How May Building Morphology Influence Pedestrians’ Exposure to PM2.5?

Abstract

Due to their sparse distribution and placement in open areas, fixed air-quality-monitoring stations fail to characterize the effect of contextual factors such as buildings on the dispersion of PM2.5. This study evaluated the effects of building morphology on PM2.5 dispersion in a pedestrian-friendly area on the University of Texas at Dallas campus, spanning approximately 0.5 km². The study collected PM2.5 data along five distinct paths exhibiting varying building morphological characteristics in terms of size, height, density, and spacing at a high spatial resolution. The interquartile range of PM2.5 levels across nine data-collection runs varied from 0.3 µg/m³ to 1.7 µg/m³, indicating relatively uniform PM2.5 levels within the study area. Furthermore, weather-related variables played a dominant role in PM2.5 distribution as temporal variation over-powered spatial variation in the PM2.5 data. The study employed a fixed-effects model to assess the effect of time-invariant morphological characteristics of buildings on PM2.5 and found that the buildings’ morphological characteristics explained 33.22% variation in the fixed effects in the model. Furthermore, openness in the direction of wind elevated the PM2.5 concentration.

1. Introduction

Information about PM2.5 concentration from the nearest air-quality-monitoring station may not represent PM2.5 in a microenvironment. Most monitoring stations are in open areas to avoid the effect of obstructions such as buildings [1]. Also, because of their fixed locations, PM2.5 observations from these stations provide limited opportunities to study spatial variabilities of PM2.5 in places where most people live, and the health effects are most impactful.

In contrast to data from fixed-ground monitoring stations, which are often sparsely located, mobile monitoring platforms provide flexibility to gather data at high spatial resolutions and temporal frequencies. Recently, several studies used mobile monitoring platforms to measure pollutant concentration [2,3,4,5,6]. These studies involved data collection across roads in urban areas using vehicles as a platform for PM2.5 monitoring [2,3,4,5,6]. Some studies used aerial platforms to study the horizontal and vertical profiles of PM2.5 [7,8].

Spatial obstructions like buildings and trees affect the dispersion and concentration of PM2.5 [3,9,10,11]. Land use regression models incorporating the buildings’ morphological characteristics to estimate PM2.5 and PM10 during stable meteorological conditions improved the model’s R² by 10% in the high-density area of downtown Hong Kong [3]. Other studies examined how urban morphology influences the flow of air pollutants in the high-density environment of Hong Kong. Some of the key variables affecting air pollution in these studies were site coverage, average building height, distance between buildings, and degree of enclosure [12,13].

In addition to physical structures or trees that might influence dispersion, PM2.5 distributions have complex relationships with geographic features. PM2.5 refers to all particles with an aerodynamic diameter smaller than 2.5 μm. The chemical components of PM2.5 vary depending on the sources of emissions in a region. Vehicle exhaustion is one of the major sources of black carbon, and black carbon correlates better with traffic patterns than with PM2.5 concentration [14]. PM2.5 observations from near-road sites appeared to be strongly correlated with nearby sites in terms of background pollution [11], but PM2.5 averages at near-road sites appeared higher than those at other nearby sites [10,11]. This indicates that on average, near-road areas are susceptible to relatively greater PM2.5 exposure. However, two studies conducted at near-road sites in the United States showed that traffic characteristics like traffic volume and traffic speed did not correlate well with PM2.5, but meteorological factors and their interaction with site characteristics had profound impacts on PM2.5 [10,11]. In both prior studies, a stable meteorological condition at night showed elevated PM2.5 concentration, whereas higher wind speed reduced the PM2.5 level. A source apportionment analysis at a roadside location in Maryland showed that on-road traffic contributed only 12.5–17% of PM2.5 [10]. High PM2.5 was associated with wind nearly perpendicular to the road, since the wind swept the maximum surface area of the road towards the monitor compared to when it blew precisely perpendicular to the road segment [11]. Moreover, the monitoring station with buildings on the windward side consistently displayed high PM2.5, suggesting that the presence of buildings trapped the pollutants coming from the roads [11].

