Nonlinear Mechanisms of PM_2.5 and O₃ Response to 2D/3D Building and Green Space Patterns in Guiyang City, China

Abstract

PM_2.5 and O₃ are now the primary air pollutants in Chinese cities and pose serious risks to human health. In particular, the two- and three-dimensional patterns of urban buildings and green spaces play a crucial role in governing the dispersion of air pollutants. Using multi-source geospatial data and 2D/3D morphology metrics, this study employs an Extreme Gradient Boosting (XGBoost) model coupled with Shapley Additive Explanations (SHAP) to analyze the nonlinear effects of 2D/3D landscape and green space patterns on PM_2.5 and O₃ concentrations in the central urban area of Guiyang City. The results indicate the following findings: (1) PM_2.5 exhibits a U-shaped seasonal pattern, being higher in winter and spring and lower in summer and autumn, whereas O₃ displays an inverted U-shaped pattern, being higher in spring and summer and lower in autumn and winter. (2) PM_2.5 concentrations are higher in suburban and industrial zones and lower in central residential areas, while O₃ concentrations increase from the urban core toward the suburbs. (3) MV, BSI, BSA, BEL, BD, FAR, and BV show significant positive correlations with both PM_2.5 and O₃ (p < 0.001), whereas TH shows a significant negative correlation with PM_2.5 (p < 0.001). (4) High-density and complex building-edge patterns intensify both PM_2.5 and O₃ pollution by hindering urban ventilation and enhancing pollutant accumulation, whereas moderate vertical heterogeneity and greater tree height effectively reduce PM_2.5 concentrations but simultaneously increase O₃ concentrations due to enhanced VOC emissions. Urban form and vegetation jointly regulate air quality, highlighting the need for integrated urban planning that balances building structures and green infrastructure. The findings of this study provide practical implications for urban design and policymaking aimed at the coordinated control of PM_2.5 and O₃ pollution through the optimization of urban morphology.

1. Introduction

The city is the densest place for human activities and is the space where air pollutants are most likely to accumulate [1,2]. Since the 1980s, China’s urbanization rate has increased from 17.9% in 1978 to 67% in 2024 [3]. This marks a historic transformation from a country dominated by an agricultural population to one where the urban population is now the majority, indicating that China has entered a critical stage of urban development [4,5,6]. However, with the rapid urbanization process and social and economic development, the long-standing extensive land use development model has given rise to a series of environmental pollution issues, particularly air pollution [7,8,9]. Developed countries such as Europe and the United States have experienced air pollution problems for more than 100 years. Pollution increased intensively in China’s economically developed regions in the past 20–30 years. In recent years, China has become one of the most polluted areas in the world in regard to PM_2.5 pollution [10,11,12]. According to national monitoring data, in 2023, the annual mean PM₂.₅ concentration (30 μg/m³) and the O₃ concentration (144 μg/m³) in China were both approximately three times higher than the global population-weighted averages. This issue is particularly prominent in urban areas, which have become the hardest-hit regions. At present, PM_2.5 and O₃ have become the primary air pollutants in Chinese cities [13,14]. They significantly affect human health [15,16], atmospheric visibility [17], climate change [18], and socio-economic development [19]. Therefore, air pollution has emerged as a major environmental challenge hindering China’s effort to achieve the urban Sustainable Development Goals (SDGs) by 2030.

As the underlying surface of the atmospheric environment, urban spatial structure has a significant impact on regional energy flow, substance circulation, and biological processes. This is an important reason for a series of urban environmental problems [20,21]. As a complex of artificial landscapes, cities have undergone substantial transformation during the rapid urbanization process. The continuous increase in urban building height and volume has significantly altered the structure of natural urban landscapes [22]. Recent studies indicate, that over the past 15 years, with the rapid rise in building heights, the gap between the growth rates of urban building volume and urban land area has gradually narrowed [23]. The expansion pattern in many cities around the world is shifting from rapid horizontal sprawl to vertical growth through increased building heights [24,25,26]. Therefore, the traditional two-dimensional representation of urban spatial form is no longer sufficient to capture the three-dimensional realities of modern cities, which in turn has made associated air pollution processes more complex [27]. By altering building height, density, and surface materials, urban expansion exerts a significant influence on local surface climatic conditions and the distribution and dispersion of air pollutants. On the one hand, the land is the underlying surface of the atmosphere. The composition and pattern of land use types can directly affect PM_2.5 and O₃, because different land use types carry different intensities of human activities, which also means different pollution emissions [28,29]. Further, changes in the landscape structure can cause changes in the local climate [30,31], with the urban heat island (UHI) effect being one of the most prominent manifestations [32,33,34]. The UHI effect can enhance atmospheric thermal inversions, while dense building clusters within cities create wind-blocking effects, both of which influence the transport and chemical transformation of PM_2.5 and O₃ [11,35]. In particular, urban spatial structural changes at regional or macro scales have a more pronounced influence on climatic conditions. Therefore, research examining the impacts of urban spatial structure on air pollution should incorporate both horizontal (two-dimensional) and vertical (three-dimensional) dimensions of urban expansion processes [36].

The three-dimensional configuration of cities serves as a crucial indicator of urban vertical structure [37,38,39]. In recent years, with the successful development of 3D urban landscape pattern indices [40], several studies have begun to focus on the relationship between urban spatial structure and atmospheric pollutants [27,41,42]. Initial findings revealed a strong correlation between building height-related indices and PM_2.5 concentrations, as demonstrated by evidence from Beijing [43]. This relationship was further validated in Zhejiang Province, where 3D urban form indicators exhibited a significantly stronger influence on PM_2.5 levels compared to traditional 2D indicators [44]. Furthermore, a broader-scale analysis involving 68 Chinese cities expanded this perspective by highlighting seasonal variations in how urban spatial structure influences ozone concentrations: while 2D urban form indicators dominated during spring and summer, 3D indicators became more influential during autumn and winter [45]. Although the relationship between three-dimensional urban landscape patterns and air pollution has been established, past urban development practices in China indicate that the issues of air pollution and urban heat island effects within multi-city clusters have not fundamentally improved. Instead, ecological and environmental problems initiated by individual cities have expanded spatially, forming interconnected regions of concentrated pollution [46,47,48].

