Akademik Veri Yönetim Sistemi
AIR QUALITY, ATMOSPHERE AND HEALTH, cilt.15, sa.12, ss.2283-2297, 2022 (SCI-Expanded)
Estimating ground-level PM2.5 concentrations by satellite-based aerosol optical depth (AOD) does not provide spatially and temporally continuous estimations due to the gaps in the AOD data. In this study, without using satellite-based AOD, spatiotemporally continuous aerosol diagnostic products such as surface mass concentrations of dust, sea salt, black carbon, organic carbon, and sulfate from the Version 2 Modern-Era Retrospective Analysis for Research and Applications (MERRA-2) were used to estimate daily PM2.5 concentrations at 94 air quality monitoring stations for 2016-2019 in the Eastern Mediterranean. The results indicated that the calculated PM2.5 concentrations from the MERRA-2 aerosol diagnostics could not sufficiently capture the spatiotemporal variation of the PM2.5 observations and nearly consistently underestimated the PM2.5 concentrations. Therefore, the non-linear autoregressive network with exogenous inputs (NARX) was used to estimate the PM2.5 concentrations with the support of the aerosol diagnostics, Angstrom exponent, and meteorological parameters from the MERRA-2 reanalysis. The NARX model provided robust and accurately estimated PM2.5 concentrations by taking advantage of the neural network approach with an R-squared of 0.73, the root means squared error of 10.6 mu g/m(3), the mean absolute error of 6.4 mu g/m(3), and the mean relative error of 15.5%. The seasonal and site-scale performances of the model were also discussed. Autumn had the highest accuracy (R-2 = 0.72), whereas spring had the lower accuracy (R-2 = 0.53). Finally, the overall performance of the developed NARX model was also compared with the other two ensemble tree-based models, such as random forest and XGBoost. The NARX model performed better than all the models with reduced estimation errors and gave the best performance with better statistical indicators and minor uncertainties overall. This study proposes a new understanding of the relationship between MERRA-2 aerosol diagnostics and meteorology. This work will also provide more accurate MERRA-2 PM2.5 data with the artificial neural network (ANN)-based calibration over the region for further studies.