Sie sind auf Seite 1von 12

Remote Sensing of Environment 206 (2018) 72–83

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage:

Satellite-based mapping of daily high-resolution ground PM2.5 in China via T

space-time regression modeling
Qingqing Hea,b, Bo Huanga,b,c,
Department of Geography and Resource Management, The Chinese University of Hong Kong, Hong Kong
Big Data Decision Analytics (BDDA) Research Centre, The Chinese University of Hong Kong, Hong Kong
Institute of Space and Earth Information Science, The Chinese University of Hong Kong, Hong Kong


Keywords: The use of satellite-retrieved aerosol optical depth (AOD) data and statistical modeling provides a promising
Aerosol optical depth approach to estimating PM2.5 concentrations for a large region. However, few studies have conducted high
Geographically and temporally weighted spatial resolution assessments of ground-level PM2.5 for China at the national scale, due to the limitations of
regression (GTWR) high-resolution AOD products. To generate high-resolution PM2.5 for the entirety of mainland China, a combined
Interior point algorithm (IPA)
3 km AOD dataset was produced by blending the newly released 3 km-resolution Moderate Resolution Imaging
Spectroradiometer (MODIS) Dark Target AOD data with the 10 km-resolution MODIS Deep Blue AOD data.
Using this dataset, surface PM2.5 measurements, and ancillary information, a space-time regression model that is
an improved geographically and temporally weighted regression (GTWR) with an interior point algorithm (IPA)-
based efficient mechanism for selecting optimal parameter values, was developed to estimate a large set of daily
PM2.5 concentrations. Comparisons with the popular spatiotemporal models (daily geographically weighted
regression and two-stage model) indicated that the proposed GTWR model, with an R2 of 0.80 in cross-validation
(CV), performs notably better than the two other models, which have an R2 in CV of 0.71 and 0.72, respectively.
The use of the combined 3-km high resolution AOD data was found not only to present detailed particle gra-
dients, but also to enhance model performance (CV R2 is only 0.32 for the non-combined AOD data).
Furthermore, the GTWR's ability to predict PM2.5 for days without PM2.5-AOD paired samples and to generate
historical PM2.5 estimates was demonstrated. As a result, fine-scale PM2.5 maps over China were generated, and
several PM2.5 hotspots were identified. Therefore, it becomes possible to generate daily high-resolution PM2.5
estimates over a large area using GTWR, due to its synergy of spatial and temporal dimensions in the data and its
ability to extend the temporal coverage of PM2.5 observations.

1. Introduction from the sparse network of surface monitoring stations.

Because satellite remote sensing has the capacity to provide data
Fine particulate matter, a complex of solid and liquid particles with large (even global) spatial coverage, satellite-derived aerosol op-
suspended in the air with aerodynamic diameters of 2.5 μm or less tical depth (AOD) information is an alternative method of inferring
(PM2.5), has been associated with adverse effects on public health by ground-level PM2.5 concentrations (Engel-Cox et al., 2004; Wang and
many epidemiological studies (Kampa and Castanas, 2008; Madrigano Christopher, 2003). To better investigate PM2.5 exposure for air pollu-
et al., 2013; Pope and Dockery, 2006). The rapid urbanization and in- tion assessment and epidemiological research, satellite-derived PM2.5
dustrialization of China has led to PM2.5 becoming a dominant factor in concentrations at high spatial resolution are needed. High-resolution
air pollution, especially over urban areas, and thus an unprecedented studies in this research field have been increasingly carried out in North
issue of public concern (Bi et al., 2014; Che et al., 2014; Guo et al., America using 1-km Multiangle Implementation of Atmospheric Cor-
2009). Given the advent of severe PM2.5 levels in China, it is urgent to rection (MAIAC) AOD (Hu et al., 2014; Kloog et al., 2014). However,
assess the health effects of PM2.5 exposure across China, using PM2.5 equivalent studies for China are scarce, due to the lack of high-resolu-
data with a high spatial resolution. However, currently, such an as- tion satellite-derived AOD products. Studies on a national scale have
sessment is seriously hindered by the limited measurements available only presented PM2.5 variations with resolutions ranging from 10 to

Corresponding author at: Department of Geography and Resource Management, The Chinese University of Hong Kong, Hong Kong.
E-mail address: (B. Huang).
Co-first author.
Received 13 March 2017; Received in revised form 2 December 2017; Accepted 15 December 2017
0034-4257/ © 2017 Elsevier Inc. All rights reserved.
Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

