Sie sind auf Seite 1von 11
Atmospheric Environment 173 (2018) 295–305 Contents lists available at ScienceDirect Atmospheric Environment

Contents lists available at ScienceDirect

Atmospheric Environment

journal homepage: www.elsevier.com/locate/atmosenv High-resolution spatiotemporal mapping of PM 2 . 5

High-resolution spatiotemporal mapping of PM 2.5 concentrations at Mainland China using a combined BME-GWR technique

Lu Xiao a , Yichao Lang a , George Christakos a , b ,

a Institute of Islands and Coastal Ecosystems, Ocean College, Zhejiang University, Zhoushan, China b Department of Geography, San Diego State University, San Diego, CA, USA

T
T

ARTICLE INFO

Keywords:

Particulate matter Space-time analysis Remote sensing Aerosol optical depth Bayesian Maximum Entropy Geographically weighted regression Meteorological elds

ABSTRACT

With rapid economic development, industrialization and urbanization, the ambient air PM 2.5 has become a major pollutant linked to respiratory, heart and lung diseases. In China, PM 2.5 pollution constitutes an extreme environmental and social problem of widespread public concern. In this work we estimate ground-level PM 2.5 from satellite-derived aerosol optical depth (AOD), topography data, meteorological data, and pollutant emis- sion using an integrative technique. In particular, Geographically Weighted Regression (GWR) analysis was combined with Bayesian Maximum Entropy (BME) theory to assess the spatiotemporal characteristics of PM 2.5 exposure in a large region of China and generate informative PM 2.5 space-time predictions (estimates). It was found that, due to its integrative character, the combined BME-GWR method o ers certain improvements in the space-time prediction of PM 2.5 concentrations over China compared to previous techniques. The combined BME- GWR technique generated realistic maps of space-time PM 2.5 distribution, and its performance was superior to that of seven previous studies of satellite-derived PM 2.5 concentrations in China in terms of prediction accuracy. The purely spatial GWR model can only be used at a xed time, whereas the integrative BME-GWR approach accounts for cross space-time dependencies and can predict PM 2.5 concentrations in the composite space-time domain. The 10-fold results of BME-GWR modeling (R 2 = 0.883, RMSE = 11.39 μg / m 3 ) demonstrated a high level of space-time PM 2.5 prediction (estimation) accuracy over China, revealing a denite trend of severe PM 2.5 levels from the northern coast toward inland China (Nov 2015 Feb 2016). Future work should focus on the addition of higher resolution AOD data, developing better satellite-based prediction models, and related air pollutants for space-time PM 2.5 prediction purposes.

1. Introduction

China's atmospheric pollution is very serious. The main pollution sources are fuel combustion, human activities and other natural pro- cesses (e.g., dust). Ambient air pollutants that have a signicant impact on human health and the environment are composed of complex ma- terials, including particulate matter, ozone, sulfur dioxide and nitrogen dioxide. Fine particle particulate matter, PM 2.5 (aerodynamic diameters smaller than 2.5 μ m ), is the most problematic among these pollutants as regards public health. PM 2.5 particles can enter into the alveoli, sub- sequently being retained in the lung parenchyma (Christakos and Hristopulos, 1998; Dockery, 2009 ), thus causing severe heart disease, cardiovascular diseases, respiratory diseases and even lung cancer ( Brook et al., 2010; Dhondt et al., 2011; Hoek et al., 2013 ). With rapid economic development, PM 2.5 pollution in China has become an ex- treme environmental and social problem having an important impact on the human body, the environment and the climate ( Song et al., 2014;

Peng et al., 2016 ). Relatively small PM 2.5 datasets are available in China. Ground- based measurements are considered as the most reliable way of col- lecting PM 2.5 concentrations ( Tao et al., 2017 ). Accordingly, most pollutant concentration information has been obtained from ground- level monitoring sites, a fact that has many limitations. Among them is the considerable information bias, the limited credibility of the ex- posure response results ( Liu et al., 2007 ), and the sparsity and uneven distribution of the monitoring stations ( Gupta and Christopher, 2008 ). These limitations a ect the geographical and demographic range of a study, so that it is usually impossible to determine the temporal and spatial variation of PM 2.5 concentrations over large geographic areas ( You et al., 2016b ), For example, PM 2.5 was not included in China's national monitoring system until 2013 ( Chu et al., 2016 ). So far, do- mestic and foreign scholars have carried out PM 2.5 variation analysis and generated ground level PM 2.5 concentration estimates using a variety of statistical models. Christakos and Serre (2000) and Christakos

Corresponding author. Institute of Islands and Coastal Ecosystems, Ocean College, Zhejiang University, Zhoushan, China. E-mail address: gchristakos@zju.edu.cn (G. Christakos).

Received 11 June 2017; Received in revised form 19 October 2017; Accepted 29 October 2017

Available online 21 November 2017 1352-2310/ © 2017 Elsevier Ltd. All rights reserved.

