Environmental Pollution 331 (2023) 121832

Contents lists available at ScienceDirect

Environmental Pollution
journal homepage: www.elsevier.com/locate/envpol

Integrating low-cost sensor monitoring, satellite mapping, and geospatial

artificial intelligence for intra-urban air pollution predictions☆
Lu Liang a, *, Jacob Daniels b, Colleen Bailey b, Leiqiu Hu c, Ronney Phillips a, John South a
a
Department of Geography and the Environment, University of North Texas, Denton, TX, 76203, USA
b
Department of Electrical Engineering, University of North Texas, Denton, TX, 76203, USA
c
Atmospheric and Earth Science Department, University of Alabama in Huntsville, Huntsville, AL, 35805, USA

A R T I C L E I N F O A B S T R A C T

Keywords: There is a growing need to apply geospatial artificial intelligence analysis to disparate environmental datasets to
Deep learning find solutions that benefit frontline communities. One such critically needed solution is the prediction of health-
Big data relevant ambient ground-level air pollution concentrations. However, many challenges exist surrounding the size
Explainable AI
and representativeness of limited ground reference stations for model development, reconciling multi-source
Inequality
PM2.5
data, and interpretability of deep learning models. This research addresses these challenges by leveraging a
MODIS AOD strategically deployed, extensive low-cost sensor (LCS) network that was rigorously calibrated through an
optimized neural network. A set of raster predictors with varying data quality and spatial scales was retrieved
and processed, including gap-filled satellite aerosol optical depth products and airborne LiDAR-derived 3D urban
form. We developed a multi-scale, attention-enhanced convolutional neural network model to reconcile the LCS
measurements and multi-source predictors for estimating daily PM2.5 concentration at 30-m resolution. This
model employs an advanced approach by using the geostatistical kriging method to generate a baseline pollution
pattern and a multi-scale residual method to identify both regional patterns and localized events for high-
frequency feature retention. We further used permutation tests to quantify the feature importance, which has
rarely been done in DL applications in environmental science. Finally, we demonstrated one application of the
model by investigating the air pollution inequality issue across and within various urbanization levels at the
block group scale. Overall, this research demonstrates the potential of geospatial AI analysis to provide
actionable solutions for addressing critical environmental issues.

1. Introduction and modeling of PM2.5 concentrations by tackling the nonlinearity of

environmental processes and improving predictive accuracy for massive
Fine particulate matter (PM2.5, <2.5 μm aerodynamic diameter) is time series data (Liu et al., 2022). However, its applications in envi­
the world’s leading environmental cause of disease and premature ronmental pollution remain in their infancy with a few challenges
death, disproportionately harming racial/ethnic minorities and lower- (Gupta et al., 2021).
income populations (Cohen et al., 2017; Bell and Ebisu, 2012; Colmer Data scarcity and data quality issues can introduce biases into DL
et al., 2020; Mikati et al., 2018; Miranda et al., 2011; Landrigan et al., models (Zhong et al., 2021). Characterizing air quality variations often
2018). Accurate estimation of PM2.5 concentrations is essential for relies on government agency-operated stations, which provide excellent
epidemiological, public health, and environmental justice analyses. data sources but lack sufficient local representation (Dimakopoulou
Much of the modeling of PM2.5 concentrations has evolved from inter­ et al., 2017; Shi et al., 2018). Given the high spatial variability of PM2.5
polation (Singh et al., 2011) and land use regression (Kerckhoffs et al., concentrations (Apte et al., 2017), using low-cost sensors (LCSs) has
2019) to machine learning (Adams and Kanaroglou, 2016; Bi et al., become a paradigm shift to supplement regulatory stations (Caubel
2020; Di et al., 2019; Lu et al., 2021). As a rapidly growing field, deep et al., 2019; Liang et al., 2019; Mao et al., 2019). However, data reli­
learning (DL) has made initial success in revolutionizing data analysis ability and site representativeness have been two contentious issues in

☆
This paper has been recommended for acceptance by Admir Créso Targino.
* Corresponding author.
E-mail address: [email protected] (L. Liang).

https://doi.org/10.1016/j.envpol.2023.121832
Received 1 February 2023; Received in revised form 12 May 2023; Accepted 13 May 2023
Available online 18 May 2023
0269-7491/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-
nc/4.0/).
L. Liang et al. Environmental Pollution 331 (2023) 121832

using LCSs for scientific discovery (Williams, 2019). LCSs typically show 2022). However, the large volume of point clouds brings additional
non-linear responsiveness to their interfering environments with drift­ challenges to data preprocessing and aggregation. The inclusion of both
ing performance (Giordano et al., 2021), making it challenging to ensure point and area contributors is promising to better reveal the complex
adequate data quality (Liang, 2021). Additionally, LCS deployment patterns of PM2.5 concentration, but the complexity of data formats and
largely depends on citizen science efforts, and participants are often varying data quality must first be addressed.
opportunistically or purposefully recruited, which may introduce bias Furthermore, the low interpretability of DL models (Chakraborty
toward certain groups with less consideration of sample representa­ et al., 2017; Meng et al., 2022) has limited their wide adoption in
tiveness (Palinkas et al., 2015). environmental research for understanding process drivers (Molnar,
Additionally, high spatial-temporal PM2.5 concentration modeling 2020) and generating new evidence-based knowledge (Mori et al.,
depends on multi-source data that may vary in data formats, resolution, 2019). In response to the long-standing black box issues of DL algo­
dimensions, and completeness, which can lead to algorithm dysfunction. rithms, explainable artificial intelligence (XAI) has emerged as a
For instance, satellite aerosol optical depth (AOD) is widely used to promising solution to improve the accountability of model behavior and
estimate PM2.5 concentrations (He et al., 2021; Van Donkelaar et al., feature importance estimation (Roscher et al., 2020; Rudin, 2019).
2019; Bi et al., 2022; Reid et al., 2015; Li et al., 2021; Yu et al., 2022), Despite its initial success in fields such as medicine, material, and geo­
but its retrievals are frequently affected by atmospheric contamination science (e.g., Ebert-Uphoff and Hilburn, 2020; Mamalakis et al., 2022;
and degrade over certain bright urban areas (Lyapustin et al., 2018; Yeung et al., 2020; Shimazaki and Tachikawa, 2022), XAI remains
Handschuh et al., 2022; Just et al., 2015; Zhang et al., 2018). Addi­ largely under-investigated in environmental science and engineering
tionally, the coarse resolution of satellite AOD (500-m to 1-km) makes practices. Given the rapid increase in both the quantity and complexity
predictions at the block-scale difficult. On the other hand, of environmental science data, the wider utilization of XAI is essential
high-resolution airborne Light Detection and Ranging (LiDAR) can for improved model interpretability and trustworthiness.
efficiently capture 3D microscale attributes related to air pollution In this study, we demonstrate our effort in this line of work by
dispersion (Xiang et al., 2020) and deposition (Ponette-González et al., integrating dense, rigorously calibrated LCS data, novel data retrievals

Fig. 1. The location of deployed sensors (as of Aug 22, 2022) overlays the sampling strata (a) in Denton County, Texas (b, c). Field photos (d–i) depict the distinct
landscape features of each stratum.