Thus, besides emission sources, geographic features, spatial forms, and their interactions with wind played a role in determining PM2.5 concentration at locations. Mobile monitoring platforms allow for studying PM2.5 distributions in different urban microenvironments, such as the effect of trees and buildings near emission sources (e.g., roads) on PM2.5 concentration [3,15]. However, PM2.5 research in low-density and non-near-road sites appears to be lacking but can offer insights into the dynamics of PM2.5 concentration in microenvironments typical of pedestrian areas or shopping plazas. This study aimed to investigate PM2.5 variations in such a built environment.

2. Materials and Methods

2.1. Study Area

We chose the University of Texas at Dallas (UTD) campus, located in the Dallas–Fort Worth Metroplex, Texas, USA (Figure 1) for the study. The Campus spanned across Dallas and Collin counties. The figure depicts the metroplex boundary, associated counties, and point source PM2.5 emissions (year 2019) in the area obtained by the Texas Commission on Environmental Quality from State of Texas Air Reporting (STAR) system [16]. Around the campus, several manufacturing facilities contributed as point sources of PM2.5 emissions. However, emissions from most sources were minimal, falling within a range of 0–3 tons per year. Among non-point-emission sources, dust from construction and roads (paved or unpaved), emissions from waste disposal, commercial cooking, and residential fuel combustion (wood) were common sources of PM2.5 in Dallas and Collin counties. In Collin County, dust from agriculture activities related to crops and livestock was also one of the major sources of non-point PM2.5 emissions [17].

A part of the University of Texas at Dallas was chosen as an example of pedestrian-dominant urban settings where city dwellers are more likely to experience direct exposure to air pollution than when indoors or in a vehicle in other urban settings. Four roads surrounded the University: W Campbell Road, Waterview Parkway, N Floyd Road, and Synergy Park Blvd. Figure 2 displays an annual average daily and peak-hour traffic count for these roads. Specifically, the study focused on five paths (A, B, C, D, and E) in the interior of the University, away from the major roads mentioned above (Figure 2). Each path was surrounded by buildings with various characteristics in terms of size, spacing, and spatial configuration. PM2.5 data collection followed these paths when traffic was light.

2.2. Data Sources

2.2.1. PM2.5 Data

The study used DR1000 from Scentroid (Stouffville, ON, USA) [19], a portable PM2.5-monitoring instrument, to collect the PM2.5 data. Before data collection (in December 2019 and February 2020), we calibrated DR1000 using a reference instrument, Fidas^® Frog (Palas GmbH, Karlsruhe, Germany), a fine dust measurement device using optical light scattering of a single particle technology. During its production, Fidas^® Frog was validated for quality by comparing it with a certified Fidas^® 200. Fidas^® Frog measures particle number, size distributions, and mass (for PM1, PM2.5, PM4, PM10, and total suspended particles) [20] and has been used in other research studies for direct measurements or calibration [21,22,23].

During the calibration, two hours of co-located measurements (on 3 September 2019) from DR1000 and Fidas^® Frog were collected. DR1000 recorded PM2.5 measurements at an interval of 3–4 s, whereas Fidas^® Frog every 5 s. After the calibration, both instruments simultaneously collected data for about an hour (on 16 September 2019). For comparison, 1-min averages were calculated for each instrument. Figure 3 displays 1-min average measurements from both instruments before and after calibration. The goodness of fit (R² value) improved from 0.55 to 0.81 after calibration.

Each collection of PM2.5 observations ran along all five paths with DR1000 mounted on a bicycle at a height of about 1.2 m above the ground. A total of nine data collection runs were completed: eight runs in December 2019 and one run in February 2020. Out of these nine runs, three runs collected data in the mornings, three in the afternoons, and three in the evenings. Each data-collection run took about an hour. Each run collected data from the north-to-south direction, was repeated from the south-to-north, and then repeated three times in each direction.

2.2.2. Weather Data

Variables such as wind direction and speed, temperature, and relative humidity could affect the PM2.5 observations. The weather station on the roof of Residential Hall West (Figure 2) on campus supplied weather data at the observation frequency of one minute. The weather station was about 320 m from path A in the northwest. The study retrieved weather data closest to the PM2.5 timestamps.

