Missing MAIAC AOD imputation with quantified uncertainty and its implications on AOD derived PM2.5 levels

Authors: Qiang Pu*, University at Buffalo, SUNY, Eun-hye Yoo, University at Buffalo, SUNY
Topics: Environmental Science, Remote Sensing, Spatial Analysis & Modeling
Keywords: Aerosol Optical Depth (AOD); fine particulate matter (PM2.5); AOD imputation; uncertainty evaluation; machine learning
Session Type: Virtual Paper
Day: 4/10/2021
Start / End Time: 3:05 PM / 4:20 PM
Room: Virtual 8
Presentation File: No File Uploaded


Satellite-derived aerosol optical depth (AOD) has been widely employed to estimate ground-level fine particulate matter (PM2.5) concentrations. A notable limitation for the use of satellite AOD is the substantial amount of missing AOD retrievals due to cloud covers and conditions of high surface reflectance. The missing AOD data thus may yield large gaps in the subsequent ground PM2.5 estimates, which usually lead to biased PM2.5 exposure assessment. Despite recent attempts to address this issue by imputing missing satellite AOD values to generate full coverage PM2.5 estimates, none has investigated the uncertainty associated with imputed AOD for PM2.5 predictions. In this study, we developed a missing satellite AOD imputation model and examined how the uncertainty associated with the imputed AOD is propagated to PM2.5 estimation using machine learning based methods. Our AOD imputation method showed good performance with a cross-validated R2 of 0.94 and RMSE of 0.017 and quantified a high level of imputation uncertainty (~49%). We further investigated the uncertainty propagation for PM2.5 predictions using multiple machine learning methods: deep neural network (DNN), random forest (RF), gradient boosting machine (GBM), and an ensemble model. The resulting annual averaged uncertainties in PM2.5 predictions from DNN, RF, GBM, and the ensemble model were 10.99%, 7.32%, 9.74%, and 9.65%, respectively, which suggested notable levels of uncertainty in the AOD-based PM2.5 estimates. This study highlights the importance of uncertainty quantification for AOD imputation and its impacts on downstream PM2.5 predictions, which are helpful for PM2.5 health effects investigation in epidemiological studies.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login