50 km, employing widely used AOD products, e.g., the Moderate Re- entirety of mainland China, in the setting of a space-time regression
solution Imaging Spectroradiometer (MODIS) Level 2 10 km AOD model using the MODIS 3 km AOD as the major predictor, and me-
product and Multi-angle Imaging SpectroRadiometer (MISR) Level 2 teorological and land use data as ancillary variables. A customized
17.6 km AOD product (Fang et al., 2016; Ma et al., 2014; Ma et al., approach to filling in missing AOD values was devised to improve the
2016). availability of daily MODIS 3 km AOD. Subsequently, the GTWR model
In early 2014, a 3-km aerosol product was included in the new improved with the parameter optimization approach was developed to
MODIS Collection 6 (C6). It is the first daily global aerosol dataset with generate ground PM2.5 concentrations at high resolution, and its in-
a relatively high spatial resolution (Remer et al., 2013). The fine aerosol ferring power also explored comprehensively. A daily GWR model and a
gradients can benefit air quality applications, i.e., improve the perfor- two-stage (LME + GWR) model were also implemented as benchmarks
mance of statistical models and provide details of spatial variation for to examine the extent to which GTWR's simultaneous spatial and tem-
fine particulate matter (Xie et al., 2015). However, the MODIS 3 km poral weighting established a reliable association between surface
AOD data were retrieved using the Dark Target (DT) algorithm that PM2.5 and satellite AOD.
derives AOD over dark surfaces, which may cause a large number of
missing values in the daily AOD image, especially for a large area with a 2. Data collection and preprocessing
complex land surface (He et al., 2017). In other words, it is impossible
to generate a daily PM2.5 prediction at high spatial resolution for the 2.1. Ground-level PM2.5 observations
entirety of China using only the MODIS 3 km AOD product as the pri-
mary predictor (Li et al., 2017; You et al., 2016). Thus, improving the Daily average PM2.5 observations from 1 January 2015 to 31
daily coverage of the 3 km MODIS AOD for ground PM2.5 estimation December 2015 were obtained from the China Environment Monitoring
over a large area with multiple types of land cover is essential. Center ( Ground PM2.5 mass con-
Various models have been developed to elucidate the quantitative centrations of mainland China are measured by the tapered-element
association between satellite-based AOD and surface PM2.5, ranging oscillating microbalance or beta-attenuation method, with calibration
from a simple linear regression model (Engel-Cox et al., 2004; Wang and quality control according to national standards GB3095-2012
and Christopher, 2003) to advanced statistical models such as the linear (China's National Ambient Air Quality Standard, or CNAAQS) (China,
mixed effect (LME) model (Lee et al., 2011; Xie et al., 2015), land use 2012) and HJ 618-2011 (Determination of Atmospheric Articles PM10
regression model (Kloog et al., 2011; Kloog et al., 2012), and geo- and PM2.5 in Ambient Air by Gravimetric Method, available at http://
graphically weighted regression (GWR) model (Hu et al., 2013; Ma At pre-
et al., 2014; Song et al., 2014; Zou et al., 2016). To date, two main sent, the observations come from 1456 monitoring stations in 329 cities
spatial and temporal regression methods have been used to establish (Fig. 1). Most are located in southeastern China, whereas the north-
the PM2.5-AOD relationship. One is the daily GWR model, which ad- western areas have little coverage.
dresses the spatial variability between PM2.5 and AOD for individual
days and achieves reasonable results (Ma et al., 2014; Song et al., 2.2. MODIS AOD data
2014). However, this model ignores the fact that the PM2.5-AOD re-
lationship could vary with time and may be dependent on previous 2.2.1. MODIS 3 km DT AOD products
days. Thus, daily GWR cannot make use of temporal autocorrelation MODIS is a sensor that has been on board NASA Terra since 1999
existing in the data and fails to build a model for days with fewer PM2.5- and Aqua since 2002, providing two daily observations of columnar
AOD samples. The other method is a two-stage model that can account aerosol properties at approximately 10:30 am (Terra) and 1:30 pm
for both spatial and temporal non-stationarities (Hu et al., 2014; Wu (Aqua) local time. The DT algorithm was originally developed to derive
et al., 2016). However, this model calibrates the spatial and temporal aerosol properties over dark surfaces (e.g., dense vegetated areas) at a
PM2.5-AOD relationship separately, i.e., using the LME model to ac- 10 km spatial resolution. To satisfy an identified need for monitoring
count for its temporal variations in the first stage, and then the GWR fine-resolution pollution, aerosol datasets with a 3 km resolution
model its spatial variations in the second stage. Consequently, this (MxD04_3K, x is O for Terra and Y for Aqua) were introduced in the
model does not simultaneously treat spatiotemporal variations of the most recently released MODIS C6 in addition to the standard 10 km
PM2.5-AOD relationship; it is thus not in a better position to predict Level 2 aerosol products (MxD04_L2). The 3 km DT aerosol retrievals
PM2.5 for days with insufficient or no PM2.5 observations. As such, it use a similar protocol to the 10 km DT aerosol products, but organize
becomes difficult to use the two methods to generate high-resolution 6 × 6 pixels into a retrieval box for inversion, rather than 20 × 20,
PM2.5 with high accuracy for a large area when there is a limited after all undesirable ones, i.e., cloud and snow/ice pixels, have been
number of PM2.5-AOD paired samples. screened out and discarded (Remer et al., 2013). MODIS 3 km AOD data
In recent years, the geographically and temporally weighted re- at 550 nm (scientific dataset name: Optical_Depth_Land_And_Ocean) in
gression (GTWR) model proposed by Huang et al. (2010), which in- 2015 (downloaded from were used
tegrates spatial and temporal distance, has gained increasing popularity to model the PM2.5-AOD relationship. Data obtained before 2015 were
in environmental studies (Bai et al., 2016; Chu et al., 2015; Guo et al., used for calibration.
2017). Unlike the two spatiotemporal models already discussed, this
space-time regression model can simultaneously incorporate temporal 2.2.2. MODIS 10 km Deep Blue (DB) AOD products
information from the estimation day or from previous days into the The MODIS DB algorithm aims to generate AOD retrievals over
spatial variability through a spatial-temporal weighting mechanism. bright land (e.g., deserts) to fill the gaps left by the DT algorithm (Hsu
Nevertheless, the predictive power of GTWR for elucidating the re- et al., 2013). Unlike the DT retrieval process, the DB algorithm first
lationship between PM2.5 and AOD, especially for days without sam- produces aerosol properties at a 1 km resolution and then averages the
ples, has yet to be explored. Optimization of parameter values for individual retrievals into a 10 × 10 km grid. As with the 3 km AOD
GTWR is also required to reduce the computational cost for a large data, we calibrated DB aerosol retrievals for 2015 using DB AOD pro-
dataset. As a corollary, there are still considerable challenges con- ducts obtained before 2015, and filled in some gaps where the 3 km DT
fronting a space-time model, such as GTWR, in exploring the PM2.5- AOD values were missing with data obtained in 2015.
AOD relationship, especially over an extensive area with a large but
insufficient sample dataset. 2.2.3. MODIS AOD combination
The objective of this study is to overcome such challenges by pre- In MODIS Collection 6, the 3 km AOD data are retrieved using the
dicting ground PM2.5 at high spatial resolution on a daily basis for the DT algorithm, and the daily availability of 3 km AOD in the desert-like

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

Fig. 1. Geographical distribution of PM2.5 monitoring stations included in the study.

area of northwestern China is very low (even no retrievals at all for Terra resampled-DB AOD datasets. These fitted parameters were
many days). The MODIS DB algorithm was originally developed to then used to calibrate AOD values for the four datasets during 2015
obtain aerosol loading for bright targets, but it only provides the AOD (SI, S1).
product at a 10 km resolution. To address this issue, a four-step cus- (3) Filling in missing values for Aqua and Terra
tomized approach of combining the MODIS 3 km DT and 10 km DB A simple linear regression was performed for the matched grid cells
AOD products was developed to improve the daily coverage of the between Aqua-3 km and Terra-3 km AOD for each day to predict
MODIS 3 km DT AOD, as follows: the missing values (i.e., estimation of the Aqua-3 km AOD using the
available Terra-3 km value, and vice versa). The same protocol was
(1) Re-gridding the 10 km DB AOD applied to the resampled-DB AODs to predict the missing AOD
To match the grid cell size of the 3 km DT AOD, the 10 km DB AOD values for Aqua and Terra (SI, S2).
were resampled to a 3-km spatial resolution grid using the cubic (4) Combining the four datasets
convolution resampling technique in ENVI/IDL5.1. To improve the daily AOD coverage, Aqua/Terra 3 km DT AOD and
(2) Calibrating the MODIS DT/DB AOD Aqua/Terra resampled-DB AOD data were merged together for
According to the global validation of MODIS aerosol products and spatiotemporal modeling. We used the same method as with the
Fig. S2 (Supplementary information [SI]), there are different re- 3 km and resampled-DB AODs to average the Aqua and Terra va-
trieval accuracies for DT and DB AODs and thus systematic bias lues. Because the 3 km DT has higher spatial resolution than the
between the DT and DB AOD values (Remer et al., 2013; Sayer 10 km DB to present finer aerosol gradients, we first averaged the
et al., 2014). Therefore, to optimize the daily AOD values and fill Aqua- and Terra-3 km DT AOD values (both retrievals and pre-
the gaps between DT and DB AOD retrievals, we calibrated AOD dicted values) for each day and then applied the averaged DB data
retrievals for 2015 using ground-level aerosol observations from the resampled to a 3-km grid to fill pixels when the 3 km AOD values
Aerosol Robotic Network (AERONET) from 2001 to 2014 (Ma et al., were missing. That is, the averaged DB-resampled AOD values were
2014). Comparison against AERONET has been widely used to va- applied as supplementary material to fill the gap when the 3 km
lidate satellite-based AOD (Sayer et al., 2014; Tao et al., 2015). A AOD values for both satellites were missing. After the combination,
linear analysis was conducted based on the long-term data before the number of daily AOD values for each pixel increased sig-
2015 to define the relationship between AERONET and satellite nificantly (see Section 5.2 for more details).
AOD. According to previous studies (Li et al., 2009; Remer et al.,
2013), in which the relationship between AERONET and MODIS When compared to the AERONET AOD observations from 2015, our
AODs showed strong seasonality, we fitted linear regressions for combined AOD (fraction of retrievals falling within the expected error
each season. These relationships were also separately established bounds (F) = 56.25% and RMSE = 0.20 for Aqua calibrated and pre-
for the Aqua 3 km DT, Terra 3 km DT, Aqua resampled-DB, and dicted AOD, and F = 55.07% and RMSE = 0.19 for Terra) shows a

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

better accuracy than the MODIS operational 3-km DT AOD (μi, υi) and day ti; and β1–β7 are the location-time-specific slopes for
(F = 46.25%, RMSE = 0.38 for Aqua and F = 49.28%, RMSE = 0.36 combined AOD, aggregated NDVI, interpolated RH, WS and T, averaged
for Terra), indicating the effectiveness of our proposed four-step ap- PBL, and aggregated DEM, respectively. The location (μi, υi) represents
proach. The detailed validation process is presented in S2 (SI). the central coordinates of a grid cell in which sample i is located. ti is
the day of the year, and only the estimation for that day and previous
2.3. Auxiliary data days (ti ≤ testimation) are used for modeling. εi is the error term for
sample i.
A number of meteorological and land-cover-related variables were To estimate the intercept of β0(μi, υi, ti) and the slopes of βi(μi, υi, ti)
used in this study. The meteorological variables were ground-based for each variable, the weight matrix W is introduced to account for the
relative humidity (RH), temperature (T), wind speed (WS), pressure (P), importance of sample j to the estimated parameters of sample i in terms
and reanalysis-produced planetary boundary layer (PBL) data; the land- of space and time (Huang et al., 2010). In this study, the widely used
cover-related variables are terrain (DEM) and Normalized Difference Gaussian distance decay-based functions and the combination method
Vegetation Index (NDVI). Details about these variables are given in S3 of spatial and temporal distance proposed in Huang et al. (2010) were
(SI). In addition, the 1 km grid population dataset of 2010 (Fu et al., applied to measure the spatiotemporal weight W. The details are pre-
2010) was used to calculate the population exposure in Section 4.4.1. sented in S4 (SI).

