Multiscale Spatial and Temporal Modelling of Fine Particulate Matter (PM2.5) from Wildfire Smoke Using Remote Sensing and Statistical Methods
Date
2020-04-22
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Wildfire smoke exposure is increasingly recognized as a critical public health problem due to the increase in the frequency and severity of wildfires in recent decades. Wildfire smoke-generated PM2.5 is considered the most concerning particulate, as it can be inhaled deep into the lungs, penetrate the human respiratory system, and enter the bloodstream. Studies assessing wildfire PM2.5 exposure and population health have traditionally employed three main approaches: (1) in situ measurements (2) satellite information of atmospheric aerosol, and (3) atmospheric models. The objective of the present study was to identify models to accurately predict PM2.5 concentration over space and time. Spatiotemporal models were built to perform a comprehensive analysis of wildfire PM2.5 concentrations, for each recent year over the study region: Land Use Regression (LUR), Linear Mixed Effect (LME), and Artificial Neural Network (ANN). Predictor variables were MODIS AOD images, ground PM2.5 measurements, and ancillary land use and meteorological data. LUR models were used to predict PM2.5 in three distinct periods: before, during, and after a wildfire. Results showed a major difference in predictors between the during-fire and the other models, due to the different contribution of traffic and industrial emissions. Daily estimation of PM2.5 concentration was derived by incorporating nested period-zone-specific random effects of the AOD-PM2.5 relationship over the province of Alberta, Canada using LME models. The LME model’s predictions also improved when additional variables were integrated with AOD measures in a multivariate framework. ANN models were used as a multivariate and non-parametric approach to empirically predict wildfire smoke using AOD along with other predictors. Daily PM2.5 concentrations were predicted using temporal and spatial ANN for the 2014 to 2017 fire seasons and each airshed zone in Alberta. The study demonstrated how MODIS AOD could be incorporated within statistical models to provide reliable predictions of daily PM2.5 concentrations over wildfire events, to feed health and epidemiological studies. The results demonstrated that mixed effect models outperformed the LUR models owing to their ability to adjust the varying relationship of AOD-PM2.5. ANN models also outperformed them owing to their advantage in modelling non-parametric and non-linear behaviours.
Description
Keywords
Fine particulate matter (PM2.5), AOD (aerosol optical depth), Wildfire Smoke, LUR (land use regression), spatial analysis, public health, Linear mixed effect (LME) model, Spatiotemporal modelling, Artificial neural network (ANN), multi-layer perceptron (MLP)
Citation
Mirzaei, M. (2020). Multiscale Spatial and Temporal Modelling of Fine Particulate Matter (PM2.5) from Wildfire Smoke Using Remote Sensing and Statistical Methods (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.