Estimation of PM2.5 Concentrations in the Southeastern U.S. Using a Two-Stage Model by Incorporating Fire Count Data

摘要

Previous studies showed that fires are significant contributors to PM2.5 levels. To date, fire data, however, have not been used as a predictor in AOD-based models for PM2.5 concentration prediction. In this paper, we developed a two-stage model by incorporating a new aerosol product with 1 km spatial resolution derived by the Multiangle Implementation of Atmospheric Correction (MAIAC) algorithm and fire count data in order to evaluate whether fire count data can improve prediction accuracy, particularly in areas with high fire occurrence. The study area is in the southeastern U.S., and the data of year 2007 was collected from various sources. Fire data were linked to each PM2.5 monitoring site by calculating the fire counts within a 75 km buffer centered at each site. Model fitting generated R2 of 0.75, MPE of 2.47 μg/m3, and RMSPE of 3.83 μg/m3, while cross validation (CV) generated R2 of 0.69, MPE of 2.75 μg/m3, and RMSPE of 4.29 μg/m3, indicating a good fit between the dependent variable and predictor variables. A comparison between models with and without the fire predictor showed that the prediction accuracy was improved from the model with the fire predictor to the one without as fire counts increased. The decrease of CV RMSPE was up to 1.5 μg/m3, representing a 13.4% improvement in prediction accuracy, and fire count data have better performance in southern Georgia and spring season due to high fire occurrence in this region and season. The results indicate that fire count data are essential in PM2.5 concentration estimation, especially in areas where fires constantly occur.