Estimation of On-Road PM2.5 Distributions by Combining Satellite Top-of-Atmosphere With Microscale Geographic Predictors for Healthy Route Planning.

How to reduce the health risks for commuters, caused by air pollution such as PM2.5 has always been an urgent issue needing to be solved. Proposed in this study, is a novel framework which enables greater avoidance of pollution and hence assists the provision of healthy travel. This framework is based on the estimation of on-road PM2.5 throughout the whole city. First, the micro-scale PM2.5 is predicted by land use regression (LUR) modeling enhanced by the use of the Landsat-8 top-of-atmosphere (TOA) data and microscale geographic predictors. In particular, the green view index (GVI) factor derived, the sky view factor, and the index-based built-up index, are incorporated within the TOA-LUR modeling. On-road PM2.5 distributions are then mapped in high-spatial-resolution. The maps obtained can be used to find healthy travel routes with less PM2.5. The proposed framework was applied in high-density Hong Kong by Landsat 8 images. External testing was based on mobile measurements. The results showed that the estimation performance of the proposed seasonal TOA-LUR Geographical and Temporal Weighted Regression models is at a high-level with an R 2 of 0.70-0.90. The newly introduced GVI index played an important role in these estimations. The PM2.5 distribution maps at high-spatial-resolution were then used to develop an application providing Hong Kong residents with healthy route planning services. The proposed framework can, likewise, be applied in other cities to better ensure people's health when traveling, especially those in high-density cities.

Introduction

In the latest handbook on sustainable urban mobility and spatial planning by United Nations (United Nations, 2020), a series of practical recommendations is outlined: (a) Active mobility is a core element of healthy cities in Mobility and Spatial Planning; (b) Both walking and cycling need to be supported not just in urban cores, but on a much larger scale; (c) Cities need to develop user‐friendly applications that support healthy cycling and walking. In the current daily health travel, outdoor air pollution related to PM2.5 has become one of the main sources of harm to people's health (Edwards et al., 2021). A recent global study published in the Bulletin of the World Health Organization estimates that 4.2 million premature deaths worldwide were caused by PM2.5 each year (Nazarenko et al., 2021). This included more than 1 million deaths in China, more than 500,000 deaths in India, nearly 200,000 in Europe, and more than 50,000 in the United States (Nazarenko et al., 2021). Most of these deaths occurred during outdoor activities. Thus, given the current “new‐found” health awareness, on how to estimate on‐road PM2.5 at a high‐spatial‐resolution and consequently heighten the ability to enable healthy travel, route planning has become one of the urgent issues relating to urban smart active mobility, especially for high‐density cities with serious air pollution.

However, due to the high maintenance costs of traditional ground monitoring systems (J. Li et al., 2020), spatial coverage to reflect microscale detailed on‐road PM2.5 variations is lacking (Lu et al., 2021). Therefore, it is crucial to accurately estimate micro‐scale PM2.5 distributions at high spatial resolution through multi‐source data imagery and analysis.

Estimating PM2.5 based on satellite remote sensing can make up for the lack of ground observation sites (T. Li et al., 2020). Remote sensing observations can provide longer‐term observational data, with global coverage, higher spatial resolution, and low cost, a matter more conducive to the study of the spread of PM2.5 (Li et al., 2021; Shogrkhodaei et al., 2021; D. Zhang et al., 2021). In the early stage, based on the aerosol optical depth (AOD) data retrieved from remote sensing data, and PM2.5 measurements at ground stations, the PM2.5 concentration value within or between regions can be effectively estimated by establishing a linear regression model between AOD and PM2.5 (Braggio et al., 2021; Meng et al., 2021). By incorporating the meteorological factors, the multiple linear regression (MLR) model was then further used to more accurately represent the relationship between AOD and PM2.5 (Ghahremanloo et al., 2021; X.Xu et al., 2021). Furthermore, the spatiotemporal geographic weighting models could estimate the PM2.5 concentration (W. Chen et al., 2020; Xue et al., 2020; Y. Zhang et al., 2020). However, because the spatial scale is limited by the resolution of remote sensing images, for example, the resolution of the most commonly used MODIS AOD products is greater than or equal to 1 km (Wang et al., 2021; Wei et al., 2021), hence it cannot effectively identify pollution hotspots and accurately reflect spatial variation in urban areas.

Some studies recently, explored, has been the retrieval of ground PM2.5 directly from the top‐of‐atmosphere (TOA) reflectance data of MODIS, while omitting the intermediate process of avoiding AOD retrieval (J. Liu et al., 2019; Yan et al., 2020; Yang et al., 2020). A multispectral empirical model has also, been proposed to estimate the concentration of PM2.5 using Landsat 8 OLI and by taking into account the corresponding meteorological parameters (Mishra et al., 2021). The results have shown a high degree of accuracy. However, the microscale environmental factors such as urban spatial morphology, and vegetation distribution in the urban canopy, both of which are more essential in impacting PM2.5 distribution in a high‐density urban environment, have not been considered in the existing TOA‐PM2.5 studies. Thus currently, most existing TOA‐PM2.5 studies cannot reflect the spatial variability of microscale PM2.5.

Previously, a framework using mobile measurement, Google Street View (GSV) data, Landsat data, and other data, has been proposed to estimate PM2.5 concentration by land‐use regression (LUR) models in the microscale urban environment (Tong et al., 2021). It has greatly improved the modeling and mapping of PM2.5 at the microscale. However, due to the limitation of the application range of mobile measurement instruments, this framework cannot be applied to the entirety of an urban area.

To overcome the above limitations, the aim of this study is to provide a fine‐scale estimation of the spatiotemporal variation of PM2.5 in the high‐density city environment for assisting healthy route planning with less PM2.5. An improved framework has been proposed based on the following steps: (a) using the Landsat‐8 TOA data to enhance LUR modeling to retrieve micro‐scale PM2.5 on roads. In this process, observation angle data of Landsat‐8 satellite, PM2.5 data of ground stations, urban environmental data (e.g., GSV images, OpenStreet‐Map (OSM) data, and land use data), and meteorological data will be integrated into the LUR modeling (Gong et al., 2018; Kang et al., 2020; Y. Liu et al., 2021). In particular, the green view index (GVI) factor of green vegetation coverage of the urban canopy, the sky view factor (SVF) related to the three‐dimensional environment of urban buildings, and the index‐based built‐up index (IBI), related to urban impervious surfaces, are considered in the modeling. (b) On‐road PM2.5 distributions of the whole city are further mapped at high‐spatial‐resolution. (c) Result maps of on‐road PM2.5 estimation are able to be used in the digital application for calculating the healthy travel route with less PM2.5, to support the planning result.