2
L. Liang et al. Environmental Pollution 331 (2023) 121832

and reconstructions for prediction variables, and XAI techniques for 2.2.2. LCS data calibration
improving the interpretability of our models and understanding patterns We utilized a previously developed neural network-based LCS cali­
and drivers of PM2.5 concentrations at the block level. Furthermore, we bration method (Liang and Daniels, 2022) and further optimized its
demonstrate the practical application of our work by showcasing the architecture through a sensitivity analysis. The number of hidden layers,
derived daily PM2.5 concentration maps in an environmental inequality neurons in those layers, and selected predictors were identified as crit­
study, with a goal of strengthening the connections between science and ical factors that control the model’s performance. We weighed the
stakeholders. trade-off between increased accuracy and computational complexity
when selecting the optimal model architecture given that increases in
2. Methods neurons and hidden layers lead to rapid increases in computational
complexity. Similarly, because the responsive rate of LCSs is controlled
2.1. Development of a dense low-cost sensor network by a range of internal and ambient environmental factors (Liang, 2021),
we created five datasets from seven PA collected variables, separately
We deployed our LCS network in Denton County, Texas, a socio- PM2.5 concentration, PM2.5 particle count (C2.5), PM1 count (C1), PM5
economically diverse county with an area size of 2275.36 km2 located count (C5), PM10 count (C10), relative humidity (RH), and temperature
on the northern edge of the Dallas–Fort Worth metroplex (Fig. 1) (US (T), to determine which covariate set possesses the most predictive
Census Bureau. nd, 2022). To ensure a good spatial representation of power for the reference PM2.5 data. The five datasets are PM2.5 solely,
PM2.5 concentration, we adopted a stratified sampling method (Matte PM2.5 + C2.5, PM2.5 + C2.5 + RH, PM2.5 + C2.5 + RH + T, and PM2.5 +
et al., 2013) that created six urban strata based on the level of urbani­ C2.5 + RH + T + C1 + C5 + C1. Each combination of model parameters
zation and traffic emission (see Sampling Design, Supporting Informa­ and input dataset was run five times for 1000 epochs and the coefficient
tion). With the goal of deploying a total of 100 sensors in Denton, 25 of determination (R2) and root mean squared error (RMSE) were aver­
sensors were allocated for rural areas, 40 in suburban, and 35 in urban. aged. The performance was evaluated using the US-wide PurpleAir
We prioritized urban and suburban areas because of their importance in correction dataset described in 2.2.1. Detailed descriptions are in SI
pollution studies and their higher spatial variability compared to rural “Calibration Algorithm and Optimization”.
areas. Within each category, the number of sensors was set propor­
tionately to their area sizes (Table S1). 2.3. Predictor variables
We employed the PM sensor PA–II–SD (dimension 3.5 × 3.5 × 5-in)
given its open data access and wide community deployment (Barkjohn In this section, we will explain the data sources of predictor variables
et al., 2021). The PA–II–SD employs laser particle counters that detect and the data processing procedure employed to minimize the discrep­
suspended particles in sizes of 0.3, 0.5, 1.0, 2.5, 5.0, and 10 μm using ancies in data formats, resolution, dimensions, and to ensure consistent
their reflectivity as they pass by laser beams. These particle counts are data completeness (Table 1).
used to infer the mass concentrations of PM1, PM2.5, and PM10 in μg/m3.
In addition to the laser particle counters, the PA–II–SD sensors are also 2.3.1. Gap-filled daily aerosol optical depth from satellite observations
equipped with temperature and humidity sensors that measure the We obtained the 1-km Moderate Resolution Imaging Spectroradi­
temperature and relative humidity of the air where the device is located. ometer Multi-Angle Implementation of Atmospheric Correction
The device uses Wi-Fi connectivity to transmit data every 2 min. The MCD19A2 AOD daily data product. To reconstruct missing data, we
sensors show strong correlations with the reference instrument both in developed a spatiotemporal long short-term memory (LSTM) convolu­
the field (R2 > 0.93) and laboratory (R2 > 0.99) (Air Quality Sensor tional autoencoder method (Daniels et al., 2022). The key steps in data
Performance Evaluation, 2023). We recruited volunteers through part­ preprocessing and model architecture are illustrated below, and a
nerships with public institutions, schools, city partners, and individuals. detailed explanation is provided in the SI.
Priority was given to locations with large vulnerable populations or To prepare the raw AOD imagery for model training, we first
underserved communities. The installation of sensors followed the field removed the cloud and water pixels using the product quality assurance
protocol described in the SI “Sensor Installation”. mask and imposed real cloud masks to cloud-free images. These cloud
masks were derived from a set of cloud-contaminated imagery to pro­
2.2. Sensor data pre-processing and calibration vide a more accurate representation of real-world cloud conditions
compared to artificial simulation. To avoid bias towards a particular
2.2.1. LCS data pre-processing corruption level, we balanced the cloud levels in the masks with an
To calibrate sensors before large-scale deployment, we collocated 10 equivalent quantity of each completeness ratio. The resulting dataset is a
sensors with a regulatory instrument at a Texas Commission on Envi­ time series of AOD retrievals that uses each complete image as the base
ronmental Quality monitoring station in Denton from April to November for training samples. The dataset consists of the current day’s complete
2020 (Fig. S1C). To make our calibration method location-independent, AOD retrieval (Dn) stacked with the previous four days of imagery (Dn-4-
we combined this local dataset with a US-wide PurpleAir correction Dn-1). Once each time series was constructed, we simulated cloudy
dataset (Barkjohn et al., 2021). To remove intra-day outliers, we per­ conditions by corrupting Dn using the previously extracted cloud masks.
formed systematic pre-processing including channel agreement checks, The spatiotemporal LSTM convolutional autoencoder method is
removal of extreme readings, and assessment of data completeness comprised of three major components: the mask-based attention mod­
(Liang and Daniels, 2022). Furthermore, we employed a modified L1 ule, preliminary reconstructions, and the convolutional autoencoder.
principal component analysis outlier detection system to remove The mask autoencoder structure extracts a dynamic attention map based
inter-day outliers (Depoian et al., 2020; Daniels et al., 2022). To ensure on the location of corrupted pixels, while the AOD autoencoder mirrors
an unbiased and equal distribution of samples, only sensors with ≥100 this functionality to match the new attention map. Both attention maps
days of pre-processed data were included. Additionally, to create an are multiplied to focus the model on the non-corrupted data, resulting in
equal representation of samples from each sensor, the number of sam­ attention-enhanced representations of Dn-4-Dn-1. These representations
ples for each sensor was truncated to match that of the sensor with the are then fed into ConvLSTM layers to enable further feature extraction
lowest sample count. Before truncation, the samples for each sensor characteristic of convolutional layers. This step preliminarily filled the
were randomized to prevent a bias towards a particular season or period corrupted Dn with the predicted AOD, which is then multiplied again by
of the year. Details are in SI pre-processing steps. the mask-based attention map to draw more weight to the known pixels.
The weighted AOD is fed into the convolutional autoencoder, which
downsamples the data into an optimized feature set and then upsamples