2.4. Data integration 3.1.2. Exploratory analysis of independent variables

Because of the relatively weak relationship between AOD and grid-
As the aim of this study was to predict daily average PM2.5 con- collocated PM2.5 across China (correlation value of 0.43), meteor-
centrations at 3 × 3 km spatial resolution, a 3-km grid was created ological and land-use data were incorporated into the GTWR model as
(1,073,351 grid cells in total) and exactly matched to the combined covariates to improve the spatial consistency of the PM2.5-AOD re-
AOD pixels. The PM2.5 observations and other variables were re- lationships (SI, S3). However, if all parameters mentioned in Section 2
processed with a consistent spatial size and temporal interval for in- were used as independent variables, the result would exist a collinearity
tegration, because these data vary in spatial and temporal resolution. problem. Variance inflation factors (VIF) were calculated as the colli-
PM2.5 measurements from multiple stations located within a 3-km grid nearity diagnostic for the independent variables (combined AOD, NDVI,
were averaged. The meteorological station data for each day were in- RH, WS, T, P, PBL and DEM). The results demonstrated that the VIF
terpolated to 3 km spatial resolution using an inverse distance weighted values of pressure and DEM were very high (VIF > 10), which in-
(IDW) technique, using the Spatial Analyst Tools in ArcGIS 10.2. dicates a strong collinearity problem if all eight independent variables
Corresponding to the daily average PM2.5 measurements, the daily are included in the PM2.5 prediction model (SI, S5). To address this
mean values obtained from averaging 3-hour PBL data were calculated. issue, we excluded pressure and kept the seven other variables to fit the
The two land-use datasets were separately aggregated to the 3 km grid final GTWR model (the equation is given in Eq. (1)).
cells. The final result was 60,810 matched samples incorporating
ground PM2.5 and all independent variables, which were distributed 3.1.3. Parameter selection and two GTWR models
over 325 days. The spatiotemporal bandwidth hST and scale factor ρ are the two key
parameters in GTWR modeling (Huang et al., 2010). In the original
3. Development and validation GTWR (Huang et al., 2010), the popular method of plotting the sum of
squared error against the parameters hST and ρ through cross-validation
Over the large area of mainland China, there is considerable spatial (Fotheringham et al., 2002), i.e., calculating the scores of CVRSS(hST; ρ)
variability in the relationship between AOD and PM2.5. In addition, for hST and ρ (Eq. (S2) in SI), was used to select the two optimal values.
temporal variation significantly affects this relationship, and PM2.5 le- However, this method is inappropriate for GTWR modeling over a large
vels on previous days may affect the concentrations on the day of area, because it requires the use of a nested-loop, which can be easily
measurement to a certain degree. Therefore, we developed a space-time implemented but consumes a large amount of computation time in
regression model, the GTWR model, to simultaneously address spatial search of the two optimal key parameter values (hST and ρ) for our large
heterogeneity and temporal influence on the PM2.5-AOD relationship. dataset. Selecting the best hST and ρ is to efficiently find the minimal
To confirm the better performance of GTWR over other relevant models values for a constrained nonlinear multivariable problem (Huang et al.,
in previous studies, a daily GWR model using the same variables and 2010). Thus, an optimization approach based on the interior point al-
matched samples as the GTWR model and a two-stage (LME + GWR) gorithm (IPA) (Byrd et al., 1999; Waltz et al., 2006) was devised in this
model (Hu et al., 2014) were also implemented for detailed compar- study.
isons. Aided by a self-concordant barrier approach, IPA can expeditiously
handle a nonlinear problem by approximately solving barrier sub-pro-
3.1. GTWR model development blems using a sequential quadratic programming iteration with trust
region techniques (Byrd et al., 1999). Compared with other optimiza-
3.1.1. General structure of the GTWR model tion methods such as the active-set algorithm (Nocedal and Wright,
Unlike the widely used GWR model, which only takes spatial var- 2006), IPA can efficiently handle large, sparse problems, which is es-
iation into account when estimating an empirical relationship sential to a large dataset of PM2.5-AOD pairs. Using the IPA approach to
(Fotheringham et al., 2002), GTWR captures spatiotemporal hetero- determine the two optimal parameter values in this study, the compu-
geneity based on a weighting matrix referencing both spatial and tational time consumed for our sample dataset was reduced by ap-
temporal dimensions. In this study, a GTWR model was fitted using the proximately 60%, from about 2.5 days for the previous nested-loop
following structure: method to about 1 day (SI, S4). In addition, to reduce the model
overfitting problem, a fixed bandwidth regime rather than adaptive
PM2.5i = β0 (μi , υi , ti ) + β1 (μi , υi , ti ) × AODi + β2 (μi , υi , ti ) × NDVIi regime was applied to derive (SI, S4).
+ β3 (μi , υi , ti ) × RHi + β4 (μi , υi , ti ) × WSi + β5 (μi , υi , ti ) × Ti With the GTWR model for the entire country using a general
bandwidth of 120 km·day, we, however, could not predict robust PM2.5
+ β6 (μi , υi , ti ) × PBLi + β7 (μi , υi , ti ) × DEMi + εi (1)
concentrations extensively in northwestern China for some days due to
where PM2.5i is the daily surface PM2.5 concentration of the sample i at the limited number of matched samples (SI, S6). To stabilize the model
location (μi, υi) on day ti; β0 denotes the intercepts at specific location performance in areas where few matched samples were available, two

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

separate GTWR models were fitted for eastern (E model) and western mean combined AOD was 0.50 with a Std. Dev of 0.35. The mean PM2.5
(W model) China with the same model structure as Eq. (1). This ap- values did not differ greatly between the two areas, but the mean
proach decreased the effect of the lower number of samples for western aerosol loading was much higher in eastern versus western China (0.50
China on model performance without degrading overall accuracy (SI, vs. 0.34). The values of the dependent and independent variables pre-
S6). Therefore, fixed bandwidths of 100 km ∙ day for eastern China and sented different seasonal variability. The maximum mean PM2.5 oc-
290 km ∙ day for the west were applied in the final model. Western curred in winter (86 μg/m3), and the minimum was reached in summer
China was defined as the provinces of Xinjiang, Qinghai, Tibet, and the (44 μg/m3). The mean combined AOD values did not differ much
western parts of Inner Mongolia, Gansu, and Sichuan; eastern China among the four seasons, with the highest in summer (0.56) and the
contained the remaining areas (SI, Fig. S4). A scale factor ρ of 0.2 with lowest in autumn (0.43). The detailed statistics for each season are
the best model performance was obtained. listed in Table S7 (SI).
A 10-fold cross-validation (CV) technique was used to examine The geographical distribution of the annual and seasonal mean
model overfitting (Rodriguez et al., 2010). To evaluate model perfor- PM2.5 measurements and combined AOD is shown in Fig. S5 (SI).
mance, several statistical indicators—R square (R2), root mean square Spatial ground PM2.5 exhibits a descending gradient from northeast to
error (RMSE), and mean prediction error (MPE)—were calculated by southwest due to differences in anthropogenic activities, land use, to-
comparing differences between the estimated PM2.5 concentrations and pography, and meteorological conditions. Generally, high levels of
surface PM2.5 measurements. The Pearson correlation coefficient was PM2.5 concentrations and AOD values occur in densely populated areas
used in this study to quantify the correlation (r). with thriving industries and low elevations (e.g., the North China Plain,
Sichuan Basin, and Central China) and in the arid and semi-arid areas of
3.2. Daily GWR model and two-stage model northwestern China (e.g., the Taklimakan Desert).