Hong Kong, a representative of high‐density cities, was selected as the study area, in which to apply the improved framework. From 2019 to 2021, more than 1,200 people died each year in Hong Kong due to outdoor air pollution such as PM2.5 (The University of Hong Kong, 2022). Therefore, how to reduce people's PM2.5 exposure and consequently better ensure their health during mobility is an urgent need in high‐density cities such as Hong Kong. In this study, the fine‐scale spatiotemporal distribution of ground‐level PM2.5 over Hong Kong was estimated by the TOA‐LUR modeling with the Landsat‐8 TOA data for the period of 2014–2021, GSV, OSM data, and other multi‐source geographic predictor data. Handheld PM2.5 measuring instruments—Airbeam2 were further used to conduct the external test with the estimation results on the selected dates of the satellite transit during the whole of 2019 and 2021. Aided by the obtained on‐road PM2.5 distribution maps at high‐spatial‐resolution, a related digital application was further developed to provide health route planning for more than 7.5 million Hong Kong residents during their active mobility.

Materials and Methods

Data Sources

Ground‐Level PM2.5 Measurements

In the headed study, hourly PM2.5 concentration data, from 1 January 2014 to 31 December 2021, were obtained from all 16 air quality monitoring stations (AQMS) in Hong Kong (Figure 1). These hourly PM2.5 concentrations are based on high‐accuracy gravimetric sampling, using the Thermal Scientific Model 1405‐DF Filter Dynamics Measurement System (FDMS) from the Hong Kong Environment Protection Department (HKEPD) (Hohenberger et al., 2021; G. Shi et al., 2021).

Figure 1

The locations of 16 air quality monitoring stations in Hong Kong.

Satellite Observations

A total of 56 Landsat 8 OLI/TIRS images during the Landsat 8 overpass time (10–11 a.m. in Hong Kong) from 1 January 2014 to 31 December 2021, was used in this above study. Landsat 8 has a spatial resolution of 30 m and a temporal resolution of 16‐day. These 56 Landsat 8 images had almost no cloud cover and could be considered “fully clear images”. In the traditional estimation of PM2.5 by AOD, the AOD products are often retrieved from TOA reflectance of the blue band, the red band, and that at 2.13 μm via a dark‐target‐based algorithm (Shen et al., 2018). Furthermore, the observation angles have been found to be important parameters regarding the physical relationship between TOA reflectance and surface PM2.5 (Shen et al., 2018). Hence, the TOA on bands 2 (Blue: 0.45–0.51 μm), 4 (Red: 0.64–0.67 μm), and 7 (SWIR 2: 2.11–2.29 μm) and observation angles (i.e., sensor azimuth, sensor zenith, solar azimuth, and solar zenith) are obtained from the Landsat 8 data set to estimate PM2.5. The technical details for converting DN values of Landsat 8 to TOA reflectance have been considered in the study of Mishra et al. (2021).

In addition, the normalized difference vegetation index (NDVI) during the same period was also estimated by band 4 and band 5 of Landsat 8 images (Ouzemou et al., 2018). The IBI during the same period was also estimated by band 3, band 4, band 5, and band 6 of Landsat 8 images (XI et al., 2019).

Meteorological Data

The PM2.5 pollution in the urban environment is affected by meteorological factors. Thus, the Goddard Earth Observing System Forward Processing (GEOS‐FP) meteorological data from 1 January 2014 to 31 December 2021, are incorporated in this study (Jiang et al., 2020; Yang et al., 2019). GEOS‐FP has been successfully used in a previous study to estimate PM2.5 from the satellite TOA reflectance (Shen et al., 2018). The meteorological data of GEOS‐FP have a spatial resolution of 0.25° latitude × 0.3125° longitude. Local time (the Landsat 8 overpass time corresponds to 10:00–11:00 a.m.) was used. (a) Wind speed at 2 m above ground (WS, m/s), (b) air temperature at a 2‐m height (TEMP, K), (c) surface pressure (PS, kPa), and (d) planetary boundary layer height (PBLH, m), were extracted from the GEOS‐FP meteorological data between 10 and 11 a.m. local time.

Google Street View Images

The contribution of urban street‐level greening in reducing PM2.5 concentrations has been proven by many studies (Cai et al., 2020; Wu et al., 2021; R. Zhang et al., 2021; Zhao et al., 2021). Neighborhoods that had higher green space coverage displayed lower PM2.5 concentrations. Hence, street‐level greenness in Hong Kong was estimated through a novel GVI using GSV images. The methodology proposed by Li et al. was employed to calculate the GVI (Li et al., 2015). To calculate the GVI of each GSV point location, the following equation is used to calculate the proportion of green pixels in four images collected (Li et al., 2015): where N indicates the number of green pixels at the ith direction, and N indicates the total number of pixels at the ith direction. The GVI indicates the ratio of total green pixels from four pictures to the total pixels of the panoramic pictures. The GVI index is then used in the assessment of street‐level greenery. To identify green pixels, the following equation is employed (Li et al., 2015): Where P indicates whether the ith pixel is a green pixel; Red , Green , and Blue denote the red, green, blue channel values of the ith pixel; t is a threshold that determines whether the pixel is a green pixel. The threshold t is chosen by the OTSU algorithm automatically. The average value of the green pixel ratios in four street view images at a specific location is used regarding the measurement of the urban‐street‐level greenery.

Land Use Data

Hong Kong land use data has been used in this study. The land‐use data includes 4 types of land use (Y. Shi et al., 2018): (a) Residential use (lu_res); (b) Commercial use (lu_com); (c) Government use (lu_gov); (d) Open space (lu_ops). In term of the total area (unit: m2) regarding the land‐use type of each of the 16 monitoring stations was calculated using the buffering analysis.