A critical gap exists in our knowledge of the complex mechanisms through which the three-dimensional urban form affects the dynamics of air pollutants, leaving a weak foundation for spatial optimization strategies. Firstly, the reliance on traditional 2D urban metrics fails to capture the vertical physical attributes, such as building height, volume, and spatial arrangement, that are critical for understanding airflow, pollutant dispersion, and microclimate formation within the urban canopy. Secondly, the analytical methods commonly employed are inadequate. The influencing factors of PM_2.5 and O₃ constitute a complex, nonlinear dynamic system [49,50,51]. However, traditional linear analysis methods [11] cannot adequately capture these intricate processes of interactions and feedback. In contrast, non-parametric machine learning models, with their flexible architectures and automated feature selection capabilities, are more suitable for characterizing these complex relationships [27]. The recent advent of interpretable machine learning methods (e.g., SHAP) further provides new tools to enhance model transparency and uncover intricate variable interactions, thereby offering refined insights for spatial planning [8,52,53,54,55]. In addition, existing research perspectives have largely been confined to the artificial built environment itself, generally overlooking green spaces as a critical component of the urban ecosystem. Cities are composites of “gray” building landscapes and “green” ecological spaces, yet prior studies have failed to examine the two as an integrated three-dimensional system to assess their synergistic impacts.

Therefore, this study focuses on the central urban area of Guiyang City to examine how 2D/3D building and green space patterns influence air pollution levels, using PM_2.5 and O₃ concentrations as representative indicators. The research aims to (1) examine the significance and marginal effects of 3D urban landscape pattern indices on O₃ and PM_2.5 concentrations, (2) reveal the underlying nonlinear mechanisms through which three-dimensional urban form influences air pollution, and (3) propose targeted spatial optimization strategies. By systematically analyzing the impact mechanisms of 3D landscape structures on O₃ and PM_2.5, this study seeks to provide more precise decision-making support for the spatial optimization of urban morphology.

2. Materials and Methods

2.1. Study Area

Guiyang, the capital of Guizhou Province in China, is situated between 106°07′–107°17′ E and 26°11′–26°55′ N. As the political, economic, and cultural center of the Guizhou province, Guiyang plays a pivotal regional role (Figure 1). The city lies in the eastern part of the Yunnan–Guizhou Plateau, which forms the second step of China’s three-tiered topographic system, with elevations ranging from 872 to 1659 m. The region features a subtropical plateau monsoon humid climate, characterized by high humidity, abundant rainfall, and distinct seasonal variations. This study focuses on six administrative districts of Guiyang: Nanming, Yunyan, Huaxi, Wudang, Baiyun, and Guanshanhu, covering a total area of 2525.46 km². In 2024, Guiyang had a resident population of 6.6025 million (https://drc.guizhou.gov.cn/xwzx/dfdt/202505/t20250512_87837828.html (accessed on 10 July 2025)), of whom 4.969 million lived in the urban districts, accounting for 75.26% of the total population. The city’s gross domestic product (GDP) reached 577.741 billion yuan, with the urban districts contributing 466.237 billion yuan, accounting for 80.07% of the total.

Figure 1. Study region.

2.2. Datasets

The independent variables covered both natural and socio-economic aspects and were divided into a training dataset (80% of the observation) and a testing dataset (20% of the observation). Table 1 lists seven types of data that were used to fit the PM_2.5 and O₃ concentration inversion model and evaluate the accuracy. The retrieval and preprocessing of these datasets in the current study are described below.

Table 1. Datasets used in this study.

Datasets	Format	Source
PM2.5, O3	Table	China Environmental Monitoring Station
	Raster	https://zenodo.org/records/6398971 (accessed on 10 July 2025)
Elevation	Raster	Geospatial data cloud of China
AOIs	Point features	Gaode Map Services, China
Building, Road and Water data	Vector	Baidu map
Boundary maps	Line features	Open Street Map
Land use	Raster	European Space Agency (ESA)
Tree Height	Raster	https://arxiv.org/pdf/2204.08322 (accessed on 10 July 2025)

2.2.1. Air Pollution Data

PM_2.5 and O₃ data were derived from hourly observations in the real-time publishing platform of urban air quality at China Environmental Monitoring Station (http://www.cnemc.cn/sssj/, accessed on 1 December 2020). There is a total of 13 monitoring stations, with the time range from 1 January 2020 to 31 December 2020. In accordance with the requirements for the validity of air pollutant concentration data in Ambient Air Quality Standards (National Standard GB 3095-2012) [56], the quality control of PM_2.5 data was performed [11,35]. Firstly, values of the hourly PM_2.5 and O₃ concentrations ≤0 and missing values were excluded. Secondly, if the measured data have been missing for more than 4 h in a day, all the data would be invalidated and excluded from the calculation of average daily PM_2.5 and O₃. Finally, a few anomalies with the hourly PM_2.5 concentrations > 900 μg/m³ were also eliminated. A monthly average of PM_2.5 was obtained based on the arithmetic mean of the daily average concentration.

Additionally, the gridded PM_2.5 and O₃ data were obtained from the ChinaHighPM_2.5 and ChinaHighO₃ products within the China High-Resolution, High-Quality Near-Surface Air Pollutants Dataset (ChinaHighAirPollutants, CHAP). These datasets employ artificial intelligence techniques to fill the spatial gaps in the MODIS MAIAC AOD satellite product using model-based data, and integrate ground-based observations, atmospheric reanalysis, and emission inventories to generate seamless nationwide near-surface PM_2.5 and O₃ data from 2000 to the present. The tenfold cross-validation results show coefficients of determination (R²) of 0.92 and 0.89, and root mean square errors (RMSE) of 10.76 µg/m³ and 15.77 µg/m³, respectively. The datasets cover the entire region of China, with a spatial resolution of 1 km and temporal resolutions at daily, monthly, and annual scales. All data are provided in units of µg/m³ and referenced to the WGS84 coordinate system. In this study, we used the data for the year 2020 and extracted PM_2.5 and O₃ data for the study area using its vector boundary in ArcGIS Pro 3.0.1.

