Improving the Accuracy of PM2.5 Concentration Prediction Using Localized Explanatory Factors: A Comparison of GWR, Multiscale GWR, GW Lasso and GW Elastic Net

Abstract

In this paper, PM_2.5 concentrations are predicted for all counties in China, using the geographically weighted regression (GWR), the geographically weighted Lasso, the geographically weighted Elastic net, and multiscale GWR models. Predictor variables include spatially localized county-level economic activities, population, road network, land use, aerosol optical depth, meteorological and topographic factors. Economic, population, road network and land use data are localize (within 8 Km from the locations studied) to improve the accuracy of the prediction. We found that incorporation of geographic weights into the Lasso and Elastic net models cannot enhance the prediction capacity of them. Multiscale GWR can partially correct the underestimation problem of the GWR model, but presents a lower cross validation R2, and proves to be a time-consuming algorithm. Among those models, GWR is the best model with the highest cross validation R2 (0.8276), and lowest RMSE (7.4752), MAE (5.3904) and MAPE (0.1127). The county-level PM_2.5 concentration map predicted by GWR is presented.