OpenStreetMap (OSM) and Bus Stop Points Data

To measure the impact of local road traffic density in Hong Kong in term of the local PM2.5 pollution, five types of road lines density have been used in this study: (a) Primary, secondary, trunk road line density (rd‐main); (b) Tertiary road line density (rd‐ter); (c) Service road line density (rd‐ser); (d) All motor vehicle roads line density (rd‐all); (e) All railway road line density (rd‐rai). Based on the related OSM data (Z. Zhang et al., 2018), each of the five types of road line density was calculated separately. Additionally, and based on the official data of bus stop points, the closest distance to bus stops (CDS) was also calculated.

Urban Surface Factors Data

The outdoor PM2.5 pollution is significantly affected by the urban form factors within the urban canopy layer, such as the mean height, the variation in building height, and the building density. Thus, to take into account the microscale urban form factor variations that affect the PM2.5 estimation in a high‐density city, a set of urban surface factors including SVF (G. Li et al., 2020), elevation, terrain slope, and terrain aspect were calculated using the buffer analysis. The buffer analysis was based on the Digital Terrain Models (DTM) data, building height data, and Google elevation data.

Commercial Restaurant and Temple Points Data

Cooking emissions from commercial restaurants are also outdoor sources for urban‐community‐level PM2.5 (Lung et al., 2020; H. Xu et al., 2020). Of these emissions, barbecues and Chinese restaurants are the main sources of PM2.5 pollution (Hsu et al., 2020). Thus, based on the points data of barbecues and Chinese restaurants, the impact of cooking emissions was calculated by using the buffering analysis, separately. In addition, more than 2 million people in Hong Kong, follow the Buddhist/Taoist religion, and during the periods of prayer incense and joss paper burnt in temples both during festival periods and as a common part of daily life (C. Chen et al., 2021; Hsu et al., 2020). Based on the points data of temples, the impact of incense/joss paper burning was calculated using buffering analysis.

TOA‐LUR Modeling and Validation Methods

The TOA derived from the Landsat 8 images was used in combination with the traditional LUR modeling to estimate the microscale PM2.5. The microscale meteorological factors and geographic predictors were incorporated in the TOA‐LUR modeling. We resampled the meteorological data to match the satellite observations. All these data were reprojected to the same coordinate system. Finally, we extracted the satellite observations and related factors for the locations where PM2.5 measurements were available.

First, the MLR method was adopted to identify the important predictor variables followed by the estimation of the microscale PM2.5 (Tong et al., 2021). The geographically and temporally weighted regression (GTWR) model was also used. The TOA‐LUR modeling process was conducted in the steps given below (Shi et al., 2018).

The Development of a Basic TOA‐PM2.5 Model

First, a basic model was developed to build the TOA‐PM2.5 correlation, with the latter mainly considering meteorological factors—Wind speed at 2 m above ground (WS), air temperature at a 2‐m height (TEMP), surface pressure (PS), and planetary boundary layer height (PBLH)： where PM 2.5 is the estimated PM2.5 concentration at the PM2.5 monitoring station i on day j. TOA , TOA , and TOA are related TOA reflectances on bands 2, 4, and 7 (Blue, Red, and SWIR 2) of Landsat 8 at the location i on day j. sat_az , sat_zn , sun_az , and sun_zn are observation angles (i.e., sensor azimuth, sensor zenith, solar azimuth, and solar zenith) of Landsat 8 at the location i on day j. b , b , b , b , b , b , and b are the slopes of TOA , TOA , TOA , sat_az , sat_zn , sun_az , and sun_zn . b , b , b , and b are the slopes for the four daily basic weather variables. β is the interception of the model. ε is the residuals that presumably vary by day.

Unlike the behavior in the summer season, air pollution in Hong Kong during the winter season is affected not only by local emissions but also by the regional impact from the Guangdong‐Hong Kong‐Macao Greater Bay Area (Y. Shi et al., 2016). Thus, to distinguish seasonal changes, seasonal MLR correlation models are proposed to predict microscale PM2.5 in each of the following periods: spring, summer, autumn, and winter.

Developing TOA‐LUR MLR Models

In this study, a total of 210 potential explanatory variables have been used for buffer analysis. Included were 14 variables of 15 different buffer sizes. In the study, a series of regression analyses were conducted on each buffer‐based variable, to further enable sensitivity testing of the different buffer sized variables in the model. Some key variables for the TOA‐LUR regression modeling were chosen to enable the prevention of over‐fitting problems in the subsequent stepwise regression modeling process. The following regression analyses were conducted using the following model structure:

Where b , b , b , b , b , b , and b are the slopes of TOA , TOA , TOA , sat_az , sat_zn , sun_az , and sun_zn . b8, b9, b10, and b11 are the slopes for the four daily basic weather variables. b is the slope for the buffer‐based variable (VAR ) using the buffer size of d. For example, the tertiary road line density calculated in 750 m buffer and 1,500 m buffer are used as two separate variables. Specifically, d is set as 10, 20, 50, 100, 200, 300, 400, 500, 750, 1,000, 1,500, 2,000, 3,000, 4,000, and 5,000 m. The model performance was indicated by using Adjusted R . Furthermore, by using the “A Distance Decay Regression Selection Strategy (ADDRESS)” (Su et al., 2009), variables at the critical buffers were chosen as candidate explanatory variables. The identification details of critical buffers have been given in a previous study (Su et al., 2009). It has been also applied in two previous studies of Hong Kong (Shi et al., 2018; Tong et al., 2021).

The forward stepwise regression modeling was used to establish the TOA‐LUR models during different seasons (Beelen et al., 2013). According to the minimum Bayesian Information Criterion (BIC), the model with the highest R 2 is selected. The variance inflation factor (VIF) regarding the total models is checked to eliminate collinearity problems in the resulting model. Variables with their p‐value greater than 0.10 and VIF greater than 5 are excluded from the model (Vienneau et al., 2013). The final model can be expressed as where PM 2.5 is the predicted PM concentration at the location of the air quality monitoring station i on day j. TOA , TOA , and TOA are related TOA reflectances on bands 2, 4, and 7 of Landsat 8 at the location i on day j. sat_az , sat_zn , sun_az , and sun_zn are observation angles (i.e., sensor azimuth, sensor zenith, solar azimuth, and solar zenith) of Landsat 8 at the location i on day j. b , b , b , b , b , b , and b are the slopes of TOA , TOA , TOA , sat_az , sat_zn , sun_az , and sun_zn . b … b are the slopes for the selected meteorological variables and the other candidate geographic variables. β is the intercept of the TOA‐LUR model. ε is the residuals.

