Investigating the effect of spatial patterns of artificial impervious surface on PM_2.5 at the intra-urban scale

ABSTRACT

Artificial impervious surface, as the hallmark of urbanization, contributes to urban development but simultaneously leads to urban PM_2.5 pollution. However, how artificial impervious surface affected PM_2.5 at the intra-urban scale was rarely investigated. By using remote-sensing techniques to derive PM_2.5 concentrations and dividing urban areas into two categories with different urban development situation (areas with drastic expansion of artificial impervious surface (DEA) and areas with slight or no changes of artificial impervious surface (SNCA)), we provided a new perspective for investigating the effects of spatial patterns of artificial impervious surface on PM_2.5. First, based on multi-source predictors, a two-staged XGBoost model was adopted to derive PM_2.5 concentrations. Then, after delineating DEA and SNCA based on the variations of artificial impervious surface and selecting indicators, the overall and local effects of spatial patterns of artificial impervious surface were respectively explored by Original Least Square regression (OLS) and Geographically Weighted Regression (GWR). The results demonstrated that, compared with in SNCA, population and economic development contributed to higher levels of PM_2.5 concentrations in DEA. Additionally, in DEA, high PM_2.5 concentrations tended to occur in areas with artificial impervious surfaces exhibiting high coverage, complex shape, and uncompact distribution. In SNCA, among all landscape spatial patterns, the effect of artificial impervious surface coverage was the strongest. Spatially, artificial impervious surface coverage and economic development had much more significant impacts on PM_2.5 than other indicators. Among these two indicators, the exacerbating effect of artificial impervious surface coverage on PM_2.5 pollution was more pronounced in SNCA, while economic development notably brought high PM_2.5 concentrations in the suburban districts of DEA. In general, the framework proposed in this study advanced the understanding of the effects of artificial impervious surface on PM_2.5 and the findings are valuable for the mitigation of PM_2.5 pollution at the intra-urban scale.

KEYWORDS:

• PM_2.5
• artificial impervious surface
• spatial patterns
• intra-urban scale
• overall and local effect

1. Introduction

Air pollution has become a severe global threat and was identified as a major cause of human mortality by World Health Organization (WHO) (World Health Organization 2018). Among all air pollutants, PM_2.5 (particulate matter aerodynamic with diameters smaller than 2.5 μm) has posed significantly negative effect on human health. It has been reported that high-concentration PM_2.5 can cause a range of respiratory and cardiovascular diseases including asthma, stroke, bronchitis, heart disease and lung cancer (Brook et al. 2010; Dominici et al. 2006; Liu, Saari et al. 2021). In 2015, PM_2.5–related deaths reached 4.2 million globally, with a large proportion occurring in developing countries such as China (Forouzanfar et al. 2016; Liu, Saari et al. 2021). Moreover, in China, diseases associated with PM_2.5 resulted in 100 billion US dollars of medical costs per year, accounting for 0.91% of national Gross Domestic Production (GDP) (Maji et al. 2018).

In urban areas, PM_2.5 pollution is particularly severe, primarily due to the presence of artificial impervious surface. Artificial impervious surface, including buildings, roads and other artificial hard surface, is the primary component of urban built-up areas and frequently harbors areas with severe air pollution (Mo et al. 2021; Qi et al. 2022). In the process of urbanization, the transformation from forest, cropland and many other natural surfaces to urban artificial impervious surface weakened the mitigating effect of ecological spaces on PM_2.5 (Chen et al. 2019a; Fan et al. 2022; Zhan, Yang, and Liu 2024). Simultaneously, high-intensity urban expansion brought numerous urban construction activities, resulting in high emission of air pollutants (Lim et al. 2020; Shi et al. 2020). Besides, accompanied with urban construction activities, the growth of urban residents and economic production in artificial impervious surface lead to significant energy consumption, releasing a substantial amount of traffic and industrial emissions into the urban atmospheric environment (Leffel et al. 2022; Pérez et al. 2010). In fact, with the expansion of artificial impervious surface, many cities in China and other developing countries are experiencing severe PM_2.5 pollution (Larkin et al. 2016; Wei et al. 2021). Therefore, understanding how artificial impervious surface affects PM_2.5 is crucial for mitigating urban PM_2.5 pollution and creating a sustainable living environment.

Urban areas with well-planned landscape spatial patterns of artificial impervious surface can exert a promoting effect on PM_2.5 pollution mitigation. Commonly, most of existing studies used landscape metrics to measure the landscape spatial patterns of artificial impervious surface or similar study objects, such as urban built-up areas (Gao et al. 2023; Liu, He et al. 2021; Shi et al. 2019). Size (coverage), shape, compactness, and many other spatial attributions of artificial impervious surface or built-up areas were quantified through landscape analysis and related influencing mechanism regarding to PM_2.5 has been preliminarily investigated. For example, it has been suggested that cities with compact urban built-up areas are beneficial for the mitigation of PM_2.5 pollution (Fan et al. 2018; Gao et al. 2023; Tao et al. 2020). In addition, Zhao et al. found that urban size played a more important role than other landscape metrics in China and larger cities tended to suffer from severer PM_2.5 pollution (Zhao et al. 2022). Certainly, it also should be pointed out that most studies were mainly conducted at large spatial scales such as national or urban-agglomeration scales, with one single city being viewed as a basic unit. Consequently, these studies tended to explore the effect of urban built-up areas rather than urban artificial impervious surface on PM_2.5 (Li and Zhou 2019; She et al. 2017). Nevertheless, intra-urban scale is closer to urban residents and how landscape spatial patterns of artificial impervious surface or urban built-up areas affect PM_2.5 at the intra-urban scale is still unclear. Moreover, it may be more meaningful to investigate the effect of landscape spatial patterns of artificial impervious surface rather than the whole urban built-up areas at the intra-urban scale. This is because urban built-up areas have clear and continuous ranges and contain many natural green and blue spaces such as trees and lakes, which can introduce some uncertainty to the investigation concerning the influencing mechanism of urban development on PM_2.5 (Kuang et al. 2013). Given that both population and economic activities are mainly distributed in artificial impervious surface (Wang and Gai 2021), for more reasonably demonstrating the effect of urban development on PM_2.5, focusing on artificial impervious surface is more appropriate at the intra-urban scale.

Apart from landscape spatial metrics, population and economic factors are also viewed as the essential parts of spatial patterns of artificial impervious surface. They are directly related with urban development and some studies have explored the effects of them on PM_2.5 at large spatial scales. It was found that there were “inversed-U” relationships between population and economic factors and PM_2.5 in many cities (Han et al. 2016; Zhao et al. 2019). This is because urban population and economic development initially tended to aggravate PM_2.5 pollution and then advanced its mitigation as the advancement of society and technology. In addition, Wu et al. found that the relationship between PM_2.5 and economy demonstrated an “inversed-N” shape in the eastern cities of China while an “inversed-U” shape in central China (J. Wu et al. 2018).

Even though the effects of spatial patterns of artificial impervious surface on PM_2.5 have been investigated to some extent, how they affect PM_2.5 at the intra-urban scale still remained to explore. From the perspective of urban development, urban areas can be broadly categorized into two typical classes: areas with drastic expansion of artificial impervious surface (DEA) and areas with slight or no changes of artificial impervious surface (SNCA) (Jiao, Mao, and Liu 2015). DEA is dominated by incremental planning and is undergoing rapid urban expansion, with most areas located in new towns. Dominated by inventory planning, SNCA generally exhibits low urban expansion intensity, as most urban constructions and road networks have already been built. The characteristics of landscape spatial metrics, population and economic development in these two types of areas are distinct (Zuo, Li, and Dong 2018), so the influence of spatial patterns of artificial impervious surface on PM_2.5 pollution may also be different. Considering these differences, for better guiding the PM_2.5 pollution mitigation at the intra-urban scale, it is worth focusing on these two typical areas with different urban development patterns and exploring the influence of the spatial patterns of artificial impervious surface on PM_2.5.