3
L. Liang et al. Environmental Pollution 331 (2023) 121832

Table 1
Variables assembled for PM2.5 concentration prediction.
Variable category Predictive variables No. of variables Description Spatial res. Frequency/year

Satellite estimates AOD 1 Aerosol optical depth 1-km daily

LULC NLCD 14 Land cover categorical classifications 30-m 2019
Lidar-derived 3D urban form Tree height 1 Mean LiDAR-derived digital tree height model within a grid 30-m 2020
Building height 1 LiDAR-derived digital building height model 30-m 2020
Traffic Road length 3 Total length of road segments within a grid 30-m 2019
Distance to roads 3 Distance to major/minor roads from the grid centroids 30-m 2019

it back to the original shape. This process retains only the most relevant 2.4. Neural network-based fine-resolution, daily prediction of PM2.5
data for the final prediction. Finally, a post-processing step replaces non- concentrations
corrupted pixels from the original Dn image in the convolutional
autoencoder’s output to create the final reconstructed AOD. 2.4.1. Modeling approach
The performance of our gap-filling method was evaluated by To address the multi-source data integration question, we developed
comparing its results to original, cloud-free images. It achieved a a multi-scale, attention-enhanced Convolutional Neural Network that
weighted average peak signal-to-noise ratio (PSNR), structural similar­ fused the sensor observations, daily satellite AOD data, and other pre­
ity index (SSIM), and R2 of 47.2, 0.992, and 0.941, respectively. This dictors to generate daily PM2.5 concentration estimates at 30-m resolu­
outperforms previous algorithms such as daily Multiple Imputation tion (Fig. 2).
(Xiao et al., 2017), weekly Non-Linear GAM, and a weekly Non-Full We first created a rough spatial distribution of PM2.5 concentration
Residual Network (Li et al., 2020) with a 22%, 16%, and 9% increase using Gaussian semivariogram-based kriging interpolation (Zhang et al.,
in R2, respectively. Fig. S2 shows a visual representation of the recon­ 2021) using calibrated LCS data. After concatenating the kriging inter­
structed AOD layers averaged by quarter. polated PM2.5 concentration with other raster data, we used a con­
volutional layer to extract features from the multitude of inputs and
2.3.2. LiDAR-derived digital tree and building height generate a refined feature set. Next, the refined feature set was processed
We obtained 1.2-TB of 1-m airborne LiDAR data to develop digital by a sequence of pooling and upsampling layers. The pooling layers
building and vegetation height models. After removing outliers from the reduced the width and height of the input image by a factor of two,
raw point clouds, we generated a Digital Surface Model (DSM) and producing a compressed sample of features that captured larger regional
Digital Elevation Model (DEM) using the kriging interpolation method. trends. The upsampling layers were then employed to restore the
We then obtained a rasterized Canopy Height Model (CHM, Fig. S3) by lower-resolution data to its original shape, matching the high-resolution
taking the difference between DSM and DEM and removing areas with data. These two sets were concatenated along the channel axis, allowing
CHM values less than 2-m (Ma et al., 2017). To further refine the CHM, the subsequent layers in the model to utilize information on both the
we used a land cover map produced using the 2022 National Agriculture hyper-localized components of the input and larger regional trends.
Imagery Program and UrbanWatch FLUTE framework (Zhang et al., After initial feature extraction and scale manipulation, we employed
2022) to mask out non-vegetation areas. Similar to CHM, the Building multiple convolutional and attention layers to refine the information
Height Model (BHM) was created by building DSM using points in the contained in the preceding outputs. Three identical blocks consisting of
building group. After subtracting the DEM from the DSM, we obtained a convolutional layer, a residual convolution block attention module
the BHM, which was polished by removing any artifacts less than 2-m (CBAM) layer, and another convolutional layer were used in series.
above the DEM. By validating against ground truth, the mean height Convolutional layers create filters capable of transforming their input
of the trees in the tree height model is 0.286-m less than the mean height into a representation that can predict the target variable. They were used
of the trees in the ground truth. to acquire characteristics from their inputs, with each of the three blocks
deriving increasingly complex relationships between the input and the
2.3.3. Other geographic variables target. Because some of the convolutional layers’ filters and their asso­
LULC variables were obtained from the 2019 National Land Cover ciated spatial patterns may possess greater predictive power than others,
Database (NLCD) which is an operational land cover monitoring pro­ we employed CBAMs between convolutional layers to emphasize the key
gram that characterizes eight level-I and 20 level-II landscape types features along both the channel and spatial axes using the channel and
using Landsat data (Dewitz, 2021). After a series of tests and trials, the spatial attention components of the CBAM, respectively. Finally, we
input of level II classes is proven to be more effective than the input of added a convolutional layer for feature extraction, followed by a single-
level I data in the DL modeling of PM2.5 concentration. In Denton filter layer that condensed the output to the desired single-channel
County, fourteen level II land cover types were identified and converted shape.
into binary layers. Two wetland Level II types were combined due to
their limited coverage (Table S2). 2.4.2. Evaluation of model performance and feature importance
The traffic variables include the proximity to the nearest road and To assess the model performance, we compiled days with more than
the total length of road within each grid cell. To calculate these vari­ four valid calibrated daily sensor records and randomly split the entire
ables, we divided the study area into a constellation of interconnected dataset with 80% of valid days during the study period being allocated
grids and calculated the distance between each grid centroid and the to the training dataset and the remaining 20% being used for testing. By
nearest road, as well as the total road length within each grid, based on performing the train/test split using days sampled from a uniform
different road categories. We obtained the road data from US Census random distribution, we avoid bias toward any subset of the study
Bureau Streets TIGER 2019 and combined the route types into three period and ensure no temporal overlap between training and test sets.
categories depending on the traffic volume, separately interstate, major, Days with fewer than four operational sensors were considered as
and minor roads (Table S3). inadequate for training and testing a DL model. To compute the error
between the predicted PM2.5 concentrations and the true values during
both training and evaluation, for each day, a leave one location out
(LOLO) schema is used. We iterate over each sensor for the current day,
leaving that sensor out of the kriging process and using it for validation.