Developing TOA‐LUR GTWR Models

The above TOA‐LUR MLR model follows the structure of the fixed effects model. It is the influence of the predictor variables in these models which is temporally fixed. However, the influence of some predictor variables changes over time. Therefore, to further improve the TOA‐LUR modeling, taking time‐dependent random effects and geographic non‐stationarity as important considerations, the TOA‐LUR GTWR is applied to estimate microscale PM2.5 (Bai et al., 2016; Huang et al., 2010): where PM 2.5 is the estimated PM2.5 concentration at the PM2.5 monitoring station i on day j. TOA , TOA , and TOA are the related TOA reflectance on bands 2, 4, and 7 of Landsat 8 at location i on day j. sat_az , sat_zn , sun_az , and sun_zn are observation angles (i.e., sensor azimuth, sensor zenith, solar azimuth, and solar zenith) of Landsat 8 at the location i on day j. b , b , b , b , b , b , and b are the slopes of TOA , TOA , TOA , sat_az , sat_zn , sun_az , and sun_zn . x , y are the geographical coordinates of the PM2.5 monitoring station i. bk are the slopes for the k meteorological variables VAR kij. b are the slopes for the m candidate geographic variables (VAR ). β and ε are the intercept and residuals.

Model Validation

and Root‐mean‐square error (RMSE) were chosen for testing both the model fit and the cross‐validation. Among them, leave‐one‐out cross‐validation (LOOCV) was used to validate these models. This validation process is repeated n times. The calculation method of and RMSE is as follows: Where R is the coefficient of determination of the models. n is the total number of measurement points. In the annual model, n is set as 699 in 56 days. In the seasonal models, n is set as (a) 164 during 14 days in spring, (b) 194 during 14 days in Summer, (c) 181 during 14 days in Autumn, and (d) 160 during 14 days in Winter. p is the total number of the selected prediction variables. is the estimated PM2.5 concentration. is the measured PM2.5 concentration.

External Model Validation

To better validate the models, the mobile measurement instruments ‐ Airbeam2 were used to test the estimated results of the above TOA‐LUR GTWR models on typical routes (a) near two roadside PM2.5 measurement stations: one in Mong Kok and the other in Causeway Bay (Figures 2a and 2b), (b) at Tsim Sha Tsui, Tuen Mun, and Tseung Kwan O (Figures 2c–2e) far from the monitor stations. The background environment of these sampled roads covers high‐density central urban areas (Mong Kok, Causeway Bay, and Tsim Sha Tsui) as well as the more open new towns (Tuen Mun, Tseung Kwan O). The testing experiments were carried out simultaneously during the Landsat‐8 transit time in Hong Kong during the period: 2019 and 2021. The prediction performance of the TOA‐LUR GTWR model was evaluated by comparing four seasonal PM2.5 estimation results with the actual PM2.5 results measured by Airbeam2. In order to reduce the accidental error of Airbeam2 in the mobile measurement, the average concentration of the measurement of two Airbeam2 was used as the actual PM2.5 concentration. In addition, the precision test of Airbeam2 in Hong Kong was conducted by comparing the mobile measurement values of Airbeam2 with the completed static measurement values of Mong Kok and Causeway Bay stations at the same locations. The results of the precision test are shown in Table 1.

Figure 2

The sampling routes used by Airbeam2 for the external data validation. (a) The sampling routes around Mong Kok Station. (b) The sampling routes around Causeway Bay Station. (c) The sampling routes at Tsim Sha Tsui. (d) The sampling routes at Tuen Mun. (e) The sampling routes at Tseung Kwan O.

Table 1

Precision Test of Airbeam2 Compared With the Two Roadside PM2.5 Monitoring Stations of Environmental Protection Department

	PM2.5(μg/m3)	PM2.5(μg/m3)
Time periods	X‐axis	Y‐axis	Linear regression	R 2
Spring	Airbeam2	Causeway Bay station	Y = 0.9144X + 1.7944	0.91
	Airbeam2	Mong Kok station	Y = 0.9261X + 1.4207	0.92
Summer	Airbeam2	Causeway Bay station	Y = 0.9136X + 1.8292	0.97
	Airbeam2	Mong Kok station	Y = 0.9868X + 1.2512	0.98
Autumn	Airbeam2	Causeway Bay station	Y = 0.9584X + 1.2205	0.92
	Airbeam2	Mong Kok station	Y = 0.9452X + 0.9935	0.92
Winter	Airbeam2	Causeway Bay station	Y = 0.8498X + 1.4676	0.90
	Airbeam2	Mong Kok station	Y = 0.9419X + 1.3123	0.90

Healthy Route Searching

An undirected graph R = (X, Z, WPM2.5), including X nodes, Z related edges, and the weights of edges ‐W . The weight W p represented the PM2.5 pollution weight of the edges. The example of the PM2.5 weight calculation method is shown in Figure 3. X1 and X2 were connected by an edge Z1 located in three grids. In three grids, the estimated PM2.5 concentrations by TOA‐LUR modeling are v 1, v 2, and v 3 respectively. The calculation formula for the weight W is (Tong et al., 2021): where d 1,1 is the length of edge Z1 in grid 1, d 1,2 is the length of edge Z1 in grid 2, and d 1,3 is the length of edge Z1 in grid 3.

Figure 3

The example of PM2.5 weight calculation method. X1 and X2 were connected by an edge Z1 located in three grids. In three grids, the estimated PM2.5 concentrations by TOA‐LUR modeling are v 1, v 2, and v 3 respectively. d 1,1 is the length of edge Z1 in grid 1. d 1,2 is the length of edge Z1 in grid 2. d 1,3 is the length of edge Z1 in grid 3.

Based on the example of the PM2.5 weight calculation method, the PM2.5 weight of all edges of the road network throughout the whole city could be obtained. The typical A* pathfinding algorithm (Szczerba et al., 2000) is then, used to extract the healthy travel route with the lowest PM2.5 cost. A* pathfinding algorithm is an acceptable heuristic search algorithm, which never overestimates the cost of reaching the destination node. Further, it can be used to quickly find the route with the lowest cost.