L. Xiao et al.

et al. (2001) have used Bayesian maximum entropy (BME) theory to represent and predict spatiotemporal particulate matter distributions in North Carolina and California, USA. Wang and Christopher (2003) used linear regression models, whereas Liu et al. (2004) proposed a Chemical Transport Model (CTM). Later, Reid et al. (2015) and Donkelaar et al. (2011) also used the CTM. Lee et al. (2011) developed the day-speci c Mixed-E ect Model (MEM), Lee et al. (2012) used a space-time geos- tatistical kriging model to estimate long-term ambient PM 2.5 exposure in U.S.A. Liu et al. (2009) and Kloog et al. (2011) proposed a two-stage generalized additive model (GAM), and Ma et al. (2014) and ( Xiao et al., 2017 ) used Geographically Weighted Regression (GWR) tools. Hence, remote sensing techniques, spatial and temporal modeling, and statistical prediction theory have been individually or in combination employed in the quantitative assessment of air pollution and environ- mental health ( Kim et al., 2015; Xiao et al., 2017 ). In view of the above considerations, the objective of the present work is to introduce and validate in terms of real data a new satellite- based technique of composite space-time modeling and estimation of PM 2.5 concentrations in China during a four-month period: this is the combined (or integrative) Bayesian Maximum Entropy-Geographically Weighted Regression (BME-GWR) method. Advantages of the combined BME-GWR method include its rigorous consideration of the physical cross-space-time dependencies of pollutant distribution, the generation of pollutant predictions in a realistic space-time domain rather than separately, the inclusion of more general cases (non-linear and non- Gaussian predictors), and its ability to jointly incorporate di erent environmental predictors, including topography data (elevation) and meteorological data (wind speed, precipitation, temperature, relative humidity and pressure), as well as pollutant emission indicators (such as NO 2 , CO, land use, population, and road network information).

2. Data and method

2.1. Study area

The present study focuses on the entire China, with the exception of

Atmospheric Environment 173 (2018) 295–305

the Xinjiang, Tibet, Qinghai, Mongolia and Heilongjiang provinces ( Fig. 1 ), which were not included in the study because the PM 2.5 monitoring sites in these areas are sparse (while these provinces cover a total area of 5.33 million Km 2 , they have only 89 monitoring sites, and this monitoring limitation would seriously aect pollutant estimation accuracy in provinces with serious pollution problems). On the other hand, the study area covers about 4.18 million Km 2 that include 1408 monitoring sites and 93% of the total population of China. For data processing and mapping purposes, the study area is covered with a grid consisting of 357,997 grid cells of 3 × 3Km 2 size ( Xiao et al., 2017 ).

2.2. Data

2.2.1. Ground-level pollutant measurements The 24 h -averaged PM 2.5 , NO 2 , and CO concentrations at nationally- referenced monitoring stations during the period November 1, 2015 to February 29, 2016 were downloaded from the China Environmental Monitoring Center (CEMC, http://106.37.208.233:20035/ ). The ob- served PM 2.5 concentrations, which served as the dependent variable of the pollutant space-time prediction (estimation) techniques used in this work, include 1408 monitoring sites ( Fig. 1 ) with a total of 3009 ob- servations in the study area, and, also, 43 monitoring sites that were evenly distributed in adjacent to the study area provinces to avoid any edge-e ects. PM 2.5 , NO 2 , and CO concentrations less than 2 μg / m 3 (5.6% of total records) were discarded since they are below the estab- lished detection limit ( EPA, 2008 ). Also, stations where data were available during less than 15 days per month were removed, according to China National Ambient Air Quality Standards (CNAAQS). All 3009 observations were distributed during four months, i.e., there exist 953, 707, 529 and 820 observations during the months of November, De- cember, January and February, respectively. Daily PM 2.5 , NO 2 , and CO data were used to calculate the monthly average pollutant concentra- tion at each site, and the monthly averages were obtained using the R programming language (R version 3.3.2, https://www.r-project.org/ ). Notice that most of the PM 2.5 monitoring sites are clustered in urban areas (rural areas have little coverage in China).

in urban areas (rural areas have little coverage in China). Fig. 1. Study area. The green

Fig. 1. Study area. The green dots represent the 1408 PM 2.5 monitoring sites within the study area and the 43 sites in neighboring provinces. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)

296

L. Xiao et al.

L. Xiao et al. Atmospheric Environment 173 (2018) 295–305 Fig. 2. The outline of BME-GWR. 2.2.2.

Atmospheric Environment 173 (2018) 295–305

Fig. 2. The outline of BME-GWR.

