Using the Mixed Effect Model as an Alternative Approach to Improve Correlation between Satellite Derived Aerosol Optical Depth (MISR & MODIS) and Ground Measured PM2.5 Data

Abstract:

The study seeks to determine the efficacy of using aerosol optical depth (AOD) data from MISR and MODIS as a surrogate for ground-based particulate matter (PM_2.5) data by using AOD as an input for various computational methods. The data set used in the study ranged from January 2011 to December 2012. The advantage of the mixed effects model is in its ability to consider temporally changing attributes through the inclusion of random effects in the regression model. The study first established that MISR and MODIS AOD has a correlation with ground measured PM_2.5 through regression analysis thereby providing rationale for further analysis. The regression analyses resulted in an R² of 0.7513 and 0.7536 for MODIS and MISR, respectively. With the rationale established, data quality improvement measures were carried out through data screening and empirical correction. The data screening process involved the removal of data entries in which the absolute difference of MODIS and MISR AOD values deviated far more than the average of the data set. On the other hand, empirical correction was done by developing correction equations through multivariate regression with ground parameters such as AERONET AOD, relative humidity, and wind speed. Both methods were found to yield marked improvement in the correlation of satellite-derived AOD with PM_2.5. After data quality had been improved, several computational methods are assessed by solving for the R² and absolute error percentage. The methods are simple linear regression with MODIS (R² = 0.7764, 18.43%) and MISR (R² = 0.7614, 17.99%), multivariate linear regression with MODIS and MISR together (R² = 0.8721, 13.63%), artificial neural network with MODIS and MISR as inputs (R² = 0.8764, 13.45%), and the mixed effects model with MODIS and MISR as predictors (R² = 0.9793, 5.20%). Among these, the mixed effects model performed the best and further error analysis showing an error that was independent on seasonality and dependent on the PM concentration value. The figure below shows the correlation between ground measured PM_2.5 concentration with computed PM_2.5 concentration derived from the application of the mixed effect model. This illustrates the capacity of the mixed effects model to compute for the actual PM2.5 data.