To evaluate the performance of GTWR based on its integrated spa- 4.2. Performance evaluation of the GTWR model
tiotemporal weighting scheme, we also fitted daily GWR (D-GWR)
models and two-stage models with the structure as in S7 (SI). The D- Fig. 2 shows the scatter plots of GTWR model fitting and CV results
GWR models separately calibrate the PM2.5-AOD relationship for in- for eastern China, western China, and the entirety of mainland China
dividual days, and could thus be regarded as a special case of GTWR if (the Seg model represents the E + W model, and the non-Seg model is
the number of estimation days equals one. The two-stage model (Hu the GTWR model using all matched samples without segmentation). For
et al., 2014) performs an LME analysis in the first stage and a GWR the model fitting, the respective overall R2 and RMSE values between
analysis in the second stage. the estimated and observed PM2.5 are 0.85 and 15.78 μg/m3 based on
We first attempted two separate D-GWR (East and West) models the results of the E + W model. For model validation, CV R2 and RMSE
using the same partitions as in the Seg-GTWR model, and the results are 0.80 and 18.58 μg/m3, which is a slight improvement compared to
show that the West D-GWR model failed to perform stably due to the the R2 and RMSE of 0.76 and 19.77 μg/m3 of the non-Seg model. The
insufficiency of samples in western China. For the two-stage model, the CV results suggest that the Seg model is slightly overfitted but the non-
results of the West model show a large overfitting problem, with R2 Seg model is significantly overfitted, i.e., CV R2 decreases by 0.05 and
decreasing from 0.82 in model fitting to 0.66 in model validation. 0.11 and RMSE increases by 2.80 and 5.21 μg/m3 for the two models,
Therefore, we only developed East D-GWR and two-stage models for respectively. It is likely that the model itself could be improved by di-
comparison. viding mainland China into two parts, because the Seg model had
comparable performance in model fitting, higher values in the CV
4. Results processes, and much less overfitting than the non-Seg model. Therefore,
we chose to fit two separate models to predict PM2.5 concentrations
4.1. Descriptive statistics over China.
The seasonal results of the Seg-GTWR model in fitting and CV
Table 1 shows the descriptive statistics of the grid-averaged PM2.5 processes based on the E and W models are shown in Fig. 3. It is clear
concentrations in 60,810 grid cells and grid-collocated independent that the model validation performance shows a similar seasonal pattern
variables distributed over the 325 days used in fitting the models. The to that of model fitting. The values of R2 reach their maximum in au-
mean PM2.5 concentration averaged for the entire study region was tumn with 0.85 for model fitting and 0.80 for CV, followed by winter
60 μg/m3 with a standard deviation (Std. Dev) of 41 μg/m3, and the with respective values of 0.84 and 0.77. Correspondingly, a lower R2 for

Table 1
Descriptive statistics of the grid-collocated data for all variables used in the GTWR model.


μg/m 3
– – % m/s °C m m

Whole (N = 60,810) Mean 60 0.50 0.35 57.01 2.22 15.32 650.99 370.49
Std. Dev 41 0.35 0.15 15.06 0.96 8.82 302.16 516.87
Min 1 0.00 0.00 9.39 0.15 − 13.90 20.39 −13.47
Max 673 3.89 0.90 97.00 13.64 34.12 2137.74 4553.23
Median 50 0.43 0.34 59.01 2.03 16.94 602.01 97.49
East (N = 58,707) Mean 60 0.50 0.35 57.58 2.23 15.30 646.02 346.66
Std. Dev 40 0.35 0.15 14.85 0.96 8.84 296.68 492.17
Min 1 0.00 0.00 9.45 0.15 − 13.90 20.39 4.09
Max 595 3.89 0.90 97.00 13.64 33.69 2137.74 3323.93
Median 51 0.43 0.34 59.65 2.03 16.95 598.48 87.58
West (N = 2103) Mean 57 0.34 0.29 40.93 1.97 16.03 789.72 1035.73
Std. Dev 62 0.32 0.11 11.80 0.70 8.22 403.83 710.44
Min 2 0.00 0.08 9.39 0.54 − 8.44 75.84 −13.47
Max 673 2.94 0.67 76.89 6.66 34.12 2095.89 4553.23
Median 37 0.25 0.29 39.60 1.87 16.69 761.45 934.51

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

Counts of points

y=0.81x+11.36 y=0.77x+12.71 y=0.81x+11.50 y=0.83x+9.86

R2=0.85 R2=0.82 R2=0.85 R2=0.87
RMSE=15.28 RMSE=26.07 RMSE=15.78 RMSE=14.56
MPE=10.16 MPE=15.09 MPE=10.33 MPE=9.53
N=58707 N=2103 N=60810 N=60810

(a) (b) (c) (d)

y=0.81x+11.69 y=0.77x+14.31 y=0.81x+11.88 y=0.79x+13.02

R2=0.80 R2=0.72 R2=0.80 R2=0.76
RMSE=18.00 RMSE=30.73 RMSE=18.58 RMSE=19.77
MPE=12.03 MPE=18.79 MPE=12.27 MPE=12.61
N=58707 N=2103 N=60810 N=60810

(e) (f) (g) (h)

Fig. 2. Scatter plots for model fitting and cross validation. (a)–(d) are the model fitting results of the E, W, Seg-GTWR (E + W), and non-Seg GTWR models, respectively. (e)–(h) are the
CV results of the E, W, Seg-GTWR, and non-Seg GTWR models, respectively. The blue line is the linear regression of the scattered dots; the dotted line is the 1-1 line as a reference. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

model fitting and validation are observed in warm seasons (R2 and CV China with higher PM2.5 concentrations than in the less populated areas
R2 are respectively 0.83 and 0.75 in spring and 0.80 and 0.71 in of western China with lower PM2.5 concentrations, which is partly at-
summer). However, contrary to R2, summer is the season with the tributable to the uneven distribution of monitoring stations. These
minimum error, with values of 12.21 and 15.61 μg/m3 for model fitting spatial and temporal variations in the mean local R2 reveal a relatively
and validation and 7.54 and 10.18 μg/m3 for the RMSE and MPE, re- stable performance for the fitted Seg-GTWR model, with some spatio-
spectively. Winter is the season with the highest RMSE values of 20.36 temporal variability for daily local regressions, and demonstrate its
and 24.82 μg/m3 (for model fitting and validation) and MPE values of ability to resolve the large spatiotemporal heterogeneities of the PM2.5-
13.83 and 16.83 μg/m3. The much higher RMSE and MPE values in AOD relationship.
autumn and winter are mainly attributable to the inclusion of more
high PM2.5 values (> 100 μg/m3) during the cold seasons. However, 4.3. Performance comparison between the models
model performance does not differ significantly over the four seasons
(i.e., < 0.04 for R2 in model fitting and 0.09 for R2 in model vali- To comprehensively understand the difference between the GTWR,
dating), which suggests that with simultaneous temporal weighting, the D-GWR, and two-stage models, D-GWR and two-stage models based on
GTWR model incorporating AOD, meteorological parameters, and land- the matched sample dataset from eastern China were implemented. It
use information can capture the spatiotemporal variability of PM2.5 to a was found that only 317 daily GWR models were fitted and models
great degree. could not be constructed for 8 days due to the limited PM2.5-AOD pairs
Following the GWR model, local R2 is also used in GTWR to examine (daily samples < 6) on these days, because D-GWR cannot capitalize
spatiotemporal heterogeneity and the performance of daily local re- on temporal dependency in the data and thus make use of samples
gression models. The results for the Seg-GTWR model are shown in before the estimation days. However, the two-stage and GTWR models
Fig. 4. In general, the overall mean value of local R2 is 0.82. The grid- could generate PM2.5 estimations for all 325 days by taking advantage
averaged local R2 values range from 0.68 to 0.93, and > 95% of the of both spatial and temporal dependency in the data. For this reason,
daily local R2 values are > 0.75. The annual mean local R2 values for the two-stage and GTWR models are much more robust in building the
each grid vary from 0.33 to 0.99, and more than about 91% of the PM2.5-AOD relationship, especially for days without sufficient matched
values are > 0.75. These results indicate that the performance of the samples. It should be noted that the remaining 40 days could not be
daily local models is relatively high. In particular, the spatial distribu- estimated by the two-stage and GTWR models mainly because of the
tion of the mean local R2 shown in Fig. 4 reveals that the values of local unavailability of AOD retrievals due to, e.g., cloud contamination (see
R2 tend to be higher in the densely populated areas of southeastern Section 2.2).

