Using a land use regression model with machine learning to estimate ground level PM_2.5

Highlights

• Estimating long-term daily PM_2.5 concentration with machine learning models.

• Land-use patterns were included in machine learning models by using land-use regression.

• Explanatory power of daily PM_2.5 concentration was increased from 0.58 to 0.94.

• XGboost outperformed random forest and deep neural network algorithms.

Abstract

Ambient fine particulate matter (PM_2.5) has been ranked as the sixth leading risk factor globally for death and disability. Modelling methods based on having access to a limited number of monitor stations are required for capturing PM_2.5 spatial and temporal continuous variations with a sufficient resolution. This study utilized a land use regression (LUR) model with machine learning to assess the spatial-temporal variability of PM_2.5. Daily average PM_2.5 data was collected from 73 fixed air quality monitoring stations that belonged to the Taiwan EPA on the main island of Taiwan. Nearly 280,000 observations from 2006 to 2016 were used for the analysis. Several datasets were collected to determine spatial predictor variables, including the EPA environmental resources dataset, a meteorological dataset, a land-use inventory, a landmark dataset, a digital road network map, a digital terrain model, MODIS Normalized Difference Vegetation Index (NDVI) database, and a power plant distribution dataset. First, conventional LUR and Hybrid Kriging-LUR were utilized to identify the important predictor variables. Then, deep neural network, random forest, and XGBoost algorithms were used to fit the prediction model based on the variables selected by the LUR models. Data splitting, 10-fold cross validation, external data verification, and seasonal-based and county-based validation methods were used to verify the robustness of the developed models. The results demonstrated that the proposed conventional LUR and Hybrid Kriging-LUR models captured 58% and 89% of PM_2.5 variations, respectively. When XGBoost algorithm was incorporated, the explanatory power of the models increased to 73% and 94%, respectively. The Hybrid Kriging-LUR with XGBoost algorithm outperformed the other integrated methods. This study demonstrates the value of combining Hybrid Kriging-LUR model and an XGBoost algorithm for estimating the spatial-temporal variability of PM_2.5 exposures.

Introduction

The health effects of ambient fine particulate matter (PM) have been broadly investigated worldwide for decades and PM ranked as the sixth leading risk factor globally for death and disability (Forouzanfar et al., 2016). It is well documented that PM_2.5 (PM with an aerodynamic diameter < 2.5 μm) is associated with various short-term and long-term adverse health effects, including increased risk of respiratory diseases, cardiovascular mortality, type 2 diabetes mellitus, hypertensive disorders of pregnancy, neurological hospital admissions, and premature mortality (Bai et al., 2020; Kioumourtzoglou et al., 2016; Pope III & Dockery, 2006). As such, a more precise method to assess PM_2.5 exposures is needed in environmental epidemiological studies. Some prior studies estimated PM_2.5 concentrations by applying data from sparse monitoring stations. This method is not feasible for large cohort studies, particularly those focused on rural areas. Hence, environmental modelling methods need to be developed in order to capture PM_2.5 spatial and temporal continuous variations at a sufficient resolution based on limited monitoring stations.

Previous studies have used aerosol optical depth (AOD) to retrieve ground-level PM estimations or have used it as a predictor variable for model improvement (Di et al., 2019; Shtein et al., 2019; Stafoggia et al., 2019). Although satellite-derived images could provide AOD properties which cover an almost global surface at a moderate spatial resolution, these remote sensing images are easily affected by cloudy weather and water/snow glint reflectance (Sayer et al., 2013; Stafoggia et al., 2019). Because of this limitation, AOD products have commonly been applied for continents and not islands or areas with cloudy weather conditions (Wu et al., 2018). For instance, Taiwan is an island and could have missing data of AOD at a rate higher than 70%. Therefore, an alternative method to estimate PM should be developed for areas such as Taiwan that may have too much missing data.

Land-Use Regression (LUR) has been used to depict intra-urban air pollution concentration variation in fine spatial-temporal resolution worldwide because of its ability to consider different types of land-use variables in assessing target pollutants (Beelen et al., 2013; Eeftens et al., 2012; Wu et al., 2017; Young et al., 2016). LUR models use a set of geographic sources as predictor variables to develop multiple linear regressions in order to estimate air pollution concentrations. A more defined land-use predictor dataset was applied because it offers the potential to obtain a LUR model with greater explanatory power. Thus, land-use/land cover dataset could play an important role in affecting model performance (Beelen et al., 2013; Eeftens et al., 2012; Hellack et al., 2017). The strength of LUR is that important predictor variables that affect air pollution concentration can be identified by a stepwise variable selection procedure, thereby reducing the dimension of predictor variables in the resultant models. Recently, more research in the field of air pollution modelling has utilized non-linear statistical methods, such as generalized additive model, to improve estimation accuracy (Li et al., 2013; Yang et al., 2018; Zou et al., 2017) and machine learning algorithms (Chen et al., 2019; Di et al., 2019; Shtein et al., 2019; Stafoggia et al., 2019; Weichenthal et al., 2016). Among them, machine learning models, with their ability to capture non-linear relationships between predictor variables and air pollution concentrations, improve a model’s prediction performance more than do traditional regression approaches. Araki et al. (2018) used random forest (RF) to estimate metropolitan NO₂ variation in Japan, and the results of RF outperformed traditional regression approaches (Araki et al., 2018). Another study developed multiple modelling approaches to estimate air pollution concentration in Europe, and it also concluded that the results of RF and artificial neural networks were better than linear regression methods (Chen et al., 2019). A limited number of studies have used extreme gradient boosting (XGBoost) algorithms to estimate air pollution concentration. Offering the ability to enable slow learners to investigate the bias of a loss function, XGBoost has potential to outperform some tree-based machine learning algorithms (Chen and Guestrin, 2016). However, one of the shortcomings of machine learning models is that most of them are unable to select important variables before training each algorithm. Training estimation models with an enormous amount of predictor variables may result in overfitting issues and may result in reduced computational efficiency. Due to the lack of variable selection procedures, predictors used in these algorithms may have limited interpretability and limited reasonable explanations for variations in air pollution. Thus, it is important to use a variable selection method before training a machine learning algorithm.

Given that the shortcoming of machine learning models in selecting proper predictor variables can be solved by applying LUR, this study aims to use an integrated approach combining LUR and machine learning models to improve the daily concentration estimation of PM_2.5 during 2000–2016 on the main island of Taiwan. In detail, important predictor variables identified by the stepwise variable selection from the LUR procedures were applied to develop the prediction models using three types of machine learning algorithms, namely deep neural network (DNN), RF, and XGBoost for improving the accuracy of PM_2.5 variation predictions. The mixed spatial prediction models that combine the strength of LUR in identifying the most influential emission predictors and the predictability of machine learning in estimating non-linear trends would be more broadly effective than techniques that rely solely on LUR or machine learning.