Effects of Landscape Patterns on the Concentration and Recovery Time of PM_2.5 in South Korea

Abstract

Landscape and urban planning efforts aimed at mitigating the risk of PM_2.5 exposure have been hindered by the difficulties in identifying the effects of landscape factors on air pollutants. To identify interactions between PM_2.5 and landscape elements, this study explored the contributions of landscape variables at multiple scales to the mean hourly PM_2.5 concentration and the duration of high PM_2.5 levels in South Korea. We found that the hourly mean PM_2.5 concentration was significantly correlated with landscape variables that explained the spatial processes contributing to fluctuations in air pollutants on a regional level while controlling the spatial autocorrelation of regression residuals. On the other hand, a constant, high PM_2.5 level was related to landscape patterns that explained relatively independent spatial processes on local levels; these processes include vegetation’s ability to reduce PM_2.5 dispersion rates and the influence of transient human activities in local buildings or heavy traffic on roadways on the emission of air pollutants. Our results highlight that urban planners looking to establish design priorities and leverage landscape factors that could reduce the negative impact of PM_2.5 on citizens’ health should consider both the more general PM_2.5 patterns that exist at regional levels as well as local fluctuations in PM_2.5.

1. Introduction

Particulate matter (PM), mainly emitted from urbanized or industrialized areas, degrades human well-being by causing low visibility to drivers and numerous health problems [1,2,3,4,5]. PM with a particle size less than 2.5 μm (PM_2.5) can cause more serious health risks than PM of a particle size less than 10 μm (PM₁₀) because PM_2.5 is more capable of travelling across a longer range, remaining for an extended period, and penetrating the respiratory tract [6,7]. Moreover, high PM_2.5 levels can result in deteriorative effects on urban ecosystems by reducing the foraging performance of honeybees and the fitness of aquatic species [8,9,10]. The environmental problems caused by PM_2.5 draw substantial attention since they can be exacerbated by population growth and changes in landscape patterns in accordance with urbanization [11,12].

Recent studies using land-use regression (LUR) models show that PM_2.5 concentration strongly depends on the surrounding land-use and landscape pattern, e.g., [12,13]. PM_2.5 levels can increase with the length and proximity of adjacent roadways, which serve as indicators of traffic volume and intensity [14,15,16]. Predictors related to residential, commercial, and industrial areas in LUR models can also contribute to a high PM_2.5 concentration; these urban land uses harbor multiple sources of PM_2.5, which is generated by household activities (cooking and incense burning), combustion from factories, and high traffic congestion [14,16]. Forested areas can lower PM_2.5 levels, as vegetation enhances the dry deposition and absorption of PM_2.5 [13,17]. Furthermore, landscape pattern indices highlight the importance of land-use configuration, rather than composition, in explaining variations in PM_2.5 concentrations. For example, a fragmented or heterogenous landscape, indicating a high degree of urbanization, can be positively correlated with high PM_2.5 levels [12,17]. Therefore, identifying relationships between PM_2.5 and landscape factors is becoming important for assessing the exposure risk of PM_2.5 as well as the temporal and spatial variability of PM_2.5.

Previous studies have suggested reasonable spatial scales to be used in determining the effects of spatial variables on PM_2.5 concentrations; however, the complex behavior of air pollutants hinders the development of clear, consistent, and process-centered conclusions [12,18]. For instance, it is difficult to determine whether the low concentration of PM_2.5 in areas with a high ratio of green space reflects that these spaces prevent long-range PM_2.5 transport or that they remove PM_2.5 from local sources. Moreover, the average concentration of PM_2.5, which does not consider temporal scales, is generally used as a dependent variable when identifying “PM_2.5 hotspots” where high exposure and health issues caused by PM_2.5 are expected [16,19,20]. Simplified metrics such as daily mean concentration can be useful for detecting and generalizing PM_2.5’s relationship with spatial factors; however, sampling PM_2.5 concentrations at different time intervals (e.g., from 1 to 3 days) reveals some spatial variability of PM_2.5, which suggests a bias in estimating the short-term impacts of PM_2.5 on human health [21]. Thus, PM_2.5 metrics reflecting fluctuations or above-threshold levels of PM_2.5 should be considered to assess the complicated interactions between PM_2.5 and landscape elements.