2.2.2. Aerosol optical depth data MODIS ( Kaufman et al., 1997 ) is an instrument aboard the Terra and Aqua satellites operated by the U.S. National Aeronautics and Space Administration ( Remer et al., 2005 ). The aerosol optical depth (AOD) data were obtained from the newly released MODIS collection of six AOD products ( http://ladsweb.nascom.nasa.gov/ ) with 3 km spatial resolution. The 3 km AOD (MYD04_3k, MOD04_3k) were all retrieved with the help of the same dark-target algorithm (a characteristic of the algorithm is that its accuracy is slightly lower over land). The Dark- Target (DT) aerosol retrieval algorithm is applied to satellite data to derive aerosol properties, including aerosol optical depth (AOD) over land and ocean, and aerosol size parameters over ocean. Products of the DT aerosol retrieval algorithm are used to develop global and regional AOD retrieval, to study the aerosols with clouds, and for air quality assessment ( Wu et al., 2016a,b; Levy et al., 2010). And the DT is better for areas covered by green plants. And this is why we have carefully selected the study area to meet this condition. Removed areas were mostly deserts, mountains covered by snow etc. such as Xinjiang province and Tibet. During the study period (November 1, 2015 to February 29, 2016) the 10:30am Terra and 1:30pm Aqua were used to synthesize the Chinese regional daily AOD products. Technically, the data processing steps were as follows: (a ) the 550 nm wavelength aerosol optical thickness data were selected following a geometric correction; ( b) after the geometric correction was embedded, the daily 550 nm wavelength aerosol optical thickness data were mosaicked to obtain the daily product and the average AOD value of the overlapping areas in the mosaic image was selected; and ( c) MODIS Terra and MODIS Aqua aerosol data were integrated in order to obtain the max- imum usable range.

2.2.3. Meteorological elds

As has been documented in the literature, weather conditions can a ect PM 2.5 distributions (e.g., Liu et al., 2014 ). Daily data at the me- teorological monitoring stations were collected, including precipita- tion- Prec (0.1 mm), temperature- Temp (0.1 °C), relative humidity- RH (%), air pressure- Pres (0.1Pa), and wind speed-WS (0.1 m/s), using the China Meteorological Data Sharing Service System ( http://cdc.cma. gov.cn ).

2.2.4. Land use information

Previous studies have showed that land-use information could a ect the relationship between PM 2.5 concentrations and satellite AOD ( Ma et al., 2014, 2016, Kloog et al., 2015a; 2015b ). For land-use informa- tion, we downloaded 500 m resolution surface classi cation data from the Global Land Cover Facility-MODIS Land Cover ( http://www. landcover.org/data/lc/ ), and we measured the total grassland (Green), water, urban, built-up (City) and forest areas covered by the

3 × 3Km 2 AOD grid (the area was measured in m 2 ).

2.2.5. Road network data monthly average

Road length data were extracted from the OpenStreetMap (http:// www.openstreetmap.org ). Major roads, primary highways and railways from all di erent layers were retained. The total road length (in m ) was

calculated on the same 3 km grid of land-use data mentioned above.

2.2.6. Elevation data and population data

The 30 m resolution digital elevation data (DEM) were downloaded from the Shuttle Radar Topography Mission (SRTM, http://srtm.csi. cgiar.org/SELECTION/inputCoord.asp ). The 1 km population data

297

L. Xiao et al.

Atmospheric Environment 173 (2018) 295–305

L. Xiao et al. Atmospheric Environment 173 (2018) 295–305 Fig. 3. Histogram and summary statistics of

Fig. 3. Histogram and summary statistics of the GWR model variables for the four months during which PM 2.5 monitoring took place (N = 3009 estimation points).

(Pop) were available from the Gridded Population of the World, Version 4 (GPWv4) ( http://sedac.ciesin.columbia.edu/data/collection/gpw- v4 ). Elevations and populations were subsequently averaged at each

3 × 3Km 2 grid cell.

2.2.7. Data integration For data integration purposes, at the data pre-processing stage: ( a) All meteorological data at a 3 km scale were interpolated using a spatial interpolation software for climatic data (ANUSPLIN), whereas for the NO 2 and CO data at the 3 km scale (i.e., matching the AOD grid size) the standard inverse distance weighted (IDW) interpolation software (ArcGIS 10.3) was used. ( b) All these data were integrated into records appropriate for model tting, validation and mapping purposes. ( c ) The collected data were re-projected onto the Asia Lambert Conformal Conic coordinate system. ( d) Two square grids with 3 km and 30 km spatial resolution were constructed, which consisted of a total number of 357,997 and 3582 grid cells, respectively.

The 30 × 30Km 2 grid data served as the soft information for BME- GWR modeling purposes (for more details on the term soft informa- tion , i.e., secondary information bases of various levels of uncertainty, see sub-section 2.3.2 below). On the other hand, the 3 × 3Km 2 grid (AOD, NO 2 , CO and meteorological) data were used, together with the

3 × 3Km 2 grid (land-cover, road length, elevation and population) data to obtain the complete 3 km resolution dataset. The PM 2.5 value at the

center of each cell of the 30 × 30Km 2 grid was calculated as the average of the values on the 3 × 3Km 2 grid nodes that fall within the 30 × 30Km 2 grid cell. Finally, the complete PM 2.5 sites dataset was extracted based on all 3 × 3Km 2 grid data in ArcGIS 10.3 (see, also, Xiao et al., 2017 ). We notice that, due to the large number and variety of environmental variables, they were standardized by using the z-score method fol- lowing multiple linear regression (MLR). The purpose of MLR is to se- lect the signi cant variables and eliminate any collinearity between variables. MLR is the most common form of linear regression analysis. As a predictive analysis, the multiple linear regression is used to explain the relationship between one continuous dependent variable and two or more independent variables. Each independent variable is given a computed VIF and Tolerance value. When VIF value is large (> 10, for example) and Tolerance < 0.1, collinearity is a problem and the of- fending variables should be removed from the model.