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

Counts of points
y=0.78x+12.01 y=0.75x+10.88 y=0.78x+12.82 y=0.80x+17.53
R2=0.83 R2=0.80 R2=0.85 R2=0.84
RMSE=13.32 RMSE=12.21 RMSE=16.46 RMSE=20.36
MPE=9.22 MPE=7.54 MPE=10.80 MPE=13.83
N=19426 N=11298 N=17187 N=12899

(a) (b) (c) (d)

y=0.78x+12.65 y=0.72x+12.39 y=0.79x+12.26 y=0.79x+18.54

R2=0.75 R2=0.71 R2=0.80 R2=0.77
RMSE=15.72 RMSE=15.61 RMSE=17.88 RMSE=24.82
MPE=10.77 MPE=10.18 MPE=11.90 MPE=16.83
N=19426 N=11298 N=17187 N=12899

(e) (f) (g) (h)

Fig. 3. Scatter plots of Seg-GTWR's model fitting and cross validation. (a)–(d) are the model fitting results for spring, summer, autumn, and winter, respectively. (e)–(h) are the CV results
for spring, summer, autumn, and winter.

To compare the overall performance of the GTWR, D-GWR, and spatiotemporally heterogeneous PM2.5-AOD relationship on a local
two-stage models in detail, the statistics of the results from the three basis. As a result, the GTWR model, with a higher R2 of 0.85 and CV R2
East models are listed in Table 2. The D-GWR model performs the of 0.80, outperforms the two-stage (LME + GWR) model, although the
worst, with a significant overfitting problem and the lowest CV R2 of difference between the model fitting and validation is slightly greater
the three models. For the two-stage model with consideration of day-to- than that generated by the two-stage model. These findings suggest
day variations in the PM2.5-AOD relationship, the first-stage LME model that, despite a slight overfitting problem, the overall estimating per-
with an R2 of 0.63 shows a notable improvement from the estimates formance of GTWR is significantly improved when spatial and temporal
generated by a multivariate linear regression model (R2 of 0.34 in S3 non-stationarities are simultaneously considered.
(SI)). Adding GWR into the modeling process to account for spatial
variability in the PM2.5-AOD relationship improves the performance of 4.4. Predicted PM2.5 concentrations over China
the two-stage model (R2 of 0.75 and CV R2 of 0.72). However, despite a
combined modeling process, the two-stage model does not capture the 4.4.1. General spatiotemporal variation of PM2.5 in China
spatiotemporal heterogeneities simultaneously within the PM2.5-AOD Surface PM2.5 maps for mainland China with a 3-km grid were de-
relationship as does the GTWR model. Through the synergy of spatial rived using the Seg-GTWR model. The overall annual mean of predicted
and temporal dimensions within the data, GTWR can calibrate the PM2.5 concentrations for the entirety of China is 44 μg/m3, which is

(a) (b)
Local R2

Fig. 4. (a) Spatial distribution of the annual mean and (b) temporal variation of spatially averaged local R2 of the Seg-GTWR model for each grid with PM2.5 monitors. Note that we use a
circle rather than a grid cell shape in the left panel because the 3 km grid located in the figure cannot display the variation of color.

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

Table 2
Comparisons of the model performance of D-GWR, two-stage and GTWR.

Model Model fitting Model validation Days can be estimated


GTWR 0.85 15.28 10.16 0.80 18.00 12.03 317

Fit line: y = 0.81x + 11.36 Fit line: y = 0.81x + 11.69
D-GWR 0.83 16.63 11.12 0.71 21.21 13.50 325
Fit line: y = 0.80x + 11.94 Fit line: y = 0.79x + 12.92
Two-stage Stage 1 0.63 24.22 16.91 0.61 24.72 17.35 325
Fit line: y = 0.62x + 22.62 Fit line: y = 0.61x + 23.16
Stage 2 0.75 19.87 13.91 0.72 21.01 14.59
Fit line: y = 0.74x + 15.97 Fit line: y = 0.73x + 16.67

(a) Spring (b) Spring (c) Summer (d) Summer

(e) Autumn (f) Autumn (g) Winter (h) Winter

(i) Annual (j) Annual

Fig. 5. Seasonal and annual mean predicted PM2.5 and surface observed PM2.5 concentrations where gridded AOD is available.

slightly less than the annual mean value of observed PM2.5 (60 μg/m3) northern China, corresponding well with the general distribution of the
because the annual mean of PM2.5 estimates is calculated based on Chinese population (i.e., dense in southeastern China and sparse in
urban, rural, and depopulated zones, but the latter only represents the northwestern China) (Hu, 1935) and inversely with the variation of
average level of urban areas. The difference between the annual pre- terrain in China (i.e., low in the southeast and high in the northwest;
dicted and measured PM2.5 may also result from the difference in the see Fig. 1). Temporally, there were still remarkable seasonal differences
days with available AOD and PM2.5 measurements (due to cloud con- in concentrations, with the most polluted season during 2015 being
tamination, the AOD images on many days do not have enough re- winter (seasonal mean value = 67 μg/m3) and the cleanest being
trievals for some areas). Fig. 5 shows the predicted maps of seasonal summer (35 μg/m3), corresponding to the seasonal variation revealed
and annual mean PM2.5 concentrations. For comparison, the ground- by Ma et al. (2014) and Zhang and Cao (2015). Winter heating and
measured PM2.5 concentrations for areas in which the gridded AOD is unfavorable weather for fine particle dispersion contribute substantially
available are also shown in Fig. 5, and the spatial and temporal patterns to the wintertime maximum (Zhang and Cao, 2015).
of both satellite-derived and ground-observed PM2.5 concentrations are It is clear that over 70% of China, corresponding to > 1.25 billion
in accordance. Fig. S6 (SI) also shows good agreement between the or 92% of the Chinese population, comprises areas with risky levels of
satellite-derived and measured PM2.5 for all days, because the differ- PM2.5 pollution according to the CNAAQS Level 2 standard (an annual
ences for ~80% of monitoring stations fall within ± 10 μg/m3. These mean value of PM2.5 concentration exceeding 35 μg/m3). Over 99% of
findings confirm that GTWR is a promising modeling method for ex- China suffers from particle pollution exceeding the CNAAQS Level 1
ploring population exposure to PM2.5. standard of 15 μg/m3.
Spatially, the general distribution of annual PM2.5 concentrations
demonstrates that high PM2.5 usually occurred in the economically and 4.4.2. PM2.5 hotspots
industrially developed areas of eastern and southern China, whereas Particularly heavily polluted areas are clustered in several centers
low PM2.5 occurred in the rural and less-developed areas of western and (Fig. 6). The North China Plain, including Beijing, Tianjin, Hebei,

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

PM2.5 (μg/m3)


Fig. 6. Zoom-in map of the predicted annual mean PM2.5 for pollution hotspots: (a) the North China Plain, (b) Sichuan Basin, and (c) Central China.