Results

TOA‐PM2.5 Models

Based on the Landsat 8 data, ground monitoring point data, and meteorological data from 2014 to 2019, MLR models of basic TOA‐PM2.5 in the annual year, spring, summer, autumn, and winter were developed. In order to illustrate the performance of these MLR models, the TOA‐only model (only TOA and related observation angles are considered to estimate PM2.5) was developed as a reference. The performance results of these models are shown in Table 2. The results show that, compared to the TOA‐only model, the basic TOA‐PM2.5 MLR models perform better in the prediction of the ground‐level PM2.5 in Hong Kong (Table 2). However, the performance of these TOA‐PM2.5 MLR models is still poor as regards the prediction of the ground‐level PM2.5 during each season throughout the annual year. The LOOCV value of the annual TOA‐PM2.5 MLR model is lower than 0.2 (Table 2). The LOOCV values of the spring, summer, autumn, and winter TOA‐PM2.5 MLR models are 0.21–0.36 (Table 2).

Table 2

The List of Resultant Baseline Models by Seasons and Improved TOA‐LUR MLR Models

		Models	Estimated performance
			R 2	R^2‾	Fitting‐RMSE (μg/m3)	LOOCV−R^2‾	LOOCV‐RMSE (μg/m3)	p
TOA‐only model	Annual	11.05*TOA2‐3.90*TOA4‐3.43*TOA7‐0.01*sat_az‐ 2.29*sat_zn+0.11*sun_az+0.26*sun_zn+231.55	0.20	0.19	18.11	0.19	18.38	<0.0001*
TOA‐PM2.5 MLR models	Annual	16.18*TOA4‐7.58*TOA2‐3.88*TOA7‐0.02*sat_az‐2.24*sat_zn+0.21*sun_az+0.49*sun_zn‐0.65*TEMP+0.15*WS‐0.03*PBLH‐0.01*PS‐1348	0.24	0.22	18.04	0.20	18.39	<0.0001*
	Spring	56.59*TOA4‐71.54*TOA2+5.51*TOA7‐0.05*sat_az‐3.60*sat_zn‐1.57*sun_az+0.31*sun_zn‐2.09*TEMP+5.51*WS‐0.02*PBLH‐0.01*PS+384.6	0.38	0.32	16.01	0.31	16.34	<0.0001*
	Summer	11.54*TOA2‐7.37*TOA4‐5.20*TOA7‐0.01*sat_az+0.002*sat_zn‐0.03*sun_az‐0.17*sun_zn+1.83*TEMP+1.32*WS‐0.01*PBLH+0.003*PS‐681	0.40	0.37	16.32	0.36	16.92	<0.0001*
	Autumn	12.23*TOA2+4.09*TOA4‐15.87*TOA7‐0.12*sat_az‐3.89*sat_zn+6.52*sun_az‐6.68*sun_zn+6.94*TEMP+14.28*WS‐0.05*PBLH‐0.01*PS‐1298	0.35	0.34	15.41	0.33	16.13	<0.0001*
	Winter	18.98*TOA2‐14.24*TOA4+3.53*TOA7‐0.002*sat_az‐0.50*sat_zn‐0.28*sun_az‐0.15*sun_zn‐0.71*TEMP‐0.40*WS‐0.001*PBLH‐0.01*PS+561	0.26	0.24	16.21	0.22	16.46	<0.0001*
TOA‐LUR MLR models	Annual	−109.6*NDVI‐0.02*PBLH‐0.03*PS+16.34*TOA2‐7.11*TOA4‐1.51*TOA7‐0.02*sat_az‐0.48*sat_zn+0.30*sun_az+0.41*sun_zn‐0.49*TEMP‐0.69*WS‐98.6*SVF‐2.13*slope‐37.14*GVI+0.01*CDS+0.13*rd‐main‐0.002*lu‐ops+4673	0.43	0.36	16.83	0.35	17.13	<0.0001*
	Spring	−45.12*NDVI‐0.02*PBLH‐0.01*PS+30.12*TOA2‐20.54*TOA4‐9.87*TOA7‐0.08*sat_az‐0.49*sat_zn‐0.43*sun_az+3.90*sun_zn+2.01*TEMP+3.23*WS+0.50*Elevation+642.1*SVF+0.32*slope‐10.86*GVI+0.01*temple+25.13*BR+0.01*CDS+0.04*rd‐main‐0.02*rd‐rai‐0.002*lu‐gov‐0.001*lu‐com‐0.002*lu‐ops‐1714	0.56	0.45	15.43	0.43	15.54	<0.0001*
	Summer	−11.88*NDVI‐0.01*PBLH+0.01*PS+13.14*TOA2‐8.34*TOA4‐6.01*TOA7 +0.04*sat_az+0.06*sat_zn+0.21*sun_az‐1.32*sun_zn+2.75*TEMP+2.32*WS+0.52 *Elevation+1104*SVF+1.73*slope‐14*GVI+0.01*BR+0.004*CR–0.01*CDS‐0.03*rd‐main+0.07*rd‐rai +0.05*rd‐ter‐0.001*lu‐ops‐8478	0.64	0.55	13.34	0.54	13.51	<0.0001*
	Autumn	0.02*PS+19.38*TOA2‐13.09*TOA4‐3.84*TOA7‐0.14*sat_az‐1.59*sat_zn+ 7.98*sun_az‐6.57*sun_zn+16.28*TEMP+19.34*WS‐897.4*SVF‐2.15*GVI‐0.002*CDS‐0.001*lu_open‐0.01*lu_res‐0.002*lu_com‐2138	0.59	0.53	13.81	0.51	14.03	<0.0001*
	Winter	−23.21*NDVI‐0.002*PS+11.32*TOA2+0.81*TOA4‐3.02*TOA7‐0.004*sat_az‐0.29*sat_zn+0.03*sun_az‐0.18*sun_zn‐0.62*TEMP‐0.58*WS‐458.6*SVF‐0.59*slope‐38.62*GVI‐0.001*CDS+0.01*CR+0.01*temple‐0.03*rd_main‐0.01*rd_rai‐0.02*rd_ter‐3715	0.50	0.44	15.68	0.41	15.82	<0.0001*