2.3. Methods

2.3.1. The GWR model Ordinary least squares (OLS) is a statistical technique for estimating the unknown parameters of a linear regression model subject to the condition of minimizing the sum of squares of the di erences between the observed responses (values of the variable being predicted) in the given dataset vs . those predicted by a linear function of a set of

298

L. Xiao et al.

Table 1 Parameterization and model tting of the GWR and OLS models.

Statistical model OLS

 

GWR

 

Variables

Coe cient

Accuracy

Value

 

indicators

November

AOD (unitless) Prec (0.1 mm) Pres (0.1Pa) RH (%) WS (0.1 m/s) Temp (0.1 °C)

2.12

R 2

0.779

0.91

RMSE

14.87

11.73

MPE

0.085

3.56

MAE

10.89

0.17

ME

0.614

17.97

NO 2 ( μg / m 3 )

8.12

CO ( μg / m 3 )

3.61

Road_Length (m) 0.5

 

DEM (m) Pop (people) Green (m 2 ) Forest (m 2 ) City (m 2 ) Water (m 2 ) Intercept R 2

0.56

0.44

1.33

0.78

1.66

0.49

41.61

0.69

December

AOD (unitless) Prec (0.1 mm) Pres (0.1Pa) RH (%) WS (0.1 m/s) Temp (0.1 °C)

3.52

R 2

0.828

2.07

RMSE

12.25

9.43

MPE

0.043

4.20

MAE

9.47

25.78

ME

0.254

20.12

NO 2 ( μg / m 3 )

10.89

CO ( μg / m 3 )

8.53

Road_Length (m) 2.07

 

DEM (m) Pop (people) Green (m 2 ) Forest (m 2 ) City (m 2 ) Water (m 2 ) Intercept R 2

4.40

1.07

0.19

1.78

-0.48

1.25

37.84

0.77

January

AOD (unitless) Prec (0.1 mm) Pres (0.1Pa) RH (%) WS (0.1 m/s)

0.58

R 2

0.803

37.83

RMSE

14.68

7.61

MPE

0.055

0.38

MAE

11.14

17.97

ME

0.135

NO 2 ( μg / m 3 )

11.18

CO ( μg / m 3 ) DEM (m) Forest (m 2 ) Intercept R 2

6.05

2.38

2.45

49.35

0.67

February

AOD (unitless) Pres (0.1Pa) RH (%) WS (0.1 m/s)

5.96

R 2

0.629

22.20

RMSE

17.67

19.00

MPE

0.044

44.02

MAE

12.64

NO 2 ( μg / m 3 )

6.56

ME

0.49

CO ( μg / m 3 ) City (m 2 ) Forest (m 2 ) Water (m 2 ) Intercept R 2

1.09

1.57

1.88

0.98

63.40

0.53

explanatory variables. Using the OLS, the Koenker (BP) Statistic (Koenker's studentized Bruesch-Pagan statistic) is a test that determines whether the explanatory variables in the model have a consistent re- lationship to the dependent variable both in geographic space and in data space. Based on OLS theory, Fotheringham et al. (1998) developed a method to explore spatial heterogeneity by means of local regression, termed the geographically weighted regression (GWR) method. In a

Atmospheric Environment 173 (2018) 295–305

sense, GWR is a re ned moving windows approach where observations within the windows are weighted based on distance from the regression point, rather than evaluated equally as in moving windows regression. Like moving windows, the GWR technique calibrates the model at multiple points in the study space, using linear combinations of neighboring data points within a window surrounding the regression point; as such, GWR is an enhanced version of moving windows re- gression in which distance-based weights are added to neighboring points, and bandwidth is allowed to vary across the study area. The GWR method embeds spatial location in the regression parameters and considers local parameter estimates ( Lin et al., 2015b ). In more words, instead of estimating global parameter values, by estimating the para- meters at each location GWR generates a continuous surface of spatially varying parameter values ( Hu et al., 2013 ). The GWR model can assess the in uence of independent variables on dependent variables in terms of the location changes and the spatial heterogeneity of the relationship between an independent and a dependent variable ( Bagheri et al., 2009 ). Before the development of the GWR method, the OLS model was used to test the spatial homogeneity of the relationship between vari- ables. The Koenker (BP) Statistic is a test used to determine if the model explanatory variables have a consistent relationship with the dependent variable (which we are trying to predict/understand) both in geo- graphic space and in data space, that is, the Koenker (BP) statistic was used to test non-stationarity (assuming that the test is statistically sig- nicant, p < 0.01), in which case the relationship is characterized as non-stationary or heteroskedastic. The GWR model can provide the spatial variance of the relationship between PM 2.5 concentrations and other environmental variables. Mathematically, the GWR is a linear regression model expressed as

P

2.5,

i

=+βs(, s )

0

1,

ii

2,

k

βs(, s ) x +ε

k

1,

i

2,

i

ik

i

(1)

are the spatial coordinates of each sample

(i = 1,2,

point i , β (,s s ) is the intercept of sample i , β (,s s ) is the re-

n ) , where (s , s )

1, i

2, i

0

1,

i

2,

i

k

1, i

2, i

gression coe cient of sample point i , x ik is an observation of the k th

environmental variable of sample i , P 2.5, i is the ground-level PM 2.5 concentration at sample i and ε i is the random error. If β = β == β n , the GWR method reduces to the earlier OLS model ( Fotheringham et al., 2002 ). Notice that, in this work separate GWR models were tted to the data of each month, with the general structure.