Additionally, artificial impervious surface, population and economic activities vary spatially (Krehl et al. 2016), which may lead to the effects of spatial patterns of artificial impervious surface on PM_2.5 within DEA or SNCA also vary spatially. To identify where the spatial patterns of artificial impervious surface affect PM_2.5 significantly, exploring local effects is also necessary. However, given the sparse distribution of ground monitoring sites of PM_2.5 at the intra-urban scale (Fan et al. 2020; Zhan, Yang, and Liu 2024), PM_2.5 concentration observations are insufficient for exploring the local effects of spatial patterns of artificial impervious surface. Fortunately, by applying remote sensing technique, spatially varied PM_2.5 concentrations can be derived and they can provide more information compared with disperse PM_2.5 ground sites (Li et al. 2020; Yin et al. 2022). Hence, deriving spatially varied PM_2.5 concentrations at the intra-urban scale and further investigating the spatially local effects of spatial patterns of artificial impervious surface on PM_2.5 is essential.

In this study, by using Wuhan, a megacity in China, as the study area, a comprehensive investigation to explore the effects of spatial patterns of artificial impervious surface on PM_2.5 were conducted. Specifically, this study aimed to 1) derive intra-urban PM_2.5 concentration maps for further investigation; 2) explore how spatial patterns of artificial impervious surface generally affect PM_2.5 in DEA and SNCA; 3) explore the local effects of spatial patterns of artificial impervious surface on PM_2.5 within DEA and SNCA.

2. Materials and methods

To achieve our goals, a research framework was developed (as Figure 1 shows). First, based on multi-source predictors, spatially varied PM_2.5 concentrations during 2015–2019 were derived based on a two-staged machine learning model. The reason for selecting the time span of 2015–2019 is that the “new urbanization strategies” of Wuhan was developed in 2014. It proposed the requirement for high-quality urban development (Zeng, Zhang, and Xu 2016). Different from the rapid development mode of the past two decades, the recent development of artificial impervious surface in Wuhan has significantly shifted its focus toward urban environmental protection, which is also aligned with China’s future urban development goals (Guan et al. 2018). Hence, selecting the time span of 2015–2019 can better represent the practical situation of the urban development in Wuhan. Then, by utilizing the data extracted from a land-use/cover dataset, annual landscape spatial patterns of artificial impervious surface were calculated by landscape analysis. Both DEA and SNCA were delineated according to the variations of artificial impervious surface coverages. Furthermore, to eliminate collinearity as much as possible, we selected appropriate influencing indicators by collinearity analysis with using Variation Inflation Factor (VIF) for measuring. Finally, Original Least Square (OLS) regression and Geographically Weighted Regression (GWR) were respectively used to explore the overall and local effects of spatial patterns of artificial impervious surface on PM_2.5 in both DEA and SNCA.

Figure 1. Flow chart of this study.

2.1. Study area

Wuhan, a megacity located in the middle of China, was selected as the study area (Figure 2). Wuhan has built-up areas of 885 km², population of over 10 million, and annual GDP of over 1800 billion yuan. With the development of social economy, urban air pollution has been a concern for this city and the annual mean PM_2.5 concentration of Wuhan has exceeded 100 μg/m³ in 2013 (Zhang et al. 2015). To address this issue, Wuhan adopted some measures such as industrial upgrades, resulting in the annual mean PM_2.5 concentration decreasing to 49 μg/m³ by 2018. However, this value was still much higher than the new guiding concentration (5 μg/m³), advised by WHO. Moreover, among all monitoring pollutants, the number of days when PM_2.5 was the primary urban pollutant was still the most in Wuhan. In our study, for focusing on the effects of urban artificial impervious surface on PM_2.5, Wuhan Urban Development Area was further selected as the study area because it contains the main urban built-up areas of Wuhan.

Figure 2. The study area of Wuhan.

2.2. PM_2.5 concentration derivation

Ground site-based PM_2.5 concentration observations cannot accurately depict the spatial variation of PM_2.5 concentrations, which can introduce some uncertainty into the investigation at the intra-urban scale. To obtain spatially varied PM_2.5 concentrations, we employed a two-staged XGBoost model that incorporated multi-source predictors and ground PM_2.5 concentration observations.

2.2.1. Data for PM_2.5 concentrations derivation

To derive PM_2.5 concentration maps, ground PM_2.5 concentration observations from 26 national sites during 2015–2019 and the data of 19 predictors were used. For increasing sampling density, in addition to the data from the 10 national sites in our study area, we also used the data from other 16 national sites located in neighboring cities with similar geographical and meteorological conditions to Wuhan. Daily mean PM_2.5 concentrations of each site were further calculated for model construction. The locations of these ground sites are shown in Figure S1. As for predictors, many temporal predictors were applied. Specifically, day of year, remote sensing predictors including Aerosol Optical Depth (AOD) and surface reflectance, meteorological predictors including relative humidity, temperature, pressure, east-west direction wind speed, north-south direction wind speed, and planetary boundary layer height were utilized. In addition, geographical predictors including elevation, road density, nightlight intensity and population density were also utilized. Among all predictors, AOD is an important property of aerosol and the main carrier of PM_2.5. It has been proved to be closely related with PM_2.5 and was also used in many previous studies for deriving PM_2.5 concentrations (Fan et al. 2020; Kahn, Banerjee, and McDonald 2001; Li et al. 2017). Multi‐Angle Implementation of Atmospheric Correction (MAIAC) AOD, an open-source AOD product with high spatial resolution of 1 km was used. In addition, PM_2.5 is also affected by meteorological conditions and land surface, so meteorological predictors, surface reflectance and some geographical predictors were also used for deriving PM_2.5 concentrations. Here, for better establishing the relationship between PM_2.5 and land surface and considering the spatiotemporal variation of land surface (Yang, Yuan, and Li 2022), 8-day averaged Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance with 500 m spatial resolution was applied. This dataset is of relatively high spatial and temporal resolution, which can provide sufficient spatio-temporal information for PM_2.5 concentration deriving. For decreasing uncertainty, reflectance with cloud covering rate less than 5% were used. Moreover, the spatial grids of reflectance were exploited for deriving PM_2.5 concentrations given that 500 m is fine enough for the investigation at the intra-urban scale. Details about the predictors are shown in Table S1.

2.2.2. Two-staged XGBoost model

Two-staged XGBoost, a machine learning model with two stages, was applied for PM_2.5 concentration deriving (Fan et al. 2020, 2022). It contains two sub-XGBoost models for AOD gap-filling and PM_2.5 concentration deriving, respectively. Due to the contamination of clouds, there are some absences in AOD data. This can render many PM_2.5 concentration observations corresponding to the AOD absences unusable and has significantly negative effect on the model construction of PM_2.5 concentration derivation (Li et al. 2023; Park et al. 2020; Yu et al. 2015). To minimize the negative effect, the first sub-model of the two-staged XGBoost integrated original AOD, day of year, meteorological predictors and geographical predictors to fill the gaps in original AOD (Fan et al. 2020). Then, with using the gap-filled AOD, ground PM_2.5 concentration observations, day of year, meteorological predictors and geographical predictors, the second sub-model derived spatially varied daily PM_2.5 concentration maps. Annual mean PM_2.5 concentration maps were further calculated. In addition, sample-based (assessing overall accuracy) and site-based (assessing spatial generalized ability) 10-fold cross-validation (CV) were used to assess the model performance (Fan et al. 2022; Gong et al. 2020). Root Mean Square Error (RMSE), Mean Predicted Error (MPE), and coefficient of determination (R²) were used as the metrics for model assessment. Details about the main tuning hyper parameters of the two-staged XGBoost model are shown in Table S2. Both the construction and assessment of the two-staged XGBoost were conducted in Python 3.7.

2.3. Analysis regarding to the spatial patterns of artificial impervious surface

2.3.1. Data for obtaining spatial patterns of artificial impervious surface

(1) Impervious surface

The artificial impervious surface data was obtained from the China’s Land-use/Cover Datasets (CLCD, https://zenodo.org/record/5816591) (Yang and Huang 2021). The dataset contains annual land cover data of China from 1990 to 2019 with spatial resolution of 30 m, which is beneficial for the investigation at the intra-urban scale. In addition, after testing more than 5000 visual interpretation samples and 5000 third-party test samples (Yang and Huang 2021), the overall classification accuracy of this dataset can reach about 80%, with high reliability. Hence, given high requirement of spatio-temporal accuracy and temporal continuity on artificial impervious surface, the CLCD dataset is suitable for the investigation.