TOA‐LUR MLR Models

Based on the above‐mentioned basic TOA‐PM2.5 MLR models, by combining the microscale geographic prediction factors, the TOA‐LUR MLR models were developed to improve the performance for ground‐level PM2.5 prediction. The testing results (Table 2) show that the prediction performances of TOA‐LUR MLR models are higher than those of the basic TOA‐PM2.5 MLR models. Among these models, the LOOCV R2 for the annual TOA‐LUR MLR models increased to 0.35 (Table 2). The prediction performances for the TOA‐LUR MLR model in spring, summer, autumn, and winter were better. Th LOOCV of these models were increased to 0.41–0.54 (Table 2). The details of the selected variables of the TOA‐LUR MLR model are shown in Table S1. It is worth noting that only a few studies have introduced the GVI index as a predictor into the LUR model, to predict street‐scale PM2.5. Among the above models in this study, the GVI factor has acted as an important predictor in each of the annual and the four seasonal TOA‐LUR MLR models. When incorporating the GVI index, the LOOCV of annual and seasonal models could be increased by 7.16%–19.29% (Table 3).

Table 3

The Performance of TOA‐LUR MLR Models With and Without the GVI Index

		Performance (with the GVI index)						Performance (without the GVI index)
		R 2	R^2‾	Fitting‐RMSE (μg/m3)	LOOCV−R^2‾	LOOCV‐RMSE (μg/m3)	p	R 2	R^2‾	Fitting‐RMSE (μg/m3)	LOOCV−R^2‾	LOOCV‐RMSE (μg/m3)	p
TOA‐LUR MLR models	Annual	0.43	0.36	16.83	0.36	17.13	<0.0001*	0.40	0.34	17.27	0.31	17.41	<0.0001*
	Spring	0.56	0.45	15.43	0.43	15.54	<0.0001*	0.52	0.41	16.39	0.41	16.48	<0.0001*
	Summer	0.64	0.55	13.34	0.53	13.51	<0.0001*	0.60	0.48	14.46	0.45	14.83	<0.0001*
	Autumn	0.59	0.53	13.81	0.51	14.03	<0.0001*	0.56	0.45	15.02	0.44	15.60	<0.0001*
	Winter	0.50	0.44	15.68	0.41	15.82	<0.0001*	0.49	0.40	16.43	0.39	16.60	<0.0001*

TOA‐LUR GTWR Models

The addition of consideration of time effects and geographic non‐stationarity features into the model, enables the annual and four seasonal TOA‐LUR GTWR models to be further developed and thereby improve the accuracy of PM2.5 prediction (Table 4). The prediction accuracy in both annual and four seasonal TOA‐LUR GTWR models has been improved significantly (LOOCV‐R2 was increased by 62.18%–70.14%). Among these models, the LOOCV R2 for the summer TOA‐LUR GTWR model reached 0.90 (Table 4). In particular, the GVI also played an important role in annual and four seasonal TOA‐LUR GTWR models by increasing the LOOCV R2 by 2.09%–4.22% (Table 4). Even when compared to other studies using the improved LUR model to estimate PM2.5, the TOA‐LUR GTWR models established in this study have achieved a better (Table 5).

Table 4

The Performance of TOA‐LUR GTWR Models With and Without the GVI Index

		Performance (with the GVI index)						Performance (without the GVI index)
		R 2	R^2‾	Fitting‐RMSE (μg/m3)	LOOCV−R^2‾	LOOCV‐RMSE (μg/m3)	p	R 2	R^2‾	Fitting‐RMSE (μg/m3)	LOOCV−R^2‾	LOOCV‐RMSE (μg/m3)	p
TOA‐LUR GTWR models	Annual	0.62	0.60	12.56	0.59	12.79	<0.0001*	0.61	0.59	13.64	0.58	13.94	<0.0001*
	Spring	0.75	0.72	11.39	0.72	11.58	<0.0001*	0.71	0.70	12.67	0.69	12.92	<0.0001*
	Summer	0.92	0.90	9.26	0.90	9.43	<0.0001*	0.89	0.88	9.71	0.86	9.90	<0.0001*
	Autumn	0.86	0.84	10.40	0.83	10.46	<0.0001*	0.83	0.81	10.85	0.81	11.13	<0.0001*
	Winter	0.74	0.72	12.02	0.70	12.16	<0.0001*	0.71	0.68	13.20	0.68	13.42	<0.0001*

Table 5

Performance Comparison of Improved TOA‐LUR GTWR Models With Other Applications of Land Use Regression Models

Study	Study area	Season	R^2‾
Mo et al. (2021)	Guangzhou	Spring	0.60
		Summer	0.56
		Autumn	0.62
		Winter	0.80
Shi et al. (2018)	Hong Kong	Spring	0.62
		Summer	0.90
		Autumn	0.71
		Winter	0.64
Chen et al. (2021)	Taiwan	Spring	0.68
		Summer	0.71
		Autumn	0.67
		Winter	0.80
Li et al. (2018)	Beijing	Spring	0.74
		Summer	0.50
		Autumn	0.68
		Winter	0.79
Lee et al. (2016)	California	Spring	0.57
		Summer	0.43
		Autumn	0.63
		Winter	0.69
This study	Hong Kong	Spring	0.72
		Summer	0.90
		Autumn	0.84
		Winter	0.72

Of additional interest, is that verification of this study has, in turn, also verified the prediction results (Table 6) of the TOA‐LUR GTWR model over four seasons concerning (a) the routes near the two roadside AQMS, one in Mong Kok and one in Causeway Bay (ultra‐high‐density central area of Hong Kong), (b) the routes without roadside monitoring stations nearby, that is, in Tsim Sha Tsui (business and leisure center area), Tuen Mun (new town), and Tseung Kwan O (new town). The above test routes cover the main high active mobility hotspots in Hong Kong. The mobile measurement instruments ‐ Airbeam2 were used to test the prediction performance of the above TOA‐LUR GTWR models on these above routes (Figures 2a–2e). From the testing results of the four seasons on the sampling roads in the above five areas, the R 2 for the four seasonal calibrations still remained at 0.67–0.89 (Table 6). This shows that seasonal TOA‐LUR GTWR models perform well in the PM2.5 estimation in different background environments.