2.2.3. Building Data

The study used building footprints released by Microsoft (version 2.0) [24] and validated in the study. However, footprints for new construction on campus were missing in the data. Such footprints were digitized manually.

2.2.4. Digital Surface Model (DSM)

A high-resolution LiDAR point cloud for the study area was downloaded from USGS Geological Survey [25]. A DSM with the 50 cm resolution was derived from the LiDAR point cloud using ArcGIS Pro 2.5 (Esri, Redlands, CA, USA)

2.3. Data Preparation

Figure 4 depicts the flowchart for data preparation, while the following subsection describes the steps involved in preparing the final dataset for the analysis.

2.3.1. Segment-Wise PM2.5 Data

The PM2.5 data came from the data collection paths shown in Figure 2. Data processing included steps to divide each path into segments of 50 m and average PM2.5 observations along each 50 m segment as a representative measure. The divide-and-average process smoothened the measurement uncertainty from the instrument and variations in movement speed during the collection, while being sufficient to capture environmental variabilities within and between all paths.

2.3.2. Building Morphology

Building morphological parameters included building coverage ratio (ratio of the area occupied by buildings to total buffer area), mean building area, number of buildings, the average distance between nearest-neighbor buildings, and mean building height. All these parameters could affect pollution dispersion [12,13]. Additionally, Shi et al. [3] considered frontal area index and sky view factor to model PM2.5 and PM10. Both these factors were related to openness. They found that the frontal area index, representing building area in the direction of wind, played an important role in determining PM2.5 concentration. However, since this study already considered openness in the direction of the wind (see directional viewsheds in the next section), the frontal area index was excluded from the analysis.

This study assumed that the 100-m buffer around each 50-m road-segment, covering an area of 41,394 m²(equivalent to approximately eight American football fields) was sufficient to study the effect of buildings’ morphology on PM2.5 in 50-m road-segments. Therefore, buffers of 100 m around each 50-m segment delineated the proximity around roads to extract buildings’ characteristics for calculating morphological parameters.

Depending on the spatial relation of buildings with the 100 m buffer, we categorized buildings into three classes:

(1)

Buildings with their centroid falling within the buffer

(2)

Part of the building/buildings intersecting with the buffer

(3)

Buildings within or touching the buffer

Figure 5 depicts buildings in all three classes highlighted in orange used to compute various morphological measures of the buildings in relation to a 100-m buffer around each 50-m road-segment. Buildings with their centroid falling within the buffer (class 1) were used to compute mean building height, mean building area, and the number of buildings per 1000 m². This was to avoid considering the buildings with only a small portion in the buffer. In order to compute the building coverage ratio, this study considered the portions of the building intersecting with the buffer (class 2). Furthermore, to determine the average nearest-neighbor distance between buildings in the buffer, this study considered all the buildings either within the buffer or touching the buffer (class 3), as a partial account of buildings would not reflect the true distance between them.

Table 1. Building morphology calculation and relevance.

Parameter	Formula	Relevance
Mean Building Area	The sum of areas of all building footprints in a 100-m buffer/Number of buildings in 100-m buffer	Measures the average size of buildings. Larger buildings provide greater horizontal enclosure, hindering PM2.5 dispersion, and vice versa.
Number of Buildings per 1000 Square Meters	Number of buildings in a 100-m buffer × 1000/Area of buffer in square meters	Standardized to number of buildings per 1000 m2 to account for buffer size variations. The number of buildings along with the mean building area indicates the amount of built-up area. A large number of large-sized buildings occupy more space and hence leave less room for dispersion, and vice versa.
Mean Building Height	The sum of heights of all buildings in a 100-m buffer/Number of buildings in a 100-m buffer	Taller buildings provide greater vertical enclosure.
Average Nearest-Neighbor Distance between Buildings	Average nearest-neighbor distance between buildings. Computed using the generate near table tool from ArcGIS Pro 2.5	Reflects building density. Tightly packed buildings hinder dispersion; greater distance between buildings allows more room for dispersion.
Building Coverage Ratio	Area of a 100-m buffer occupied by buildings/Area of a 100-m buffer	Indicates the proportion of buffer area covered by buildings. A greater coverage ratio limits room for dispersion; and vice versa.

provides the relevance and computation of each of these parameters:

Table 2. Summary of building morphological characteristics around data collection paths.