1

2

PM 2.5 AOD+Prec+Pres+RH+WS+Temp+NO 2 +CO+DEM+Roa-

(2)

d_Length + Pop + Green + Forest + City + Water,

where PM 2.5 here refers to the monthly-averaged ground-level PM 2.5 concentrations ( μg / m 3 ), and the symbol denotes that the PM 2.5 is related to/can be predicted by the following environmental factors (seen as independent variables): the MODIS aerosol optical depth value AOD (unitless), the precipitation Prec (0.1 mm ), the air pressure Pres (0.1 Pa), the relative humidity- RH (%), the WS is wind speed (0.1 m/s ), Temp denotes temperature (0.1° C ), NO 2 is nitrogen dioxide ( μg / m 3 ), the carbon monoxide CO ( μg / m 3 ), the elevations DEM (m) , the sum of road lengths Road_Length ( m ), the count of people Pop , and the land-use cover the Green , Forest , City , and Water areas (m 2 ). In this study, adaptive bandwidths were used due to the uneven distribution of the PM 2.5 data points. These bandwidths were obtained by minimizing the corrected Akaike Information Criterion value (AICc). There are other bandwidth selection criteria such as the standard Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Compared to AIC, the AICc is less biased and can avoid the large variability and under-smoothness tendency of AIC ( Hurvich et al., 1998 ). Unlike AIC, the BIC is not an estimator of the Kull- back Leibler information distance that is used to decide which model is closest to reality ( Hu et al., 2013 ). Also, how the BIC could be extended to non-parametric models with variable bandwidth and e ective de- grees of freedom is yet not clear ( Bekara et al., 2005 ).

299

L. Xiao et al.

Atmospheric Environment 173 (2018) 295–305

L. Xiao et al. Atmospheric Environment 173 (2018) 295–305 Fig. 4. The spatiotemporal empirical covariance (in

Fig. 4. The spatiotemporal empirical covariance (in (μg / m ) 2 ) and the tted theoretical model used by the BME-GWR technique.

3

2.3.2. The BME technique

Bayesian Maximum Entropy (BME, Christakos, 1990, 2000 ) is a spatiotemporal modeling and prediction theory with very general fea- tures (e.g., it provides non-linear estimators and allows non-Gaussian probability laws, it incorporates information from many di erent sources, as well as core and site-speci c knowledge bases-KB). Its im- plementation is made possible in practice in terms of various software libraries, like the one used here, namely, the Spatiotemporal Epistemic Knowledge Synthesis Graphical User Interface software library (SEKS- GUI, Yu et al., 2007 ). The basic set of BME equations of space-time PM 2.5 estimation used in the present study are ( Christakos, 2000, 2010 ),

dG g

dSe

(

μ g

)

g e

Af

μ g

=

PM 2.5

0,

= 0

(3a-b)

where g is a vector of functions expressing mathematically the available core G -KB, including theoretical space-time covariance models, popu- lation exposure laws, and scienti c theories, including the GWR model;

g denotes the mean value of g ; S denotes the available site-speci c KB about the pollutant in the specic study region as described earlier ( S- KB may include AOD, meteorological monitoring data, road networks information, land-use data, DEM and population); μ is a vector of coe cients representing the relative importance of each g -function ( μg denotes the inner product of the vectors g and μ , which are both functions of space-time); f PM 2.5 is the probability law of the PM 2.5 dis- tribution in space-time, where the distribution is considered as a random eld model ( Christakos, 2017 ); and A is a normalization parameter. Technically, S-KB may represent both hard and soft data:

hard data include site-speci c measurements with negligible or no uncertainty associated with them, whereas soft data include site-spe- cic information in the form of uncertain observations, auxiliary vari- ables, and probabilistic assessments (e.g., intervals of possible values and probability distribution functions of any shape; He and Kolovos, 2017 ). This allows BME to rigorously integrate any non-Gaussian soft data, such as soft data with a truncated Gaussian distribution ( Reyes and Serre, 2014 ). Eq (3a-b) can be solved with respect to the PM 2.5 probability law f PM 2.5 at all pollutant mapping points (i.e., space-time points at which predictions of the PM 2.5 concentrations are sought). Also, it has been proven in theory that Kriging is a special case of BME under limiting conditions linear estimation, Gaussian distribution and hard only site- speci c data are considered ( Christakos, 2000 ). More technical details concerning the BME approach above can be found in the relevant lit- erature. Software libraries have been developed dealing with the so- lution of Eq (3a-b) in real-world conditions, including BMElib, SEKS- GUI, QuantumBME, and StarBME ( Yu et al., 2007 ).

