Visibility-Based PM2.5 Concentrations in China - ACS Publications

Subscriber access provided by University of Florida | Smathers Libraries

Article

Visibility-based PM2.5 concentrations in China: 1957-1964 and 1973-2014 Miaomiao Liu, Jun Bi, and Zongwei Ma Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.7b03468 • Publication Date (Web): 24 Oct 2017 Downloaded from http://pubs.acs.org on October 24, 2017

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Environmental Science & Technology is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 24

Environmental Science & Technology

1

Visibility-based PM2.5 concentrations in China: 1957-1964 and

2

1973-2014

3

Miaomiao Liu a, Jun Bi a,b, Zongwei Ma a, c,*

4

a

5

Nanjing University, Nanjing, Jiangsu, China

6

b

7

Technology (CICAEET), Jiangsu, China

8

c

9

China

State Key Laboratory of Pollution Control and Resource Reuse, School of the Environment,

Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment

School of Geographic and Oceanographic Sciences, Nanjing University, Nanjing, Jiangsu,

10 11

*Correspondence to:

12

Dr. Zongwei Ma, School of the Environment, Nanjing University, 163 Xianlin Avenue,

13

Nanjing 210023, P. R. China. Tel.: +86 25 89681526. E–mail address: [email protected].

14

1

ACS Paragon Plus Environment


15

ABSTRACT

16

China established ground PM2.5 monitoring network in late 2012 and hence the long-term and

17

large-scale PM2.5 data were lacking before 2013. In this work, we developed a national-scale

18

spatiotemporal linear mixed effects model to estimate the long-term PM2.5 concentrations in China

19

from 1957 to 1964 and from 1973 to 2014 using ground visibility monitoring data as the primary

20

predictor. The overall model-fitting and cross-validation R2 is 0.72 and 0.71, suggesting that the model

21

is not over-fitted. Validation beyond the model year (2014) indicated that the model could accurately

22

estimate historical PM2.5 concentrations at the monthly (R2 = 0.71) level. The historical PM2.5

23

estimates suggest that air pollution is not a new environmental issue that occurs in the recent decades

24

but a problem existing in a longer time before 1980. The PM2.5 concentrations have reached 60-80

25

µg/m3 in the north part of North China Plain during 1950s-1960s and increased to generally higher

26

than 90 µg/m3 during 1970s. The results also show that the entire China experienced an overall

27

increasing trend (0.19 µg/m3/yr, P 90% to avoid the

97

influence of potential fog or rain events based on the RH and precipitation data 2. The sources and purposes

98

of all datasets are listed in Supporting Information (SI, Table S1). The hourly data were then aggregated

6


Page 6 of 24

Page 7 of 24

99


into daily average for further analysis. Since the NCEI meteorological data from 1965 to 1972 are totally

100

missing, we did not modeled and analyzed the historical PM2.5 for this period. It should be noted that there

101

was still some missing data for the period from 1957 to 1964 and from 1973 to 2014. There could be more

102

missing data for some sites than others.

103 104 105

Figure 1. Spatial distribution of ground PM2.5 (A) and meteorological (B) monitoring sites involved in this study.

106

Data integration

107

To obtain more matched data points and develop a robust statistical model, we first created a radius

108

buffer zone for each single meteorological site to match PM2.5 and meteorological data. For each day,

109

the daily PM2.5 concentrations within the buffer zone of a single meteorological site were averaged and 7



110

assigned to the corresponding single meteorological site. Then the buffer-averaged PM2.5 was matched

111

to the meteorological variables in that meteorological site. We conducted a sensitivity analysis to find

112

the optimal buffer radius for data matching. We created a series of radius buffer zones ranged from

113

5km to 70km and compared the cross-validation (CV) R2 results for models developed using datasets

114

from different radius buffer matching strategy (SI, Text S1 and Figure S1). Results show that 20 km is

115

the optimal buffer radius which can achieve the best model performance.