In this study, we aimed to examine the effects of landscape patterns at multiple scales on PM_2.5 by adopting a novel metric called “PM_2.5 recovery hour” (PRH). PRH describes the duration of high PM_2.5 levels as well as the hourly mean PM_2.5 concentration during the high pollution season using spatial regression models. Specifically, this study had three objectives: (1) examine the general relationship between PM_2.5 levels, landscape metrics, and other factors at different scales; (2) compare the effects of landscape patterns on two different PM_2.5 metrics; and (3) offer new urban planning perspectives regarding high-PM_2.5 scenarios and PM_2.5’s spatial behavior to facilitate the evaluation of health risks posed by PM_2.5.

2. Materials and Methods

2.1. Study Area and PM_2.5 Data

The study area encompassed much of South Korea, which is a peninsular country located in East Asia (126°–132° E, 33°–38° N) with an area of 100,210 km² (Figure 1) and a population of 51.74 million (as of 2021) [22]. More than 70% of the land is composed of mountains, mainly stretching from north to south on the eastern side of the peninsula. For this geographical reason, the urban infrastructures and agricultural areas are generally developed on the western side of the Korean peninsula. In particular, the northwestern region contains dense metropolitan areas with populations over 20 million, including Seoul, Gyeonggi, and Incheon. PM_2.5 concentrations are generally higher in winter and spring (from December to March), when northwesterly winds carry transboundary fine dust sources, and lower in summer (from June to August) because of the relatively unpolluted air brought by summer monsoons from the northwestern Pacific [23].

PM_2.5 data from February to March in 2021 were retrieved through an Application Programming Interface (API) service provided by the National Institute of Forest Science of Korea (https://www.data.go.kr/data/15078005/openapi.do (accessed on 20 June 2022)). The PM_2.5 concentration (µg/m³) was measured every 10 min by a GRIMM Environmental Dust Monitor (GRIMM EDM-SVC 365) at 40 atmosphere monitoring sites, which cover the major cities (Seoul, Busan, etc.) in different regions of South Korea. The temporal coverage of the collected PM_2.5 data includes the seasons of annual peak PM_2.5 levels. We calculated the hourly mean concentration of PM_2.5 and PRH after removing outliers and missing values using Python v. 3.8 [24]. PRH is the duration, in hours, from when an hourly PM_2.5 concentration rises above a threshold to when it falls below this threshold. Each occurrence of PRH was determined with a minimum interval time of 1 h. Although World Health Organization (WHO) updated its health guideline for 24 h of PM_2.5 exposure from 25 µg/m³ to 15 μg/m³, we set the threshold for calculating PRH at 35 μg/m³. We selected this higher threshold both because there is no consensus on a threshold for fine dust exposure, and because the Environment Ministry of Korea sets the 24 h average concentration at which the ultra-fine dust alert is released at 35 μg/m³. Furthermore, considering WHO’s interim PM_2.5 concentration targets for countries struggling with high levels of air pollution (15 μg/m³ annual mean and 37.5 μg/m³ 24 h mean), 35 μg/m³ is a reasonable threshold for South Korea, for which the annual mean concentration of PM_2.5 is about 20 μg/m³ [25]. Both the mean PRH values and the hourly mean concentration of PM_2.5 at each monitoring site were used as the dependent variables for PM_2.5 regression models.

2.2. Landscape Pattern Indices and Other Variables

Table 1. List of independent variables used for regression analyses. The format of this table is adapted from Wu et al. [12].