2.3.3. The combined (integrative) BME-GWR technique

In this work, we developed a combined (integrative) BME-GWR technique to estimate space-time PM 2.5 concentrations in the most

300

economically developed area of China. The role of GWR in this in- tegration was to generate soft data. The monthly GWR models involve di erent variables after they are ltered by OLS ( Table 2 ), where Eq (2) describes the general multi-variable structure of monthly GWR. And, the role of BME is to use these soft data together with hard data to produce space-time maps of PM 2.5 distribution during four months (3 × 3 Km 2 resolution). Speci cally, the hard data that served as input to the BME-GWR technique included the measured PM 2.5 concentration values at mon- itoring stations for all eligible station-days during the period November

1, 2015 to February 28, 2016 (all predictions are about monthly-

averaged data). On the other hand, the soft data generated by the GWR model and used in the GWR-BME technique consisted of PM 2.5 esti-

mates and an associated variance at the center of each 30 × 30 Km 2 grid cell. The soft data obtained by the GWR model (in the form of probability distributions with mean and variance estimated by GWR) serve as useful auxiliary information that can improve the accuracy of

the predictions generated by BME at the unsampled points of the space-

time grid. Notice that, four separate GWR models were tted to the data

during each month, and the BME-GWR method was applied for the entire four-month period. An outline of the combined BME-GWR technique is shown in Fig. 2 .

2.3.4. 10-Fold cross-validation between spatiotemporal estimates and ground observations In order to assess the performance of the combined BME-GWR technique, the coe cient of determination (R 2 ), the mean prediction error (MPE), the mean error (ME), the mean absolute prediction error (MAE), and the root mean squared prediction error (RMSE) were the accuracy indicators calculated in terms of the dierences between the space-time PM 2.5 predictions generated by the combined BME-GWR technique and the ground-level PM 2.5 observations at a set of control points. Since the sample-based, 10-fold cross validation (CV) procedure ( Diego Rodriguez et al., 2010; Lee et al., 2011; Hu et al., 2014; Ma et al., 2014 ) has been more widely tested in previous PM 2.5 -AOD

modeling studies than other procedures (such as, e.g., the site-based CV technique, Chang et al., 2014 ), in this study we also chose the 10-fold

CV method to test the potential over-tting of the BME-GWR technique.

Previous GWR studies that used the 10-fold cross validation method include Hu et al. (2013), Ma et al. (2014), Fang et al. (2016), You et al. (2016b) , and Zou et al. (2016) . In the present study, the data set was divided into k = 10 folds, a classi er learned using k -1 = 9 folds, and

an error value was calculated by testing the classi er in the remaining

fold. Then, the k-fold cross-validation classi cation ( k-cv) error esti-

mator was calculated as the average value of the errors committed in each fold. Note that the k-cv error estimator depends on two factors: the training set and the partition into folds ( Diego Rodriguez et al., 2010 ).

L. Xiao et al.

Atmospheric Environment 173 (2018) 295–305

L. Xiao et al. Atmospheric Environment 173 (2018) 295–305 Fig. 5. Scatter plots of model fi

Fig. 5. Scatter plots of model tting and validation result. The solid line denotes the trend line: (a) CV results for the BME-GWR technique (N = 3009 estimation points, t = 4 months); (b) (e) are GWR model tting results during the four months considered.

301

L. Xiao et al.

Table 2 10-fold cross-validation results of space-time BME-GWR estimation of monthly PM 2.5 averages (k = 1408 monitoring stations, N = 3009 estimated points, t = 4 months).

Atmospheric Environment 173 (2018) 295–305

where c = 0.8, c = 0.2, aa==1,

1

2

s

1

s

2

14 (in °C), and a = 3.8, a = 13

t

1

t

2

(in months). Using the above covariance model, we obtained BME es- timates of PM 2.5 that are representative of the actual distribution.

Method

R 2

RMSE

MPE

MAE

ME

 

3.3.

Cross validation results

 

BME-GWR

0.883

11.39

0.067

7.90

1.065

3. Results

3.1. Descriptive statistics

Various simulations of the 16-variable model were initially gener- ated using the GWR and OLS techniques. The histograms and summary statistics of GWR variable tting for the four months (Nov Dec 2015, Jan Feb 2016) are plotted in Fig. 3 . These plots show that all the variables are roughly lognormally and unimodally distributed. The geometric mean, standard deviation, maximum, and minimum for all variables for all days are also presented in Fig. 3 . The mean PM 2.5 concentration over all monitoring sites is 63.75 μg / m 3 , and the overall mean of the MODIS generated AOD is 0.41. On the basis of the OLS results, the Koenker (BP) Statistic tests were found to be statistically signi cant ( p < 0.01) during the four months considered in the present study, indicating that there is a non-stationary relationship between dependent variables and independent variables, that is, the local sublinearity between environmental variables and PM 2.5 concentrations in a wide geographic area can be better inter- preted by the GWR model. This result is consistent with previous stu- dies ( Hu et al., 2013; Song et al., 2014; Zou et al., 2016 ). Table 1 provides a summary of the parameters of the tted GWR and OLS models. The overall mean adjusted R 2 = 0.75 (GWR model) is greater than the overall mean adjusted R 2 = 0.67 (OLS model). This result conrms the superiority of the GWR model over the OLS model in simulating site-based PM 2.5 concentrations. These ndings are con- sistent with previous GWR applications ( Hu et al., 2013; Zou et al., 2016 ). The GWR result for December was the most accurate among the GWR results for the four months considered, with the highest R 2 (= 0.828) and the lowest RMSE (=12.25 μg / m 3 ). As regards the OLS re- sults, the model with the highest R 2 (=0.77) corresponds to December. Di erent monthly OLS models have di erent variables, and the same variable has di erent relationships (positive or negative) with PM 2.5 . For example, the AOD coecients for December, November, January, and February are, respectively, 3.52, 2.12, 0.58, and 5.96.