Henan, Anhui, Shandong, and Jiangsu provinces, is one of the most level of PM2.5 compared with 2013 levels.
severely particle-polluted areas in China. The annual mean PM2.5 esti-
mate is 62 μg/m3, and mean values for various subareas generally range 5. Discussion
from 20 to 104 μg/m3. The heavy pollution is mainly attributable to
activities related to rapid economic development and accelerated ur- 5.1. Predictive power of the GTWR model
banization and is aggravated by the unfavorable topography of the
Taihang mountains located in the west of Hebei province, which pre- Comparisons between the GTWR, D-GWR, and two-stage models
vent fine particle dispersion (Fu et al., 2014; Quan et al., 2011; Tao show that GTWR outperforms the latter two models in expressing the
et al., 2012). The Taklimakan Desert, located in the southern Xinjiang PM2.5-AOD relationship for days with PM2.5-AOD paired samples. Given
Autonomous Region, is another heavily polluted region. The annual that GTWR involves simultaneous consideration of spatiotemporal ef-
mean PM2.5 level of this desert (the largest in China) is generally 62 μg/ fects within the data, it is reasonable to assume that this model is
m3, with a maximum of 88 μg/m3 and minimum of 38 μg/m3, due to capable of predicting PM2.5 concentrations for some days without
the large amount of airborne dust and sand originating from the desert samples. Therefore, we carried out two separate 10-fold CVs that ran-
surface. Central China, including Hubei and Hunan provinces, is an- domly dropped 10% of monitoring stations (space CV) and days (time
other hotspot with high pollution, with an annual mean predicted PM2.5 CV) for East China to examine the predictive power of GTWR.
of 55 μg/m3, which is related to low elevation, abundant bodies of Compared to the model fitting results in Fig. 2(a), the CV results in
water, thriving industries, and intense human activities (e.g., fossil fuel Fig. 7(a–b) suggest that some overfitting exists when the CV technique
combustion) (Guo et al., 2012; Wang et al., 2015). The Sichuan Basin, is changed. However, because the distributions of PM2.5-AOD paired
which contains Chongqing and eastern Sichuan province, often suffers samples differ from day to day, the space and time CV not only examine
from episodes of heavy air pollution with an annual mean value of overfitting in space and time separately, but also reveal the predictive
49 μg/m3, which are related to stagnant air circulation and high power of GTWR modeling, especially for time CV. Despite the presence
mountains at its margins (Li et al., 2015). Influenced by dust storms of overfitting in time CV, GTWR can explain 58% of variation in daily
from the Taklimakan Desert (Han et al., 2009), the northern Tibetan ground PM2.5 for days without samples (Fig. 7(b)), which indicates that
Plateau has annual mean values of 25–55 μg/m3, much higher than the GTWR model has a better predictive ability than other statistical
those in the southern part. The areas with low PM2.5 concentrations are models (i.e., generating relatively reliable PM2.5 predictions for days
located mainly in less densely populated areas, including northern without PM2.5 observations).
Xinjiang, western Sichuan, southern Tibet, eastern Inner Mongolia, Furthermore, we collected the required data in 2014 and processed
Hainan, and Fujian, with PM2.5 values generally lower than 35 μg/m3. them using the methods presented in Section 2 to examine GTWR's
According to Ma et al. (2014), the general spatial pattern of PM2.5 ability to predict historical PM2.5 concentrations. The relationships
pollution in China is similar, except that Central China is becoming a built in 2015 were used to predict PM2.5 in 2014 with the same as-
PM2.5 hotspot, and the northern Tibetan Plateau has a relatively high sumption as in Ma et al. (2016) that the daily PM2.5-AOD relationship is

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

Counts of points
y=0.79x+12.38 y=0.61x+25.13 y=0.54x+25.84
R2=0.75 R2=0.58 R2=0.47
RMSE=20.73 RMSE=28.24 RMSE=37.57
MPE=13.31 MPE=19.31 MPE=24.51
N=58707 N=58707 N=25601

(a) (b) (c)

Fig. 7. Scatter plots of model CV results for eastern China: (a) 10-fold CV results with 10% of monitoring sites dropped; (b) 10-fold CV results with 10% of days dropped; (c) evaluation of
historical PM2.5 estimations in 2014 on a daily basis.

consistent for the same day of the preceding year. The results in availability of satellite-derived AOD, because model performance can
Fig. 7(c) show an improved accuracy at higher spatial resolution as be substantially improved with greater inclusion of PM2.5-AOD pairs for
compared to the study of Ma et al. (2016) at a coarse scale (with CV R2 GTWR modeling. As a result, the combined AOD not only presents
of 0.41 for the daily mean of historical PM2.5 predictions in China). better spatial details of PM2.5 (compare Fig. S9(a) and (b)) and makes it
These findings reveal that GTWR can extend the temporal coverage of possible to generate high-resolution PM2.5 for the entirety of China, but
PM2.5 observations and holds great potential in predicting historical also enhances the model performance of GTWR.
PM2.5 concentrations.
In addition, the model performance of GTWR is affected by the in- 5.3. Comparisons with previous studies
tegration method of spatial and temporal distances. Fig. S7 (SI) shows
that the CV R2 values vary with the scale factor ρ. Therefore, in addition In previous studies, satellite-based ground PM2.5 estimates for the
to the widely examined parameter of bandwidth in GWR, the scale entirety of China could only reveal coarse gradients, with grid sizes
factor related to the metric system of spatiotemporal distance is also a of > 10 km (Fang et al., 2016; Li et al., 2017; Ma et al., 2014; Ma et al.,
key parameter for GTWR. 2016; You et al., 2016). Using the newly released MODIS 3-km AOD
data, Li et al. (2017) and You et al. (2016) generated ground PM2.5
5.2. Effect of combined AOD predictions at a coarse resolution (≥ 10 km). Our study has shown how
the combined AOD dataset could improve spatial coverage for each day
The coverage of our combined AOD data is improved (Fig. S8, SI). A and maintain the spatial details of the MODIS 3 km AOD, directly
comparison of Fig. S8(a–c) (SI) shows that merging 3 km DT AOD by benefiting the generation of high-resolution PM2.5 predictions on a
averaging Aqua and Terra AOD values (including the retrievals and daily basis.
predicted values) during the whole year (Fig. S8(c), SI) increased data By considering the estimating power of modeling daily PM2.5-AOD
availability temporally by 10%–50% over the original 3-km Aqua/Terra relationships, our Seg-GTWR model accounts for 80% of the daily
data (Fig. S8(a–b), SI), and the improvements were mainly located over variations in the 10-fold model validation process for high-resolution
eastern areas. Correspondingly, the matched PM2.5-AOD samples in- PM2.5 modeling, surpassing the performance of previous models, such
creased from 2917 (for the AOD dataset that only fused Aqua and Terra as the daily GWR models developed for coarse-resolution modeling over
aerosol retrievals) to 31,118 (for the merged 3 km AOD dataset that the entirety of China by Ma et al. (2014) with a CV R2 of 0.64 and You
fused the Aqua and Terra retrieved and predicted values). et al. (2016) with a CV R2 of 0.79, and the mixed effects model (CV
Subsequently, daily coverage of the combined AOD further improved R2 = 0.79) (Xie et al., 2015) and two-stage spatiotemporal model (CV
after the combination step (Step 4 in Section 2.2.3) of consolidating the R2 = 0.72) (Wu et al., 2016) developed for fine-resolution PM2.5
merged 3 km DT AOD with the resampled-DB AOD (Fig. S8(d), SI). mapping over small regions. In recent years, Bai et al. (2016) and Guo
Comparison with Fig. S8(c) (SI) shows that the sampling frequency for et al. (2017) used the GTWR model to estimate PM2.5 concentrations,
densely populated eastern and southern China increased by 10%–80% but restricted their study areas to small regions and did not tap the
(mean value ~ 35%). The data availability for each pixel over the arid possibility of optimizing GTWR parameters or examine its estimating/
and semi-arid areas of northwestern China showed the most improve- predictive power in elucidating the PM2.5-AOD relationship for days
ment, from null value to ~150 valid values during the whole year; the without samples or in historical years. In addition, the GTWR model
Tibetan plateau also showed significant improvement (20%–300%). developed in the present study also achieved a performance comparable
Lastly, the samples for GTWR modeling increased to 60,810. to those achieved in the U.S. using MODIS 1 km AOD retrieved with the
According to our statistics, the model using the dataset of 2917 MAIAC algorithm (CV R2 = 0.62–0.84) (Chudnovsky et al., 2014 and
samples showed poor performance (R2 = 0.59, CV R2 = 0.32), in- Hu et al., 2014).
dicating that it is impossible to estimate PM2.5 for the entirety of China
using the original MODIS AOD retrievals without coverage improve- 6. Conclusion
ment. However, the model performance (R2 = 0.77, CV R2 = 0.60)
showed an improvement using the dataset of 31,118 samples compared Despite the increasing number of studies of satellite-derived PM2.5
with the results of the model using the dataset of 2917 samples, but still for China, there is still little detailed information on the spatial pattern
performed much worse than the GTWR model using the combined AOD of fine particles due to the limitations of the spatial resolution of sa-
dataset. These findings indicate that it is necessary to improve the data tellite AOD. In using combined high-resolution MODIS AOD data, this