The dataset includes eight land cover types and the data of impervious surface from 2015 to 2019 was used for the investigation of influencing mechanism. Furthermore, to explore the influence of artificial impervious surface on PM_2.5, natural impervious surfaces such as beach and bare land were removed manually.

(2) Population and economic data

In addition to impervious surface, population density and nightlight intensity were also used as the indicators for the investigation. In the process of urban development, after the expansion of artificial impervious surface, population and economic activities will increase, resulting in air pollutants emitted to urban atmosphere. Therefore, both population and economy are also the important attributes of artificial impervious surface spatial patterns (Han 2018). For depicting the economic development in artificial impervious surface, night intensity with spatial resolution of about 500 m was obtained from NPP-VIIRS satellite and downloaded in Google Earth Engine (https://earthengine.google.com/). Annual mean nightlight intensity during 2015–2019 was further calculated. Additionally, annual population with spatial resolution of 1 km was downloaded from the world population database (https://www.worldpop.org/).

(3) Meteorological data

In the derivation of PM_2.5 concentrations, measured by variable importance, temperature, relative humidity and wind speed were found to have the most significant impact on deriving PM_2.5 concentrations among all meteorological factors (Figure S2). In addition, these three indicators were also commonly used in previous studies for investigating the influencing mechanism regarding PM_2.5 (Fan et al. 2018; Gao et al. 2023; Zhao et al. 2021). Hence, they were used as the auxiliary indicators for investigating the influencing mechanism. In the investigation of influencing mechanism, wind speed is the root of the arithmetically squared sum of the east-west wind speed and the north-south wind speed, which were used in PM_2.5 concentration deriving. The annual average values of these three meteorological factors were calculated for the investigation.

2.3.2. Landscape analysis

Landscape spatial patterns of artificial impervious surface have been proved to be related with PM_2.5 concentrations at national or urban-agglomeration scales. To reduce information redundancy, we selected candidate indicators from three aspects: size, shape and aggregation after referring some existing studies (Liang and Gong 2020; Shi et al. 2019). Specifically, they contain: Percentage of Landscape (PLAND, size), Area-weighted Mean Shape Index (AWMSI, shape), Area-weight Perimeter-area Fractal Dimension Index (AWFRAC, shape), Contagion (CO, aggregation), Aggregation Index (AI, aggregation), Dispersion Index (DI, aggregation).

Among them, PLAND represents the coverage of artificial impervious surface and its equation is Equation (1)(1) $\mathrm{PLAND}=A_i/\mathit{A}\ast 100\mathrm{\%}$(1) . Here, $A_i$ is the area of artificial impervious surface in the i^th grid and $\mathit{A}$ is the area of a grid.

(1) $\mathrm{PLAND}=A_i/\mathit{A}\ast 100\mathrm{\%}$(1)

AWMSI and AWFRAC can represent the shape complexity of artificial impervious surface. The larger the two indicators, the more complex and irregular the shape of artificial impervious surface is. The range of AWMSI and AWFRAC are (1,+∞) and (1, 2) respectively. The equations of AWMSI and AWFRAC are Equation (2)(2) $\mathrm{AWMSI}=\mathit{\hspace{0.17em}}\sum _{\mathit{j}=1}^n\left[0.25\ast p_{\mathit{ij}}/a_{\mathit{ij}}\ast a_{\mathit{ij}}/\sum _{\mathit{j}=1}^n{a}_{\mathit{ij}}\right]$(2) and Equation (3)(3) $\mathrm{AWFRAC}=\sum _{\mathit{j}=1}^n\left[2\ast ln\left(0.25\ast p_{\mathit{ij}}\right)/ln\left(a_{\mathit{ij}}\right)\ast a_{\mathit{ij}}/\sum _{\mathit{j}=1}^n{a}_{\mathit{ij}}\right]$(3) , respectively. Here, $a_{\mathit{ij}}$ is the area of the j^th artificial impervious surface patch in the i^th grid, and $p_{\mathit{ij}}$ is the perimeter of the j^th artificial impervious surface patch in the i^th grid.

(2) $\mathrm{AWMSI}=\mathit{\hspace{0.17em}}\sum _{\mathit{j}=1}^n\left[0.25\ast p_{\mathit{ij}}/a_{\mathit{ij}}\ast a_{\mathit{ij}}/\sum _{\mathit{j}=1}^n{a}_{\mathit{ij}}\right]$(2)

(3) $\mathrm{AWFRAC}=\sum _{\mathit{j}=1}^n\left[2\ast ln\left(0.25\ast p_{\mathit{ij}}\right)/ln\left(a_{\mathit{ij}}\right)\ast a_{\mathit{ij}}/\sum _{\mathit{j}=1}^n{a}_{\mathit{ij}}\right]$(3)

CO, AI and DI can represent the aggregation of artificial impervious surface. Higher levels of CO and AI or lower levels of DI represent more compact artificial impervious surface. The equations of CO and AI are Equation (4)(4) $\mathrm{CO}=\left[1+\frac{{\sum }_{i=1}^m{\sum }_{k=1}^m\left(P_i\ast \frac{{\mathit{g}}_{\mathit{ik}}}{{\sum }_{k=1}^m{g}_{\mathit{ik}}}\right)\ast \left(\ln P_i\left(\frac{{\mathit{g}}_{\mathit{ik}}}{{\sum }_{k=1}^m{g}_{\mathit{ik}}}\right)\right)}{2ln\left(\mathit{m}\right)}\right]\ast 100$(4) and Equation (5)(5) $\mathrm{AI}=\left[\sum _{\mathit{j}=1}^n\left(g_{\mathit{ij}}/max(g_{\mathit{ij}})\right)\right]\ast 100$(5) , respectively. In Equation (4)(4) $\mathrm{CO}=\left[1+\frac{{\sum }_{i=1}^m{\sum }_{k=1}^m\left(P_i\ast \frac{{\mathit{g}}_{\mathit{ik}}}{{\sum }_{k=1}^m{g}_{\mathit{ik}}}\right)\ast \left(\ln P_i\left(\frac{{\mathit{g}}_{\mathit{ik}}}{{\sum }_{k=1}^m{g}_{\mathit{ik}}}\right)\right)}{2ln\left(\mathit{m}\right)}\right]\ast 100$(4) , $P_i$ is the percentage of the i^th patch, $g_{\mathit{ik}}$ is the number of adjacent edges among all artificial impervious surface patches in a grid, and m is the number of patches. In Equation (5)(5) $\mathrm{AI}=\left[\sum _{\mathit{j}=1}^n\left(g_{\mathit{ij}}/max(g_{\mathit{ij}})\right)\right]\ast 100$(5) , $g_{\mathit{ij}}$ is the number of the adjacent edges between two patches, $max\left(g_{\mathit{ij}}\right)$ is the maximum value of the number of adjacent edges, and n is the number of all patches. DI is the combination of the number of patches (NP) and the Largest Patch Index (LPI), which can represent the fragmentation and integrity of patches respectively (Taubenböck et al. 2019). NP is the number of artificial impervious surface patches, and LPI is the proportion of the largest patch’s area in the total artificial impervious surface area. The combination of them can represent the aggregation of artificial impervious surface patches. The equation of DI is Equation (6)(6) $\mathrm{DI}=\frac{N{P}_n+\left(100-\mathit{L}{P}_n\right)}{2}$(6) . Here, $\mathit{N}{P}_n$ and $\mathit{L}{P}_n$ are the normalized NP and LPI, respectively. The ranges of them and DI are all in [0,100] and lower DI can depict more compact artificial impervious surface. More details about DI are represented in the supplementary file.