3.2. Covariance model tting

The covariance function ( Kolovos et al., 2004; Ma, 2008 ) is a sta- tistical tool that oers information about the variation of PM 2.5 con- centrations across space-time, when is mathematically represented by a spatiotemporal random eld (Christakos, 2017 ). Concerning the em- pirical space-time covariance presented in Fig. 4 , we noticed that the PM 2.5 covariance used by the combined BME-GWR technique had a linear shape at the space-time origin (in random eld modeling terms, this covariance shape is interpreted as characterizing considerable PM 2.5 concentration changes both in space and in time), followed by a sharp PM 2.5 covariance drop along space, and a very slow decline along the time axis (this dual behavior indicates a short PM 2.5 correlation range in space and a long PM 2.5 correlation range in time). Given that four months are considered in this study, the temporal correlation of adjacent months is strong. In light of these empirical PM 2.5 covariance features, the following theoretical covariance model was tted to the empirical one of Fig. 4 ,

c

PM

2.5

(, )

= c

1

1

3 h

2

a

s

1

+

3 ⎞ ⎛

h

2 a

3 ⎠ ⎝

s

1

⎟⎜

1

3 τ

2

a

t

1

+

1

3

τ

2 a

3

t

+

c 2 e

3 (

h +

a

s

2

τ

a t

2

)

(4)

302

Fig. 5 and Table 2 show the cross-validation (CV) results of the BME- GWR estimation technique. The accuracy indicators of technique, in- cluding RMSE, MPE, MAE, and ME, are listed in Table 2 . For model validation purposes, the R 2 , RMSE, MPE, MAE and ME values of the BME-GWR technique were 0.883, 11.39, 0.067, 7.90, and 1.065 μg / m 3 , respectively.

3.4. Spatial distributions of PM 2.5 predictions

Fig. 6 presents ground-level PM 2.5 measurements and the corre- sponding monthly-averaged PM 2.5 predictions using the space-time BME-GWR technique. The rst observation is that the fact that most of the PM 2.5 monitoring sites are clustered in urban areas (rural areas have little coverage in China), may impact the performance of a space-time prediction technique that covers the entire China. Provinces (such as Xinjiang, Tibet, Qinghai, Mongolia and Heilongjiang) with sparse and unevenly distributed PM 2.5 monitoring sites were removed from the present study. Also, the considerable number of missing PM 2.5 data in the Tibet and Xinjiang provinces can lead to substantial errors in PM 2.5 concentration prediction ( Ma et al., 2014; Fang et al., 2016 ). These provinces, also, were not included in the present study. The second observation is that, while the Beijing, Tianjin and Hebei provinces were always highly polluted areas, in southern provinces the air quality has been much better. The BME-GWR technique provided considerably detailed maps of the space-time pollutant distribution due to its in- corporation of soft data. As the maps of Fig. 6 reveal, the PM 2.5 concentrations were high in the northern part of the study area, especially in the Beijing, Tianjin, Hebei and Shandong provinces. The spatial gradients of PM 2.5 con- centrations showed a signicant change during the four months con- sidered, and the overall PM 2.5 concentration trend was high in the north and low in the south part of the study region. The temporal PM 2.5 concentrations in the study region showed a clear monthly variation, and there was a PM 2.5 concentration trend from the northern coast toward inland during the period November of 2015 to February of 2016. A possible explanation may be the monsoon weather conditions that caused the dispersion of the coastal PM 2.5 pollution towards inland China ( Ainslie et al., 2008). The average PM 2.5 concentration during December 2015 (reaching 76.1 μg / m 3 ) was signi cantly higher than that during the other three months, while the most widely distributed PM2 2.5 concentrations were observed during January 2016 (with spa- tially distributed values that exceeded the 85 μg / m 3 level).

4. Discussion

From a methodological perspective, this work presented a combined BME-GWR technique of space-time pollutant modeling and mapping. In this setting, the role of GWR was to generate soft data, and that of BME to use these data together with the available harddata to pro- duce space-time maps of PM 2.5 distribution over China during a four- month period. In particular:

(a) The BME approach is able to integrate general or core KBs (theo- retical covariance models, trend functions, physical laws, and sci- enti c relationships) about the attribute of interest with site-spe- cic KBs (hard or accurate data, such as monthly-averaged PM 2.5

concentrations, and soft data in the form of uncertain measure-

ments, probabilistic assessments and auxiliary information) to

generate space-time PM 2.5 prediction maps and the associated prediction accuracy maps.

L. Xiao et al.

L. Xiao et al. 303 Atmospheric Environment 173 (2018) 295–305 Fig. 6. (a) – (d) Ground

303

Atmospheric Environment 173 (2018) 295–305

Fig. 6. (a) (d) Ground measurements and (e) (h) spatial distribu- tions of monthly PM 2.5 concentrations generated by the BME-GWR space-time technique during November 2015, December 2015, January 2016, and February 2016.