116

China established ground PM2.5 monitoring network in late 2012, thus we have relatively complete

117

PM2.5-visibility matched dataset covering entire China in 2013 and 2014 to fit the model. When using

118

the model developed from the two-year data to estimate the historical PM2.5 concentrations, we need to

119

evaluate the historical estimates using ground measurements before 2013. Although we have ground

120

PM2.5 measurements from 2004 to 2012, most of them located in Hong Kong and Taiwan. Therefore,

121

using ground monitoring data from 2004 to 2012 cannot represent the performance of historical

122

estimates for the whole China. Thus, we used matched data of 2014 to fit and cross validate the model

123

in the modeling year. We matched and compared the visibility-based PM2.5 historical estimates with

124

the ground PM2.5 measurements using the same buffer matching strategy as mentioned above from

125

2004 to 2013 to evaluate the accuracy of historical PM2.5 estimates.

126

Model development and validation

127

We developed a spatial-temporal linear mixed effects (LME) model to fit the month- and

128

province-specific and random intercepts and random slopes for the log-transformations of visibility

129

using the matched daily PM2.5 and daily visibility data. In other words, we established the relationship

130

between daily average PM2.5 concentration and daily mean visibility range for each month in a 8


Page 8 of 24

Page 9 of 24


131

province. The model structure is shown as follow:

132

PM2.5,st = (µ +µ’ij) + (β1 +β1’ij) lnVISst + εst (µ’ij,β1’ij)~N[(0,0), Ψ]

133

where PM2.5,st is the daily ground PM2.5 concentration at site s on day t; lnVISst is log-transformation of

134

visibility range at site s on day t; µ is the fixed intercept; µ’ij is the month- and province-specific

135

random intercept, i.e., the regression intercept in province i in month j; β1-β2 are fixed slopes for

136

independent valuables; β1’ij is the random slope for log-transformation of visibility range in province i

137

in month j; εst is the error term at site s on day t. Ψ is the unstructured variance-covariance matrix for

138

the random effects. The selection of lnVIS as the only independent variable is based on the comparison

139

of models using different combinations of independent variables (SI, Text S2 and Table S3).

140

Statistical indicators of coefficient of determination (R2) and root mean squared prediction error

141

(RMSE) were calculated to assess the performance of model fitting. We used 10-fold CV to test

142

potential model over-fitting, that is, the model might have better predictive performance in the model

143

dataset than in the rest of the data from the study area. First, all samples in the model dataset were

144

randomly and equally divided into ten subsets. For the first round, one subset was selected as the

145

testing dataset while the remaining nine subsets were used as the model dataset. We fitted the model

146

using the nine subsets and applied the model on the visibility data in the testing dataset to predict

147

PM2.5 concentrations. For the next round, another subset was used as the testing dataset and the

148

remaining nine subsets were used to fit the model. This process was repeated ten rounds until each

149

subset was used exactly once for validation. The agreement between CV-predicted PM2.5 and ground

150

observations was evaluated using R2 and RMSE. Statistics of R2 and RMSE between model fitting and

151

CV were compared to assess the potential model over-fitting.

152 9


[1]


153

Estimation, evaluation, and time series analysis of historical PM2.5

154

The historical daily PM2.5 concentrations (1957-1964, 1973-2013) were estimated by applying the

155

model developed in 2014 on the historical visibility data. Note that data from 1965 to 1972 are missing

156

and therefore we cannot estimate the PM2.5 concentrations during this period. Although the model has

157

been cross validated in the modeling year of 2014, it does not mean that historical (before 2014) PM2.5

158

estimates would have similar accuracy as the model CV results. Therefore, we evaluated historical

159

PM2.5 estimations at the daily and monthly scales using ground-measured PM2.5 from 2004-2013.