2.2.2. Land Use Data

The land use data employed in this study were obtained from the World Cover product developed by the European Space Agency (ESA), available at: https://worldcover2021.esa.int/ (accessed on 10 July 2025). The dataset is generated using Sentinel-1 and Sentinel-2 satellite imagery and provides a high spatial resolution of 10 m. It adopts the WGS84 coordinate reference system and classifies global land cover into 11 categories: cropland, shrubland, tree cover, grassland, wetland, built-up areas, bare/sparse vegetation, snow/ice, permanent water bodies, mangroves, and moss/lichen. These classifications cover the major land use and land cover types, offering detailed and globally consistent information for land surface analysis.

2.2.3. Tree Height Dataset

The Global Sentinel-2 10 m Canopy Height (2020) dataset, developed by ETH Zurich (Lang et al., 2023) [57], provides a high-resolution global map of vegetation canopy height at 10 m spatial resolution, representing conditions for the year 2020. This dataset was generated by fusing GEDI LiDAR observations with Sentinel-2 optical imagery using a probabilistic deep learning model. The approach enables reliable estimation of vertical vegetation structure worldwide, overcoming saturation issues commonly found in optical data alone. The dataset supports applications in forest monitoring, carbon stock assessment, and biodiversity conservation. We accessed the dataset by Google Earth Engine (https://gee-community-catalog.org/projects/canopy/ (accessed on 10 July 2025)) which is provided in the WGS84 geographic coordinate reference system. Full methodological details are provided in https://arxiv.org/pdf/2204.08322 (accessed on 10 July 2025).

2.2.4. Building, Road and Water Data

The building, road, and water body data used in this study were all sourced from Baidu Maps (https://lbsyun.baidu.com (accessed on 10 July 2025)). The building vector dataset of Guiyang in 2020 contains detailed information, including the number of floors for each building. It includes 50,555 individual buildings, with a total building footprint area of approximately 33.72 km², covering a spatial extent of 1540.55 km² within the Guiyang region. The road and water body data were also obtained via the Baidu Maps API, which provides high-precision vector information on transportation networks and hydrological features. The road dataset covers a wide range of road types across the country, including highways, urban primary and secondary roads, and rural roads. The water body data includes rivers, lakes, reservoirs, and ponds. All datasets are available in multiple standard vector formats, such as GeoJSON, WKT, and Shapefile, which enables their direct integration into geographic information systems for spatial analysis and mapping. A Python 3.12-based program was developed in this study to access and retrieve these datasets through the Baidu Maps API.

2.3. Methods

2.3.1. Two- and Three-Dimensional Urban Landscape Pattern Indices

Urban spatial expansion refers not only to the horizontal two-dimensional growth of built-up areas but also to the spatial differentiation of the 3D urban landscape pattern. To measure the structural characteristics and heterogeneity of urban building landscapes in three-dimensional space, and to analyze the spatiotemporal features of 2D/3D urban landscape patterns, this study constructs nine 2D/3D urban landscape pattern indices from four perspectives: building density, building morphology, height, and spatial distribution. In addition, to reflect the impact of greenery on air quality, tree canopy height and the proportion of woodland within each land block were selected as indicators of the vertical and horizontal distribution of vegetation, respectively. The distribution of water bodies within land blocks was used to construct three indices reflecting the influence of water bodies on air quality. The specific calculation formulas and definitions are provided in Table 2.

Table 2. Two- and three-dimensional landscape indices and description [31,49,50].

Landscape Metric	Formula		Description
Mean Architecture Height (MAH)	$MAH= {\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{H}_i$, $H_i$: the height of individual building; $n$: number of buildings.	(1)	Mean building height in the analytical unit.
Height standard deviation (HSD)	$HSD=\sqrt{{\displaystyle \frac{{{\sum }_{i=1}^n(H_i-MAH)}^2}{n}}}$	(2)	Variation in the buildings in the analytical unit.
Mean Volume (MV)	$MV={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{a}_i\times H_i$, $a_i$: building cover area.	(3)	Reflects the average building volume within the region.
Building Structure Index (BSI)	$BSI={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{\displaystyle \frac{a_i}{H_i}}$	(4)	Variation in building structural characteristics analyzed by the ratio of building footprint area to height.
Building surface area (BSA)	$BSA= {\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(a}_i+P_i\ast H_i)$, $P_i$: the perimeter of building.	(5)	Reflects the average building surface area within the region.
Building density (BD)	$BD={\sum }_{i=1}^n{a}_i/A_{i }$, $A_i$: land parcel cover area.	(6)	Reflects the intensity of building within the analytical unit.
Floor area ratio (FAR)	$FAR={\displaystyle \sum _{i=1}^n}({\displaystyle \frac{H_i}{3}})\times a_i)/A_i$	(7)	The ratio between the sum of gross floor area and analytical unit area.
Building edge length (BEL)	$BEL={\displaystyle \sum _{i=1}^n}{(a}_i+P_i+H_i)$	(8)	Reflects the morphological complexity and spatial development of building edges within the analytical unit.
Building Evenness Index (BEI)	$BEI=\sqrt{{\left[{\displaystyle \sum _{i=1}^n}{(a}_i\times H_i)-\mathrm{M}\mathrm{V}\right]}^2/}{A}_i$	(9)	Reflects the uniformity of building distribution in three-dimensional space.
Tree height (TH)	$TH= {\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{TH}_i$, ${TH}_i$: height of tree.	(10)	Total tree height in the analytical unit.
Greenery density (GD)	$GD={\sum }_{i=1}^n{G}_i/A_i$,$G_i$: green cover area.	(11)	Reflects the intensity of green spaces within the analytical unit.
Building-to-Vegetation Volume Ratio (BV)	$BV= {\displaystyle \frac{V_{building}}{{ V}_{vegetation }}}$, $V_{building}$: building volume; $V_{vegetation }$: vegetation volume.	(12)	Reflects the spatial balance between built-up and vegetated volumes, indicating the degree of urban development intensity and ecological greenness within the analytical unit.