(4) $\mathrm{CO}=\left[1+\frac{{\sum }_{i=1}^m{\sum }_{k=1}^m\left(P_i\ast \frac{{\mathit{g}}_{\mathit{ik}}}{{\sum }_{k=1}^m{g}_{\mathit{ik}}}\right)\ast \left(\ln P_i\left(\frac{{\mathit{g}}_{\mathit{ik}}}{{\sum }_{k=1}^m{g}_{\mathit{ik}}}\right)\right)}{2ln\left(\mathit{m}\right)}\right]\ast 100$(4)

(5) $\mathrm{AI}=\left[\sum _{\mathit{j}=1}^n\left(g_{\mathit{ij}}/max(g_{\mathit{ij}})\right)\right]\ast 100$(5)

(6) $\mathrm{DI}=\frac{N{P}_n+\left(100-\mathit{L}{P}_n\right)}{2}$(6)

2.3.3. Delineation of DEA and SNCA

This study focused on the investigation at the intra-urban scale and explored how spatial patterns of artificial impervious surface affected PM_2.5 in two typical urban areas, i.e. DEA and SNCA. To characterize these two abstract concepts, we made definitions for them based on the spatiotemporal coverages of artificial impervious surface during 2015–2019. For DEA, the definition is: from 2015 to 2019, the coverage of artificial impervious surface remained unchanged or was higher than that in the previous year, and the increase of artificial impervious surface coverage from 2015 to 2019 exceeded a certain threshold. For SNCA, the definition is: from 2015 to 2019, the inter-annual variations of artificial impervious surface coverages were less than 1%, and the artificial impervious surface coverage in 2015 exceeded a certain threshold. These two definitions are represented by Equation (7)(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8) and Equation (8)(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8) , respectively. Here, $A_1$and $A_2$ represent the type of DEA and SNCA respectively, $A_{2015-2019}^i$ is the increase of artificial impervious surface coverage in the i^th grid from 2015 to 2019, $A_y^i$ is the artificial impervious surface coverage in the i^th grid in a year, $A_{y-1}^i$ is the artificial impervious surface coverage in the i^th grid in the previous year, $A_{2015}^i$ is the artificial impervious surface coverage in the i^th grid in 2015, and $T_1$ and $T_2$ are the thresholds to identify DEA and SNCA respectively. Additionally, for focusing on the investigation within urban built-up areas, the identified areas outside the urban built-up areas would be removed.

(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8)

2.3.4. Indicators selection

Although the candidate indicators comprehensively reflected the spatial patterns of artificial impervious surface, the information they contained still had some redundancy. Therefore, before data analysis, we conducted collinearity analysis by adopting Variation Inflation Factor (VIF) to detect the collinearity among all indicators in both DEA and SNCA, and removed the indicators with high collinearity. The higher the VIF value is, the greater the collinearity between the corresponding indicator and other explanatory variables. The indicator with the highest VIF was removed in every VIF calculation until all VIFs of the remaining indicators were lower than a threshold. After referring some studies related to the influencing mechanism of PM_2.5 (Chen et al. 2019b; G. Gao et al. 2023), the threshold was set as 10 in our study.

2.4. Statistical methods

2.4.1. Correlation analysis

Pearson Correlation Coefficient was used to measure the relationships between the spatial patterns of artificial impervious surface and PM_2.5 concentrations. Its equation is Equation (9)(9) $R_{\mathit{xy}}=\sum _{\mathit{i}=1}^n\left(x_i-\bar{x}\right)\left(y_i-\bar{y}\right)/\surd \sum _{\mathit{i}=1}^n{\left(x_i-\bar{x}\right)}^2\surd \sum _{\mathit{i}=1}^n{\left(y_i-\bar{y}\right)}^2$(9) . Here, $\mathit{x}$ represented one indicator of artificial impervious surface spatial patterns, $\mathit{y}$ represented PM_2.5 concentration, and $R_{\mathit{xy}}$ is the correlation coefficient between them, with value ranging from $-$1 to 1. If the value is closer to 1, it indicates a stronger positive correlation. While if it is closer to $-$1, it suggests a stronger negative correlation.

(9) $R_{\mathit{xy}}=\sum _{\mathit{i}=1}^n\left(x_i-\bar{x}\right)\left(y_i-\bar{y}\right)/\surd \sum _{\mathit{i}=1}^n{\left(x_i-\bar{x}\right)}^2\surd \sum _{\mathit{i}=1}^n{\left(y_i-\bar{y}\right)}^2$(9)

2.4.2. OLS regression

To investigate the overall effects of spatial patterns of artificial impervious surface on PM_2.5 concentrations, OLS regression was used. Compared with correlation analysis, which only examines the correlation between one single indicator and PM_2.5 concentration, OLS regression can simultaneously explore the effects of multiple indicators on PM_2.5 concentration. The equation of OLS regression is Equation (10)(10) $\mathit{Y}={\beta }_0+\sum _{\mathit{k}=1}^m{\beta }_k{X}_k+\mathit{\epsilon }$(10) . Here, X is the indicator of the spatial patterns of artificial impervious surface, ${\beta }_k$ is the corresponding regression coefficient, ${\beta }_0$ is the intercept, Y is PM_2.5 concentration, and ε is the random error. In addition, to facilitate the comparison between different years and between DEA and SNCA, all indicators of each year were normalized.

(10) $\mathit{Y}={\beta }_0+\sum _{\mathit{k}=1}^m{\beta }_k{X}_k+\mathit{\epsilon }$(10)

2.4.3. GWR

GWR was applied to investigate the local effects of spatial patterns of artificial impervious surface on PM_2.5 concentrations. It constructs many OLS regression models in different locations and the local coefficients can represent the local effects of indicators on PM_2.5 concentrations (Brunsdon, Fotheringham, and Charlton 1998; Fotheringham, Brunsdon, and Charlton 2003). Its equations are Equations (11)(11) $\mathit{Y}\left(u_i,v_i\right)={\beta }_0\left(u_i,v_i\right)+\sum _{\mathit{k}=1}^m{\beta }_k\left(u_i,v_i\right)X_k\left(u_i,v_i\right)+{\epsilon }_i$(11) –(13(13) $\mathit{W}\left(\mathit{i},\mathit{j}\right)=\left\{\begin{array}{c}{\left[1-{\left(\frac{{\mathit{d}}_{\mathit{ij}}}{b_i}\right)}^2\right]}^2,d_{\mathit{ij}}<b_i\\ 0,d_{\mathit{ij}}\ge b_i\end{array}\right.$(13) ), respectively. Here, Equation (11)(11) $\mathit{Y}\left(u_i,v_i\right)={\beta }_0\left(u_i,v_i\right)+\sum _{\mathit{k}=1}^m{\beta }_k\left(u_i,v_i\right)X_k\left(u_i,v_i\right)+{\epsilon }_i$(11) is similar with Equation (10)(10) $\mathit{Y}={\beta }_0+\sum _{\mathit{k}=1}^m{\beta }_k{X}_k+\mathit{\epsilon }$(10) but all parameters have their own locations. By Equation (12)(12) $\widehat{\beta \left(u_i,v_i\right)}=[X^{\mathrm{T}}\mathit{W}\left(u_i,v_i\right)\mathit{X}{]}^{-1}[X^{\mathrm{T}}\mathit{W}\left(u_i,v_i\right)\mathit{Y}$(12) , GWR can estimate the local coefficients. $\mathit{W}\left(u_i,v_i\right)$ is the weighted matrix that considers the relationship between neighboring samples and the local sample. In this study, we used Bi-square function (Equation (13)(13) $\mathit{W}\left(\mathit{i},\mathit{j}\right)=\left\{\begin{array}{c}{\left[1-{\left(\frac{{\mathit{d}}_{\mathit{ij}}}{b_i}\right)}^2\right]}^2,d_{\mathit{ij}}<b_i\\ 0,d_{\mathit{ij}}\ge b_i\end{array}\right.$(13) ) to generate the weighted matrix. $d_{\mathit{ij}}$ is the distance between the i^th sample and the j^th sample and $b_i$ is the bandwidth in the i^th sample.