Category	Sub-Category	Abbreviation	Buffer Radius (m)	Variable Names
Landscape pattern indices	Urban	PLAND	100; 300; 500; 1000; 2000; 3000; 5000	U_PLAND_radius
		PD		U_PD_radius
		ED		U_ED_radius
	Forest	PLAND		F_PLAND_radius
		PD		F_PD_radius
		ED		F_ED_radius
	Crop	PLAND		C_PLAND_radius
		PD		C_PD_radius
		ED		C_ED_radius
	Bare ground	PLAND		B_PLAND_radius
		PD		B_PD_radius
		ED		B_ED_radius
	Water	PLAND		W_PLAND_radius
		PD		W_PD_radius
		ED		W_ED_radius
Population density	-	Pop	1000; 2000; 5000	Pop_radius
Length of major roadways	-	MR	100; 200; 300; 500; 750; 1000	MR_radius
WGS-84 coordinates	Longitude	Lon	-	Lon
	Latitude	Lat		Lat

shows the predictors representing landscape patterns and expected proxies of PM_2.5 sources, which were selected for spatial regression analyses based on previous studies [12,17,20,26]. We used five land cover categories, specifically urban, forest, cropland, bare ground, and water areas, to calculate landscape pattern indices. The 2020 land cover data with a spatial resolution of 1 m were obtained from the Environmental Geographic Information Service [27]. Percentage of landscape (PLAND), Patch density (PD), and edge density (ED) of each land cover type were used as landscape pattern indices [28] (Supplementary Materials Table S1). The landscape pattern indices were calculated after adopting buffers with radii of 100, 300, 500, 1000, 2000, 3000 and 5000 m at the 40 atmospheric monitoring stations. Human population density data were retrieved as a raster image with a spatial resolution of 100 m from an open API service provided by the National Geographic Information Institute [29]. The sum of the population in each buffer with radii of 1000, 2000, and 5000 m was calculated as a predictor. The total length of major roadways (trunk, motorway, primary, and secondary roads) was calculated in buffers with radii of 100, 200, 300, 500, 750, and 1000 m after retrieving the national scale data from OpenStreetMap (OSM) (http://download.geofabrik.de/asia.html (accessed on 10 July 2022)). Lastly, the WGS-84 geographical coordinates, longitude and latitude in decimal degrees, were used as independent variables.

2.3. Statistical Analysis

To analyze the contribution of landscape pattern indices and other variables to hourly mean PM_2.5 concentrations and PRH, we followed the steps illustrated in Figure 2. Before building regression models, the independent variables with a variance inflation factor (VIF) higher than 3 and an absolute value of Pearson’s correlation coefficient higher than 0.7 were removed using SPSS software v. 25.0 [30], to control multicollinearity. We selected key independent variables based on ordinary least square (OLS) models by using a bidirectional stepwise method, which considers Akaike information criterion (AIC) values at each step. The OLS model for hourly mean PM_2.5 concentrations and PRH is described as follows:

$$y={\beta }_0+{\sum }_{i=1}^k{\beta }_i{x}_i+\epsilon$$

(1)

where $y$ denotes the dependent variables; ${\beta }_0$ is an intercept; and ${\beta }_i$ represents the regression coefficient of an explanatory variable $x_i$ ($i=1, \dots , k)$. Variable selection for OLS models was performed using the stepreg package [31] in R (v. 4.1.3) [32].

The data collected in this study cover distant regions of South Korea; however, the atmosphere monitoring sites sometimes occur in groups of two or three, located within a radius of approximately 200 m to 2 km. In addition, about half of the total sampling sites are concentrated in a northwestern region where Seoul and its neighboring cities are located. We found significant spatial autocorrelation between the dependent variables by using the global Moran’s index (MI) in the R package ape [33,34] (Table S2). Spatial autocorrelation caused by adjacency effects can complicate interpretation of the contribution of spatial variables to the dependent variables in OLS models. For this reason, we conducted a spatial autocorrelation analysis to determine whether regression residuals of OLS models show potential interdependencies regarding geographical position. MI was adopted to identify the remaining influence of locational adjacency on regression residuals from OLS models. The MI of OLS regression residuals was calculated as follows:

$$I_d=\frac{{\sum }_{i=1}^n{\sum }_{i=1}^n{W}_{ij}\left(r_i-\overline{r}\right)\left(r_j-\overline{r}\right) }{({\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{ij})S^2}\hspace{1em}\left(i\ne j\right)$$

(2)

where $I_d$ is an MI calculated with a spatial weight matrix, $W_{ij}$, which provides spatial weights assigned by the group of neighbors identified within a specific Euclidean distance (d) between atmosphere monitoring stations; $n$ is the total number of atmosphere monitoring stations; $r_i$ and $r_j$ are regression residuals of two sampling sites, $i$ and $j$; and $\overline{r}$ and $S^2$ are the mean and variance of the regression residuals, respectively. Because spatial autocorrelation is scale dependent, we assessed how MI changes by adjusting $d$ by 500 m up to 50,000 m when calculating$W_{ij}$. We used the value of $d$ for which |MI| shows the maximum values for making spatial weight matrices for further analysis, including an MI test of regression residuals, Lagrange multiplier (LM) tests, and spatial regressions. The spdep package [35] in R (v. 4.1.3) was utilized to calculate MI values.