2.3.2. XGBoost Model

XGBoost is a decision tree-based ensemble learning algorithm that iteratively constructs weak learners (regression trees) and optimizes the objective function using gradient boosting, thereby significantly enhancing the model’s predictive performance and generalization ability [51]. It is one of the most widely used regression algorithms and is known for its high regression accuracy [5,52,53,54]. XGBoost also offers advantages such as parallel computation, strong regularization mechanisms, and automatic handling of missing values, making it particularly well-suited for modeling high-dimensional and complex nonlinear data [55]. Therefore, this study applies the XGBoost model to predict the impact of built environment variables on congestion levels.

The structure score of XGBoost is defined by its objective function, which explains training loss and regularization and contributes to its effective performance.

$$Obj\left(\theta \right)=L\left(\theta \right)+\Omega \left(\theta \right)$$

(13)

$L\left(\theta \right)$ is used to evaluate the model performance on the training data, and $\Omega \left(\theta \right)$ is used for regularization to control overfitting of the model. Additionally, the complexity of the tree is computed as follows:

$$\Omega \left(f\right)=\gamma T+0.5\lambda {\displaystyle \sum _{j=1}^T}{w}_j^2$$

(14)

where $T$ is the number of leaves and ω is the vector of scores on leaves. Finally, the structural score of XGBoost is calculated using the following formula:

$$Obj={\displaystyle \sum _{j=1}^T}\left[G_j{\omega }_j+{\displaystyle \frac{1}{2}}\left(H_j+\lambda \right){\omega }_j^2\right]+\gamma T$$

(15)

where ${\mathsf{\omega }}_j$ are independent of each other; The form $G_j{\omega }_j+{\displaystyle \frac{1}{2}}\left(H_j+\lambda \right){\omega }_j^2$ is quadratic and the best ${\omega }_j$ for a given structure.

2.3.3. SHAP Model

To quantify the interactive effects between key indicators at each time point, this study proposes a hybrid model that integrates an improved XGBoost algorithm with SHAP. SHAP, derived from the Shapley value principle in cooperative game theory [58], calculates the marginal contribution of each characteristic factor to the model’s prediction. The SHAP value (i.e., attribution value) for each feature reflects its relative importance in contributing to the predicted outcome [34,59], thereby revealing both threshold effects and synergistic interactions between explanatory and response variables. The Shapley value formula for feature $i$ is defined as follows:

$${\varphi }_i={\sum }_{S\subseteq N\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(n-\left|S\right|-1\right)!}{n!}}\left[f\left(S\cup \left\{i\right\}\right)-f\left(S\right)\right]$$

(16)

where ${\varphi }_i$ represents the contribution of the feature $i$, N denotes the set with n features, $f\left(S\cup \left\{i\right\}\right)$ and $f\left(S\right)$ represent the model results with or without the feature $i$.

3. Results

3.1. Spatiotemporal Patterns of PM_2.5 and O₃

3.1.1. Time Variation Characteristics of PM_2.5 and O₃ Concentrations

Figure 2 shows the temporal variation in daily average concentrations of PM_2.5 and O₃ in Guiyang City during 2020. As can be seen from Figure 2, Mean, maximum, and minimum concentrations of PM_2.5 and O₃ were 23.55/53.94, 72.02/126.63, and 5.12/11.21 µg/m³, respectively. The seasonal distribution of PM_2.5 reveals a clear variation over the seasons, forming a U-shaped pattern, with the highest concentrations occurring during winter in the following sequence: winter (31.48 µg/m³)> spring (28.7 µg/m³) > autumn (20.9 µg/m³) > summer (15.23 µg/m³). This pattern is related to Guiyang’s topography, because during winter and spring stable atmospheric stratification is unfavorable for convective transport and pollutants tend to accumulate. In contrast to PM_2.5, The seasonal distribution of O₃ showed a clear inverted “U” shape, with the highest level of 92 µg/m³ in spring, followed by 52.47 µg/m³ in spring, and lower levels of 48.93 µg/m³ in autumn and 44.27 µg/m³ in winter, respectively. This is mainly due to O₃’s strong sensitivity to solar radiation, which is more intense in spring and summer in the study area, thereby accelerating photochemical ozone formation, and weaker in autumn and winter. The concentrations of PM_2.5 and O₃ also display an alternation of peaks and troughs over time, these phenomena corroborate that there is some correlation between the two pollutants [48].

Figure 2. Temporal variation of PM_2.5 and O₃ concentrations in Guiyang city.

3.1.2. Spatial Distribution Characteristics of PM_2.5 and O₃

Figure 3 shows the spatial distribution of the station-level annual and seasonal mean concentrations of PM_2.5 and O₃ in Guiyang City during 2020. From annual figure, the annual mean PM_2.5 concentrations (station-level means) ranged from 20.05 to 31.28 µg/m³; the annual mean O₃ concentrations ranged from 49.21 to 57.85 µg/m³. From seasonal figure, PM_2.5 concentrations (station-level seasonal means) ranged from 24.31 to 36.42 µg/m³ in spring, 10.74 to 21.76 µg/m³ in summer, 17.86 to 27.44 µg/m³ in autumn, and 27.30 to 39.51 µg/m³ in winter; O₃ concentrations (station-level seasonal means) ranged from 62.66 to 74.13 µg/m³ in spring, 50.88 to 56.55 µg/m³ in summer, 46.70 to 52.90 µg/m³ in autumn, and 41.43 to 47.84 µg/m³ in winter.