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

study is the first to assess high-resolution PM2.5 exposures across the Che, H., Xia, X., Zhu, J., Li, Z., Dubovik, O., Holben, B., Goloub, P., Chen, H., Estelles, V.,
entirety of China on a daily basis using a spatiotemporal modeling Cuevas-Agulló, E., 2014. Column aerosol optical properties and aerosol radiative
forcing during a serious haze-fog month over North China Plain in 2013 based on
method. ground-based sunphotometer measurements. Atmos. Chem. Phys. 14, 2125–2138.
A space-time regression model was developed to generate high-re- China, M., 2012. Ambient Air Quality Standards. GB 3095-2012. China Environmental.
solution PM2.5 with high accuracy even with limited PM2.5-AOD mat- Science Press, Beijing.
Chu, H.-J., Huang, B., Lin, C.-Y., 2015. Modeling the spatio-temporal heterogeneity in the
ched samples. The advantage of the GTWR model developed in this PM10-PM2.5 relationship. Atmos. Environ. 102, 176–182.
project is that it calibrates AOD-retrieved PM2.5 models using the clo- Chudnovsky, A.A., Koutrakis, P., Kloog, I., Melly, S., Nordio, F., Lyapustin, A., Wang, Y.,
sest days and simultaneously accounts for the space-varying PM2.5-AOD Schwartz, J., 2014. Fine particulate matter predictions using high resolution Aerosol
Optical Depth (AOD) retrievals. Atmos. Environ. 89, 189–198.
relationship. To reduce the computational cost for a large dataset, we Engel-Cox, J.A., Holloman, C.H., Coutant, B.W., Hoff, R.M., 2004. Qualitative and
devised an optimization approach to selecting optimal parameter values quantitative evaluation of MODIS satellite sensor data for regional and urban scale air
rather than using a simple method as in the original GTWR modeling quality. Atmos. Environ. 38, 2495–2509.
Fang, X., Zou, B., Liu, X., Sternberg, T., Zhai, L., 2016. Satellite-based ground PM 2.5
process. Comparisons with widely-used spatiotemporal models (daily
estimation using timely structure adaptive modeling. Remote Sens. Environ. 186,
GWR and two-stage models) were also carried out, and the GTWR 152–163.
model was found to perform significantly better than other models. Fotheringham, A.S., Brunsdon, C., Charlton, M., 2002. Geographically Weighted
Making use of the temporal dependency existing in the data, the GTWR Regression: The Analysis of Spatially Varying Relationships. Wiley, New York.
Fu, J., Jiang, D., Huang, Y., 2010. KM Grid Population Dataset of China
model can generate PM2.5 estimations for less sampled days to over- (PopulationGrid_China). Global Change Research Data Publishing & Repository.
come the model construction failure when the daily GWR model is used. Fu, G., Xu, W., Yang, R., Li, J., Zhao, C., 2014. The distribution and trends of fog and haze
With its simultaneous consideration of spatial and temporal non-sta- in the North China Plain over the past 30 years. Atmos. Chem. Phys. 14,
tionarities within the PM2.5-AOD relationship, the GTWR model also Guo, J.-P., Zhang, X.-Y., Che, H.-Z., Gong, S.-L., An, X., Cao, C.-X., Guang, J., Zhang, H.,
outperforms the two-stage (LME + GWR) model by improving the Wang, Y.-Q., Zhang, X.-C., 2009. Correlation between PM concentrations and aerosol
overall accuracy with an increased CV R2 of 0.80 at high spatial re- optical depth in eastern China. Atmos. Environ. 43, 5876–5886.
Guo, Y., Hong, S., Feng, N., Zhuang, Y., Zhang, L., 2012. Spatial distributions and tem-
solution. Furthermore, we found that GTWR modeling improves on poral variations of atmospheric aerosols and the affecting factors: a case study for a
other spatiotemporal statistical models in another way: it can extend region in central China. Int. J. Remote Sens. 33, 3672–3692.
the temporal coverage of measured PM2.5 by predicting relatively re- Guo, Y., Tang, Q., Gong, D.Y., Zhang, Z., 2017. Estimating ground-level PM 2.5 con-
centrations in Beijing using a satellite-based geographically and temporally weighted
liable PM2.5 concentrations for days without PM2.5 measurements, and regression model. Remote Sens. Environ. 198, 140–149.
has great potential to generate historical PM2.5 concentrations. Thus, Han, Y., Fang, X., Zhao, T., Bai, H., Kang, S., Song, L., 2009. Suppression of precipitation
GTWR is a promising technique for exploring population exposure to by dust particles originated in the Tibetan Plateau. Atmos. Environ. 43, 568–574.
He, Q., Zhang, M., Huang, B., Tong, X., 2017. MODIS 3 km and 10 km aerosol optical
PM2.5, especially for generating high-resolution data on PM2.5 exposure
depth for China: evaluation and comparison. Atmos. Environ. 153, 150–162.
over an extensive area. Hsu, N., Jeong, M.J., Bettenhausen, C., Sayer, A., Hansell, R., Seftor, C., Huang, J., Tsay,
In addition, the spatial resolution of ground PM2.5 data for China is S.C., 2013. Enhanced Deep Blue aerosol retrieval algorithm: the second generation. J.
improved to a fine scale of 3 km resolution, which can be attributed to Geophys. Res.-Atmos. 118, 9296–9315.
Hu, Y., 1935. The distribution of population in China, with statistics and maps. Acta
improvement in daily coverage of the combined 3-km AOD through the Geograph. Sin. 2 (2), 33–72 (In Chinese).
fusion of MODIS 3-km DT and 10-km DB AODs. This improvement is of Hu, X., Waller, L.A., Al-Hamdan, M.Z., Crosson, W.L., Estes, M.G., Estes, S.M., Quattrochi,
significant value, especially for high resolution epidemiological studies. D.A., Sarnat, J.A., Liu, Y., 2013. Estimating ground-level PM2.5 concentrations in the
southeastern U.S. using geographically weighted regression. Environ. Res. 121, 1–10.
In addition, the combined AOD presents better spatial details of PM2.5 Hu, X., Waller, L.A., Lyapustin, A., Wang, Y., Al-Hamdan, M.Z., Crosson, W.L., Estes,
and enhances the performance of GTWR with the inclusion of more M.G., Estes, S.M., Quattrochi, D.A., Puttaswamy, S.J., Liu, Y., 2014. Estimating
samples for modeling. However, the segmentation of China in this study ground-level PM2.5 concentrations in the Southeastern United States using MAIAC
AOD retrievals and a two-stage model. Remote Sens. Environ. 140, 220–232.
(i.e., dividing East and West) is not only easy to implement but also Huang, B., Wu, B., Barry, M., 2010. Geographically and temporally weighted regression
worth further examination; it may be possible to optimize the partition for modeling spatio-temporal variation in house prices. Int. J. Geogr. Inf. Sci. 24,
approach using an advanced spatial statistical method. In addition, it is 383–401.
Kampa, M., Castanas, E., 2008. Human health effects of air pollution. Environ. Pollut.
possible to improve the method for deriving high-resolution AOD by
151, 362–367.
making use of a higher sampling frequency. Kloog, I., Koutrakis, P., Coull, B.A., Lee, H.J., Schwartz, J., 2011. Assessing temporally
and spatially resolved PM2.5 exposures for epidemiological studies using satellite
aerosol optical depth measurements. Atmos. Environ. 45, 6267–6275.
Kloog, I., Nordio, F., Coull, B.A., Schwartz, J., 2012. Incorporating local land use re-
gression and satellite aerosol optical depth in a hybrid model of spatiotemporal PM2.
This study is supported by National Natural Science Foundation of 5 exposures in the Mid-Atlantic states. Environ. Sci. Technol. 46, 11913–11921.
China ( 41371417), National Key R&D Program of China Kloog, I., Chudnovsky, A.A., Just, A.C., Nordio, F., Koutrakis, P., Coull, B.A., ... Schwartz,
J., 2014. A new hybrid spatio-temporal model for estimating daily multi-year PM 2.5
(2017YFB0503605), and Research Grants Council of Hong Kong concentrations across northeastern USA using high resolution aerosol optical depth
(14652016). We also acknowledge Qingyue Open Environmental Data data. Atmos. Environ. 95, 581–590.
Center ( for providing part of the air pollution Lee, H.J., Liu, Y., Coull, B.A., Schwartz, J., Koutrakis, P., 2011. A novel calibration ap-
proach of MODIS AOD data to predict PM2.5 concentrations. Atmos. Chem. Phys. 11,
data. 7991–8002.
Li, X., Zhang, P., Zhang, X., Sun, L., Qi, J., Zhang, Y., 2009. Validation of aerosol optical
Appendix A. Supplementary information thickness product over China with MODIS data operated at NSMC. J. Appl. Meteorol.
Sci. 20 (2), 147–156 (In Chinese).
Li, Y., Chen, Q., Zhao, H., Wang, L., Tao, R., 2015. Variations in PM10, PM2.5 and PM1.0
Supplementary data to this article can be found online at https:// in an urban area of the Sichuan Basin and their relation to meteorological factors. Atmosphere 6, 150–163.
Li, T., Shen, H., Zeng, C., Yuan, Q., Zhang, L., 2017. Point-surface fusion of station
measurements and satellite observations for mapping PM 2.5 distribution in China:
References methods and assessment. Atmos. Environ. 152, 477–489.
Ma, Z., Hu, X., Huang, L., Bi, J., Liu, Y., 2014. Estimating ground-level PM2.5 in China
using satellite remote sensing. Environ. Sci. Technol. 48, 7436–7444.
Bai, Y., Wu, L., Qin, K., Zhang, Y., Shen, Y., Zhou, Y., 2016. A geographically and tem-
Ma, Z., Hu, X., Sayer, A.M., Levy, R., Zhang, Q., Xue, Y., Tong, S., Bi, J., Huang, L., Liu, Y.,
porally weighted regression model for ground-level PM2. 5 Estimation from satellite-
2016. Satellite-based spatiotemporal trends in PM2.5 concentrations: China,
derived 500 m resolution AOD. Remote Sens. 8, 262.
2004–2013. Environ. Health Perspect. 124, 184–192.
Bi, J., Huang, J., Hu, Z., Holben, B., Guo, Z., 2014. Investigating the aerosol optical and
Madrigano, J., Kloog, I., Goldberg, R., Coull, B.A., Mittleman, M.A., Schwartz, J., 2013.
radiative characteristics of heavy haze episodes in Beijing during January of 2013. J.
Long-term exposure to PM2.5 and incidence of acute myocardial infarction. Environ.
Geophys. Res.-Atmos. 119, 9884–9900.
Health Perspect. (Online) 121, 192.
Byrd, R.H., Hribar, M.E., Nocedal, J., 1999. An interior point algorithm for large-scale
Nocedal, J., Wright, S.J., 2006. Numerical Optimization, 2nd ed. Springer-Verlag, Berlin,
nonlinear programming. SIAM J. Optim. 9 (4), 877–900.
New York (ISBN 978-0-387-30303-1).

