the monthly average
PM2.5
concentrations are lowest
in summer (June, July, August) while highest in winter
(December, January, February). The annual trend of
PM2.5
concentrations exhibits significant seasonal-
ity, which is consistent with the previous studies
(Chu
et al.,
2015; Li
et al.,
2017; Shi
et al.,
2012).
The improved
GTWR
model can capture the seasonal
variability to estimate the
PM2.5
more accurately. To
better understand the correlations between
PM2.5
and
all predictors (
AOD
, meteorological and land use data),
the correlation coefficients (r) of all variables are cal-
culated and shown in Table 2. It can be seen that
PM2.5
concentrations have a positive correlation with
AOD
,
wind direction, pressure and road length, whereas
negative correlations are observed between
PM2.5
and
temperature, wind speed,
RH
, and
NDVI
. In addition,
the correlation coefficient of
PM2.5
and
AOD
is 0.619,
showing a stronger correlation over other variables.
Model Fitting
After pairing up
PM2.5
measurements with other
predictors and eliminating the samples with cloud
contaminated
AOD
data and missing
PM2.5
observa-
tions, a total of 1,478 samples (369, 444, 665 samples
for 2012, 2013, 2014, respectively) remain for model
fitting. The bandwidths (
h
ST
) of
GWR
,
GTWR
, and
IGTWR
are determined as 1.358, 0.540, and 0.573, respective-
ly, indicating the different rate of weights decay. The
temporal and seasonal scale factors in
IGTWR
model
are determined to be 121 and 90,000 by
CV
when the
spatial scale factor is set to be 1.
Table 3~6 lists the summary of parameters and the
statistics of model fitting (R
2
and
AIC
). Since
OLS
is a
global model, the parameters are fixed in space and
time. As the parameters estimated by
GWR
-based mod-
els vary geographically and temporally, five columns
of minimum (Min), lower quartile (
LQ
), median (Med),
upper quartile (
UQ
), and maximum (Max) are listed to
show the variations of parameters. From Tables 3 to 6
we can see that the parameters in
OLS
are within the
minimum and the maximum values of those in
GWR
-
based models. Additionally, although the parameters
of
GWR
-based models vary with locations and time,
they have similar magnitude. According to the signs of
all parameters, for most of the data in this study,
AOD
,
pressure, wind direction, and road length have gener-
ally positive relationships with
PM2.5
concentrations,
while temperature,
RH
, wind speed,
NDVI
have negative
impacts on
PM2.5
, which is consistent with the correla-
tions between
PM2.5
and other variables in Table 2.
This can be explained as follows. First,
AOD
reflects
the air pollution level and increases with the
PM2.5
concentrations rising. When pressure is low, the
Table 2. Correlation Coefficients (r) of all variables used in this study.
r
PM 2.5 AOD T
P
RH WD WS NDVI
RL
PM2.5 1
0.619 -0.162 0.119 0.137 0.037 -0.112 -0.059 0.042
AOD 0.619
1
-0.037 0.027 -0.074 0.013 -0.054 -0.023 0.022
T -0.162 -0.037
1
-0.851 0.169 0.101 -0.006 -0.098 0.065
P 0.119 0.027 -0.851
1
-0.250 -0.243 0.103 0.007 -0.002
RH -0.137 -0.074 0.169 -0.250
1
0.102 -0.142 0.058 -0.039
WD 0.037 0.013 0.101 -0.243 0.102
1
-0.134 0.172 -0.265
WS -0.112 -0.054 -0.006 0.103 -0.142 -0.134
1
-0.196 0.142
NDVI
-0.059 -0.023 -0.098 0.007 0.058 0.172 -0.196
1
-0.662
RL 0.042 0.022 0.065 -0.002 -0.039 -0.265 0.142 -0.662
1
Table 4. Summary of
GWR
model parameter estimation
(Bandwidth=1.358).
Parameter
Min
LQ
Med
UQ
Max
Intercept
155.831 186.945 221.816 305.225 384.595
AOD
72.188
72.575 73.195 73.424 74.624
Temperature
-0.845
-0.747
-0.638
-0.602
-0.575
Pressure
-0.339
-0.265
0.114
0.147
0.181
RH
-0.138
-0.115
-0.110
-0.106
-0.084
Wind speed -0.473
-0.396
-0.371
-0.362
-0.337
Wind direction 0.003
0.004
0.004
0.005
0.009
NDVI
-12.305 -11.887 -11.521 -10.440 -7.517
Road length -6.582×10
-5
1.328×10
-5
3.967×10
-5
6.423×10
-5
1.803×10
-4
R
2
0.609
AIC
7389.136
Table 5. The summary of
GTWR
model parameter estimation
(Bandwidth=0. 540).
Parameter
Min
LQ
Med
UQ
Max
Intercept
-2552.822 -523.203 587.868 859.875 1644.485
AOD
28.685
63.813
70.721 83.488 87.299
Temperature
-3.522
-0.979
-0.569
0.205
1.159
Pressure
-1.575
-0.811
0.548
0.573
2.495
RH
-0.574
-0.081
-0.032
0.137
0.568
Wind speed -1.575
-0.342
-0.208
0.098
0.852
Wind direction -0.038
-0.002
0.008
0.014
0.026
NDVI
-55.116 -14.313 -10.000 -4.955
45.454
Road length -7.906×10
-4
-1.495×10
-4
1.448×10
-5
2.629×10
-4
7.067×10
-4
R
2
0.748
AIC
6737.456
Table 6. The summary of
IGTWR
model parameter estimation
(Bandwidth=0. 573).
Parameter
Min
LQ
Med
UQ
Max
Intercept
-3096.400 -0.101
0.100 879.978 3957.903
AOD
10.311
51.732
69.535 84.526 90.826
Temperature
-8.194
-1.049
-0.195
0.587
5.496
Pressure
-3.808
-0.835
0.031
0.070
3.022
RH
-1.941
-0.205
0.019
0.158
1.253
Wind speed -6.014
-0.582
-0.181
0.074
1.983
Wind direction -0.113
-0.005
0.010
0.016
0.128
NDVI
-50.114 -15.461
-8.514
-2.515
59.712
Road length -9.281×10
-4
-1.522×10
-4
4.052×10
-5
2.971×10
-4
1.1×10
-3
R
2
0.801
AIC
6397.054
Table 3. Summary of
OLS
model
parameter estimation.
Parameter
Values
Intercept
208.813
AOD
73.468
Temperature
-0.669
Pressure
0.168
RH
-0.115
Wind speed -0.371
Wind direction 0.008
NDVI
-9.200
Road length 2.679×10
-5
R
2
0.554
AIC
7584.192
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2018
765