Figure 3. Spatial distribution of PM_2.5 and O₃ concentrations in Guiyang city.

The spatial distribution patterns of PM_2.5 and O₃ were consistent between the annual mean and the four seasons (spring, summer, autumn, and winter). The highest PM_2.5 concentration occurred at the Yanzichong suburban station in Wudang District, whereas the lowest occurred at the Jianhu Road suburban station in Guanshanhu District, indicating the lightest pollution. Spatially, PM_2.5 tended to be higher in parts of the suburbs and industrial zones and lower in central residential areas, suggesting an outward shift in urban development toward the suburbs. The highest annual mean O₃ concentration was observed at the Jianhu Road suburban station in Guanshanhu District, while the lowest was recorded at the Taiciqiao residential station in Huaxi District. Accordingly, O₃ generally increased from the urban core to the suburbs, a hierarchical pattern similar to that reported for other cities. This is primarily because residential and industrial areas have elevated NO concentrations due to motor vehicle emissions and other factors; high NO not only inhibits O₃ formation but also titrates existing O₃ [35]. In addition, suburban areas around Tongmuling, Jianhu Road, and Yanzichong have high vegetation cover, and emissions of biogenic VOCs, which are important and highly reactive precursors for photochemical reactions, promote O₃ formation.

3.2. Analysis of Urban 2D/3D Landscape Patterns

3.2.1. Spatial Distributions of Buildings Height

We set the height of each floor to 3 m according to the Chinese construction standard, so that the building height distribution can be calculated. The number and footprint area of buildings with different floors are shown in Table 3. Statistics indicate that the study area contains 50,555 buildings, of which 28,366 (56.11%) are low-rise, 12,992 (25.70%) are multistory, 8867 (17.54%) are high-rise, and 330 (0.65%) are super high-rise. The building height ranges from 3 m to 180 m, with an average height of 18.23 m and a standard deviation of 23.18 m, reflecting notable vertical variability in the urban landscape. Overall, the building-height structure is characterized by “predominantly low-rise with a small fraction of super high-rise”; the shares of high-rise and super high-rise buildings are relatively small, indicating an urban morphology dominated by low- and multistory development, with super high-rises concentrated in a limited number of locations.

Table 3. Statistics of numbers and areas of buildings with different floors.

Building Type	Low-Rise (<10 m)	Multistory (11–24 m)	High-Rise (25–100 m)	Super High-Rise (>100 m)	Total Building
Quantity	28,366	12,992	8867	330	50,555
Proportion	0.56%	0.26%	0.18%	0.01%	100%

Note: Buildings are divided into five categories according to the National Standards of China of Code for Design of Civil Buildings (GB 50352-2019) [60]: http://www.jianbiaoku.com/webarbs/book/404/1476906.shtml (accessed on 10 July 2025).

Figure 4 shows the spatial pattern of building height distribution in the central urban area of Guiyang. The figure indicates a multi-centered pattern of concentration with height and density declining from the core outward. The traditional core is concentrated in Yunyan District and Nanming District, while a new administrative and financial core has formed in Guanshanhu District. Buildings in Yunyan, Nanming, and Guanshanhu together account for 67% of the building stock in the central urban area. In Yunyan, large clusters of high-rise and super high-rise buildings are concentrated around the Jiaxiu Tower precinct and toward the northwest, primarily along major arterials, reflecting strong functional mixing and business-service agglomeration. In Nanming, overall building heights are slightly lower than in Yunyan, but the Huaguoyuan complex stands out as the tallest cluster citywide, forming a contiguous residential and commercial fabric. Guanshanhu is dominated by multistory and high-rise buildings, with a concentration of office and public-service facilities and a relatively balanced spatial distribution. From the core through the near suburbs to the periphery, building height distribution and density diminish markedly, and peripheral areas are predominantly low-rise, indicating outward expansion of urban development from the center.

Figure 4. Spatial distributions of buildings with varying height in Guiyang.

3.2.2. Spatial Patterns of 2D/3D Landscape Indices

This study calculated 12 2D/3D landscape pattern indices, including MAH, HSD, MV, BSI, BSA, BD, FAR, BEL, BEI, TH, GD, and BV. After standardizing them to the [0, −1] range, their spatial distributions were analyzed across a total of 1336 land blocks. As showed in Figure 5, The results indicate that in densely built-up areas such as Yunyan District, Nanming District, and Guanshanhu District, most urban two-and three-dimensional landscape pattern indices exhibit higher values, revealing a distinct spatial clustering pattern. The high-value areas of MAH are mainly concentrated in Yunyan District, Nanming District, and Guanshanhu District, with some areas of Huaxi District also showing certain high-value zones. The spatial distributions of HSD, MV, BSI, and BSA are relatively uniform. Although their high-value areas are mainly located in the central urban districts, some localized high values also appear in the peripheral areas of the city. The high-value areas of the BD, FAR, NB, and BEI indicators are concentrated within the main urban districts, particularly in Yunyan District, Nanming District, and Guanshanhu District. TH and GD show low values in densely built-up areas such as Yunyan District and Nanming District, whereas in peripheral urban areas with abundant green spaces and lower construction density, the TH and GD indicators have higher values. The high-value areas of TH are distributed in Nanming District, Guanshanhu District, and Baiyun District; those of GD are concentrated near the boundary between Yunyan District and Nanming District. The overall distribution of BV is relatively low, but relatively prominent high-value areas appear in the southern part of Nanming District’s main urban area and in some local areas of Qingzhen City.

Figure 5. Spatial distribution of standardized 2D/3D landscape pattern indices in Guiyang City.