(11) $\mathit{Y}\left(u_i,v_i\right)={\beta }_0\left(u_i,v_i\right)+\sum _{\mathit{k}=1}^m{\beta }_k\left(u_i,v_i\right)X_k\left(u_i,v_i\right)+{\epsilon }_i$(11)

(12) $\widehat{\beta \left(u_i,v_i\right)}=[X^{\mathrm{T}}\mathit{W}\left(u_i,v_i\right)\mathit{X}{]}^{-1}[X^{\mathrm{T}}\mathit{W}\left(u_i,v_i\right)\mathit{Y}$(12)

(13) $\mathit{W}\left(\mathit{i},\mathit{j}\right)=\left\{\begin{array}{c}{\left[1-{\left(\frac{{\mathit{d}}_{\mathit{ij}}}{b_i}\right)}^2\right]}^2,d_{\mathit{ij}}<b_i\\ 0,d_{\mathit{ij}}\ge b_i\end{array}\right.$(13)

3. Results

3.1. PM_2.5 concentration deriving

The two-staged XGBoost performed well to derive PM_2.5 concentrations. Both sample-based and site-based assessment represented great performance with R², RMSE, MPE of 0.94, 7.88 μg/m³, 5.26 μg/m³ and 0.90, 10.55 μg/m³, 7.40 μg/m³ respectively. High R² values in both sample-based and site-based assessment indicated good ability of the two-staged XGBoost among all samples and spatially for PM_2.5 concentration deriving.

Annual mean PM_2.5 concentration maps were further derived and were as Figure 3 shows. It was demonstrated that Wuhan suffered from the most serious PM_2.5 pollution in 2015, with concentrations ranging from 59.56 μg/m³ to 78.75 μg/m³. Then, by 2019, PM_2.5 concentrations generally decreased, ranging from 39.96 μg/m³ to 57.71 μg/m³. Additionally, it is worth noting that the decline rate slowed down during 2016–2019. Spatially, the distributions of all annual mean PM_2.5 concentration maps represented similar patterns, with high PM_2.5 concentrations in some urban central areas such as Jianghan district, and low PM_2.5 concentrations in some suburban areas such as Jiangxia district. Moreover, some areas with numerous industrial parks, such as Qingshan district, also suffered from serious PM_2.5 pollution.

Figure 3. Annual mean PM_2.5 concentration maps.

3.2. Extractions of DEA and SNCA and indicators selection

3.2.1. Extraction of DEA and SNCA

The spatial distribution of artificial impervious surface coverages during 2015–2019 and the increasing rates of artificial impervious surface in each administrative district of the study area are shown in Figure 4 and Table 1 respectively. Based on the natural breaks (Jenks), artificial impervious surface coverages in each year were divided into three levels: low coverage, medium coverage and high coverage. In general, it can be seen that high artificial impervious surface coverages were mainly located in the urban central areas within the third ring road and low coverages were mainly found in suburban areas. From the perspective of administrative districts, as shown in Table 1, the artificial impervious surface coverages in urban central districts were very high, with values exceeding 50%. The artificial impervious surface coverages in suburban districts were about 20%. However, from 2015 to 2019, the increasing rates of artificial impervious surface in suburban districts were much higher than those in urban central districts. Except for Hongshan District and Hanyang District, the increasing rates of artificial impervious surface in urban central districts were less than 3%, while suburban districts experienced increasing rates exceeding 15%.

Figure 4. Coverages of artificial impervious surface.

Table 1. Coverages and increasing rates of artificial impervious surface in administrative districts of the study area.

	District	Year					Increasing rate
		2015	2016	2017	2018	2019
Urban central districts	Jiang’an	61.98%	62.41%	62.80%	63.48%	63.62%	2.65%
	Jianghan	91.81%	91.93%	92.06%	92.15%	92.15%	0.37%
	Qiaokou	85.10%	85.84%	86.35%	86.63%	86.71%	1.88%
	Hanyang	57.34%	58.65%	59.35%	60.51%	60.96%	6.30%
	Qingshan	71.28%	71.54%	71.76%	71.99%	72.44%	1.63%
	Wuchang	51.01%	51.02%	51.03%	51.08%	51.09%	0.16%
	Hongshan	32.42%	33.41%	34.24%	35.38%	35.92%	10.79%
Suburban districts	Dongxihu	21.07%	21.99%	22.94%	23.89%	24.68%	17.16%
	Hannan	31.56%	32.76%	34.14%	35.37%	37.56%	19.00%
	Jiangxia	15.97%	16.96%	17.80%	18.67%	19.05%	19.28%
	Xinzhou	12.46%	12.99%	13.49%	14.10%	14.47%	16.12%
	Caidian	22.50%	23.76%	24.66%	25.65%	26.26%	16.71%
	Huangpi	16.39%	17.17%	18.01%	18.97%	19.42%	18.48%

DEA was extracted mainly based on the variations of artificial impervious surface (as Equation (7)(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8) shows). The spatial variations of artificial impervious surface are shown in Figure 5(a). It can be found that most high increasing rates of artificial impervious surface were located in areas near to the fourth ring road and areas between the third and fourth ring road. Identified by Jenks, the threshold in Equation (7)(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8) was set as 7.3%, which was the threshold between “low increasing area” and “medium increasing area”. By setting this threshold, we can not only filter out the areas with low increasing rates but also capture rapidly expanding urban areas. In addition, for ensuring that the final results can more accurately represent the situation of urban context, few areas that were not located within the urban built-up areas (as shown in Figure 5(b)) were excluded from DEA. The data of urban built-up area was obtained from the Globeland30 product in 2020 (http://www.globallandcover.com).

Figure 5. (a) Spatial variations of artificial impervious surface during 2015.

The extraction of SNCA was mainly based on the coverages of artificial impervious surface in 2015 and the variations of artificial impervious surface during 2015–2019 (as Equation (8)(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8) shows). Since SNCA already had high artificial impervious surface coverages in 2015, the threshold in Equation (8)(8) $\mathrm{Are}{\mathrm{a}}^i=\left\{\begin{array}{c}{A}_1,A_{2015-2019}^i>T_1\mathit{and}{A}_y^i\ge A_{y-1}^i(7)\\ A_2,A_{2015}^i>T_2\mathit{and}\left(A_y^i-A_{y-1}^i\right)<1\mathrm{\%}\end{array}\right.$(8) was set as 55% by Jenks, which was between “Medium coverage” and “High coverage” (as Figure 4 shows). Also, similar with the extraction of DEA, areas that were located outside the boundaries of urban built-up areas were also excluded from SNCA.

Final extracted DEA and SNCA are as Figure 6 shows. DEAs were mainly distributed in some new urban areas outside the third ring road, while SNCAs were mainly distributed in the urban central areas within the third ring road and some industrial parks in suburban areas. These two types of areas basically contained the majority of urban built-up regions, thereby providing benefits for the comprehensive investigation at the intra-urban scale. Furthermore, some statistical metrics of PM_2.5 concentrations in both DEA and SNCA were calculated and are shown in Table 2. All mean PM_2.5 concentrations in DEA were lower than that in SNCA during 2015–2019. Moreover, both the maximum and variance of PM_2.5 concentrations in DEA were the highest in 2015 and subsequently decreased, representing a reduction of PM_2.5 pollution in DEA after 2015. However, it is also noteworthy that the mean PM_2.5 concentration in DEA was gradually approaching that in SNCA, indicating that the mitigation of PM_2.5 pollution in DEA were weakening.

Figure 6. The spatial distribution of DEA and SNCA.

Table 2. Statistical values of PM_2.5 concentrations in DEA and SCNA.