Path	Building Coverage Ratio	Mean Building Area (m2)	Number of Buildings per 1000 m2	Average Distance between Nearest-Neighbor Buildings (m)	Mean Building Height (m)
Path A	0.19	886	0.23	12	6.70
Path B	0.33	2837	0.12	9	9.70
Path C	0.38	4132	0.07	13	14.80
Path D	0.36	5661	0.06	14	12.00
Path E	0.16	2520	0.04	16	8.40

summarizes the building morphological characteristics along the data collection paths. Among all the paths, the building coverage ratios of paths A and E were smaller. Many small buildings surrounded path A. In contrast, a few bigger and taller buildings surrounded path E. Path C had the largest site coverage ratio, followed by path D, and followed by path B. Still, on average, buildings around paths C and D were bigger in size, taller in height, and fewer in number compared to path B. In short, path A constituted smaller, shorter, and denser buildings with intermediate building coverage. Path B consisted of moderately dense, medium-sized buildings with median height and median building coverage. Path C was composed of relatively low-density, large-sized, and taller buildings but had the highest building coverage. Path D contained the largest-sized low-density and distantly placed buildings with medium height and relatively high building coverage. Finally, path E was surrounded by small-sized, sparsely placed and short buildings. Path E also had the lowest building coverage among all paths.

2.3.3. Directional Viewshed

The amount of open area around a location could impact the PM2.5 value observed at a location. Brown et al. [11] provided evidence for the effect of the interaction between wind direction and site characteristics on PM2.5. Thus, this study calculated the open area in the direction of the wind for each segment centroid (Figure 6a) at varying distances of 100 m, 200 m, 400 m, 800 m, and 1500 m by applying viewshed analysis to determine the amount of area visible from any given location in all directions. The viewshed analysis considered 12 different wind directions, starting with 0° in an increment of 30°, resulting in viewsheds at 15°, 45°, 75°, …, 345°. These directional viewsheds served as the basis to analyze the interactions of PM2.5 and openness based on the corresponding wind direction.

2.3.4. Data Integration

In each data collection run, the mean PM2.5 for each 50 m segment was obtained by averaging all the PM2.5 data points associated with the respective segment. Based on the timestamp of the segment-wise PM2.5, the weather data closest in time was combined with the segment-wise PM2.5 data. Further, based on the wind direction, directional viewshed data were integrated with PM2.5 and weather data. In combination, the study datasets consisted of the segment-wise PM2.5, weather-related variables, and directional viewshed. The resulting dataset was further combined with the corresponding building morphological data.

2.4. Modeling

The data collected in the study formed spatial panel data because it consisted of the same set of locations observed multiple times (n = 9). The study included time-varying variables to explain temporal variations in the PM2.5 observations. Location-specific variables, such as building morphological characteristics, would not change over time, and they would have fixed effects on the response variable. Therefore, the study built a fixed-effects panel model to assess the effect of time-varying variables on PM2.5 and extract the individual or location-specific fixed effects of time-invariant variables.

The equation for the fixed-effects panel model can be expressed as follows:

Y_it = α_𝑖 + β × X_it + e_it

(1)

where _i refers to a fixed location; _t refers to the time at which an observation was collected; Y_it is an observed PM2.5 at location _i and time _t; X_it is a k × 1 vector where k is the number of dependent variables specified in Equation (3); β is a 1 × k vector of parameters; $e_{it}$ is the error term. Intercept ${\alpha }_i$ represents a fixed effect that is time-invariant but varied across locations. Following fixed-effects models, the study estimated coefficient β by “de-meaning”, which removed the average over time from each observation:

$$\left(Y_{it}-\overline{Y_i}\right)=\left({\alpha }_i-\overline{{\alpha }_i}\right)+\beta \times \left(X_{it}-\overline{X_i}\right)+\left(e_{it}-\overline{e_i}\right)$$

(2)

where $\overline{Y_i}=\frac{1}{T}{\sum }_{t=1}^T{Y}_{it}$, $\overline{X_i}=\frac{1}{T}{\sum }_{t=1}^T{X}_{it}$, and $\overline{e_i}=\frac{1}{T}{\sum }_{t=1}^T{e}_{it}$.