Spatial regression models, the spatial lag model [SLM, Equation (3)] and the spatial error model [SEM, Equation (4)], are described as follows:

$$y={\beta }_0+\rho Wy+\sum {\beta }_i{x}_i+\epsilon$$

(3)

$$y={\beta }_0+\sum {\beta }_i{x}_i+u, u={\rho }^{\prime }W\epsilon +\xi$$

(4)

where $\rho$ and ${\rho }^{\prime }$are the spatial lag and spatial error parameters, respectively; $W$ is the spatial weight matrix; $\epsilon$ represents the error term; $u$ is the residual; and $\xi$ is the white noise of the SEM. The SLM hypothesizes that variation among neighbors influences the dependent variable. In contrast, the SEM hypothesizes that spatial dependence in the error term arises, which violates the basic hypothesis of the OLS. An LM test was conducted to diagnose whether the SLM or SEM would be statistically more appropriate than OLS. OLS-based variable selection and construction of spatial regression models were performed using stepreg and spatialreg packages [31,36] in R (v. 4.1.3).

3. Results

3.1. OLS Regression Models

OLS-based regressions were applied to investigate potential spatial factors contributing to hourly mean PM_2.5 concentration and PRH (

Table 2. Results of the OLS models based on AIC stepwise model selection.

Dependent Variable	Independent Variable	B	Standardized B	T	p-Value	Model Performance
Hourly mean PM2.5 concentration (µg/m3)	Constant	251.230	-	2.217	0.035	Adj. R2 = 0.718
	Lon	−2.319	−0.306	−2.941	0.006	F = 25.835
	Lat	2.155	0.306	2.993	0.005	AIC = 226.16
	MR_1000	0.265	0.259	2.700	0.011
	F_PLAND_5000	−0.103	−0.330	−3.459	0.001
PM2.5 recovery hour	Constant	−17.981	-	−1.174	0.248	Adj. R2 = 0.674
	Pop_5000	0.408	0.531	4.381	0.000	F = 17.140
	MR_1000	0.163	0.311	3.079	0.004	AIC = 179.54
	U_PD_300	0.037	0.372	3.032	0.005
	F_ED_100	−0.009	−0.258	2.913	0.032
	Lat	0.749	0.207	2.439	0.090

). Long, Lat, MR_1000, and F_PLAND_5000 were selected as significant spatial factors influencing hourly mean PM_2.5 concentration (p < 0.05). In the OLS model of hourly mean PM_2.5 concentration, MR_1000 and Lat showed positive regression coefficients, while Lon and F_PLAND_5000 had negative coefficient values. In the stepwise-selected model for PRH, four independent variables, including Pop_5000, MR_1000, U_PD_300, and F_ED_100 were significant factors at p < 0.05. The increments of MR_1000 and U_PD_300 as well as Pop_5000 were correlated with higher PRH values. On the other hand, F_ED_100 was the only independent variable with a negative regression coefficient in the stepwise-selected model for PRH.

3.2. Spatial Autocorrelation and LM Tests

The independent variables were screened to control the multicollinearity effect before building OLS models; however, spatial autocorrelation of OLS residuals can hinder interpretation of the effects of predictors on dependent variables. |MI| values of both OLS residuals were mostly below 0.3, regardless of distance and the type of dependent variables (Figure 3). However, MI presented the opposite sign of its value depending on the type of dependent variable, and overall |MI| values computed from hourly mean PM_2.5 concentrations were smaller than those calculated from PRH. The OLS residuals of hourly mean PM_2.5 concentration exhibited insignificant spatial autocorrelation at the maximum value of |MI| (MI = −0.263, distance = 3500 m, p = 0.822). In contrast, the maximum value of |MI| in the PRH OLS model was statistically significant (MI = 0.306, distance = 5000 m, p < 0.05). In line with this result, the LM statistics indicated that neither spatial error nor spatial lag were significant in the OLS model for hourly PM_2.5 concentration (

Table 3. Results of the LM tests.