L. Xiao et al.

Table 3 Summary of previous GWR, improved GWR, and Bayesian models PM 2.5 results in China.

Atmospheric Environment 173 (2018) 295–305

Publication

Study period

Model

10-fold CV- R 2 and RMSE

Ma et al. (2014) Song et al. (2014) You et al. (2016b) Lv et al. (2016) You et al. (2016a) Zou et al. (2016) Fang et al. (2016)

2012

2013

GWR

R 2 = 0.64, RMSE = 32.98(N = 835), China. R 2 = 0.738, RMSE = 32.98(N = 37), PRD. R 2 = 0.760, RMSE = 20.85 (MODIS, N = 943); R 2 = 0.850, RMSE = 24.86 (MISR, N = 943), China. R 2 = 0.780, RMSE = 33.39 (N = 298), North China. R 2 = 0.79, RMSE = 18.6 (N = 943), China. R 2 = 0.750, RMSE = 10 (GWR, N = 52), R 2 = 0.530, RMSE = 16 (OLS, N = 52), BTH R 2 = 0.80, RMSE = 22.75 (TSAM, N = 509 945), China

2012

2013

GWR

2014

GWR

2014

Bayesian

2014

GWR

2013

GWR, OLS

2013/12014/6

TSAM

(1) N denotes the number of PM 2.5 monitoring sites. (2) Model: GWR (Geographically Weighted Regression); Bayesian model, Bayesian hierarchical framework; OLS (ordinary least square); TSAM (timely structure adaptive modeling) based on GWR. (3) PRD (Pearl River Delta region); BTH (Beijing City, most of Tianjin City, and parts of Hebei Province).

(b) The GWR model, on the other hand, explores spatial heterogeneity by means of local regression. Instead of estimating global parameter values, by estimating the parameters at each location GWR gen- erates a continuous surface of spatially varying parameter values. Also, GWR can assess the inuence of independent variables on dependent variables in terms of location changes and spatial het- erogeneity of the relationship between an independent and a de- pendent variable.

As regards ( b), many earlier studies have used various versions of GWR-based models, as wells as Bayesian models. Each one of these studies exhibited di erent characteristics because of the di erent con- ditions, PM 2.5 concentration levels, and geological and meteorological environments considered ( Table 3 ). By accounting for cross space-time correlations, allowing more general assumptions (non-linear and non- Gaussian predictors), and incorporating a su cient number of key environmental factors (capable of explaining a large proportion of the PM 2.5 concentration variance) the model used in this work performed better than the earlier ones, because of its advantages as outlined in the Introduction section and in other parts of the paper. In particular, it provided the best 10-fold cross-validating results (R 2 = 0.883, RMSE = 11.39). As regards ( a), it was shown that the combination of the BME technique with the GWR model o ers certain improvements in the space-time prediction (estimation) of PM 2.5 concentrations over China. In particular, we compared the combined BME-GWR technique pro- posed in this work with seven previous studies of satellite-derived PM 2.5 concentrations in China. The combined BME-GWR technique generated realistic maps of space-time PM 2.5 distribution, and its performance was superior to these techniques, in terms of prediction accuracy. Other important di erences and advantages of BME-GWR over the previous techniques include the following. First, due to sparse sites and AOD missing data, Ma et al. (2014) obtained higher PM 2.5 concentra- tion estimates in rural regions and slightly lower concentration esti- mates in the Sichuan basin. The study by Fang et al. (2016) was char- acterized by a considerable number of missing PM 2.5 data in the Tibet and Xinjiang provinces, which generated substantial errors in the PM 2.5 concentration estimates they derived. Second, in the present study we collected and processed more data of environmental factors than in previous studies, such as 3 km AOD, pollutant emissions, meteor- ological data and land use. Furthermore, for the rst time, we in- corporated a GWR model into the BME estimation technique in a high spatial/temporal resolution setting. The calculated accuracy indicators of the combined GWR-BME technique showed a considerable im- provement compared to the existing techniques. As was mentioned throughout the paper, the limited number of monitoring sites is a limitation of this work, as well as the lack of information about other environmental variables. When more monitoring stations are estab- lished in certain rural areas, the BME-GWR can be used to improve the study of air pollution in these areas. And, we hope that data about more

304

environmental variables will become available, such as wind directions, in this case. The ndings of this study are useful for exposure assessment and health risk management purposes, as well as for air pollution control strategies and environmental protection related studies. Accordingly, future work on the combined GWR-BME technique should consider ( i ) adding higher resolution data sources, such as 1 km AOD data ( Lin et al., 2015a; Wu et al., 2016a,b ), ( ii ) focusing on developing satellite- based models for the prediction of historical PM 2.5 and other air pol- lutants, ( iii) visualizing exposure estimates in ArcEngine 10 in order to properly assess the long-term e ects of PM 2.5 exposure on human health and facilitate better prevention and control of PM 2.5 exposure in China.

Acknowledgments

This research was supported by the National Science Foundation of China (Grant No. 41671399).

References

L. Xiao et al.

409

Atmospheric Environment 173 (2018) 295–305

305