4
L. Liang et al. Environmental Pollution 331 (2023) 121832

Fig. 2. Multi-scale, attention-enhanced convolutional neural network architecture for predicting daily PM2.5 concentrations.

This provides Ns sensor-data/target-data pairs to be used for model been proven to be a useful indicator of health outcomes and performs
input and error calculation, respectively, where Ns corresponds to the similarly to multidimensional indices (Berkowitz et al., 2015; Lynch
number of sensors reporting valid data on any given day. This allows et al., 2004). We classified the urbanization levels based on population
evaluation of the model using real sensor data while maximizing the density, with areas having more than 2534.5 persons/mile2 classified as
number of sensor-data inputs that can be used for kriging. Combined urban, areas with less than 500 persons/miles2 classified as rural, and
with the aforementioned day-based partitioning between the training areas in between labeled as suburban (US Census Bureau. nd, 2022). We
and testing datasets, this ensures complete independence for all possible then calculated the annual mean PM2.5 concentration for each block
sensor reporting combinations between the training data on which the group using the daily prediction and linked the concentration data with
model learns and the testing data on which error metrics are reported. the family income data for each urbanization level.
MAE, RMSE, and R2 are computed between the predictions and the left-
out ground-truth sensor to provide a robust understanding of the 3. Results and discussion
model’s performance.
We further investigated the roles of several key variable sets that 3.1. ML-based data calibration and sampling strategy
influence daily PM2.5 concentration variations using the permutation
feature importance method. This involved permuting the feature values The best sensor calibration performance was achieved using a neural
in the test dataset to determine the extent to which they contributed to network consisting of a single hidden layer of 128 neurons with all seven
the model’s predictive power, both alone and in combination with other input predictor variables. The R2 and RMSE values for this model were
variables. To ensure the integrity of the dataset and avoid unrealistic 0.94 and 1.34 μg/m3, respectively. After calibration, the mean MAE
value distributions, we replaced the permuted values with correspond­ significantly dropped towards zero (Fig. 3a) and the comparison against
ing values for the selected variable from the other samples. In the case of the reference data demonstrated good distribution along the 1:1 line
static layers that remain constant across all samples, we also shuffled the (Fig. 3b). Further details regarding the effects of algorithm selection and
intra-sample values to facilitate feature importance detection. This predictor variable combination on calibration performance can be found
permutation and performance computation was repeated five times for in SI “Calibration algorithm and optimization”. Since the usage of deep
each input feature to ensure reliable results and minimize any bias due learning in LCS calibration remains limited, the dataset features and
to randomization seeds. neural network architecture detailed here provide a baseline model for
such purposes.
2.5. Case study: assessing air pollution disparities We addressed the site representativeness issue through a stratified
sampling approach to mitigate sampling errors when the sample size is
Using the predicted PM2.5 concentration maps, we further evaluated insufficient. However, determining the optimal number of sites within
how air pollution varies across neighborhoods with different socio­ the designed strata is challenging for a county that is three times larger
demographic characteristics. Although income inequality between than New York City. Fortunately, our research team has established
urban and rural areas is well documented (Thiede et al., 2020), we long-term partnerships with a local city and a school district, which aid
specifically focused on assessing air pollution disparities in the urban in the recruitment of sensor hosts. Additionally, the use of a solar panel-
context and its suburban and rural counterparts. powered system enables the set up in rural areas without access to Wi-Fi
We retrieved sociodemographic data from the American Community or power, which also proves the superiority of LCSs in serving histori­
Survey, a demographics survey program conducted by the U.S. Census cally underserved regions (Pope et al., 2018; Raheja et al., 2022).
Bureau annually. We used the 2020 five-year estimates instead of the
one-year estimates to obtain block group level data and ensured that the 3.2. Integration and interpretation of multi-source features in the black-
acquisition year matched with other datasets (U.S. Census Bureau, box DL model predicting for PM2.5 concentration
2022). To measure socioeconomic status, we selected the average family
income variable (variable A14011_001 in Social Explorer), which has The proposed multi-scale, residual-attention-enhanced

5
L. Liang et al. Environmental Pollution 331 (2023) 121832

Fig. 3. The performance of sensor calibration by comparing against the reference data. A) the distributions of mean absolute error (MAE) of the raw PurpleAir data
(in red) and calibrated PurpleAir data (in blue); B) the correlation plot between the reference data and the calibrated PurpleAir data of PM2.5 concentration, with
their corresponding density plots. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