Dependent Variables	Test Type	LM Statistics	p-Value
Concentration (µg/m3)	Spatial error	2.693	0.101
	Spatial lag	0.032	0.856
PRH (hour)	Spatial error	3.819	0.051
	Spatial lag	6.911	0.009

). Thus, the OLS model was used to interpret the effects of predictors on hourly mean PM_2.5 concentration. In the case of PRH, the effect of the spatial lag process in causing spatial autocorrelation of regression residuals was statistically more certain than that of the spatial error process. However, since the p-value of the LM test for the spatial error process was very close to 0.05, the model performance of the SEM and SLM should be compared thoroughly to confirm the appropriate spatial regression model for PRH.

3.3. Spatial Regression Models for PRH

We built the SLM and SEM to more precisely estimate the effects of spatial factors on PRH while handling spatial autocorrelation in the regression residuals (

Table 4. Results of the spatial regression models for PRH (* p < 0.05, ** p < 0.01).

Spatial Model	Independent Variables	Coefficient	z-Value	p-Value	Model Performance
SEM	Constant	−15.771	−1.060	0.289	Wald statistic: 4.809 *
	Pop_5000	0.442	4.506	0.000	Log likelihood: −80.936
	MR_1000	0.150	3.004	0.003	AIC: 177.87
	U_PD_300	0.039	3.336	0.001
	F_ED_100	−0.008	−2.364	0.018
	Lat	0.679	1.631	0.103
SLM	Constant	−14.207	−1.130	0.259	Wald statistic: 9.120 **
	Pop_5000	0.286	3.277	0.001	Log likelihood: −78.627
	MR_1000	0.112	2.422	0.015	AIC: 173.26
	U_PD_300	0.027	2.487	0.013
	F_ED_100	−0.007	−1.897	0.058
	Lat	0.604	1.706	0.088

). In building the SLM, the AIC value was significantly reduced from 179.54 (the AIC of the OLS model) to 173.26. On the other hand, the AIC value of the SEM was not significantly smaller (a decrease of less than 2) than that of the OLS model. Therefore, the statistical significance and regression coefficients for spatial factors were interpreted based on the SLM. Compared to the OLS model, the p-values of independent variables generally increased in the SLM model while the absolute value of regression coefficients decreased. However, the statistical significance of predictors did not vary noticeably except for in the case of F_ED_100, which negatively affected the duration of high PM_2.5 levels according to the SEM and OLS model.

4. Discussion

4.1. Contribution of Spatial Factors to Hourly Mean PM_2.5 Concentration

The OLS model identified the effects of spatial factors on hourly mean PM_2.5 concentrations. The significant influences of longitude and latitude on the dependent variable may be explained by two factors. First, the contribution of air pollutants emitted from industrial and urban areas on China’s inland and eastern coasts to air quality can rise due to northwesterly winds carrying dust sources [37,38]. Second, our study had many air quality monitoring stations in metropolitan areas located in the northwestern regions of South Korea. Therefore, the significance of these two terms may reflect the importance of anthropogenic activities, such as the operation of thermal power plants and heavy traffic in the metropolitan area, as sources of PM_2.5 [37,39]. The total length of major roadways within a 1000 m radius can serve as a proxy for the traffic density or volume as well as the emission of exhaust fumes. This is consistent with previous studies that found that PM_2.5 is mainly derived from motor vehicles in the major cities of South Korea [40,41].

