Augmenting limited background monitoring data for improved performance in land use regression modelling: Using support vector regression and mobile monitoring. (15th March 2019)

• Record Type: Journal Article
• Authors: Basu, Bidroha; Alam, Md Saniul; Ghosh, Bidisha; Gill, Laurence; McNabola, Aonghus
• Is Part Of: Atmospheric environment. Volume 201(2019)
• Page Start: 310
• Page End: 322
• Publication Date: 2019-03-15
• Subjects: Land-use regression -- PM2.5 -- Mobile monitoring -- Support vector regression; Air -- Pollution -- Periodicals; Air -- Pollution -- Meteorological aspects -- Periodicals; 551.51
• DOI: 10.1016/j.atmosenv.2018.12.048
• Languages: English
• ISSNs: 1352-2310

Abstract

Abstract: Land Use Regression (LUR) models was developed using fixed site and mobile station data in Dublin Ireland. The uncertainty associated with short term, mobile station data was accounted for using a weighted function, considering the differences in record duration of data across the study. A systematic buffer distance decay curve-based approach was also considered in this study to identify appropriate predictors and their associated buffer distance. The analysis was performed in Dublin city for fine particulate matter (PM2.5 ) where 10 fixed monitoring stations data were available, with average daily concentration data records of 1–7 years in length. 105 mobile station data were also collected at several locations in the city for durations of 12 h to 3 days. The performance of a developed multiple linear regression based LUR model was evaluated by using the leave-one-out-cross-validation procedure. The model explained variances (R 2 ) of the LUR model were found to be 0.219 for winter and 0.688 for summer months. One of the primary reasons for poor performance of the LUR model was that the relationship between the predictors and the mean PM2.5 concentration were found to be non-linear which cannot be modelled by using multiple linear regression approach. To address this, a weighted support vector regression (WSVR) approach was also considered that can account for non-linearity while developing the LUR model. The model performance in terms of R 2 was found to improve to 0.912 for winter and 0.916 for summer using this method. The developed WSVR model can be used to predict PM2.5 concentration at any ungauged locations in the study area with considerable accuracy, which can be the basis for future epidemiological studies. Highlights: Weighted Support Vector Regression based nonlinear land use regression were developed. The model effectively capture spatial variation of PM2.5 air pollution in Dublin. Combination of fixed and mobile station data were used in the analysis. Predictors were identified based on systematic buffer distance decay curve approach.