convolutional network model achieved an MAE, RMSE, and R2 of 1.15 prediction model due to its lower resolution compared to the target
μg/m3, 1.74 μg/m3, and 0.792, respectively. In contrast, the kriging- variable, providing less information for small-scale effects, and the AOD
based predictions only achieve a mean MAE of 2.58 μg/m3 and R2 of layer has inherent error from the required data reconstruction needed to
0.105 (Fig. S6). The model demonstrated good predictive power and achieve full satellite coverage. The daily temporal resolution of the AOD
stability, as evidenced by the small deviation from true PM2.5 values layer generally provides necessary inter-day change detection, but the
across a broad range of concentrations (Figs. S7 and S8). By comparing kriging-interpolated surface using PA sensor data already contributes
against the daily mean PM2.5 measured by the only TCEQ station in significantly to this requirement, as evidenced by its correlation with
Denton County, our predicted results reached an RMSE of 1.64 μg/m3, AOD (Fig. S8). Building heights provide little to no useable variance as
an MAE of 1.25 μg/m3, and R2 of 0.80 (Fig. S9). Furthermore, unlike Denton County contains limited high-story buildings and is relatively
land-use regression, this model does not impose any requirements on a flat. The tree heights, on the other hand, provide information on the
statistical distribution of the input variables or other model assumption. relative distribution and density of trees and likely can serve as a proxy
Instead, it recognizes sophisticated aggregate interactions between in­ for the level of development known to be associated with higher PM2.5
puts and the target PM2.5 concentration distribution. In comparison to concentrations and, therefore, augment the land use data.
previous studies, our work advanced this line of work by using the
kriging method to generate a baseline pollution pattern and incorpo­
rating a multi-scale residual approach. This allows the model to identify 3.3. Air pollution patterns and disparity
both regional patterns and localized events, enabling high-frequency
feature retention. There is no obvious disparity between air quality and income status
Table 2 presents the permutation test results in descending order of when comparing urban, rural, and suburban block groups in Denton
feature importance, equivalent to the descending error-increase post- County. There are 298 urban block groups with a mean family income of
permutation. The baseline model, which includes the entire set of input $120,501 ± $57,925. As expected, the annual mean PM2.5 concentra­
variables, has the highest accuracy due to its most comprehensive in­ tion in urban areas is the highest (4.28 μg/m3) with the largest variation
formation from the predictors. The traffic, LULC, and PA features are the (±1.32 μg/m3). In contrast, the rural area occupies the majority area of
most informative, followed by LiDAR tree height. AOD contributes the county’s area (69%), with the lowest mean PM2.5 concentration
marginally while LiDAR building height provides no relevant prediction (3.50 ± 1.19 μg/m3) across all block groups. The average family income
power. in rural areas is $120,455 ± $62,592, which is close to its urban coun­
The prominence of traffic variables in the top rank of feature terpart. The suburban block groups constitute 19% of the whole county
importance aligns with previous studies that have identified road and have the highest average family income of $146,129 ± $63,349.
emissions contribute greatly to PM2.5 concentrations in the growing The annual mean PM2.5 concentration is 3.75 ± 1.27 μg/m3 and is
Denton County (Hart et al., 2020; Luce et al., 2020). The importance of ranked middle.
land cover layers is also expected since human activity is a major source However, at the intra-group level, the disparities in exposure to air
of PM2.5 concentrations. In contrast, AOD contributes less to the pollution among lower-income groups become prominent, especially in
urban (Fig. 4a and d) and suburban areas (Fig. 4b and e). In rural areas,
the air pollution concentration is more even across the income range
Table 2 (Fig. 4c, f, Fig. 5). The air pollution in the suburbs decreases as the
Permutation feature importance sorted in descending importance order. Scores average family income increases (Fig. 5). This negative relationship
are given as MAEmean ± MAEstd and ΔMAEmean ± MAEstd, respectively. becomes more evident in the urban areas, where many low-income
block groups are exposed to a high PM2.5 concentration.
Permuted Feature MAE ΔMAE
Although particulate air pollution in the U.S. has declined consid­
Baseline 1.146 –
erably in recent decades, our findings suggest that relative disparities
Traffic 4.240 ± 0.011 3.095 ± 0.011
LULC 3.707 ± 0.024 2.562 ± 0.024 persist for most block groups. This agrees with the conclusion of some
Purple Air (kriging) 2.586 ± 0.060 1.441 ± 0.060 recent work (e.g., Bell and Ebisu, 2012). While previous studies have
LiDAR Tree Height 1.293 ± 0.010 0.148 ± 0.010 typically focused on cross-comparisons among income or racial/ethnic
AOD 1.164 ± 0.004 0.019 ± 0.004 groups (e.g., Jbaily et al., 2022), we investigated relative disparities
LiDAR Building Height 1.146 ± 0.001 0.000 ± 0.001
within rural, urban, and suburban areas. This division was made

6
L. Liang et al. Environmental Pollution 331 (2023) 121832

Fig. 4. The visual comparison between average family income (a–c) and predicted annual mean PM2.5 concentration (d–f) at the urban, suburban, and rural scales.
For better graphical illustration, both income and PM2.5 concentration are classified into five quantiles within their range.

primarily because between-group inequality in environmental health 4. Conclusions

risk is not always measurable (Harper et al., 2013). For instance, the cost
of living in rural areas is typically lower than urban or suburban areas. In summary, this study demonstrates a multi-disciplinary way of
Using the family average income or other survey metrics alone cannot using deep learning in air quality applications. We overcame the black
accurately reflect the poverty status rate. Evaluating the environmental box issue of deep learning models in environmental applications using
inequality at the intra-group level can lead to more accurate results. This the explainable machine learning approach. In this case study, traffic,
research framework is also adaptable to other regions that are experi­ LULC, and sensor data were proven to be most useful in PM2.5 concen­
encing constantly reshaped socio-economic profiles of these trartion prediction. Surprisingly, AOD that has been widely used pre­
communities. viously was not informative. The new multi-scale, attention-enhanced

7
L. Liang et al. Environmental Pollution 331 (2023) 121832

Fig. 5. The correlation between average family income and annual mean PM2.5 concentration at the three different urbanization levels.

