Evaluation of linear, nonlinear, and hybrid models for predicting PM2.5 based on a GTWR model and MODIS AOD data

Ground monitoring station data of PM2.5 are not available for each day and all places in urban areas. In this research, taking Tehran as an example, a geographically and temporally weighted regression (GTWR) model was utilized to investigate the spatial and temporal variability relationship between PM2.5 concentrations measured at ground monitoring stations and satellite aerosol optical depth (AOD) data. The Moderate Resolution Imaging Spectroradiometer (MODIS) sensor produced AOD values with 3-km spatial resolution. Using meteorological variables and land use information as additional predictors utilized in the GTWR model, the AOD was converted to PM2.5 at ground level for warm (October to March) and cold seasons (April to September) from 2011 to 2017. To improve the accuracy of the correlation coefficient between converted PM2.5 from the GTWR model and PM2.5 concentrations measured at ground monitoring station, the results of a linear model (LR), a nonlinear model (artificial neural network (ANN)), and hybrid models including general regression neural network (GRNN) and adaptive neuro-fuzzy inference system (ANFIS) were compared. The results of the linear, nonlinear, and hybrid models for the cold season display higher accuracy compared with the results for the warm season. Among the used models, the GRNN model has higher accuracy compared with the other models. This study reveals that AOD conversion to particulate matter by the GTWR model and its simulation to PM2.5 at ground level using a hybrid model such as the GRNN can be used to study air quality.