How does convolution in neural networks improve hourly PM 2.5 estimates

Authors: Yogita Karale*, University of Texas At Dallas, May Yuan, University of Texas at Dallas
Topics: Geographic Information Science and Systems, Remote Sensing
Keywords: PM2.5; MAIAC; AOD; CNN
Session Type: Virtual Paper
Day: 4/10/2021
Start / End Time: 3:05 PM / 4:20 PM
Room: Virtual 43
Presentation File: No File Uploaded


Convolutional neural networks combine inputs from adjacent units to determine the activation functions to the next layer, which provides a means to consider surrounding influences on PM2.5 concentration at locations. This study investigates the effects of the varying sizes of convolutional layers on hourly PM2.5 estimates using high resolution MODIS Aerosol Optical Depth (AOD) and the spatial extent to which predictors from the surrounding area impact the PM2.5 at a location. Many studies use machine learning methods to estimate 24-hr averages of PM2.5, but research is lacking for hourly PM2.5 to improve measures of human exposure and health outcome. Furthermore, no study has investigated the effect of varying size of convolutional layers on PM2.5 estimation. This study uses AOD and meteorological parameters as predictors to estimate hourly PM2.5 in Dallas Fort-Worth and surrounding area by using 10 years of data from 2006-2015. The study analyzes convolutional layers of sizes 3 by 3, 5 by 5,……,19 by 19. Models without considerations of predictors from surrounding area perform more poorly than the other models with surrounding predictors. The study concludes that the model performance, in terms of correlation coefficients and root mean squared error (RMSE), improves as the size of the convolutional layers increases. The convolution layer of size 19 by 19 gives best results with correlation coefficients of 0.87 and RMSE of 2.4.

Abstract Information

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

To access contact information login