3.3. Relationship Between PM_2.5, O₃ and 2D/3D Landscape Indices

Air quality is closely related to urban morphology. Due to the sparse distribution of monitoring stations, even using kriging interpolation to obtain gridded PM_2.5 and O₃ data across the study area resulted in limited accuracy. Therefore, in the correlation and subsequent regression analyses, the mean concentrations of PM_2.5 and O₃ were derived from the ChinaHighAirPollutants dataset, which provides high-resolution gridded data validated against national monitoring stations in China [61,62]. As shown in Figure 6, MV, BSI, BSA, BEL, BD, FAR, and BV exhibit significant positive correlations with PM_2.5 and O₃ (p < 0.001), indicating that when the urban spatial form is characterized by higher building density, larger building volumes, and more complex building edge features, wind speed is reduced, air circulation is hindered, and the heat island effect and pollutant accumulation are intensified, thereby exacerbating PM_2.5 and O₃ pollution levels. In addition, higher building density often implies greater traffic load and vehicle emissions. By contrast, MAH, HSD, TH, and GD are negatively correlated with PM_2.5 and O₃, among which TH is significantly negatively correlated with PM_2.5 (p < 0.001). This suggests that a more dispersed building layout facilitates pollutant dispersion, while taller trees and higher green space density can reduce atmospheric pollutant concentrations through shading, deposition, and absorption effects. Therefore, it is recommended to adopt low-rise building designs in combination with large-scale green infrastructure development (including rooftop gardens and vertical greening systems) to improve microclimatic conditions and effectively mitigate pollution levels.

Figure 6. The relationship between PM_2.5, O₃ concentration and 2D/3D landscape indices.

3.4. Relationship Between PM_2.5, O₃ and 2D/3D Landscape Indices

3.4.1. Relative Impacts of 3D Landscape Indices on PM_2.5 and O₃

To examine the influence of urban 2D/3D landscape metrics on PM_2.5 and O₃ concentrations, this study employed the XGBoost–SHAP model to quantify the relative contribution of each metric. The results indicate that the XGBoost model achieved reliable predictive performance for both pollutants, showing particularly strong accuracy in estimating PM_2.5 concentrations (R² = 0.77, RMSE = 18.49, MAE = 11.01), while the corresponding values for O₃ were R² = 0.571, RMSE = 13.99, and MAE = 4.54. The relative importance of the twelve-landscape metrics is presented in Figure 7. The results indicate that BEL exerts the strongest explanatory power for both pollutants, contributing 29.5% to PM_2.5 and 34.9% to O₃, highlighting the critical role of building edge complexity in shaping pollutant accumulation and dispersion. For PM_2.5, BD, TH, and BSI also exhibit substantial contributions (27.0%, 10.5%, and 10.4%, respectively), suggesting that higher building density and volumetric structural features markedly restrict urban ventilation and intensify pollutant buildup. In contrast, FAR, GD, BV, HSD, and BEI each account for less than 2%, indicating relatively minor effects. For O₃, MV, BSA, BD, and BSI contribute 17.6%, 11.2%, 10.5%, and 9.0%, respectively, whereas FAR, BEI, TH, BV, and HV each explain less than 3%. Notably, TH makes a considerable contribution to PM_2.5 (10.5%) but a negligible one to O₃ (<2%), underscoring the differential sensitivity of particulate matter and photochemical pollutants to landscape attributes: tree height plays a significant role in reducing particulate concentrations but has limited influence on ozone. Overall, the results emphasize that building morphological characteristics (e.g., edge complexity, density, and volume) dominate the determination of urban air pollution levels, while green infrastructure and vertical heterogeneity, though relatively weaker, offer potential benefits for mitigating PM_2.5 pollution.

Figure 7. Relative importance of 2D/3D landscape indices on PM_2.5 and O₃.

3.4.2. Marginal Effect of 2D/3D Landscape Indices on PM_2.5 and O₃

After analyzing the correlation coefficients and relative importance of urban 3D landscape metrics with PM_2.5 and O₃, this study further examined the marginal effects of 12 indicators to reveal their nonlinear impacts on pollutant concentrations. As shown in Figure 8, BD and BEL exert strong positive effects on both PM_2.5 and O₃, when the BD value exceeds 0.03, the O₃ concentration rises sharply; when the BD value exceeds 0.04, both PM_2.5 and O₃ concentrations stabilize, reflecting a critical threshold at which high building density significantly suppresses urban ventilation. With increasing building density and edge complexity, pollutant concentrations rise sharply, indicating that high-density and morphologically complex building patterns significantly hinder urban ventilation and reinforce pollutant retention and accumulation. In addition, higher building density often implies greater traffic load and vehicle emissions, which further explains its highest contribution observed in the relative importance analysis. MV, BSI, and BSA are also positively associated with pollutants; notably, when MV exceeds 0.03, BSI exceeds 0.02, or BEL exceeds 0.02, PM_2.5 and O₃ concentrations increase markedly, further confirming the aggravating role of large building volumes and complex structural features in air pollution. In contrast, MAH and HV show negative effects on PM_2.5 and O₃, with concentrations dropping sharply as their values increase from 0 to about 0.2 and then stabilizing. This suggests that moderate vertical heterogeneity facilitates the formation of effective ventilation corridors, thereby improving pollutant dispersion. Similarly, HSD exhibits a gradual decline in both PM_2.5 and O₃ with increasing values, reinforcing the importance of height heterogeneity for enhancing airflow. The effect of TH (tree height) shows clear differentiation. For PM_2.5, a threshold effect is observed: in the low-value range (<0.05), PM_2.5 concentrations increase, indicating that low and sparse vegetation provides limited suppression of particulate pollution; however, once TH exceeds a certain height, PM_2.5 concentrations decrease significantly, suggesting that taller trees can effectively reduce particulate matter through shading, deposition, and absorption. By contrast, TH shows a positive effect on O₃, with concentrations generally increasing as tree height rises. This may be linked to enhanced emissions of volatile organic compounds (VOCs) from taller vegetation and the promotion of photochemical reactions. Some indicators exhibit differentiated effects between pollutants. FAR has a positive effect on PM_2.5 but a slight negative effect on O₃, implying that higher floor area ratios are often associated with increased traffic-related particulate emissions, while their influence on O₃ formation remains relatively weak. BEI shows a gradual increase in PM_2.5, while O₃ displays a nonlinear pattern: concentrations are high at low BEI, decrease significantly at intermediate levels, and rise sharply again at high BEI, suggesting that uneven building distribution may either trap pollutants or enhance dispersion under certain conditions. GD is positively correlated with both pollutants, with concentrations increasing rapidly when GD < 0.1 and then leveling off, indicating that higher green space density does not linearly improve air quality and its effect strongly depends on spatial configuration. BV exhibits the weakest effect, showing only slight increases in PM_2.5 at low levels and negligible changes in O₃.