160

We deseasonalized the historical PM2.5 data by calculating the monthly PM2.5 anomaly time series for

161

each site to remove the influence of the seasonal effect. The monthly anomaly was calculated by

162

subtracting the long-term average PM2.5 concentration of the corresponding month from the monthly

163

PM2.5 estimate. We then applied least squares regression to the monthly PM2.5 anomaly time series for

164

each site to analyze the PM2.5 trends 21, which has been applied to analyzed the satellite AOD or

165

satellite-based PM2.5 trends 5, 22, 23. We also calculated the national and regional (including Jing-Jin-Ji

166

region, Yangtze River Delta, and Pearl River Delta) trends. We first obtained the national or regional

167

averaged monthly PM2.5 anomalies and then calculated the national and regional trends.

168

RESULTS AND DISCUSSION

169

Descriptive statistics of model fitting dataset

170

Descriptive statistics of the model fitting dataset are shown in Table S3 (SI). Using the 20-km radius

171

buffer matching strategy, we obtained 28,098 matched data records to fit the model. The overall mean

172

PM2.5 concentration is 56.74 µg/m3 and the mean value of visibility is 13.18 km. The high standard

10


Page 10 of 24

Page 11 of 24


173

deviations of PM2.5 (47.27µg/m3) and visibility (7.88 km) indicate the strong spatial and temporal

174

variations of PM2.5 pollutions in China. The values exhibit strong seasonality. The highest mean PM2.5

175

concentration is in winter (81.18 µg/m3), with the lowest mean visibility (11.37 km). Summer

176

possesses the lowest mean PM2.5 (40.37 µg/m3) and the highest mean visibility (14.55 km). Spring and

177

autumn presents similar PM2.5 and visibility levels. The seasonal variations are consistent with

178

previous studies 8, 24, 25. Central heating with coal burning combined with frequent stagnant weather

179

and temperature inversion is the main reason for high PM2.5 pollution levels in winter China24, 26.

180

Model fitting and cross validation

181

Figure 2 summarized the model fitting and CV results. For the model fitting, the R2 value is 0.72 and

182

the RMSE is 24.86 µg/m3. The CV R2 value is 0.71 and the CV RMSE is 25.62 µg/m3. The differences

183

between model fitting and CV R2 (or RMSE) values are often used to assess the degree of model

184

overfitting. Compared to the model fitting, the CV R2 only decreases by 0.01 and CV RMSE only

185

increases by 0.76 µg/m3, indicating that our model is not substantially over-fitted. The model CV

186

results show that the model has satisfactory performance when predicting PM2.5 concentrations in

187

those meteorological sites which are not included in the model dataset in the model year of 2014.

11



188 189 190

Figure 2. Density scatterplots of model fitting (A) and cross-validation (B) (N = 28,098).

191

Table S4 and S5 summarizes the fixed effects and variance-covariance matrix of random effects for

192

the LME model, respectively. Results show that the log-transformed visibility is significant predictor

193

for PM2.5 concentrations. Higher PM2.5 concentrations would cause lower visibility in the days with

194

relative humidity less than 90% and without rains. This finding is consistent with our common sense

195

and previous studies 13, 27, suggesting that our model can be physically explained.

196

We also fitted models only with month-specific or province-specific random effects for comparison.

197

For model only with month-specific random effects, the model fitting and CV R2 values are 0.45 and

198

0.44, respectively. For the province-specific model, the model fitting and CV R2 values are 0.57 and

199

0.56, respectively. Results showed that combing month- and province-specific random effects can

200

greatly improve the model performance. Besides, we also fitted the geographically weighted

201

regression (GWR) model for comparison. The overall model fitting R2 of the GWR model is 0.83,

202

which is much higher than the model fitting R2 of our LME model. However, the model CV R2

203

decreased to 0.64 for GWR model, indicating substantial over-fitting problem.