After removing spatial variables causing multicollinearity, various candidate independent variables representing landscape pattern metrics regarding land cover and spatial scales remained before building OLS models. However, among landscape pattern indices, only F_PLAND_5000 showed a significant effect on hourly mean PM_2.5 concentration and the OLS model performance. PLAND of land cover classes have shown more consistent effects on PM_2.5 compared to other landscape metrics, such as PD and ED [12,17,26,39,42]. Furthermore, our results are consistent with those of Wu et al. [12], who found that percent vegetation cover within a radius of 5000 m was the most consistently significant factor in LUR models, each of which was computed for PM_2.5 concentration in different seasons. Green space is well known as a PM_2.5 sink where plants facilitate the absorption and dispersion processes of air pollutants [43,44]; thus, the inclusion of a variable representing green space in the OLS model is unsurprising. However, since F_PLAND_5000 represents the total area of forested land on a regional level, landscape metrics representing more complex configurations or fragmentation occurring over small spatial scales may have relatively low contributions to average PM_2.5 levels that represent the general seasonal patterns of PM_2.5.

4.2. Spatial Autocorrelation of OLS Residuals

Significant global spatial autocorrelation (MI > 0.4, p < 0.001) of both PM_2.5 metrics as dependent variables emphasizes the need for caution in interpreting the results of OLS-based regressions. High and positive MI statistics (>0.7) are commonly reported when using remote sensing data to present city-level PM_2.5 concentrations [39,45,46]. On the other hand, studies using data from monitoring stations that cover multiple regions report global MI statistics between 0.1 and 0.4 [47,48]. Mirroring the results of these larger-scale studies, this study showed significant but relatively low MI values, likely because of the nationwide extent of data collection combined with the partially biased location of sampling sites. However, the more pressing problem to be addressed is the spatial autocorrelation of residuals and its impact on robust interpretation of regression models. MI and LM tests for hourly mean PM_2.5 concentration demonstrated that its OLS model accepted the null hypothesis of residuals free of spatial autocorrelation. This result indicates that explanatory variables are sufficient to explain spatial autocorrelation of the variance in the OLS model [49]. The OLS model for hourly mean PM_2.5 concentration contained geographic coordinates (latitude and longitude) as significant independent variables, which could reflect the spatial variation of sampling sites. For this reason, the OLS model for hourly mean PM_2.5 concentration may not require additional terms or parameters in order to remove spatial autocorrelation of residuals.

The OLS model for PRH and its regression residuals were somewhat similar to those of hourly mean PM_2.5 concentration. Regarding Pearson’s correlation coefficient (r = 0.841, p < 0.001) of the two dependent variables, it is acceptable that they share the same two landscape factors (MR_1000 and Lat) as independent variables in the stepwise-selected models. However, the OLS model for PRH not only included different independent variables but also violated the assumption of independent residuals. Thus, the selected independent variables in the OLS model for PRH did not adequately explain the underlying spatial processes inducing spatial autocorrelation. We found that spatial lag rather than spatial error process was more appropriate for removing spatial autocorrelation of regression residuals for PRH. The SEM assumes a situation in which incorrect or erroneous observations occur during the PM_2.5 monitoring [50]. Since monitoring sites are run with a standardized method, the SLM more plausibly explains spatial autocorrelation, as the lagged dependent variable and its correlated variables are likely to show the attributional dependence caused by neighboring effects.

4.3. Contribution of Spatial Factors to PRH

Pop_5000, U_PD_300, and MR_1000 can reflect intense human activity, elevated energy consumption, densely built-up areas, and high traffic volumes. Therefore, high PM_2.5 concentrations can be caused by increasing impacts of those landscape factors, which may be associated with higher emissions of air pollutants from cooking or motor vehicles and more densely organized buildings that impede the diffusion of air pollutants [12,39,51]. Contrasting with the results of the PM_2.5 concentration regression, PRH was correlated with the edge density of forested areas within a 100 m buffer. Edge and patch density generally represent the complexity of a landscape and the interaction between vegetated areas and adjacent patches, which might increase PM_2.5 deposition and absorption and thus locally decrease PM_2.5 levels [52]. With respect to U_PD_300, PRH can be highly variable because of local and transient events such as high traffic volume during rush hours or other sources of a surge in PM_2.5 concentrations [53]. Moreover, local air stagnation, which occurs easily in dense clusters of buildings [54], can increase the recovery time of PM_2.5. On the other hand, vegetation can act as a barrier to PM_2.5 emitted near roads while preventing frequent spikes and rapid dispersion of PM_2.5 [43,55,56]. The significance of U_PD_300 and F_ED_100 may point to the variability of PRH, which can be attributed to the contrasting spatial processes in a relatively small area or short term.