The variables with bars (i.e., ($\overline{Y_i}, \overline{{\alpha }_i},..$) represent temporal means. De-meaning removes the fixed effect ${\alpha }_i$ since a constant equates its mean over time ($\overline{{\alpha }_i}$). Thus, only time-varying effects remain in the equation. The study was subjected to the drawback of the fixed-effects model that the model could not include individual time-invariant covariates in the model because de-meaning the estimated ${\alpha }_i$ (i.e., the sum of all fixed effects) effectively removed the time-invariant observations.

Specifically, this study used the fixed-effects model to assess the effect of weather-related variables and the openness in the direction of wind on PM2.5. Although openness in the wind direction was not a time-varying variable, this variable might not exhibit a full time-invariant nature since its value changed according to the wind direction. Wind direction was a circular variable; therefore, it was used in the form of sine and cosine components in the model to account for its cyclical nature. Another wind-related variable resulted from the interaction of wind direction and the inlet of the DR1000 instrument. When the wind blew into the instrument’s inlet, DR1000 recorded an elevated PM2.5 compared to the situation when the wind blew away from the inlet. Therefore, in addition to weather-related variables and directional viewshed, this fixed-effects model included the angle between the instrument travel direction and wind direction. When the wind blew into the inlet, this angle was 180°, whereas when it blew away from the inlet, it was 0°. In other cases, the angle varied between 0° to 180°. Consequently, PM2.5 maxima coincided with the angle at 180° and minima at 0°, confirming the cosine of this angle as the appropriate measure to include in the model.

Considering all the variables explained above, the fixed-effects model in Equation (1) was specified as Equation (3) to fit the panel data model using the R package ‘plm’ (version 2.6.4) [26]. Further, this study also tested spatial autocorrelation in the errors for each data collection run. The distance used to determine neighbors was set to 51.5 m, ensuring each data point had at least one neighbor. The weight matrix for computing spatial autocorrelation was derived using the ‘spdep’ package in R (version 1.3.4) [27], and spatial autocorrelation (Moran’s I) using the ‘ape’ package (version 5.8) [28].

$$\begin{array}{cc}\hfill {PM2.5}_{it}={\alpha }_i& +{Wind Speed}_{it}+{Cos Wind Dir}_{it}+{Sin Wind Dir}_{it}+{Temperature}_{it}+{Relative Humidity}_{it}\hfill \\ & +{Cosine of the angle between travel direction and wind direction}_{it}\hfill \\ & {+Directional Viewshed }_{it}\hfill \end{array}$$

(3)

The fixed effects in the model represented location-specific characteristics that were not included in the model. The study related these fixed effects with building morphological characteristics: building coverage ratio, mean building area, number of buildings, average distance between nearest-neighbor buildings, and mean building height. The study assumed that the key time-invariant factor was building morphology and correlated the estimated fixed effects with time-invariant building morphological parameters. In order to assess the contribution of building morphological characteristics, the study developed a regression model relating building morphological characteristics with the fixed effects.

3. Results

3.1. PM2.5 and Weather Data

Figure 7a shows the distribution of PM2.5 across all segments and all data collection runs. The PM2.5 data were positively skewed and adjusted with a Box–Cox transformation (Figure 7b) in order to use them in the proposed fixed-effects panel model. The Box–Cox transformation parameter, lambda, was −0.6057.

Table 3. Details of the date and time of data collection rounds.