CNN architecture has been proven effective, which can be applicable for org/10.1016/j.envpol.2023.121832.
dynamically understanding granular-scale pollution gradients. The air
pollution disparity is most prominent among lower-income groups in References
urban and suburban areas, while rural areas have less evident air quality
inequality issue. Our reproducible data-science approach offers a new Adams, M.D., Kanaroglou, P.S., 2016. Mapping real-time air pollution health risk for
environmental management: combining mobile and stationary air pollution
perspective on the associated disparities within urban, rural, and sub­ monitoring with neural network models. J. Environ. Manag. 168, 133–141.
urban communities. Air quality sensor performance evaluation center. URL. https://www.aqmd.gov/docs/de
fault-source/aq-spec/summary/purpleair-pa-ii-summary-report.pdf?sfvrsn=4.
(Accessed 14 March 2023).
Funding sources Apte, J.S., Messier, K.P., Gani, S., Brauer, M., Kirchstetter, T.W., Lunden, M.M.,
Marshall, J.D., Portier, C.J., Vermeulen, R.C., Hamburg, S.P., 2017. High-resolution
This project is supported by the National Science Foundation (BCS- air pollution mapping with Google street view cars: exploiting big data. Environ. Sci.
Technol. 51 (12), 6999–7008.
2117505). Barkjohn, K.K., Gantt, B., Clements, A.L., 2021. Development and application of a United
States-wide correction for PM2.5 data collected with the PurpleAir sensor. Atmos.
Meas. Tech. 14 (6), 4617–4637.
Credit author statement
Bell, M.L., Ebisu, K., 2012. Environmental inequality in exposures to airborne particulate
matter components in the United States. Environ. Health Perspect. 120 (12),
Lu Liang: Conceptualization, Methodology, Writing – review & 1699–1704.
editing, Project administration. Jacob Daniels: Formal analysis; Inves­ Berkowitz, S.A., Traore, C.Y., Singer, D.E., Atlas, S.J., 2015. Evaluating area-based
socioeconomic status indicators for monitoring disparities within health care
tigation, Writing – original draft. Colleen Bailey: Methodology; Ron­ systems: results from a primary care network. Health Serv. Res. 50 (2), 398–417.
ney Phillips: Investigation, Visualization. John South: Investigation, Bi, J., Wildani, A., Chang, H.H., Liu, Y., 2020. Incorporating low-cost sensor
Visualization. Leiqiu Hu: Data curation, Writing – review & editing. measurements into high-resolution PM2.5 modeling at a large spatial scale. Environ.
Sci. Technol. 54 (4), 2152–2162.
Bi, J., Knowland, K.E., Keller, C.A., Liu, Y., 2022. Combining machine learning and
numerical simulation for high-resolution PM2.5 concentration forecast. Environ. Sci.
Declaration of competing interest Technol. 56 (3), 1544–1556.
Caubel, J.J., Cados, T.E., Preble, C.V., Kirchstetter, T.W., 2019. A distributed network of
The authors declare the following financial interests/personal re­ 100 black carbon sensors for 100 days of air quality monitoring in West Oakland,
California. Environ. Sci. Technol. 53 (13), 7564–7573.
lationships which may be considered as potential competing interests:
Chakraborty, S., Tomsett, R., Raghavendra, R., Harborne, D., Alzantot, M., Cerutti, F.,
Lu Liang reports financial support was provided by National Science Srivastava, M., Preece, A., Julier, S., Rao, R.M., Kelley, T.D., 2017. Interpretability of
Foundation. Deep Learning Models: A Survey of Results. In 2017 IEEE Smartworld, Ubiquitous
Intelligence & Computing, Advanced & Trusted Computed, Scalable Computing &
Communications, Cloud & Big Data Computing, Internet of People and Smart City
Data availability Innovation (smartworld/SCALCOM/UIC/ATC/CBDcom/IOP/SCI). IEEE, pp. 1–6.
Cohen, A.J., Brauer, M., Burnett, R., Anderson, H.R., Frostad, J., Estep, K.,
Data will be made available on request. Balakrishnan, K., Brunekreef, B., Dandona, L., Dandona, R., Feigin, V., 2017.
Estimates and 25-year trends of the global burden of disease attributable to ambient
air pollution: an analysis of data from the Global Burden of Diseases Study. Lancet
Acknowledgment 389 (10082), 1907–1918, 2015.
Colmer, J., Hardman, I., Shimshack, J., Voorheis, J., 2020. Disparities in PM2.5 air
pollution in the United States. Science 369 (6503), 575–578.
The authors would like to thank many student volunteers at the Daniels, J., Bailey, C., Liang, L., 2022. Filling cloud gaps in satellite AOD retrievals using
University of North Texas for setting up the calibration system at the an LSTM CNN-Autoencoder model. In: Proceedings of International Geoscience and
Texas site, deploying sensors, and collecting data. Special thanks to Dr. Remote Sensing Symposium, pp. 2758–2761. https://doi.org/10.1109/
IGARSS46834.2022.988448.
Alexandra Ponette-González for her constructive feedback on the Depoian II, A.C., King, E., Bailey, C.P., Guturu, P., 2020. Meteorological Data Outlier
manuscript and the project. Detection: a Principal Component Approach. In Remote Sensing for Agriculture,
Ecosystems, and Hydrology XXII., vol. 11528. SPIE, p. 159.
Dewitz, J., 2021. 2021. National land cover database (NLCD) 2019 products (ver. 2.0. U.
Appendix A. Supplementary data S. Geological Survey data release. https://doi.org/10.5066/P9KZCM54.
Di, Q., Amini, H., Shi, L., Kloog, I., Silvern, R., Kelly, J., Sabath, M.B., Choirat, C.,
Supplementary data to this article can be found online at https://doi. Koutrakis, P., Lyapustin, A., Wang, Y., 2019. An ensemble-based model of PM2.5

