Predicting Fine Particulate Matter (PM2.5) in the Greater London Area: An Ensemble Approach using Machine Learning Methods

Abstract

Estimating air pollution exposure has long been a challenge for environmental health researchers. Technological advances and novel machine learning methods have allowed us to increase the geographic range and accuracy of exposure models, making them a valuable tool in conducting health studies and identifying hotspots of pollution. Here, we have created a prediction model for daily PM2.5 levels in the Greater London area from 1st January 2005 to 31st December 2013 using an ensemble machine learning approach incorporating satellite aerosol optical depth (AOD), land use, and meteorological data. The predictions were made on a 1 km × 1 km scale over 3960 grid cells. The ensemble included predictions from three different machine learners: a random forest (RF), a gradient boosting machine (GBM), and a k-nearest neighbor (KNN) approach. Our ensemble model performed very well, with a ten-fold cross-validated R2 of 0.828. Of the three machine learners, the random forest outperformed the GBM and KNN. Our model was particularly adept at predicting day-to-day changes in PM2.5 levels with an out-of-sample temporal R2 of 0.882. However, its ability to predict spatial variability was weaker, with a R2 of 0.396. We believe this to be due to the smaller spatial variation in pollutant levels in this area.

A satellite-based spatio-temporal machine learning model to reconstruct daily PM2.5 concentrations across Great Britain

“…The spatio-temporal ML models presented here demonstrated comparable predictive performance to similar methods applied in other countries, based either on single-learner ML models [de Hoogh et al, 2018; Stafoggia et al, 2019], ensemble ML models [Chen et al, 2018; Di et al, 2019; Yazdi et al, 2020], or generalised additive models (GAM) [Kloog et al, 2015]. Yazdi et al (2020) developed an ensemble ML model (composed by RF, Deep Neural Network, GAM, Gradient Boosting, K-nearest Neighbour) to estimate PM 2.5 for Greater London, reaching mean 2005-2013 CV spatial-R 2 of 0.396. Using 2008-2013 period, this study reached CV spatial-R 2 of 0.637 for the whole Great Britain.…”

“…Country-wide maps generated by emission-dispersion models are usually available at coarser spatial or temporal resolution [DEFRA, 2020; EMEP4UK, 2020; Savage et al, 2013, and generally they show lower small-scale accuracy when tested against observed monitoring data [Hood et al, 2018; Lin et al, 2017]. The spatio-temporal ML models presented here demonstrated comparable predictive performance to similar methods applied in other countries, based either on single-learner ML models [de Hoogh et al, 2018; Stafoggia et al, 2019], ensemble ML models [Chen et al, 2018; Di et al, 2019; Yazdi et al, 2020], or generalised additive models (GAM) [Kloog et al, 2015]. Yazdi et al (2020) developed an ensemble ML model (composed by RF, Deep Neural Network, GAM, Gradient Boosting, K-nearest Neighbour) to estimate PM 2.5 for Greater London, reaching mean 2005-2013 CV spatial-R 2 of 0.396.…”

“…Such studies used various analytical methods, including multiple linear regression [Gupta and Christopher, 2009], land-use regression [Beckerman et al, 2013; Vienneau et al, 2010], and mixed effect models [de Hoogh et al, 2018; Kloog et al, 2011; Kloog et al, 2015; Lee et al, 2011; Staffogia et al, 2016]. The last development in this research area is represented by the application of machine learning (ML) algorithms, including various architectures such as random forests [Chen et al, 2018; Di et al, 2019; Staffogia et al, 2019; Yazdi, et al, 2020], neural network [Di et al, 2019; Yazdi, et al, 2020], and gradient boosting [Chen et al, 2019; Di et al, 2019; Just et al, 2018; Zan et al, 2017]. These have demonstrated higher performances, linked with the ability to model any kind of predictor(s)-response association and to deal better with potentially complex relationship between PM 2.5 , spatial, and spatio-temporal predictors [Di, et al, 2019; Polley et al, 2010].…”

Prediction and assessment of change in PM2.5 during COVID-19 lockdown using remote sensing and deep learning approach: A case study of Kanpur city

“…As all the methods used are data driven model may be inclusion of more number of data may have provided better accuracy. Even though the results obtained by using this hybrid method for prediction of PM2.5 is not too high as 0.97 obtained by (Amanollahi and Ausati 2020) still, comparatively better than some of the results obtained by (Joharestani et al 2019;Yazdi et al 2020). Thus, this method can easily be implemented because of its robustness as it has can be used for prediction as well as spatial representation of PM2.5 compared to other approaches used.…”

“…(Joharestani et al 2019) has implemented random forest, extreme gradient boosting, and machine learning (ML) approaches along with 23 features and obtained better prediction results. (Yazdi et al 2020) has used and compared various machine learning approach and found out that the random forest (RF) technique performs the best in comparison with all other method used. (Shahriar et al 2020) has used the linear-support vector machine (L-SVM), medium Gaussian-support vector machine (M-SVM), Gaussian process regression (GPR), arti cial neural network (ANN), random forest regression (RFR), and a time series model namely PROPHE and found out that GPR and ANN perform better in (Ventura et al 2019) has applied two models, Holt-Winters (HW) and arti cial neural network (ANN), using PM 2.5 concentration time series and found ANN performs better with less RMSE error.…”