	Minimum (μg/m^3)	Maximum (μg/m^3)	Mean (μg/m^3)	Variance (μg/m^3)
DEA in 2015	59.56	78.75	68.43	2.59
SNCA in 2015	61.76	77.15	70.15	2.17
DEA in 2016	50.08	63.28	55.81	2.49
SNCA in 2016	49.98	63.45	57.05	2.29
DEA in 2017	48.23	58.35	52.75	1.92
SNCA in 2017	48.62	62.13	54.06	2.18
DEA in 2018	40.18	53.99	46.42	1.80
SNCA in 2018	41.86	59.01	47.78	2.18
DEA in 2019	37.77	50.35	44.50	1.82
SNCA in 2019	38.77	57.71	45.70	2.65

3.2.2. Indicators selection

After extracting DEA and SNCA, we conducted a collinearity analysis by using VIF to select influencing indicators for the further investigation. The results are shown in Figure 7. Compared with the indicators in DEA, the indicators in SNCA exhibited stronger collinearity, with many VIFs exceeding 10. Therefore, we selected indicators according to the result of SNCA first. The indicators with the highest VIF for each year were sequentially removed, including AI, AWMSI, CO and temperature. Then, a collinearity detection was conducted again after the selection, and it was found that all VIFs of the remaining indicators were less than 10, which met the requirements. Similarly, in DEA, after removing AI, AWMSI, CO and temperature, it was found that all VIFs of the remaining indicators were also less than 10. Therefore, the remaining indicators, i.e. PLAND, AWFRAC, wind speed, relative humidity, DI, population density and nightlight intensity, were applied to explore the influencing mechanism. Among them, PLAND, AWFRAC, and DI represented the size, shape and aggregation of the artificial impervious surface, respectively. Population density and nightlight intensity represented population and economy, respectively.

Figure 7. VIFs in DEA and SNCA before and after indicators selection. Here, PLAND represents the percentage, AWMSI represents the area-weighted mean shape index, AWFRAC represents the area-weight perimeter-area fractal dimension index, CO represents the contagion, AI represents the aggregation index, WS represents the wind speed, RH represents the relative humidity, TEM represents the temperature, DI represents the dispersion index, POP represents the population density, NL represents the nightlight intensity.

3.3. Effects of artificial impervious surface spatial patterns in DEA and SNCA

3.3.1. Overall effects in DEA and SNCA

The overall correlation relationships between spatial patterns of artificial impervious surface and PM_2.5 in DEA were explored by correlation analysis and the results are shown in Figure 8(a). It can be found that the PLAND of artificial impervious surface exhibited a strong positive correlation with PM_2.5 concentrations. Among other patterns, DI showed a negative correlation with PM_2.5 concentrations. In terms of population and economic factors, both population density and nightlight intensity showed positive correlations with PM_2.5 concentrations. Moreover, the positive correlation between nightlight intensity and PM_2.5 concentration was the strongest among all spatial patterns, indicating that in DEA, economic development may have the strongest association with PM_2.5 pollution. Two meteorological factors showed negative correlations with PM_2.5 concentrations, indicating that higher wind speed and relative humidity may provide benefits for decreasing PM_2.5 concentrations in DEA.

Figure 8. The correlation coefficients between spatial patterns of artificial impervious surface and PM_2.5 in DEA (a) and SNCA (b). The blank indicates that the coefficient didn’t pass the significance level test of p < 0.05. Here, PLAND represents the percentage, AWFRAC represents the area-weight perimeter-area fractal dimension index, WS represents the wind speed, RH represents the relative humidity, DI represents the dispersion index, POP represents the population density, NL represents the nightlight intensity.

In comparison, the correlations between the spatial patterns of artificial impervious surface and PM_2.5 concentrations in SNCA are stronger than that in DEA (Figure 8(b)). The positive correlation between the PLAND of artificial impervious surface and PM_2.5 concentration was the strongest. As for other landscape spatial patterns, AWFRAC displayed a significantly negative correlation with PM_2.5 concentrations, which was different from the situation in DEA. In SNCA, when the edges of artificial impervious surface are complex, there are usually some irregular natural water bodies or woodlands around artificial impervious surface, and they can contribute to PM_2.5 pollution mitigation. In addition, DI also had a negative correlation with PM_2.5. In terms of meteorological factors, wind speed and relative humidity, similar to the correlations in DEA, showed negative associations with PM_2.5 concentrations in most years.

To analyze the overall effects of the influencing indicators on PM_2.5, OLS was applied. Table 3 showed the results of model assessment and regression coefficients in DEA. The metrics of model assessment contain R², adjusted R² (Adj.R²) and RMSE. Higher R², adjusted R² and lower RMSE represent better model performance. It can be found that both R² and adjusted R² increased and RMSE decreased during 2015–2019, indicating an improvement temporally. Among all spatial patterns, economy (represented by NL, nightlight intensity) had the greatest effect on aggravating PM_2.5 pollution, which was similar with the results in correlation analysis. The regression coefficient of NL did not consistently increase but displayed a trend of initially increasing, then decreasing, and finally slowly increasing during 2015–2019. In addition, the aggravating effect of population on PM_2.5 pollution became stronger and stronger after 2016. In comparison, the coefficients of landscape spatial patterns demonstrated decreasing trends and the effects of them on PM_2.5 were weaker than that of population after 2016. This indicated that after the drastic expansion of artificial impervious surface, accompanied increasing population may bring serious PM_2.5 pollution. The regression coefficients of PLAND, AWFRAC and DI were all positive, indicating that the artificial impervious surface with large coverage, complex shape and uncompact distribution can aggravate PM_2.5 pollution in DEA. Also, it should be noted that the regression coefficients of DI were positive but the corresponding coefficients in correlation analysis were negative. This suggested that correlation analysis may not accurately reflect the practical impact when multiple indicators were considered. In terms of meteorological factors, except for the wind speed in 2015, both high wind speed and relative humidity can mitigate PM_2.5 pollution. The reason for the positive coefficient of wind speed in 2015 May be related with the extensive construction of buildings and roads in 2015 and wind speed generally promoted the spread of PM_2.5 from construction sites to atmospheric environment (Xie et al. 2022).

Table 3. The OLS coefficients of spatial patterns of artificial impervious surface in DEA.

	Year
	2015	2016	2017	2018	2019
PLAND	0.05	0.10	0.10	0.09	0.08
AWFRAC	0.05	0.06	0.03	0.02	–
DI	0.06	0.09	0.07	0.06	0.05
POP	–	–	0.12	0.17	0.21
NL	0.35	1.05	0.88	0.82	0.83
WS	0.09	−0.06	−0.03	−0.07	–
RH	−0.05	−0.09	−0.04	−0.02	−0.06
Model assessment	R^2 = 0.10	R^2 = 0.12	R^2 = 0.17	R^2 = 0.21	R^2 = 0.28
	Adj.R^2 = 0.10	Adj.R^2 = 0.12	Adj.R^2 = 0.17	Adj.R^2 = 0.21	Adj.R^2 = 0.28
	RMSE = 0.13μg/m^3	RMSE = 0.16μg/m^3	RMSE = 0.11μg/m^3	RMSE = 0.08μg/m^3	RMSE = 0.07μg/m^3

Note: “-” indicates that the coefficient didn’t pass the significance level test of p<0.05. Here, PLAND represents the percentage, AWFRAC represents the area-weight perimeter-area fractal dimension index, WS represents the wind speed, RH represents the relative humidity, DI represents the dispersion index, POP represents the population density, NL represents the nightlight intensity.

The overall effects of spatial patterns of artificial impervious surface on PM_2.5 in SNCA were also explored by OLS and the results were as Table 4 shows. All OLS models performed better than those in DEA during 2015–2019. Temporally, the changed trends of the effects of PLAND, population and economy were basically the same as those in DEA and the effect of economy was also the strongest. However, the regression coefficients of economy in SNCA were much lower than those in DEA, indicating that the economic development in SNCA had weaker effect on PM_2.5 pollution. A similar phenomenon was also reflected in population. In terms of the landscape spatial patterns of artificial impervious surface, the coefficients of PLAND in SNCA were significantly higher than those of other landscape spatial patterns and were also higher than those of the PLAND in DEA. This indicated that artificial impervious surface coverage was more important than other landscape attributes in affecting PM_2.5 in SNCA. AWFRAC had negative impacts on PM_2.5 concentrations in some years, which were much weaker than that of PLAND (lower absolute values). In addition, DI had no significant effect on PM_2.5 in most years. In terms of meteorological factors, higher wind speed and relative humidity could bring stronger effects on mitigating PM_2.5 pollution when compared with those in DEA.

Table 4. The OLS coefficients of spatial patterns of artificial impervious surface in SNCA.

