Prediction of PM2.5 daily concentrations for grid points throughout a vast area using remote sensing data and an improved dynamic spatial panel model. (15th September 2020)

• Record Type: Journal Article
• Authors: Fu, Miao; Kelly, J. Andrew; Clinch, J. Peter
• Is Part Of: Atmospheric environment. Volume 237(2020)
• Publication Date: 2020-09-15
• Subjects: PM2.5 concentration -- Out-of-sample prediction -- Temporal-spatial correlation -- Dynamic spatial panel model -- Remote sensing; Air -- Pollution -- Periodicals; Air -- Pollution -- Meteorological aspects -- Periodicals; 551.51
• DOI: 10.1016/j.atmosenv.2020.117667
• Languages: English
• ISSNs: 1352-2310

Abstract

Abstract: The incorporation of spatial and temporal correlations can significantly improve the accuracy of PM2.5 concentration prediction models. However, the dynamic spatial panel model which explicitly deals with these two correlations remains absent from current approaches to out-of-sample concentration prediction. An issue is that the prediction of daily concentrations for grid points across a vast area may well overwhelm existing algorithms, as it requires an enormous amount of computational resources and an enlarged spatial weight matrix. This paper develops improved algorithms that address this issue. The dynamic spatial panel approaches used in this paper predict daily series PM2.5 concentrations for grid points covering Mainland China, using daily aerosol, vegetational and meteorological remote sensing data as the explanatory variables. The predicted concentration maps offer more realistic detail in areas where monitoring stations are sparse. Indeed, the error map for the out-of-sample prediction shows that MAPE is less than 30% in most regions, and the average MAPE is 24.28%, which is relatively low compared with similar studies. In contrast to methods which cannot provide coefficients of variables, the developed method offers coefficients to assess the contributions of explanatory variables and temporal-spatial correlation terms, allows quantification of convergence effects, and can distinguish between spillover effects and local effects. A performance comparison of models with various spatial weight matrices shows that model achieves the optimal fitting levels by using the neighbouring unit number threshold of 18 or the distance threshold of 150 km. The case analysis in this paper finds that spillover effects are about three times larger than local effects, and the spatial correlation is greater than the cumulative effects of earlier concentrations. This finding adds further weight to the notion that management of PM2.5 pollution and associated impacts requires multi-regional and even multi-national coordination and effort. Highlights: Increased prediction accuracy by including the temporal-spatial correlation of the pollutant. Improved dynamic spatial panel approaches that can handle N = 1462 and T = 363 data set. Out-of-sample prediction with extended spatial weight matrices and optimal thresholds. Increased frequency and coverage of predicted concentrations using remote sensing data. Coefficients, convergence effects, local and neighbouring effects for policy analysis.