12


Page 12 of 24

Page 13 of 24


204

Validation of historical PM2.5 estimates

205

Figure 3 summarizes the evaluations of historical PM2.5 estimates at daily and monthly levels. We

206

required at least 20 daily PM2.5 estimates to calculate the monthly PM2.5. Compared to the accuracy of

207

daily PM2.5 estimates in the modeling year (CV R2=0.71), the historical estimates demonstrate poorer

208

accuracy (R2=0.60, N=44,884) at daily level. However, when aggregating the daily estimates to

209

monthly mean values, the accuracy improved greatly. The evaluation R2 at monthly level is 0.71

210

(N=1,545). The poor historical daily PM2.5 estimates were probably due to the strong model

211

assumption that the relationships (i.e., month- and province-specific random effects) between daily

212

PM2.5 and visibility derived from the 2014 model remained constant for the same month and province

213

in each year. Nonetheless, the accuracy of historical monthly mean PM2.5 estimates from 2004 to 2013

214

is satisfactory. These results are consistent with our previous satellite AOD-PM2.5 modeling in China

215

23

216

concentration from 2004-2012, we found poor performance for daily level estimates (R2 = 0.41) but

217

fairly accurate performance for monthly mean estimates (R2 = 0.73) 23. Due to the lack of ground

218

PM2.5 observations before 2004, we cannot assess the historical daily and monthly PM2.5 estimates

219

before 2004, which is a limitation and uncertainty in this study.

. When using the PM2.5-AOD statistical model developed in 2013 to estimate the historical PM2.5

13



220 221 222

Figure 3. Evaluation of historical PM2.5 estimates at daily (A) and monthly (B) levels from 2004 to 2013.

223

Spatial variations of model estimated PM2.5 concentrations

224

Figure 4 presents the spatial distributions of PM2.5 concentrations for different periods. For each period,

225

we required at least 60% of days with available PM2.5 data to calculate the average PM2.5 for each site.

226

Similar spatial distribution patterns were found across different periods. Higher PM2.5 pollution levels

227

always appeared in the north part of North China Plain, Northeast China, Central China, Sichuan

228

Basin, and Northwest China. These areas have been suffering relatively high air pollution levels since

229

1950s. For example, during 1950s and 1960s, the PM2.5 concentrations have reached 60-80 µg/m3 in

230

the north part of North China Plain (Shandong Province and the south part of Hebei Province), which

231

were much higher than Level 2 annual PM2.5 standard (35 µg/m3) of China’s National Ambient Air

232

Quality Standard (CNAAQ). And the PM2.5 concentrations increased to generally higher than 90

233

µg/m3, with maximum values higher than 110 µg/m3 during and after 1970s. The PM2.5 concentrations

234

in late 1970s had reached to a similar pollution level as in 2000s except East China (Jiangsu, Anhui,

235

and Zhejiang Provinces) and Southeast China (Fujian Province). For Xinjiang Province in Northwest 14


Page 14 of 24

Page 15 of 24


236

China, the PM2.5 levels were generally 50-60 µg/m3, which were mainly come from the dust impact of

237

Taklimakan Desert and relatively stable over decades.

238 239 240

Figure 4. Spatial distributions of long-term average PM2.5 estimations for 1957-1964 (A), 1973-1980 (B), 1981-1995 (C), 1996-2005 (D), 2006-2014 (E), and 1957-2014 (F).

241

Our findings suggest that air pollution is not a new environmental issue that occurs in the recent 2-3

242

decades but a problem existing in a longer time before 1980 28, 29, especially in the areas mentioned

243

above. To further confirm these findings, we reviewed the limited papers which studied the air

244

pollution in China from1960s-1980s. Although there were no PM2.5 studies in this period, there had

245

been a handful of PM10 pollution studies in China, which are summarized in Table 1. Results show

246

that these major Chinese cities have suffered severe PM10 pollution since 1970s. Given the fact that the

247

proportion of PM2.5 in PM10 ranges from 50% to 85% in Chinese cities30-32, the results from previous

15



Page 16 of 24

248

PM10 studies suggested that Chinese cities have suffered severe PM2.5 pollutions since 1970s.