8
L. Liang et al. Environmental Pollution 331 (2023) 121832

concentration across the contiguous United States with high spatiotemporal patterns of multiple combustion air pollutants in New York City: design and
resolution. Environ. Int. 130, 104909. implementation. J. Expo. Sci. Environ. Epidemiol. 23 (3), 223–231.
Dimakopoulou, K., Gryparis, A., Katsouyanni, K., 2017. Using spatio-temporal land use Meng, C., Trinh, L., Xu, N., Enouen, J., Liu, Y., 2022. Interpretability and fairness
regression models to address spatial variation in air pollution concentrations in time evaluation of deep learning models on MIMIC-IV dataset. Sci. Rep. 12 (1), 1–28.
series studies. Air quality. Atmosphere & Health 10 (9), 1139–1149. Mikati, I., Benson, A.F., Luben, T.J., Sacks, J.D., Richmond-Bryant, J., 2018. Disparities
Ebert-Uphoff, I., Hilburn, K., 2020. Evaluation, tuning, and interpretation of neural in distribution of particulate matter emission sources by race and poverty status. Am.
networks for working with images in meteorological applications. Bull. Am. J. Publ. Health 108 (4), 480–485.
Meteorol. Soc. 101 (12), E2149–E2170. Miranda, M.L., Edwards, S.E., Keating, M.H., Paul, C.J., 2011. Making the environmental
Giordano, M.R., Malings, C., Pandis, S.N., Presto, A.A., McNeill, V.F., Westervelt, D.M., justice grade: the relative burden of air pollution exposure in the United States. Int.
Beekmann, M., Subramanian, R., 2021. From low-cost sensors to high-quality data: a J. Environ. Res. Publ. Health 8 (6), 1755–1771.
summary of challenges and best practices for effectively calibrating low-cost Molnar, C., 2020. Interpretable machine learning. URL. https://christophm.github.io/i
particulate matter mass sensors. J. Aerosol Sci. 158, 105833. nterpretable-ml-book/.
Gupta, S., Aga, D., Pruden, A., Zhang, L., Vikesland, P., 2021. Data analytics for Mori, N., Debeljak, B., Škerjanec, M., Simčič, T., Kanduč, T., Brancelj, A., 2019.
environmental science and engineering research. Environ. Sci. Technol. 55 (16), Modelling the effects of multiple stressors on respiration and microbial biomass in
10895–10907. the hyporheic zone using decision trees. Water Res. 149, 9–20.
Handschuh, J., Erbertseder, T., Schaap, M., Baier, F., 2022. Estimating PM2.5 surface Palinkas, L.A., Horwitz, S.M., Green, C.A., Wisdom, J.P., Duan, N., Hoagwood, K., 2015.
concentrations from AOD: a combination of SLSTR and MODIS. Remote Sensing Purposeful sampling for qualitative data collection and analysis in mixed method
Applications: Society and Environment 26, 100716. implementation research. Adm. Pol. Ment. Health 42 (5), 533–544.
Harper, S., Ruder, E., Roman, H.A., Geggel, A., Nweke, O., Payne-Sturges, D., Levy, J.I., Ponette-González, A.G., Chen, D., Elderbrock, E., Rindy, J.E., Barrett, T.E., Luce, B.W.,
2013. Using inequality measures to incorporate environmental justice into Lee, J.H., Ko, Y., Weathers, K.C., 2022. Urban edge trees: urban form and
regulatory analyses. Int. J. Environ. Res. Publ. Health 10 (9), 4039–4059. meteorology drive elemental carbon deposition to canopies and soils. Environ.
Hart, R., Liang, L., Dong, P., 2020. Monitoring, mapping, and modeling spatial–temporal Pollut. 314, 120197.
patterns of PM2.5 for improved understanding of air pollution dynamics using Pope, F.D., Gatari, M., Ng’ang’a, D., Poynter, A., Blake, R., 2018. Airborne particulate
portable sensing technologies. International Journal of Environmental Eesearch and matter monitoring in Kenya using calibrated low-cost sensors. Atmos. Chem. Phys.
Public Health 17 (14), 4914. 18 (20), 15403–15418.
He, Q., Gao, K., Zhang, L., Song, Y., Zhang, M., 2021. Satellite-derived 1-km estimates Raheja, G., Harper, L., Hoffman, A., Gorby, Y., Freese, L., O’Leary, B., Deron, N.,
and long-term trends of PM2.5 concentrations in China from 2000 to 2018. Environ. Smith, S., Auch, T., Goodwin, M., Westervelt, D.M., 2022. Community-based
Int. 156, 106726. participatory research for low-cost air pollution monitoring in the wake of
Jbaily, A., Zhou, X., Liu, J., Lee, T.H., Kamareddine, L., Verguet, S., Dominici, F., 2022. unconventional oil and gas development in the Ohio River Valley: empowering
Air pollution exposure disparities across US population and income groups. Nature impacted residents through community science. Environ. Res. Lett. 17 (6), 65006.
601 (7892), 228–233. Reid, C.E., Jerrett, M., Petersen, M.L., Pfister, G.G., Morefield, P.E., Tager, I.B.,
Just, A.C., Wright, R.O., Schwartz, J., Coull, B.A., Baccarelli, A.A., Tellez-Rojo, M.M., Raffuse, S.M., Balmes, J.R., 2015. Spatiotemporal prediction of fine particulate
Moody, E., Wang, Y., Lyapustin, A., Kloog, I., 2015. Using high-resolution satellite matter during the 2008 northern California wildfires using machine learning.
aerosol optical depth to estimate daily PM2.5 geographical distribution in Mexico Environ. Sci. Technol. 49 (6), 3887–3896.
City. Environ. Sci. Technol. 49 (14), 8576–8584. Roscher, R., Bohn, B., Duarte, M.F., Garcke, J., 2020. Explainable machine learning for
Kerckhoffs, J., Hoek, G., Portengen, L., Brunekreef, B., Vermeulen, R.C., 2019. scientific insights and discoveries. IEEE Access 8, 42200–42216.
Performance of prediction algorithms for modeling outdoor air pollution spatial Rudin, C., 2019. Stop explaining black box machine learning models for high stakes
surfaces. Environ. Sci. Technol. 53 (3), 1413–1421. decisions and use interpretable models instead. Nat. Mach. Intell. 1 (5), 206–215.
Landrigan, P.J., Fuller, R., Acosta, N.J.R., Adeyi, O., Arnold, R., Basu, N.N., Baldé, A.B., Shi, X., Zhao, C., Jiang, J.H., Wang, C., Yang, X., Yung, Y.L., 2018. Spatial
Bertollini, R., Bose-O’Reilly, S., Boufford, J.I., Breysse, P.N., Chiles, T., Mahidol, C., representativeness of PM2.5 concentrations obtained using observations from
Coll-Seck, A.M., Cropper, M.L., Fobil, J., Fuster, V., Greenstone, M., Haines, A., network stations. J. Geophys. Res. Atmos. 123 (6), 3145–3158.
Hanrahan, D., Hunter, D., Khare, M., Krupnick, A., Lanphear, B., Lohani, B., Shimazaki, T., Tachikawa, M., 2022. Collaborative approach between explainable
Martin, K., Mathiasen, K.V., McTeer, M.A., Murray, C.J.L., Ndahimananjara, J.D., artificial intelligence and simplified chemical interactions to explore active ligands
Perera, F., Potočnik, J., Preker, A.S., Ramesh, J., Rockström, J., Salinas, C., for cyclin-dependent kinase 2. ACS Omega 7 (12), 10372–10381.
