PE&RS December 2018 Full - page 761

Estimating Spatio-Temporal Variations of PM2.5
Over Hong Kong Using an Improved GTWR Model
and SARA AOD Retrievals
Xin Li and Yongjun Feng
Abstract
The
PM
2.5 (particulate matters with aerodynamic diameter
2.5μm) distribution with large scale and high resolution is
required for the protection of atmospheric environment and
epidemiology studies. In this study, an improved geographi-
cally and temporally weighted regression (
IGTWR
) model
incorporating seasonal characteristics is proposed to estimate
the
PM
2.5 concentrations over Hong Kong. The
IGTWR
model
is established using
MOD
erate resolution Imaging Spectro-
radiometer (
MODIS
) aerosol optical depth (
AOD
) derived by
SARA
(simplified aerosol retrieval algorithm), meteorologi-
cal data and land use data. A comparison among ordinary
least squares (
OLS
), geographically weighted regression (
GWR
),
geographically and temporally weighted regression (
GTWR
)
and the proposed
IGTWR
indicate that the
IGTWR
model has
the best prediction performance for
PM
2.5 concentrations
with highest R2 (0.752) and smallest
RMSE
(9.651μg/m3). The
annual mean
PM
2.5 distribution at 500 m resolution over
Hong Kong from 2012 to 2014 is derived using
IGTWR
, dem-
onstrating its potential of satellite-based
PM
2.5 monitoring.
Introduction
In recent decades, atmospheric pollution has become more
serious due to the rapid development of industry and urban-
ization. As a main source of air pollution, the particulate mat-
ters with aerodynamic diameter less than 2.5 μm (
PM2.5
) not
only reduce the air visibility but also have adverse impacts on
human health, leading to respiratory disease, cardiovascular
disease, and even death (Ghosh
et al.,
2012; Liu
et al.,
2009;
Peng
et al.,
2009; Pope
et al.,
2011). Given the urgent demand
for assessing the effects of
PM2.5
exposure and further pollu-
tion prevention, accurate
PM2.5
monitoring and prediction
are essential, particularly for the protection of atmospheric
environment and epidemiology studies.
Conventionally, the ground-based
PM2.5
monitoring is
considered as the most reliable way to retrieve the
PM2.5
concentrations with high accuracy. However, the unevenly
and sparsely spatial distribution of stations cannot satisfy the
requirement of
PM2.5
monitoring and further application in
large scale and fine resolution (Wang
et al.,
2010).
On the contrary, remote sensing technique can predict
ground-level
PM2.5
concentrations at a large coverage with de-
tailed spatial distributions based on satellite-retrieved aerosol
optical depth (
AOD
) data. The
AOD
is a quantitative measure
of aerosol defined as the integrated extinction coefficient over
a vertical column of unit cross section (Malakar
et al.,
2012).
It reflects the air pollution degree and the particles content
in the air. The
AOD
has been demonstrated to have a strong
correlation with
PM2.5
concentrations due to its sensitivity to
particular matters (aerodynamic diameter 0.1-2 µm) (Liu
et
al.,
2009; Mao
et al.,
2012). Currently, the
AOD
products have
been retrieved from several satellite sensors including Geosta-
tionary Operational Environmental Satellite (
GOES
), MEdius
Resolution Imaging Spectraoradiometer (
MERIS
), MODerate
resolution Imaging Spectroradiometer (
MODIS
), etc. (Bilal
et
al.,
2013; Wang
et al.,
2010).
In the past few years, various statistical models have been
developed for the
PM2.5
prediction using satellite
AOD
data.
Mao
et al.
(2012) revised the land-use regression (
LUR
) model
to predict
PM2.5
concentrations in Florida by integrating
AOD
,
land-use data, population, and traffic data. Liu
et al.
(2009)
used the
AOD
data from
GOES
and other meteorological data
to establish a generalized additive model (
GAM
) to predict the
PM2.5
distributions in the USA. Xie
et al.
(2015) considered
the daily variation of the relationship between
AOD
and
PM2.5
and proposed a mixed effect model for
PM2.5
predictions in
Beijing. These studies presume constant correlations between
PM2.5
concentrations and influencing factors, whereas they ig-
nore the fact that the relationships varies over space and time
(Yap and Hashim, 2012).
To capture the variation of
PM2.5
-
AOD
relation over space
and time, some studies incorporate the heterogeneous spatial
and/or temporal effects in
PM2.5
prediction modeling. A
spatiotemporal statistical model was developed by Wu
et al.
(2016) to estimate
PM2.5
concentrations in Beijing-Tianjin-
Hebei using
VIIRS
(Visible Infrared Imaging Radiometer Suite)
AOD
and other dependent variables. Zhan
et al.
(2017) pro-
posed a Geographically-Weighted Gradient Boosting Machine
model to predict the daily
PM2.5
concentrations in China.
As typical localized modeling techniques, geographically
weighted regression (
GWR
) and its extension, geographically
and temporally weighted regression (
GTWR
), are also applied
to estimate
PM2.5
concentrations. You
et al.
(2016) employed
the
GWR
model for predicting
PM2.5
variation in China. Bai
et al.
(2016) derived the
PM2.5
distribution of eastern China
using the
GTWR
model with high resolution
AOD
data and me-
teorological data. Guo
et al.
(2017) applied the
GTWR
model to
estimate
PM2.5
concentrations in Beijing by using 3-km
MODIS
AOD
, meteorological and land-use variables as predictors. The
results showed that
GTWR
model performed better than other
global and local regression models.
However, as shown in previous studies (Chu
et al.
2015;
Li
et al.
2017; Zhang and Cao 2015; Zhao
et al.,
2009),
PM2.5
concentrations exhibit significant seasonal variations, which
has been ignored by the aforementioned
GWR
-based and
GTWR
-based methods. Although some authors attempted to
overcome this limitation by adding a cosine function when
Department of Resources and Environment, Shandong
Agricultural University, No. 61 Daizong Street, Tai’an City,
Shandong Province, China 271018 (
).
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 12, December 2018, pp. 761–769.
0099-1112/18/761–769
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.12.761
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2018
761
743...,751,752,753,754,755,756,757,758,759,760 762,763,764,765,766,767,768,769,770,771,...814
Powered by FlippingBook