249

Table 1. Ground measured PM10 concentrations in several major cities during 1970s City Beijing

33, 34

Tianjin 35 Shenyang

33

Harbin 36

37

Chongqing

Shanghai 33, 38

PM10 concentrations (µg/m3) a

Year

Site

1974-1976

Shijingshan

1,860

1979

Urban area

510

1973-1974

/

1,480

1979

Urban area

760

1978-1979

Daoli

1,605

948

Daowai

2,127

1,598

Nangang

1,811

1,341

Reference site

285

182

Winter

Spring

Summer

Autumn

Annual 600

1,010

640

670

950

1977

Urban center

510

Jiulongpo

680

1973-1981

Urban area

300-450

1978-1981

Suburban area

140-270

1979

Urban area

350

Lanzhou 33

1979

Urban area

1,290

Dukou (now called Panzhihua) 39

1979-1980

Nongnongping

950

1,150

300

520

730

Bingcaogang

540

370

60

150

280

Renhe

190

220

50

70

130

250

The East China seems to be the area which experienced the biggest drop in air quality in China. The

251

PM2.5 concentrations in East China were generally lower than 25 µg/m3 during 1950s and 1960s and

252

increased to 30-50 µg/m3 in around 1995, and increased to greater than 60 µg/m3 in around 2005. The

253

rapid social economic development with a lack of effective environmental regulation after the reform

254

and opening started from late 1970s was the major driver for the degradation of PM2.5 concentrations

255

in East China. Shanghai City is one exception in East China. The PM2.5 pollution level had reached

256

60-70 µg/m3 in 1970s, which was consistent with previous studies shown in Table 1. This is probably

257

because Shanghai had been a highly industrialized city before 1980s. Tibet and Hainan Provinces were

258

always the cleanest areas in China, and the pollution levels were approximately 15-25 µg/m3 over the

259

past decades. Taiwan, Yunnan, and Inner Mongolia were also relatively clean areas, where PM2.5 16


Page 17 of 24


260

concentrations were generally 30-40 µg/m3 over the past decades.

261

Historical trends of estimated PM2.5 concentrations

262

We further quantitatively analyzed the temporal trends of PM2.5 concentrations from 1957 to 2014

263

using the monthly PM2.5 anomaly time series. Figure S2 (SI) shows the site-specific annual trends of

264

visibility-derived PM2.5 concentrations. Most of sites had the increasing trends during the whole

265

studied periods, especially sites in Jiangsu, Anhui, Henan, and Hubei Provinces. The highest

266

site-specific trends in these provinces were greater than 1.2 µg/m3/yr, indicating that the PM2.5

267

concentrations have increased by more than 60 µg/m3 for those sites during the past 50 years. There

268

were also some sites presenting decreasing trends (e.g., sites in Tibet and Qinghai Provinces, and north

269

part of Inner Mongolia).

270

Figure 5 and Table 2 summarize the overall PM2.5 trends for entire China and four major regions, i.e.,

271

Jing-Jin-Ji region (JJJ, including Hebei Province, Tianjin City, and Beijing City), the Yangtze River

272

Delta (YRD, including Jiangsu Province, Zhejiang Province, and Shanghai City), and the Pearl River

273

Delta (PRD) in Guangdong Province. Overall, the entire China and all three regions experienced

274

significant increasing trends from 1957 to 2014. The overall trends for entire China, JJJ, YRD, and

275

PRD were 0.19, 0.36, 0.68, and 0.32 µg/m3/yr.

17



276 277 278 279 280 281

Page 18 of 24

Figure 5. Time series of visbility-derived monthly PM2.5 anomaly (µg/m3) for entire China (A), Jing-Jin-Ji region (B), the Yangtze River Delta (C), and the Pearl River Delta (D). Table 2. Historical trends and 95% confidence intervals (CI) of estimated PM2.5 concentrations for entire China and Jing-Jin-Ji, Yangtze River Delta, and Pearl River Delta Regions. Period

Trend

Entire China 3

1957-2014

0.19

0.36

0.68

0.32

(0.18, 0.20)

(0.30, 0.41)

(0.65, 0.71)

(0.31, 0.34)

Significance

P

Visibility-Based PM2.5 Concentrations in China - ACS Publications

Recommend Documents