Run	Date and Time		PM2.5 (µg/m3)		Wind Speed (mph)	Wind Direction	Temperature (Degrees Fahrenheit)	Relative Humidity (%)
		Mean	Range	Interquartile Range
1	10 December 2019 13:32:16	4.21	4.00	0.60	3.85	90	44.50	39
2	10 December 2019 17:42:52	5.20	2.00	0.90	4.42	58	50.00	42
3	12 December 2019 11:02:42	6.20	2.00	0.60	7.00	174	51.00	45
4	12 December 2019 14:23:48	7.40	2.75	1.10	3.40	187	42.50	53
5	13 December 2019 11:21:26	13.10	3.00	1.00	2.00	296	75.60	51
6	13 December 2019 17:22:52	11.50	4.50	1.70	4.40	270	65.00	58
7	14 December 2019 11:26:38	5.00	1.20	0.40	7.80	87	49.00	54
8	15 December 2019 16:35:22	8.40	3.40	0.80	5.60	297	40.00	70
9	21 February 2020 15:24:03	5.00	2.00	0.30	2.60	250	20.00	47

provides the time, average PM2.5, and weather data for each data collection run. Across the runs, the maximum and minimum observed PM2.5 values were 13.10 μg/m³ and 4.21 μg/m³, respectively. There was a slight variation between the maximum and minimum PM2.5 recorded in each run, ranging from 1.2 μg/m³ to 4.5 μg/m³.

3.2. Variation in Runs across Paths

Figure 8 shows the PM2.5 distribution during each run. While the study focused on five data-collection paths, some runs covered additional nearby segments. All data collection runs encountered low variations in PM2.5 measurements. Considering the interquartile range, runs five and six exhibited comparably higher variations. Figure 9 shows PM2.5 distributions at individual data collection paths during all data collection runs. When the wind came from the south, the PM2.5 values along all paths were highly similar (runs three and four). All the paths oriented in the north–south direction with open passages in the same orientation could facilitate dispersion and explain the similar PM2.5 observations.

While variations in PM2.5 observations across all paths appeared low, there were a few exceptions. During northwest winds (runs five, six, and eight), path E collected higher PM2.5 values compared to the other paths. A possible explanation was that the open area to the north of path E allowed the efficient dispersion of pollutants from surrounding areas, and the pollutants were subsequently trapped by large buildings along this path. In other cases, when the wind came from different directions, these buildings shielded the incoming airflow. In general, path A had low PM2.5 values, except when the wind came from the northeast (run two) across a sizable parking lot. A higher wind speed led to better dispersion. At the wind speed of 7.8 mph (run seven), the highest wind speed observed among all the data collection runs, PM2.5 values were uniform across all paths.

3.3. Results from the Fixed-Effects Model

This study calculated the directional viewshed at multiple distances to account for the unknown effective distance over which openness in a viewshed would influence PM2.5 dispersion using Equation (3). An experiment using five fixed-effects models, each with an effective distance, examined the distance effect on model performance. All other weather-related variables remained the same in all five models. The results showed comparable model performances (R²) at varying viewshed distances (

Table 4. Fixed-effects model results with varying viewshed distance.

Directional Viewshed Distance	R2	Is Directional Viewshed Significant?
100 m	0.82496	No
200 m	0.82516	No
400 m	0.82658	Yes
800 m	0.82812	Yes
1500 m	0.82778	Yes

).

Nevertheless, small viewshed distances up to 200 m were not statistically significant in the model. The model R² value increased up to 800 m, and then it started decreasing. Besides the directional viewshed, all other weather-related variables used in the model appeared to be statistically significant.

Table 5. Fixed-effects model results.

Variable	Coefficient	p-value
Wind speed	−0.0003894	2.358 × 10−6
Cosine of wind direction	−0.014296	0.0001239
Sine of wind direction	−0.062830	<2.2 × 10−16
Temperature	0.006214	<2.2 × 10−16
Relative humidity	0.0040963	<2.2 × 10−16
Cosine of angle between travel direction and wind direction	−0.01895	0.0024263
Directional viewshed (800 m)	3.3831 × 10−7	0.0013893

summarizes the detailed results of the model (Equation (3)) built with a viewshed of 800 m.

The small transformed PM2.5 values (0.84–1.32) in the dependent variable contributed to the small coefficients for all explanatory variables. Wind speed had a negative impact on PM2.5 as the model asserted a negative coefficient. Both the components of the wind direction, sine and cosine, had negative coefficients in the model, whereas both temperature and humidity had positive coefficients. This study used a digital surface model with a 50-cm resolution to determine the directional viewshed of a location, and 50 × 50 cm² cells in a viewshed up to 800 m were used in the regression model. Depending on the location’s visibility, the number of cells in the viewshed varied from hundreds to thousands, which explained the small coefficient for directional viewsheds. The amount of open area in the wind direction had a small but positive impact on PM2.5.