	Year
	2015	2016	2017	2018	2019
PLAND	0.15	0.39	0.36	0.35	0.33
AWFRAC	–	–	−0.03	−0.02	−0.05
DI	0.02	–	–	–	–
POP	0.12	0.13	–	0.05	0.14
NL	0.24	0.56	0.50	0.52	0.55
WS	−0.05	–	−0.10	–	−0.12
RH	−0.18	–	−0.11	−0.11	−0.25
Model assessment	R^2 = 0.27	R^2 = 0.23	R^2 = 0.24	R^2 = 0.27	R^2 = 0.35
	Adj.R^2 = 0.27	Adj.R^2 = 0.23	Adj.R^2 = 0.24	Adj.R^2 = 0.27	Adj.R^2 = 0.35
	RMSE = 0.10μg/m^3	RMSE = 0.14μg/m^3	RMSE = 0.12μg/m^3	RMSE = 0.10μg/m^3	RMSE = 0.08μg/m^3

Note: “-” indicates that the coefficient didn’t pass the significance level test of p<0.05. Here, PLAND represents the percentage, AWFRAC represents the area-weight perimeter-area fractal dimension index, WS represents the wind speed, RH represents the relative humidity, DI represents the dispersion index, POP represents the population density, NL represents the nightlight intensity.

3.3.2. Local effects in DEA and SNCA

By GWR, the local effects of spatial patterns of artificial impervious surface on PM_2.5 were explored. Compared to OLS, GWR performed better, with both R² and Adj.R² exceeding 0.5 in most years (Table 5). To ensure the validity of the investigation, only the local coefficients passing the significance level test of p < 0.05 were retained and the proportions were as Table 6 shows. It can be found that, compared with those of AWFRAC, DI and POP, much more local coefficients of PLAND and NL passed the significance level test. This indicated that the PLAND and economic factor of artificial impervious surface may play more important roles in affecting local PM_2.5 concentrations. Therefore, in the further analysis, we would focus on the local effects of PLAND and NL on PM_2.5.

Table 5. The performance of GWR.

	Year
	2015	2016	2017	2018	2019
R^2	0.57	0.58	0.62	0.64	0.73
Adj.R^2	0.52	0.52	0.57	0.60	0.69

Table 6. Proportions that the local coefficients passed the significance level test of p < 0.05.

	Year
	2015	2016	2017	2018	2019
PLAND	23%	35%	43%	37%	33%
AWFRAC	5%	7%	6%	5%	3%
DI	9%	11%	10%	6%	4%
POP	8%	6%	9%	9%	13%
NL	19%	35%	37%	45%	53%

Note: PLAND represents the percentage, AWFRAC represents the area-weight perimeter-area fractal dimension index, DI represents the dispersion index, POP represents the population density, NL represents the nightlight intensity.

The local coefficients of PLAND are shown in Figure 9 and their spatial variations were demonstrated. In general, the significant coefficients showed similar spatial patterns during 2015–2019 and most of them were positive, indicating that the increase of PLAND led to more serious PM_2.5 pollution in these areas. Spatially, most coefficients with high values were found in the SNCA within the second ring road, especially in the boundary region between Jianghan district and Qiaokou district, which contains many urban villages and industrial parks with dense buildings. In these areas, though urban construction intensity was lower than that in DEA and industrial activities were less than those in some large industrial parks located in suburban areas, densely distributed buildings can block air flow and make it difficult for PM_2.5 to disperse (Yuan et al. 2019; Y. Zhou et al. 2019). In addition, some DEAs were near to these SNCAs and the intensive construction activities from the DEAs may also bring some air pollutants.

Figure 9. The local coefficients of PLAND.

The local coefficients of NL were depicted in Figure 10. Temporally, the spatial distributions of the local coefficients of NL in these 5 years were similar. Spatially, the spatial distribution of the local coefficients of NL were quite different from that of PLAND. Most local coefficients within the second ring road were insignificant and high NL local coefficients were more prevalent in the DEAs outside or near the third ring road. Most of these DEAs were distributed in suburban districts such as Caidian, Jiangxia and Huangpi, which are suffering from relatively rapid urban expansion. Compared with the districts in urban central areas, these areas had more factories with high emissions, making them more susceptible to PM_2.5 pollution. In addition, given lower standards and attention, air pollution control related with economic activities in these areas was still not as strict as that in urban central areas. Certainly, it should be noted that in some areas of SNCA, economic develop also brought negative effects on the mitigation of PM_2.5 pollution. For example, there were many high local coefficients in the Qingshan industrial park, located in Qingshan district. This industrial park is an important heavy industrial base for steel and chemical manufacturing and has been established for decades. The economic activities here have resulted in significant emissions and serious PM_2.5 pollution.

Figure 10. The local coefficients of NL.

4. Discussion

4.1. The effects of artificial impervious surface on PM_2.5 at the intra-urban scale

The effects of the spatial patterns of artificial impervious surface on PM_2.5 at the intra-urban scale were explored in this study. Many previous studies have extensively discussed how artificial impervious surface affected PM_2.5 at large scales such as within a country or an urban agglomeration (Gao et al. 2023; Shi et al. 2020; Wang and Gai 2021), while neglecting the influencing mechanism at the intra-urban scale. In this study, we provided a new perspective from DEA and SNCA to explore the effects of spatial patterns of artificial impervious surface. In fact, they represented two types of urban environment. DEA just underwent the transformation from “rural areas” to “urban areas” and consisted of many urban construction activities. In addition, given ample land availability and low rental costs, many industrials tended to gather here (Gao, Liu, and Dunford 2014). As for SNCA, after many years of urban development, it typically had fewer activities of industrial production and urban construction but more population and traffic activities. Therefore, because of the differences in urban environment, distinct effects of spatial patterns of artificial impervious surface on PM_2.5 at the intra-urban scale can be found.

For landscape spatial patterns of artificial impervious surface, the coverage of artificial impervious surface played a more significant role than other landscape attributes in affecting the PM_2.5 in SNCA while the influencing degrees of all landscape spatial patterns were almost at a same level in DEA. In some previous studies, scholars have found that both urban size and compactness were important in affecting PM_2.5 pollution (Fan et al. 2018; Gao et al. 2023) at national and urban-agglomeration scales. In our study, at the intra-urban scale, the important role of the coverage of artificial impervious surface was also confirmed. However, we found that compactness brought significant effects on PM_2.5 only in DEA but had insignificant effects in SNCA. There are many intensive urban construction activities in DEA. Compact artificial impervious surface can not only limit the spread of these construction activities (Jabareen 2006), but also reduce urban traffic commuting intensity, thus reducing PM_2.5 emission (Fan et al. 2018). However, the artificial impervious surface in SNCA contains higher and denser buildings. Although compact artificial impervious surface can mitigate traffic air pollution to some extent, it can also weaken the dispersion of PM_2.5 (Yang et al. 2020), thus rendering the overall effects of artificial impervious surface compactness in SNCA insignificant.

Population and economic factors are also the essential parts of the spatial patterns of artificial impervious surface. Both of them caused more PM_2.5 pollution in DEA than in SNCA, indicating that the PM_2.5 emission from human economic activities in SNCA was lower. On one hand, this is because many factories with high-energy consumption have relocated from urban central areas to the suburbs in recent years. On the other hand, urban residents in SNCA are mainly involved in industries with low energy consumption, such as service and high-tech industry, which generate lower PM_2.5 emissions. Hence, the industries with high-energy consumption in DEA still require more attention in air pollution control.

Temporally, this study analyzed the effects of artificial impervious surface during 2015–2019. It can be seen from Tables 3 and 4 that the aggravating effect of most artificial impervious surface spatial patterns on PM_2.5 pollution had weakened since 2016 in both SNCA and DEA. This may be linked with the intensified measures proposed by Wuhan government in 2016, which enhanced the supervision of highly polluted industries thus contributing to PM_2.5 pollution mitigation. However, the aggravating effects of population and economy slightly increased during 2018–2019. Such a temporal change demonstrated that although Wuhan can reduce the PM_2.5 emissions in a short time period by adopting strong temporary measures, population and economic development may still aggravate PM_2.5 pollution in a long run (Chinese Academy of Science 2021). Hence, the quality of urban development in Wuhan still needs improvement and some fundamental measures regarding energy consumption should be further developed.