Figure 8. Nonlinear effects of 2D/3D landscape indices on PM_2.5 and O₃ concentrations.

In summary, high-density and complex building-edge patterns substantially exacerbate air pollution, whereas moderate vertical heterogeneity and taller trees can mitigate particulate matter concentrations by enhancing ventilation and deposition. However, taller trees may also promote O₃ formation due to increased VOC emissions. These findings are consistent with the results of correlation and relative importance analyses and further highlight the differentiated impacts of urban landscape structures on PM_2.5 and O₃ through mechanisms of ventilation, traffic emissions, and vegetation-related ecological functions.

4. Discussion

Previous studies have demonstrated that variations in building morphology, particularly changes in building height, play a crucial role in shaping the local climate, which in turn influences the spatial distribution of urban air pollutants [44]. A higher building height distribution density can create a shielding effect, which delays longwave radiation within street canyons and increases both the average temperature and atmospheric turbulent energy, thereby facilitating the vertical dispersion of PM_2.5 concentrations [14]. Our study found that among the twelve constructed indicators, building density (BD) and building edge length (BEL) have the greatest impact on PM_2.5 concentrations. BD and BEL more directly reflect the spatial clustering and boundary complexity of buildings, and they are closely related to near-surface ventilation conditions in urban areas. The longer the building edge length, the more complex the urban structure and the narrower the streets, which severely impairs natural ventilation, reduces wind speed, and hinders the dilution and transport of PM_2.5, resulting in elevated local concentrations. Meanwhile, high building density implies fewer gaps between buildings, obstructing airflow and contributing to increased PM_2.5 concentrations. Regions with greater building edge length and higher building density typically correspond to densely built-up areas with heavy traffic, which are also hotspots for concentrated PM_2.5 emissions. In contrast, indicators such as MV, BSI, BSA, FAR, and BEI primarily emphasize the comprehensive measurement of building volume, functional structure, or spatial distribution characteristics. Although these metrics can reflect urban development intensity or vertical spatial morphology, their direct impact on near-surface aerodynamics is relatively limited. For example, MV primarily reflects building height distribution and bulk, but does not necessarily indicate the continuity of wind corridors between buildings. BSI and BSA are mainly used to assess the complexity of building surfaces, but do not fully represent the wind-blocking effect of buildings in the horizontal urban plane. FAR, defined as the ratio of total floor area to land area, is more of a planning-oriented metric rather than a direct indicator of ventilation performance. BEI, on the other hand, focuses on the spatial evenness of building distribution and is not directly related to the “diffusion resistance due to pollutant aggregation.”

The concentration of buildings reflects the process of urbanization, leading to high population density, all of which substantially the impact urban ozone concentration [24]. Built environment factors represent a key component of anthropogenic influences. The three-dimensional configuration of urban structures affects ventilation and heat dissipation within the city, thereby modulating the rate of ozone formation and related chemical reactions [27]. This study found that BD and BEL significantly increase O₃ concentrations. High-density urban areas, with extensive impervious surfaces and high thermal inertia, tend to retain heat during the night, intensifying the urban heat island effect. The resulting increase in local temperature and solar radiation enhances the photochemical production of the ozone. However, our study also found that when building density exceeds a certain threshold, O₃ concentrations tend to decrease. The increased urban spatial enclosure deteriorates ventilation conditions, hindering the diffusion of O₃ near the ground. Meanwhile, NO_x concentrations may rise sharply due to dense traffic and population, enhancing the titration reaction with O₃ and thereby suppressing its concentration increase.

Beyond elucidating the influence of urban building morphology on PM_2.5 and O₃, our findings reveal a divergent impact of tree height on these two pollutants. Specifically, we observed that increased tree height contributes to a marked reduction in PM_2.5 concentrations but exerts only a marginal influence on O₃ levels. Taller trees undoubtedly enhance the capture and reduction in particulate matter by providing a greater leaf area for dry deposition. However, the interaction with ozone is more complex. While trees facilitate ozone removal through dry deposition, taller and healthier trees, which typically possess greater biomass, also exhibit a higher potential for emitting biogenic VOC, the key precursors to O₃. Under certain meteorological conditions, this BVOC-driven O₃ formation potential may partially offset or even outweigh the removal via deposition. This mechanism is corroborated by existing literature; studies quantifying urban tree biogenic emissions’ impact on PM_2.5 and O₃ have reported consistent findings [63], confirming that certain tree species’ BVOC emissions can indeed compromise their net air quality benefits.