Q. He, B. Huang Remote Sensing of Environment 206 (2018) 72–83

Pope, C.A., Dockery, D.W., 2006. Health effects of fine particulate air pollution: lines that nonlinear optimization that combines line search and trust region steps. Math.
connect. J. Air Waste Manage. Assoc. 56, 709–742. Program. 107 (3), 391–408.
Quan, J., Zhang, Q., He, H., Liu, J., Huang, M., Jin, H., 2011. Analysis of the formation of Wang, J., Christopher, S.A., 2003. Intercomparison between satellite-derived aerosol
fog and haze in North China Plain (NCP). Atmos. Chem. Phys. 11, 8205–8214. optical thickness and PM2.5 mass: implications for air quality studies. Geophys. Res.
Remer, L., Mattoo, S., Levy, R., Munchak, L., 2013. MODIS 3 km aerosol product: algo- Lett. 30.
rithm and global perspective. Atmos. Meas. Tech. 6, 1829–1844. Wang, L., Gong, W., Xia, X., Zhu, J., Li, J., Zhu, Z., 2015. Long-term observations of
Rodriguez, J.D., Perez, A., Lozano, J.A., 2010. Sensitivity analysis of k-fold cross vali- aerosol optical properties at Wuhan, an urban site in Central China. Atmos. Environ.
dation in prediction error estimation, IEEE transactions on pattern analysis and 101, 94–102.
machine intelligence. 32 (3), 569–575. Wu, J., Yao, F., Li, W., Si, M., 2016. VIIRS-based remote sensing estimation of ground-
Sayer, A., Munchak, L., Hsu, N., Levy, R., Bettenhausen, C., Jeong, M.J., 2014. MODIS level PM 2.5 concentrations in Beijing–Tianjin–Hebei: a spatiotemporal statistical
Collection 6 aerosol products: comparison between Aqua's e-Deep Blue, Dark Target, model. Remote Sens. Environ. 184, 316–328.
and “merged” data sets, and usage recommendations. J. Geophys. Res.-Atmos. 119, Xie, Y., Wang, Y., Zhang, K., Dong, W., Lv, B., Bai, Y., 2015. Daily estimation of ground-
13,965–913,989. level PM2.5 concentrations over Beijing using 3 km resolution MODIS AOD. Environ.
Song, W., Jia, H., Huang, J., Zhang, Y., 2014. A satellite-based geographically weighted Sci. Technol. 49 (20), 12280–12288.
regression model for regional PM 2.5 estimation over the Pearl River Delta region in You, W., Zang, Z., Zhang, L., Li, Y., Pan, X., Wang, W., 2016. National-scale estimates of
China. Remote Sens. Environ. 154, 1–7. ground-level PM2.5 concentration in China using geographically weighted regression
Tao, M., Chen, L., Su, L., Tao, J., 2012. Satellite observation of regional haze pollution based on 3 km resolution MODIS AOD. Remote Sens. 8, 184.
over the North China Plain. J. Geophys. Res.-Atmos. 117 (D12). Zhang, Y.-L., Cao, F., 2015. Fine particulate matter (PM2.5) in China at a city level. Sci.
Tao, M., Chen, L., Wang, Z., Tao, J., Che, H., Wang, X., Wang, Y., 2015. Comparison and Rep. 5.
evaluation of the MODIS Collection 6 aerosol data in China. J. Geophys. Res.-Atmos. Zou, B., Pu, Q., Bilal, M., Weng, Q., Zhai, L., Nichol, J.E., 2016. High-resolution satellite
120, 6992–7005. mapping of fine particulates based on geographically weighted regression. IEEE
Waltz, R.A., Morales, J.L., Nocedal, J., Orban, D., 2006. An interior algorithm for Geosci. Remote Sens. Lett. 13 (4), 495–499.