4.2. Local effects on PM_2.5

PLAND and NL, representing the coverage of artificial impervious surface and economic development respectively, demonstrated the most significant local effects on PM_2.5 in many areas. In essence, these two factors are directly related with the emission and dispersion of PM_2.5 pollutants, thus significantly impacting PM_2.5 concentrations (Li et al. 2016; D. Zhang, Zhou, and He 2022). In addition, it can be found that there are some similarities between the overall effects and local effects. In these two effects, the coverage of artificial impervious surface significantly affected the PM_2.5 in SNCA while economic development had more significant effects in DEAs. Meanwhile, some specific and significant effects that were neglected by overall investigation can further be discovered by local investigation. For instance, in our overall investigation, economic development resulted in higher PM_2.5 concentrations in DEA than in SNCA. However, in some SNCAs such as the Qingshan industrial park, economic development was also found to cause serious PM_2.5 pollution. Therefore, local effects are not only the projections of overall effects in space but also can disclose some distinct results (Lu et al. 2022). These findings are crucial for the precise mitigation of PM_2.5 pollution, and local investigation using spatial regression approaches such as GWR, are highly recommended to be applied at the intra-urban scale for exploring the effects of artificial impervious surface on PM_2.5.

4.3. Implication for urban PM_2.5 pollution mitigation

By exploring the effects of landscape spatial patterns of artificial impervious surface on PM_2.5 in DEA and SNCA, measures for optimizing artificial impervious surface can be considered. In DEA, high PM_2.5 concentrations tended to occur in areas with high-coverage and uncompact artificial impervious surface. Generally, DEA consists of many new urban areas, usually surrounded by some rural areas with high-quality ecological environment. Therefore, delineating ecological controlling lines and prohibiting drastic urban construction activities occurring outside the controlling lines can effectively hinder the extension of the artificial impervious surface. After controlling the range of urban constructions, planning large green parks is also worth being conducted. It can not only make the configuration of artificial impervious surface in DEA more compact but also decrease PM_2.5 concentrations by the “depositing effect” of vegetation (Abhijith and Kumar 2020). For SNCA, the coverage of artificial impervious surface played the most significant role among all landscape spatial patterns in affecting PM_2.5. However, most SNCAs have been constructed for many years and it is not practical to eliminate existing artificial impervious surface. Hence, some urban renewal measures such as the greening of roofs and roads can be considered for decreasing the PM_2.5 concentrations in SNCA. Also, exploiting urban street canyon as the wind tunnel can enhance the dispersion of PM_2.5, which is also beneficial for the mitigation of PM_2.5 pollution in the areas with high-density buildings (He, Ding, and Prasad 2019; Wu et al. 2021).

From the perspective of population and economic factors, some measures can also be considered. It can be found from Section 3.3.2 that the local effects of both population and economic factors on PM_2.5 in many SNCAs were insignificant. Hence, relocating industries with high emissions from urban central areas can be an effective way to reduce the PM_2.5 pollution in SNCA. For some industrial base that can’t be moved, such as the Qingshan park in our study, it is crucial to emphasize the increased use of clean energy and the phased-out of energy-intensive processes in industrial production. For DEA, increasing population density and economic activities can individually lead to jobs-housing imbalance as well as low-quality industrial development, both of which can exacerbate PM_2.5 pollution (Atash 2007). For satisfying the commuting needs of residents and further reducing traffic emissions caused by congestion, it is necessary to construct convenient public transportation networks in DEA (Lu et al. 2021). Moreover, developing green economy and high-tech industrials can also bring benefits to the air pollution mitigation in DEA.

4.4. Limitations

Some works can further be conducted to investigate the effects of spatial patterns of artificial impervious surface on PM_2.5 at the intra-urban scale. First, we only analyzed the effects during 2015–2019 while the long-term overall and local effects are also worth being explored, which may be helpful for understanding more comprehensive spatiotemporal effects of the spatial patterns of artificial impervious surface on PM_2.5. In the future, with the derivation of PM_2.5 concentrations and the obtainment of artificial impervious surface in a long period, more annual and seasonal findings are expected to be found. Second, different types of cities, such as small and medium-sized cities, have not been discussed in this study. They own different development modes (Cai et al. 2024), which may cause varied effects on PM_2.5. Therefore, in the future, we will select more cities, compare different models for PM_2.5 concentration deriving and investigate how the spatial patterns of artificial impervious surface affect PM_2.5 pollution among them.

5. Conclusion

In this study, by using Wuhan as a case, the effects of the spatial patterns of artificial impervious surface on PM_2.5 at the intra-urban scale were explored. We provided a new perspective from urban development and divided urban areas into areas with drastic expansion of artificial impervious surface and areas with slight or no changes of artificial impervious surface for investigating the effects of artificial impervious surface on PM_2.5. The findings are as follows: (1) By the two-staged XGBoost model, the annual spatial distribution of PM_2.5 concentrations during 2015–2019 were derived with good performance. PM_2.5 pollution was the most serious in 2015 with concentrations ranging from 59.56 μg/m³ to 78.75 μg/m³ and PM_2.5 concentrations gradually decreased after 2015 with declining rate slowing down. (2) Based on the variations of artificial impervious surface during 2015–2019, both DEA and SNCA were delineated and appropriate influencing indicators were selected by collinearity analysis. DEAs were mainly distributed in the new urban areas outside the third ring road, while SNCAs were mainly distributed in the urban central areas within the third ring road and some suburban industrial zones. (3) Overall, population and economic factors caused more serious pollution in DEA than that in SNCA. Among all landscape spatial patterns, artificial impervious surface with high coverage, complex shape and uncompact distribution can bring high PM_2.5 concentrations in DEA. While in SNCA, the coverage of artificial impervious surface (represented by PLAND) dominantly affected the aggravation of PM_2.5 pollution. (4) Spatially, PLAND and economic development affected PM_2.5 more significantly. PLAND had more significant effects on aggravating the PM_2.5 pollution in SNCA, especially in the boundaries between Jianghan district and Qiaokou district. Whereas, economic development played a more important role in aggravating the PM_2.5 pollution in DEA, especially in some suburban districts such as Caidian, Jiangxia and Huangpi, which are suffering from relatively rapid urban expansion. In general, this study developed a new perspective to investigate the effect of artificial impervious surface on PM_2.5 at the intra-urban scale and the findings can provide differentiated guidance for the mitigation of intra-urban PM_2.5 pollution.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/10095020.2024.2350179

Additional information

Funding

This work is supported by the National Natural Science Foundation of China [grant numbers: 51878515 and 52078389] and the Key Laboratory of Monitoring, Evaluation and Early Warning of Territorial Spatial Planning Implementation of Ministry of Natural Resources [grant number: 401147].

Notes on contributors

Zhiyu Fan

Zhiyu Fan received his PhD degree at Wuhan University and is currently a postdoctoral fellow at Wuhan Natural Resources Conservation and Utilization Center. His research interests include urban ecological pattern and effect, urban sustainability and the application of remote sensing and GIS in urban planning.

Qingming Zhan

Qingming Zhan is a professor at Wuhan University. His research interests include digital and smart cities, the application of spatial information in urban planning and planning support systems.

Yun Wang

Yun Wang is a professorate senior engineer at Wuhan Natural Resources Conservation and Utilization Center. Her interest includes ecological planning and practice of urban planning and management.

Zhenfeng Shao

Zhenfeng Shao is a professor at Wuhan University. His research interest includes high-resolution remote sensing image processing and analysis, key technologies and applications of digital and smart cities.

Yuli Fan

Yuli Fan is an assistant professor at China University of Geosciences. His research interest includes urban wind and thermal environment, urban air pollution and regional development research.

Yifeng Zhang

Yifeng Zhang is working at Wuhan Natural Resources Conservation and Utilization Center. Her interest includes urban governance and practice of urban planning and management.