以大氣頂反射率應用統計模型預測地面細懸浮微粒與成分濃度

Application of Satellite Top of Atmosphere Reflectance with Statistical Modelling for Ground-Level Fine Particulate Matter and Composition Concentration Estimation

葉卓衛(Cheuk Wai Yip)

指導教授 ： 吳章甫

國立臺灣大學/公共衛生學院/環境與職業健康科學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202101566

摘要

空氣中的細懸浮微粒 (PM2.5)被認為與不良健康結果有關。近年來，台灣的PM2.5濃度雖逐漸下降，但目前仍高於世界衛生組織訂定的空氣品質標準10 μg/m3。由於暴露於PM2.5的人數眾多，因此此議題須更加受到重視。基於多年的研究基礎，科學界對PM2.5已有充分的認識，研究指出特定的 PM2.5成分，如鈣、鎘、鎳、釩和鋅，對人體的危害比其他成分更大。因此本研究的目的之一為基於台灣六個空氣品質監測站的監測資料透過統計模型建模的方式推估PM2.5以及PM成分濃度。
本研究選擇六處空氣品質監測站（板橋、忠明、斗六、嘉義、小港、花蓮）的監測資料作為PM2.5和PM成分模型建立基礎，這六處測站每六日進行二十四小時採樣。利用來自 Himawari-8 衛星的大氣頂反射率與其他變項以Stepwise AIC的方式建立迴歸模型。本研究中使用四種不同模型情境，包括添加氣象變項、氣體污染物與加入主成分分析的應用。
本研究結果顯示，在所有四種方法中，六個空氣品質監測站的 PM2.5模型都一致顯現出使用大氣頂反射率波段、氣象變項、氣體污染物、與主成分分析為最佳模型（本研究的方法 4 ）。小港測站的PM2.5濃度推估模型預測能力較強（CV R2 = 0.70），其餘測站濃度推估模型則有中等偏強的預測能力（CV R2 >= 0.50）。
在PM成分預測模型中，方法4也顯示出最佳的模型表現，各測站以小港和忠明測站的預測模型表現較佳，而花蓮測站預測模型表現較差。銅、鎘、鐵、錳和鋅，在不同測站的預測模型中多有良好的表現，這些成分都有至少三至四個測站的CV R2 >= 0.40。本研究也呈現透過模型預測2017至2020年各測站的錳推估濃度，以達到在無量測日期填補推估值之目的。
本研究發現大氣頂反射率、氣體污染物和氣象變項有助於推估 PM 成分濃度，大氣頂反射率波段可用於不同預測模型中。但本研究的模型未能把推估模型應用在六處測站以外的測站和地區，未來研究可考慮在更多地點及更頻繁地量測台灣不同地區的PM 成分濃度，讓大氣頂反射率推估模型能預測和了解台灣不同地方的PM 成分濃度。

關鍵字

PM2.5 ； PM 成分 ； 統計模型 ； 大氣頂反射率 ； 主成分分析

並列摘要

Particulate matter with an aerodynamic diameter <2.5 μm (PM2.5) has been suggested to be associated with adverse health outcomes. The annual concentration of PM2.5 has been gradually declining in Taiwan, but it is still exceeding the World Health Organization’s air quality guideline of 10 μg/m3. Due to the amount of people being exposed to PM2.5, more attention must be placed on this issue.
Numerous estimation models from studies have been developed to predict PM2.5 concentrations, offering a lot of PM2.5 concentration information. While PM2.5 has been sufficiently addressed for years, it has been pointed out that specific PM compositions, like cadmium, vanadium, and zinc, have more harm on the human body than others. Thus, one of objectives of this study was to estimate PM2.5 and PM composition concentration at six air quality monitoring stations (AQMS) in Taiwan through statistical modelling.
A total of six AQMSs (Banqiao, Zhongming, Douliu, Chiayi, Xiaogang, and Hualien), which had a 6-day sampling interval, were selected for PM2.5 and PM composition model building. This study used the top of atmosphere reflectance (TOAR) from the satellite Himawari-8 and other variables using stepwise Akaike Information Criterion (AIC) regression models. This study introduced four different methods for model building, including the addition of meteorological factors (wind speed, ambient temperature, and relative humidity), gaseous pollutants (SO2, NO2, NOx, NO, and CO), and the application of principal component analysis (PCA) in stepwise AIC models.
Out of all the four methods, all PM2.5 models for each station had the best models using TOAR bands, meteorological factors and gaseous pollutant with PCA in a stepwise AIC model (Method 4 of this study). The PM2.5 model at Xiaogang station had a strong predictive ability (CV R2 = 0.70), while rest of the stations had moderate to moderately strong predictive ability.
Likewise, PM composition models had the most satisfactory model performances (CV R2 > 0.40) using Method 4. Zhongming station had the most number of satisfactory model performances (11 models), and Hualien station had the least (4 models). Several PM compositions, including Cu, Cd, Fe, Mn and Zn, were able to perform with a CV R2 above 0.40 across stations, with at least three to four stations showing satisfactory performances. For local application, estimated Mn concentrations at different stations from 2017 to 2020 were used to fill-in information on unsampled days at different AQMSs.
This study showed that TOAR band information along with gaseous pollutants, and meteorological were helpful for both PM2.5 and PM composition concentration estimation. Prediction models suggested the potentials of TOAR bands to be employed in different modelling techniques. However, a limitation was that the models built in this study could not be interpreted and applied beyond the six locations, due to fixed AQMS locations. Future research shall consider to have more frequent samplings at extra locations to have better understandings of the PM composition levels at different regions of Taiwan.

並列關鍵字

PM2.5 ； PM composition ； statistical model ； top of atmosphere reflectance ； principal component analysis

參考文獻

1.Abdi, H., Williams, L. J. (2010). Principal component analysis. WIREs Computational Statistics, 2(4), 433-459. doi:https://doi.org/10.1002/wics.101

2.Aryal, A., Harmon, A. C., Dugas, T. R. (2021). Particulate matter air pollutants and cardiovascular disease: Strategies for intervention. Pharmacology Therapeutics, 223, 107890. doi:https://doi.org/10.1016/j.pharmthera.2021.107890

3.Baccarelli, A., Martinelli, I., Zanobetti, A., Grillo, P., Hou, L. F., Bertazzi, P. A., . . . Schwartz, J. (2008). Exposure to particulate air pollution and risk of deep vein thrombosis. Arch Intern Med, 168(9), 920-927. doi:10.1001/archinte.168.9.920

4.Bai, H., Zheng, Z., Zhang, Y., Huang, H., Wang, L. (2021). Comparison of Satellite-based PM2.5 Estimation from Aerosol Optical Depth and Top-of-atmosphere Reflectance. Aerosol and Air Quality Research, 21(2), 200257. doi:10.4209/aaqr.2020.05.0257

5.Barnett, A. G., Williams, G. M., Schwartz, J., Best Trudi, L., Neller Anne, H., Petroeschevsky, A. L., Simpson, R. W. (2006). The Effects of Air Pollution on Hospitalizations for Cardiovascular Disease in Elderly People in Australian and New Zealand Cities. Environmental Health Perspectives, 114(7), 1018-1023. doi:10.1289/ehp.8674