Table 6

Prediction Test From Seasonal TOA‐LUR GTWR Models and Measurements by Airbeam2 on the Sampling Routes of Hong Kong

	PM2.5(μg/m3)	PM2.5(μg/m3)	Linear regression	R 2
Time periods	X‐axis	Y‐axis
Spring	Estimations in Mong Kok	Airbeam2 in Mong Kok	Y = 0.8463X + 1.8336	0.71
	Estimations in Causeway Bay	Airbeam2 in Causeway Bay	Y = 0.8264X + 2.9915	0.69
	Estimations in Tsim Sha Tsui	Airbeam2 in Tsim Sha Tsui	Y = 0.8246X + 2.3258	0.69
	Estimations in Tuen Mun	Airbeam2 in Tuen Mun	Y = 0.8422X + 1.8336	0.70
	Estimations in Tseung Kwan O	Airbeam2 in Tseung Kwan O	Y = 0.5715X + 5.6394	0.68
Summer	Estimations in Mong Kok	Airbeam2 in Mong Kok	Y = 0.991X + 0.0165	0.89
	Estimations in Causeway Bay	Airbeam2 in Causeway Bay	Y = 0.9244X + 1.1708	0.88
	Estimations in Tsim Sha Tsui	Airbeam2 in Tsim Sha Tsui	Y = 0.921X + 0.9719	0.87
	Estimations in Tuen Mun	Airbeam2 in Tuen Mun	Y = 0.9255X + 1.1793	0.88
	Estimations in Tseung Kwan O	Airbeam2 in Tseung Kwan O	Y = 0.9621X + 0.32	0.88
Autumn	Estimations in Mong Kok	Airbeam2 in Mong Kok	Y = 0.8102X + 2.9568	0.83
	Estimations in Causeway Bay	Airbeam2 in Causeway Bay	Y = 0.8556X + 2.2829	0.82
	Estimations in Tsim Sha Tsui	Airbeam2 in Tsim Sha Tsui	Y = 0.8397X + 3.3487	0.81
	Estimations in Tuen Mun	Airbeam2 in Tuen Mun	Y = 0.9046X + 0.6963	0.83
	Estimations in Tseung Kwan O	Airbeam2 in Tseung Kwan O	Y = 0.8642X + 1.6912	0.80
Winter	Estimations in Mong Kok	Airbeam2 in Mong Kok	Y = 0.7167X + 10.256	0.70
	Estimations in Causeway Bay	Airbeam2 in Causeway Bay	Y = 0.6617X + 12.904	0.68
	Estimations in Tsim Sha Tsui	Airbeam2 in Tsim Sha Tsui	Y = 0.6928X + 6.6171	0.68
	Estimations in Tuen Mun	Airbeam2 in Tuen Mun	Y = 0.834X + 1.8524	0.69
	Estimations in Tseung Kwan O	Airbeam2 in Tseung Kwan O	Y = 0.7324X + 5.0851	0.67

Obtaining On‐Road PM2.5 Maps at High Spatial Resolution

The relevant seasonally on‐road PM2.5 spatial distribution maps were averaged, based on the daily PM2.5 estimations derived from above TOA‐LUR GTWR models (Figure 4). The on‐road PM2.5 distribution maps cover 70,788 roads in Hong Kong. It can be seen from the PM2.5 distribution maps that the roads in Central, Causeway Bay, Mong Kok, and Sham Shui Po, all of which are in the central urban area, are more polluted by PM2.5. The above‐mentioned areas are also key areas for air pollution prevention and control in Hong Kong (Li & Feng, 2018). In addition, this study also shows that Tuen Mun had severe PM2.5 pollution. The traffic volume on the roads in Tuen Mun area was usually high. However, this area is vulnerable to the air pollution of the Pearl River Delta, due to its geographical location. Thus, from the maps, it can be seen that the on‐road PM2.5 concentration in this area was particularly high in the winter. In summer, when local air pollution was dominated, on‐road PM2.5 concentration there was relatively low. The above results show that the TOA‐LUR GTWR models proposed by this study can not only identify the hot spots of PM2.5 pollution, but also show the seasonal differences in PM2.5 pollution. Therefore, the TOA‐LUR GTWR models proposed in this study can provide a reliable estimation of on‐road PM2.5 in the complex environment of high‐density cities.

Figure 4

PM2.5 distribution maps on 70, 788 roads of Hong Kong obtained by seasonal TOA‐LUR GTWR models. (a) PM2.5 distribution in Spring. (b) PM2.5 distribution in Summer. (c) PM2.5 distribution in Autumn. (d) PM2.5 distribution in Winter.

Developing the Healthy Route Planning App

Based on the mapping results of on‐road PM2.5 in Hong Kong and the health route searching method, a healthy route planning app was further developed, to provide the healthy travel route (including walking, running, and cycling) with less PM2.5 in Hong Kong. A typical example is shown in Figure 5. The example indicates the healthy travel route and the shortest route from the HONG KONG Polytechnic University to the Hong Kong Metropolitan University. The healthy travel route is 100m more than the shortest route but the PM2.5 pollution on the healthy travel route showed a decrease of 5.12%–10.58% over the four seasons. Furthermore, according to the PM2.5 exposure reduction of 15 typical route planning examples in Hong Kong, healthy route planning could reduce the PM2.5 exposure by 4.79%–25.08%, compared with the shortest route planning (Table 7). As a result, people, during their daily travel, can reduce their exposure to PM2.5 pollution.

Figure 5

The example of the healthy travel route planning from the HONG KONG Polytechnic University to the Hong Kong Metropolitan University.