Samson, L.D., Sandilya, K., Sly, P.D., Smith, K.R., Steiner, A., Stewart, R.B., Suk, W. Singh, V., Carnevale, C., Finzi, G., Pisoni, E., Volta, M., 2011. A cokriging based approach
A., van Schayck, O.C.P., Yadama, G.N., Yumkella, K., Zhong, M., 2018. The Lancet to reconstruct air pollution maps, processing measurement station concentrations
Commission on pollution and health. Lancet 391 (10119), 462–512. https://doi.org/ and deterministic model simulations. Environ. Model. Software 26 (6), 778–786.
10.1016/S0140-6736(17)32345-0. Thiede, B.C., Butler, J.L., Brown, D.L., Jensen, L., 2020. Income inequality across the
Li, L., Franklin, M., Girguis, M., Lurmann, F., Wu, J., Pavlovic, N., Breton, C., rural-urban continuum in the United States, 1970–2016. Rural Sociol. 85 (4),
Gilliland, F., Habre, R., 2020. Spatiotemporal imputation of MAIAC AOD using deep 899–937.
learning with downscaling. Rem. Sens. Environ. 237, 111584. Retrieved from The American Community Survey 2020 5-year Estimates, 2022. U.S.
Li, H., Yang, Y., Wang, H., Li, B., Wang, P., Li, J., Liao, H., 2021. Constructing a Census Bureau. https://www.socialexplorer.com/blog/post/the-american-communi
spatiotemporally coherent long-term PM2.5 concentration dataset over China during ty-survey-2020-5-year-estimates-are-live-on-social-explorer-12373. (Accessed 31
1980–2019 using a machine learning approach. Sci. Total Environ. 765, 144263. March 2023).
Liang, L., 2021. Calibrating low-cost sensors for ambient air monitoring: techniques, US Census Bureau. nd. Urban area facts. URL. https://www.census.gov/programs-sur
trends, and challenges. Environ. Res. 197, 111163. veys/geography/guidance/geo-areas/urban-rural/ua-facts.html#:~:text=2%
Liang, L., Daniels, J., 2022. What influences low-cost sensor data calibration?- a 2C534.4%20persons%20per%20square%20mile%3A%20Overall%20urbanized,
systematic assessment of algorithms, duration, and predictor selection. Aerosol Air population%20density%20in%20the%20U.S. (Accessed 20 September 2022).
Qual. Res. 22, 220076. Van Donkelaar, A., Martin, R.V., Li, C., Burnett, R.T., 2019. Regional estimates of
Liang, L., Gong, P., Cong, N., Li, Z., Zhao, Y., Chen, Y., 2019. Assessment of personal chemical composition of fine particulate matter using a combined geoscience-
exposure to particulate air pollution: the first result of City Health Outlook (CHO) statistical method with information from satellites, models, and monitors. Environ.
project. BMC Publ. Health 19 (1), 1–2. Sci. Technol. 53 (5), 2595–2611.
Liu, X., Lu, D., Zhang, A., Liu, Q., Jiang, G., 2022. Data-driven machine learning in Williams, D.E., 2019. Low cost sensor networks: how do we know the data are reliable?
environmental pollution: gains and problems. Environ. Sci. Technol. 56 (4), ACS Sens. 4 (10), 2558–2565.
2124–2133. Xiang, Y., Lv, L., Chai, W., Zhang, T., Liu, J., Liu, W., 2020. Using Lidar technology to
Lu, T.J., Marshall, J.D., Zhang, W.W., Hystad, P., Kim, S.Y., Bechle, M.J., Demuzere, M., assess regional air pollution and improve estimates of PM2.5 transport in the North
Hankey, S., 2021. National empirical models of air pollution using microscale China Plain. Environ. Res. Lett. 15 (9), 94071.
measures of the urban environment. Environ. Sci. Technol. 55 (22), 15519–15530. Xiao, Q., Wang, Y., Chang, H.H., Meng, X., Geng, G., Lyapustin, A., Liu, Y., 2017. Full-
https://doi.org/10.1021/acs.est.1c04047. coverage high-resolution daily PM2.5 estimation using MAIAC AOD in the Yangtze
Luce, B.W., Barrett, T.E., Ponette-González, A.G., 2020. Student cyclists experience River Delta of China. Rem. Sens. Environ. 199, 437–446.
PM2.5 pollution hotspots around an urban university campus. Geogr. Bull. 61 (2). Yeung, C., Tsai, J.M., King, B., Kawagoe, Y., Ho, D., Knight, M.W., Raman, A.P., 2020.
Lyapustin, A., Wang, Y., Korkin, S., Huang, D., 2018. MODIS collection 6 MAIAC Elucidating the behavior of nanophotonic structures through explainable machine
algorithm. Atmos. Meas. Tech. 11 (10), 5741–5765. learning algorithms. ACS Photonics 7 (8), 2309–2318.
Lynch, J., Smith, G.D., Harper, S.A., Hillemeier, M., Ross, N., Kaplan, G.A., Wolfson, M., Yu, X., Wong, M.S., Liu, C.H., Zhu, R., 2022. Synergistic data fusion of satellite
2004. Is income inequality a determinant of population health? Part 1. A systematic observations and in-situ measurements for hourly PM2.5 estimation based on
review. Milbank Q. 82 (1), 5–99. hierarchical geospatial long short-term memory. Atmos. Environ. 286, 119257.
Ma, Q., Su, Y., Guo, Q., 2017. Comparison of canopy cover estimations from airborne Zhang, X.Y., Chu, Y.Y., Wang, Y.X., Zhang, K., 2018. Predicting daily PM2.5
LiDAR, aerial imagery, and satellite imagery. IEEE J. Sel. Top. Appl. Earth Obs. Rem. concentrations in Texas using high-resolution satellite aerosol optical depth. Sci.
Sens. 10 (9), 4225–4236. Total Environ. 631–632, 904–911. https://doi.org/10.1016/j.
Mamalakis, A., Barnes, E.A., Ebert-Uphoff, I., 2022. Investigating the fidelity of scitotenv.2018.02.255.
explainable artificial intelligence methods for applications of convolutional neural
networks in geoscience. Artificial Intelligence for the Earth Systems 1 (4), e220012.
Mao, F., Khamis, K., Krause, S., Clark, J., Hannah, D.M., 2019. Low-cost environmental
sensor networks: recent advances and future directions. Front. Earth Sci. 7, 221.
Matte, T.D., Ross, Z., Kheirbek, I., Eisl, H., Johnson, S., Gorczynski, J.E., Kass, D.,
Markowitz, S., Pezeshki, G., Clougherty, J.E., 2013. Monitoring intraurban spatial

9
L. Liang et al. Environmental Pollution 331 (2023) 121832

Zhang, H., Zhan, Y., Li, J., Chao, C.Y., Liu, Q., Wang, C., Jia, S., Ma, L., Biswas, P., 2021. Zhong, S., Zhang, K., Bagheri, M., Burken, J.G., Gu, A., Li, B., Ma, X., Marrone, B.L.,
Using Kriging incorporated with wind direction to investigate ground-level PM2.5 Ren, Z.J., Schrier, J., Shi, W., 2021. Machine learning: new ideas and tools in
concentration. Sci. Total Environ. 751, 141813. environmental science and engineering. Environ. Sci. Technol. 55 (19),
Zhang, Y., Chen, G., Myint, S.W., Zhou, Y., Hay, G.J., Vukomanovic, J., Meentemeyer, R. 12741–12754.
K., 2022. UrbanWatch: a 1-meter resolution land cover and land use database for 22
major cities in the United States. Rem. Sens. Environ. 278, 113106.

10