3.3.1. Error Analysis

Moran’s I was calculated to evaluate the spatial autocorrelation in the errors for each data collection run. All but run three had a significant moderate to strong spatial autocorrelation in the errors (

Table 6. Spatial autocorrelation in errors of the fixed-effects models.

Data Collection Run	Number of Observations	Moran’s I (Panel Data)	Significance	Moran’s I (Spatial Panel Data Model)	Significance
1	57	0.47	Significant	0.68	Significant
2	57	0.87	Significant	0.90	Significant
3	57	0.14	Insignificant	0.30	Significant
4	57	0.31	Significant	0.26	Significant
5	57	0.31	Significant	0.55	Significant
6	57	0.48	Significant	0.80	Significant
7	57	0.43	Significant	0.75	Significant
8	57	0.40	Significant	0.68	Significant
9	57	0.48	Significant	0.59	Significant

), suggesting the possible omission of variables responsible for the PM2.5 variations across the study area. To account for this spatial dependence, this study considered both spatial lag models and spatial error models. Spatial lag models examine the existing autocorrelation in the response variable [29]. As the study intended to mediate the model bias due to spatially autocorrelated errors and investigate the effect of predictors on PM2.5 estimates, the study opted for the spatial error modeling approach. Spatial autocorrelation in the errors increased with spatial specification in each panel. Moran’s I values and their significance to the errors observed in each data collection round are outlined in Table 6, evaluated within both the panel model (aspatial specification) and the spatial panel model. The results suggested that aspatial models suffered less than spatial models from spatial dependence in the errors, as indicated by Moran’s I. Therefore, the study selected the fixed-effects panel model over the fixed-effects spatial panel data model for further analysis.

3.3.2. Relation between Fixed Effects and Building Morphology

The correlation between the fixed effects and building morphological characteristics varied from 0.79 to −0.51 (Figure 10). The fixed effects were weak to moderately correlated with all building morphological characteristics, except for the average nearest-neighbor distance between buildings. Mean building area, mean building height, and building coverage ratio were strongly positively correlated with each other. On the other hand, while the average nearest-neighbor distance between buildings was only weakly correlated with all the other building morphological characteristics, the number of buildings was moderately negatively correlated with mean building height.

As such, the study excluded the average nearest-neighbor distance between buildings in regression modeling, as it was almost uncorrelated with the fixed effects.

Table 7. Results of regression between fixed effects and building morphological characteristics.

Variable	Coefficient	p-Value
Mean building area	−2.33 × 10−6	0.1939
Building coverage ratio	0.0731	0.0046 **
Mean building height	−0.00162	0.1329
Number of buildings per 1000 m2	−0.1099	0.0001 ***
Model R2: 0.3322

** and *** denote significance at the level of 0.01 and 0.0001, respectively.

presents the regression results. The study found only the building coverage ratio and the number of buildings to be significant variables that explained variations in the fixed effects. Overall, building morphological characteristics explained 33.22% variation in the fixed effects.

4. Discussion

Despite the moderate variation in building morphology, the study observed relatively uniform PM2.5 levels within an area of about 0.5 km² across all data collection runs. The consistent PM2.5 levels observed in this study agree with the findings of a study by Harrison et al. [2] aimed at understanding spatial variability at a neighborhood scale. Harrison et al. [2] collected daily on-road PM2.5 measurements over a month within an area of 100 km² (equivalent to MODIS aerosol product’s single pixel), encompassing the University of Texas at Dallas. Their study found that depending on the weather condition, the spatial scale of PM2.5 variation in the area varied from 0.8 km to 5.2 km, suggesting uniform PM2.5 levels over a distance of 0.8 km or smaller.