The dominant role of BEL and BD in exacerbating both PM_2.5 and O₃ concentrations aligns with observations in cities like Beijing [43], indicating that morphological complexity and high-density development universally impede ventilation and aggravate pollution. However, Guiyang’s basin topography intensifies certain mechanisms. The pronounced influence of building volume-related metrics (e.g., BSI, MV) reflects an amplified pollution-trapping effect within this naturally ventilation-limited environment. Furthermore, the muted role of FAR and the distinctive nonlinear response of TH, which reduces PM_2.5 but slightly elevates O₃, further underscore the context-dependent nature of these interactions. The observed TH and O₃ relationship, wherein taller trees are associated with a slight increase in O₃, may be attributed to several mechanisms specific to Guiyang’s humid and low-wind basin climate. Thus, while the influence of key morphological features such as BEL and BD is fundamental across cities [44,45], their magnitude and the behavior of specific indicators (e.g., BV, TH) are strongly modulated by local topography and climate. Under the backdrop of global climate change and rapid urbanization, air pollution issues have become increasingly prominent. PM_2.5 and O₃, as major urban environmental pollutants, not only pose serious threats to public health but also place higher demands on sustainable urban development. Especially in mountainous cities with complex terrain, urban morphology plays a significant regulatory role in the dispersion, accumulation, and transformation of pollutants. Research has found a significant correlation between the spatial morphological characteristics of buildings and pollutant concentrations. During the construction of new urban areas and the renewal of old cities, priority should be given to controlling excessive building concentration. By optimizing the arrangement and spatial spacing of buildings, the risk of localized pollutant accumulation can be reduced. In addition, for PM_2.5, vertical greening and the arrangement of tall trees should be strengthened to enhance turbulent exchange and vertical diffusion efficiency; for O₃, the clustering of buildings with high volume and large surface area in high-temperature areas should be restricted to reduce the formation of thermal conditions necessary for photochemical reactions.

5. Policy Recommendations and Limitations

Under the dual pressures of global climate change and rapid urbanization, air pollution has emerged as an increasingly critical environmental issue. As the dominant urban air pollutants, PM_2.5 and O₃ not only pose severe threats to public health but also impose higher demands on sustainable urban development. Numerous cities worldwide have sought to mitigate air pollution through the deliberate design of urban landscape patterns. However, for cities that are already built-up, implementing an entirely new urban form is unrealistic. Urban designers and planners generally focus on redesigning and retrofitting the existing built environment and ecological spaces. Therefore, identifying effective approaches to improve urban air quality by adjusting and optimizing existing three-dimensional urban morphology is of substantial practical importance. It is important to recognize that mitigation strategies are not universally applicable; measures effective in one city may not yield the same results in another. Hence, the development of locally adapted strategies that account for site-specific topographic and climatic characteristics is essential. In particular, in mountainous cities characterized by complex terrain, the diffusion of air pollutants is far more intricate than in flat urban areas. Urban morphology exerts a critical regulatory influence on the diffusion, accumulation, and chemical transformation of pollutants. Our findings reveal a significant correlation between the spatial morphological characteristics of buildings and pollutant concentrations. During the construction of new districts and the renewal of existing ones, it is crucial to avoid excessive building concentration. Optimizing building layout and spatial spacing can effectively reduce the risk of localized pollutant accumulation. Furthermore, for PM_2.5 mitigation, the enhancement of vertical greening and the incorporation of tall trees are recommended, as they can strengthen turbulent exchange, promote vertical dispersion, and improve the adsorption and purification of PM_2.5 by vegetation. For O₃ control, the clustering of large-volume buildings with extensive surface areas in high temperature industrial zones should be limited, so as to reduce the thermal conditions conducive to photochemical reactions. Overall, these findings provide theoretical support and practical guidance for three-dimensional landscape planning and design aimed at improving air quality in mountainous urban environments.

Although this study revealed the relationship between urban 2D/3D landscape metrics and the concentrations of PM_2.5 and O₃, there are several limitations. First, the analysis was constrained by the limited spatial resolution of the available datasets and the sparse distribution of air quality monitoring stations, which affected data accuracy and spatial representativeness. These constraints prevented the inclusion of several important factors, such as meteorological conditions, socioeconomic status, land use, and landscape morphological indicators. In particular, the lack of explicit human activity data is a notable limitation, as anthropogenic emissions are major drivers of air pollution. The omission of these control variables introduces some uncertainty into the results. Future research should integrate additional influencing factors, especially detailed human activity patterns, and improve data coverage and resolution to enhance model accuracy and provide a more complete understanding of the interactions between urban form and air quality.

6. Conclusions

Based on the construction of twelve 2D/3D urban landscape metrics covering building density, morphology, height, spatial distribution, and greening-related indices, and the application of the XGBoost–SHAP model, this study analyzed the relative importance and marginal effects of these indicators on PM_2.5 and O₃ concentrations. The following main conclusions are reached:

(1)

PM_2.5 shows a clear U-shaped seasonal cycle, with concentrations peaking in winter and spring and reaching their lowest in summer and autumn. In contrast, O₃ exhibits an inverted U-shaped cycle, with high values in spring and summer and low levels in autumn and winter. These opposite seasonal dynamics highlight the need for differentiated pollution control strategies across seasons.

(2)

PM_2.5 concentrations are generally higher in suburban and industrial areas and lower in central residential districts, reflecting the combined influence of emission sources and local ventilation conditions. By contrast, O₃ concentrations increase gradually from the urban core toward suburban areas, indicating that photochemical processes and precursor transport play stronger roles outside the city center.

(3)

Among the constructed indicators, MV, BSI, BSA, BEL, BD, FAR, and BV show significant positive correlations with both PM_2.5 and O₃ (p < 0.001), confirming that higher building density, volume, and structural complexity exacerbate pollution accumulation. In contrast, TH exhibits a significant negative correlation with PM_2.5 (p < 0.001), emphasizing the role of urban vegetation in particulate matter removal.

(4)

High-density and complex building-edge patterns intensify both PM_2.5 and O₃ pollution by hindering air circulation and promoting pollutant retention. Conversely, moderate vertical heterogeneity and taller trees can mitigate PM_2.5 by improving ventilation and enhancing deposition, yet they also promote O₃ formation through increased VOC emissions and photochemical activity. These findings underscore the dual role of urban form and vegetation in air quality regulation and call for integrated planning strategies that balance building configuration and green infrastructure to achieve co-benefits for public health and sustainable urban development.