Landscape pattern metrics representing relatively fine-scale impacts might help to explain fluctuations in the PM_2.5 metric; however, those landscape metric factors can be more spatially independent because of the lack of overlapping spatial characteristics. Therefore, unlike the relationship between hourly mean PM_2.5 and F_PLAND_5000, F_ED_100 possessed a nearly significant contribution to PRH in the SLM (p = 0.058). Similarly, Lat and Long were not significant independent variables in the SLM for PRH, possibly because they act as indicators of broad-scale impacts caused by transboundary air pollutants and locational effects explaining the variance of spatial autocorrelation.

4.4. Implications for Urban Design

Compared to previous research analyzing the effects of various landscape factors on PM_2.5 with spatial regression, this study is unique in two major regards. First, the actual measurement data from atmosphere monitoring stations, rather than satellite-derived data, were used [39,45,46]. Although remote sensing can be advantageous in allowing for the collection of sufficient samples to cover a large study area, actual measurements can have higher spatiotemporal resolution and accuracy. Relying on actual measurements enables us to utilize landscape pattern metrics representing spatial heterogeneity occurring within a radius of a few hundred meters. Second, we used PRH to determine if variability in PM_2.5 was governed by spatial processes not captured by the hourly mean PM_2.5 concentration, which represents a broad-scale pattern. The PRH metric can help to reveal what spatial variability lies under the short-term impacts of PM_2.5 on health.

Numerous studies have reported that acute exposure or peak concentrations of PM_2.5 can increase the incidence of negative health conditions such as cardiovascular diseases, heart distress, and cerebrovascular diseases as well as mortality [57,58]. However, we lack sufficient recommendations for urban design features that alleviate the adverse health effects of high-intensity PM_2.5 exposure at fine spatiotemporal scales. To address this gap in understanding, we can use the results of this study to infer urban planning priorities for improving public health. Urban areas in which a consistently high PM_2.5 concentration is found would benefit from increasing the total amount of green space on a regional level. Although the average PM_2.5 level in an urban area is below 35 μg/m³ at a municipal level, high PM_2.5 concentrations can persist locally due to heavy traffic or human activities. Our results also highlight that the hourly fluctuation of PM_2.5 tends to be governed by the land cover types and their landscape patterns at a small spatial extent rather than the area of land cover types at a large spatial extent. Thus, the implementation of more sophisticated urban designs that consider the vulnerable populations to PM_2.5 exposure and landscape patterns that may ameliorate these risks is recommended. For example, elderly populations, who are susceptible to short-term exposure to high PM_2.5 concentrations [57], require a safe walking environment to keep their health [59]. Senior citizens will benefit from compact infrastructure that reduces vehicle usage [60] as well as street trees, which facilitate the removal of air pollutants by wind flowing parallel to the roadways [43,61].

5. Conclusions

The contributions of landscape factors at multiple spatial scales to hourly mean PM_2.5 concentrations and the duration of high PM_2.5 concentration levels were identified using OLS and spatial regression models. The hourly mean PM_2.5 concentration was significantly correlated with landscape variables explaining its spatial autocorrelation pattern as well as representing spatial processes related to the input or removal of PM_2.5 on a regional level [12,23]. On the other hand, a constant, high PM_2.5 level was related to landscape patterns that explain relatively independent spatial processes on a local level; these processes include the reduction of PM_2.5 dispersion rates by roadside vegetation and air pollutants emitted by transient human activities in nearby buildings or heavy traffic on roadways [53,55,56,57]. Our results highlight that urban planners who want to reduce the impact of PM_2.5 exposure on citizens’ health should consider both the more general PM_2.5 patterns that exist at regional levels, as well as local fluctuations in PM_2.5 in order to establish urban design priorities regarding vulnerable populations. In future studies, more advanced models that consider climate factors, compare spatial scales, and reflect different levels of PM_2.5 concentration will be required to clarify the effects of landscape patterns on PM_2.5, and thus to provide practical recommendations for PM_2.5 quality management.