Table 7

The PM2.5 Exposure Reduction of the Healthy Travel Route Compared With the Shortest Travel Route

		PM2.5 reduction (%)
Area	Route	Spring	Summer	Autumn	Winter
Wan Chai	Kew Green Hotel Wanchai ‐ Wan Chai Sports Ground	11.52	13.02	11.79	8.68
	Hong Kong Arts Centre–Tung Shing Building	12.19	14.01	11.95	9.46
	United Centre ‐ Wan Chai Market	13.07	13.95	13.45	10.23
	Cheong Ip Building ‐ JEMS Character Academy	10.43	11.06	10.67	6.51
	Bakehouse ‐ Harbour Centre	15.36	16.59	15.78	12.17
Tsim Sha Tsui	Kowloon Park ‐ The Royal Garden	13.15	14.87	13.47	9.95
	iSQUARE ‐ The Luxe Manor	9.78	10.34	9.43	5.17
	Hyatt Regency Hong Kong ‐ East Ocean Centre	11.23	11.78	11.36	8.52
	InterContinental Grand Stanford Hong Kong ‐ Hong Kong Observatory	8.22	10.05	8.54	4.79
	Tsim Sha Tsui Baptist Church ‐ Harbour City	10.96	11.52	11.23	6.98
Tuen Mun	On Ting Estate ‐ Tin Hau Temple Plaza	19.01	21.65	19.42	17.02
	Chelsea Heights Plaza ‐ Tsing Tin Playground	24.17	25.08	24.63	22.96
	Tai Hing Estate Hing Yiu House ‐ New Life Farm	22.47	24.01	23.75	19.12
	Tuen Mun San Hui Market ‐ Hotel COZi Resort	18.69	19.42	18.96	17.25
	Tin's Centre ‐ Waldorf Garden Tuen Mun	21.35	21.89	21.17	18.04

Conclusions

No matter where we live or how we travel, everyone begins and ends each trip as a pedestrian. In cities across the world, many people rely almost exclusively on walking and cycling to get to their destinations. The UN Environment Programme is calling for the development of new solutions that ensure that people who walk and cycle are able to experience improved physical health, and reduced air pollution (United Nations Environment Programme, 2020). Therefore, proposed is a framework that combines satellite observations and microscale geographic predictors to predict on‐road PM2.5 to assist healthy travel planning with less pollution to promote active mobility.

The results of the application of the proposed framework in Hong Kong reflect its characteristics in the following aspects:

The assurance of people's needs regarding healthy travel. The current existing route planning applications are mainly based on periods of time or road conditions for planning, and lastly lack of environmentally friendly route planning. Based on the microscale PM2.5 estimation and high‐resolution mapping of more than 70,000 roads in Hong Kong, a digital application providing healthy route planning has been developed. As a result, ensured has been the traveling health during the active mobility of all residents in Hong Kong. Furthermore, a feasible reference for healthy travel route planning has been provided for other regions.

Improved the correlation of TOA‐PM2.5 for the LUR models of spatiotemporal PM2.5 estimation in the high‐density city. The previous PM2.5 estimation research related to the LUR model (including the model combining LUR and AOD) can cover a large spatial range (such as national or regional scale), but cannot support a sufficiently high spatial resolution (especially below 100 m) to reflect the spatial variation of micro PM2.5 pollution within the city. In this study, by using TOA‐LUR modeling to consider micro‐scale geographic predictors including new factors such as GVI, the spatiotemporal correlation between TOA and ground PM2.5 was effectively enhanced. This makes the spatial estimation of PM2.5 more accurate, thus further supporting the acquisition of high spatial resolution micro‐scale PM2.5 distribution maps in the entire city.

Provides a fine‐scale mapping of the spatiotemporal distribution of on‐road PM2.5 based on TOA‐LUR modeling. These maps provide useful information for healthy route navigation and air pollution prevention by identifying more greatly refined PM2.5 hotspots. The latter is critical for high‐density cities with further complex environments such as Hong Kong. This also complements the shortcomings of the only three roadside inspection stations in the central area of Hong Kong. Thus, TOA‐LUR modeling and geographic mapping technologies remote sensing data can be used to fill in the gaps of existing monitoring networks and help public health applications on a finer spatial scale. This study is the first application case in which TOA‐LUR modeling is used to estimate on‐road PM2.5 in an ultra‐high‐density city, even though the established TOA‐LUR model cannot be directly transferred to other places. However, it is clear and of value that the framework proposed in this study provides a generalizable method for the refined estimation of air pollution in high‐density cities. Thus, relevant policy makers can make direct reference directly to it and use local data to adjust and further develop the model, to better enable a more comprehensive understanding of the differences in PM2.5 in different areas of the high‐density city. In addition, the applicability of the framework in other cities with less traffic needs to be further explored.

In order to assist people to travel more healthily during the active mobility, proposed is a framework to assist in the planning of healthy routes exposed to less PM2.5, by estimating PM2.5 at the high‐spatial‐resolution. First, based on the long‐term observational data from 2014 to 2021 and the data on micro‐geographical predictors, TOA‐LUR models were developed to predict the spatiotemporal PM2.5 distribution. Based on TOA‐LUR models, the daily fine‐scale spatiotemporal distribution of on‐road PM2.5 in Hong Kong was mapped. Further, by a healthy route search algorithm, an application that provides healthy travel route planning was developed. As a result, the framework has been used to assist Hong Kong's more than 7.5 million residents in healthy active mobility.

This study does have some limitations. Firstly, due to limitations in the availability of the data, especially the Landsat 8 data, this study can currently only estimate on‐road PM2.5 in the time frame of satellite transit from 10 to 11 a.m. each morning. The route planning services provided by the related healthy travel route planning applications are mainly for the morning rush hour travel in Hong Kong. After that, this study will further combine the satellite and ground‐based multi‐source pollution monitoring data with the higher spatiotemporal resolution to perform spatiotemporal fusion to estimate more refined PM2.5, to serve for the healthy route traveling during more time periods. Secondly, the direct estimation of PM2.5 from TOA reflectance is a very complicated process, especially for the nonlinear physical process. This study will consider applying the deep learning model to PM2.5 estimation based on TOA reflectance at the urban fine‐scale scale to further improve the estimation accuracy.

Every year, exposure to outdoor air pollution is estimated by World Health Organization (World Health Organization, 2021) to cause 7 million premature deaths and result in the loss of millions more healthy years of life. This mortality is due to exposure to PM2.5 in outdoor activities. As a result, the WHO calls on countries around the world to establish planning tools to support people's healthier and cleaner daily travel. Although there are some limitations, the original framework proposed in this study provides an estimation and a mapping of on‐road PM2.5 at the fine scale, providing people with healthier travel choices. It may become a reference and guide for high‐density cities, for smart travel planning, and refined monitoring of air pollution in urban areas. It is hoped, therefore, that this study assists more people in urban areas to reduce their exposure to outdoor air pollution, both to protect their health, and assist in the construction of healthy cities, especially for the high‐density city.

Conflict of Interest

The authors declare no conflicts of interest relevant to this study.

Table S1

Click here for additional data file.