A spatially structured adaptive two-stage model for retrieving ground-level PM2.5 concentrations from VIIRS AOD in China

Abstract

While the aerosol optical depth (AOD) product from the Visible Infrared Imaging Suite (VIIRS) instrument has proven effective for estimating regional ground-level particle concentrations with aerodynamic diameters less than 2.5 μm (PM 2.5 ), its performance at larger spatial scales remains unclear. Despite the wide application of statistical models in building ground-level PM 2.5 satellite remote sensing retrieval models, a limited number of studies have considered the spatiotemporal heterogeneities for model structures. Taking China as the study area, we used the VIIRS AOD, together with multi-source auxiliary variables, to develop a spatially structured adaptive two-stage model to estimate ground-level PM 2.5 concentrations at a 6-km spatial resolution. To this end, we first defined and calculated a dual distance from the ground-level PM 2.5 monitoring data. We then applied the unweighted pair-group method with arithmetic means on dual distances and obtained 13 spatial clusters. Subsequently, we combined the time fixed effects regression (TEFR) model and geographically weighted regression (GWR) model to develop the spatially structured adaptive two-stage model. For each spatial cluster, we examined all possible combinations of auxiliary variables and determined the best model structure according to multiple statistical test results. Finally, we obtained the PM 2.5 estimates through regression mapping. At least seven model-fitting data records per day made a good threshold that could best overcome the model overfitting induced by the second-stage GWR model at the minimum price of losing samples. The overall model fitting and ten-fold cross validation (CV) R 2 were 0.82 and 0.60, respectively, under that threshold. Model performances among different spatial clusters differed to a certain extent. High-CV R 2 values always exceeded 0.6 while low-CV R 2 values less than 0.5 also existed. Both the size of the model-fitting data records and the extent of urban-industrial characteristics of spatial clusters accounted for these differences. The PM 2.5 estimates agreed well with the PM 2.5 observations with correlation coefficients all exceeding 0.5 at the monthly, seasonal, and annual scales. East of Hu's line and north of the Yangtze River were characterized by high PM 2.5 concentrations. This study contributes to the understanding of how well VIIRS AOD can retrieve ground-level PM 2.5 concentrations at the national scale and strategies for building ground-level PM 2.5 satellite remote sensing retrieval models. © 2019