The building morphology accounted for 33.22% of the variations in the fixed effects, suggesting that the built environment could affect PM2.5. Regression results from the fixed effects showed that the building coverage ratio had a positive impact on PM2.5. Thus, the greater built-up area likely hindered the dispersion of the pollutants. Despite the negative impact of the number of buildings per 1000 m² on PM2.5, a similar inference could be drawn. A negative correlation between the number of buildings per 1000 m² and mean building area, mean building height, and building coverage ratio in the study area denoted the contrast between areas with many smaller buildings and areas with fewer larger buildings, suggesting areas with smaller and shorter buildings experienced better dispersion than other areas. Nevertheless, the study could not assess each variable in isolation, as an inference about it could change based on its relationships with other variables in the model. The insignificant impact of mean building height and mean building area could be attributed to limited variability among building morphological characteristics in the study area.

The negative effect of wind speed on PM2.5 in the model reaffirmed findings from the previous studies that increasing wind speed improved dispersion and hence reduced the PM2.5 concentration [10,11]. Increased temperature promoted air circulation and was negatively related to PM2.5 measures [10]. Therefore, the temperature was expected to have a negative effect on PM2.5 contrary to the positive effect found in this study. Several factors could contribute to this discrepancy. One possible explanation is that elevated temperature contributed to chemical reactions that might have formed new particles including PM2.5 [30]. Other reasons could be the blocking of the airflow by surrounding buildings or the desorption of particles from different surfaces due to increased temperature. As for the relative humidity, depending on its value, it could increase or decrease in PM2.5 mass. Increased humidity promoted particle size growth, but after a certain threshold, very high humidity could lead to particle deposition due to heavy particle growth [30]. Moreover, Lou et al. [31], in their 3-year study in the Yangtze River delta, described the relationship between relative humidity and PM2.5 as an inverted U-shape, where the hygroscopic growth of particles continued with an increase in relative humidity until it reached 70%, and then it started decreasing. The relative humidity observed in the study ranged from 39 to 70%—not high enough to cause particle deposition—and resulted in a positive correlation with PM2.5. The positive coefficient associated with directional viewshed suggested that a more open area in the direction of the wind was consistent with additional incoming pollution from the other areas, as confirmed by Brown et al. [11]. Of the different viewshed distances used in the model, model performance consistently increased up to 800 m distance, and then it decreased, suggesting openness up to 800 m distance affected PM2.5. However, increasing distance led to only marginal improvements in the model, and openness within 400 m also yielded satisfactory results.

5. Conclusions

The study collected the PM2.5 data using a mobile monitoring platform at a non-near-road site with different building morphological characteristics. Building morphological characteristics varied from high-density, small-sized, shorter buildings to low-density, medium to large-sized, taller buildings. The data collection strategy included multiple runs across paths representing varied building morphology. The fixed-effects panel model was used to investigate the effect of weather-related variables and building morphological characteristics on PM2.5. Unlike meteorological variables, building morphological characteristics did not change with time. A regression model was developed to find the contribution of building morphological characteristics to fixed effects extracted from the panel data model. While the weather-related variables explained variations in PM2.5, building morphological characteristics also showed positive effects on PM2.5. Furthermore, openness in the direction of the wind allowed pollutants from other areas and raised PM2.5 concentration in the area.

The model presented in the study carried spatially autocorrelated errors even with spatial specifications, possibly due to the small study area with low PM 2.5 variability. A larger study extent and coarser unit of spatial analysis might mediate the issues of spatially autocorrelated errors. Moreover, the fixed-effects model applied the same spatial weight matrix across all runs. Depending on the weather conditions and its interactions with the surrounding built environment, the nature of spatial dependence could change. Therefore, a dynamic spatial weight matrix would be more appropriate to account for this spatial dependence. Alternatively, space–time convolution could also address the complex space–time dependence in the data overlooked in the current model specification. Nevertheless, this study showed small spatial variation in PM2.5 over a small area (<1 km²), typical of non-near-road sites with moderate variation in building morphology. The study considered interactions between wind and openness in the wind direction and overall building morphological characteristics within a 100 m buffer. Future studies can incorporate interactions between wind and building heights and measures related to the spatial arrangement of buildings to further understand the effects of buildings on PM2.5. Furthermore, variation in PM2.5 depends on many factors like weather parameters and emission sources that vary across different parts of the year. All the scenarios cannot be explored at a single site with a limited number of runs. More sites with different urban settings need to be studied over an extended period to improve our understanding of the microscale dynamics of PM2.5.