones located in the southern and northeastern peripheries feature intact green space systems but currently underdeveloped built environments, representing significant potential for future urban expansion. However, the coupling coordination in the northeastern corridors and southwestern patches still requires improvement. These zones should be designated as strategic urban expansion belts, adopting an “ecological first, orderly development” approach to guide urban function spillover and population redistribution, while simultaneously strengthening green space preservation and green infrastructure construction. This approach aims to avoid repeating the mistakes of “build first, compensate later,” thereby facilitating a shift from incremental urban growth towards ecologically coordinated urbanization.

It is important to recognize that the coordinated relationship between the built environment and green space patterns is not only determined by physical spatial attributes but is also profoundly influenced by underlying institutional and planning mechanisms. For example, zoning regulations, ecological red lines, and strategic urban planning initiatives (such as green corridors and development boundaries) can significantly impact land use intensity and green space protection, thereby altering the spatial manifestation of coordinated integration. In some areas with higher coordination levels, this may not only reflect physical spatial matching but also the results of deliberate planning and control. Conversely, areas lacking systematic policy support may be more prone to green space fragmentation or ecological mismatch. Therefore, future research should consider incorporating institutional variables into coupling models to more comprehensively reveal the underlying drivers of spatial coordination mechanisms.

5. Conclusions

5.1. The Main Conclusions of This Sdy

This study first utilizes the DNN-Attention model to analyze the spatial characteristics of composite indicators of built environment and green space patterns in urban spaces. Secondly, it introduces the coupling coordination model and K-Means clustering method to classify and identify urban spatial units and analyze their coordination levels, thereby revealing the heterogeneity of urban spaces. Finally, by combining the LightGBM and SHAP analysis models, the dominant driving factors of coupling coordination are identified. It is found that structural indicators of the built environment (such as average compactness, weighted average height, and land use diversity) have a significant impact on the degree of coupling coordination, while the influence of green space indicators is relatively minor. The study draws the following main conclusions:

(1)

There are significant spatial differences between the built environment and green space patterns. High values of the Built Environment Index (BEI) are concentrated in central urban areas, while high values of the Green Space Pattern Index (GLPI) are predominantly distributed in peripheral ecological zones, revealing the typical contradiction between the expansion of urban grey infrastructure and the retreat of green spaces.

(2)

The coupling coordination degree (CCD) exhibits obvious spatial heterogeneity. Areas with high built environments and moderate green space patterns (such as urban core areas) demonstrate higher coordination levels, while areas rich in ecological resources but underdeveloped built environments show lower coordination levels. Peripheral areas with fragmented infrastructure and incomplete green space layouts exhibit the lowest coupling coordination degree.

(3)

K-Means clustering analysis identified three main spatial types, which were analyzed in conjunction with coupling coordination degree and the spatial heterogeneity of built environment and green space patterns to examine the characteristics of different spatial types. These are: first, the Urban Dominant Coordination Zone, characterized by dense built environment and well-developed green space configuration, exhibiting mature coordination mechanisms; second, the Green-Oriented Transition Zone, rich in ecological resources but with weak built environment, offering significant development potential but low coordination levels; Third, the Dispersed Natural Zone: both built environment and green space pattern indices are low, with minimal ecological intervention, presenting a natural preservation state.

(4)

Machine learning analysis based on LightGBM and SHAP indicates that structural indicators of the built environment (such as average compactness, weighted height, and land use diversity) are the primary drivers of coupling coordination, while green space pattern indicators, though ecologically significant, have a relatively minor influence in the model.

Finally, based on the K-Means clustering results, we analyzed the built environment, green space patterns, and coupling coordination characteristics of the clustered areas, analyzed the development directions of the regions, and proposed corresponding optimization policies in conjunction with the driving mechanisms.

5.2. Limitations and Future Research

This study is based on static cross-sectional data, which limits our ability to capture the temporal dynamics of urbanization and the evolving interactions between the built environment and green space patterns. Urbanization is inherently a long-term and gradual process, and the spatial relationships between the built environment and green space patterns are also in a state of constant change. Relying on cross-sectional static data may lead to the overestimation or underestimation of the coordination degree of certain areas. For example, areas undergoing urban renewal or ecological restoration may have phased spatial patterns, and data from a single year may only reflect temporary improvements in coordination rather than long-term stable trends. Conversely, some areas that are gradually advancing green space patterns may have low CCD values at present, but their coordination levels are on an improving trajectory. Future research should employ multi-temporal datasets to develop dynamic coupling-coordination models that trace the evolutionary trajectories of different spatial zones, thereby enabling more precise urban intervention strategies. Although qualitative discussions on urban policy directions are included, institutional, policy, or governance variables were not incorporated into the modeling process, which may introduce biases in understanding urban evolution and policy implications. Subsequent studies are encouraged to integrate spatial governance variables—such as land-use regulations, urban renewal projects, and planning intensity—to explore the real regulatory effects of urban management actions on coupling relationships, thereby enhancing the policy relevance and decision-making value of the research.

In addition, although this study selected a 300 m × 300 m grid as the basic spatial unit to achieve a balance between accuracy and computational efficiency, the selection of this fixed scale may still obscure spatial heterogeneity at finer (e.g., neighborhood level) or coarser (e.g., administrative district-level) scales. Spatial processes such as green space fragmentation, land use intensity, and policy interventions often operate through multi-scale mechanisms, and some local coupling relationships may be diluted or exaggerated as a result, thereby affecting the interpretability of analysis results. Future research could consider incorporating multi-scale comparative analysis, integrating street-level or neighborhood-level data on top of the existing grid scale to obtain higher-resolution local features, while also combining administrative district scales to align with actual policy and governance units. This approach could explore the sensitivity of coupled coordination relationships to scale changes, thereby enhancing the robustness and practical value of research conclusions.