Satellite Based Mapping of Ground PM2.5 Concentration Using Generalized Additive Modeling

Abstract

Satellite-based PM 2.5 concentration estimation is growing as a popular solution to map the PM 2.5 spatial distribution due to the insufficiency of ground-based monitoring stations. However, those applications usually suffer from the simple hypothesis that the influencing factors are linearly correlated with PM 2.5 concentrations, though non-linear mechanisms indeed exist in their interactions. Taking the Beijing-Tianjin-Hebei (BTH) region in China as a case, this study developed a generalized additive modeling (GAM) method for satellite-based PM 2.5 concentration mapping. In this process, the linear and non-linear relationships between PM 2.5 variation and associated contributing factors, such as the aerosol optical depth (AOD), industrial sources, land use type, road network, and meteorological variables, were comprehensively considered. The reliability of the GAM models was validated by comparison with typical linear land use regression (LUR) models. Results show that GAM modeling outperforms LUR modeling at both the annual and seasonal scale, with obvious higher model fitting-based adjusted R 2 and lower RMSEs. This is confirmed by the cross-validation-based adjusted R 2 with values of GAM-based spring, summer, autumn, winter, and annual models, which are 0.92, 0.78, 0.87, 0.85, and 0.90, respectively, while those of LUR models are 0.87, 0.71, 0.84, 0.84, and 0.85, respectively. Different to the LUR-based hypothesis of the "straight line" relations, the "smoothed curves" from GAM-based apportionment analysis reveals that factors contributing to PM 2.5 variation are unstable with the alternate linear and non-linear relations. The GAM model-based PM 2.5 concentration surfaces clearly demonstrate their superiority in disclosing the heterogeneous PM 2.5 concentrations to the discrete observations. It can be concluded that satellite-based PM 2.5 concentration mapping could be greatly improved by GAM modeling given its simultaneous considerations of the linear and non-linear influencing mechanisms of PM 2.5 .

Deriving Hourly PM2.5 Concentrations from Himawari-8 AODs over Beijing–Tianjin–Hebei in China

“…Meanwhile, overestimated PM2.5 concentrations existed with slightly polluted levels, which was similar to those in the same region according to another study [18]. This result could be attributed to the nonlinear relationship between PM2.5 concentrations and AODs at different aerosol loadings [9]. Moreover, the predicted PM2.5 concentrations using the average AOD for the PM2.5 monitoring sites within the 5 km radius may not fully represent the site measurements.…”

“…The satellitederived population-weighted average of PM2.5 in Beijing was 51.2 μg/m 3 during the study period (March 2013 to April 2014) [17]. A one-year study on the PM2.5 estimations in the BTH region using a generalized additive model presented an annual mean value of 69.4 µ g/m 3 with values ranging from 13.3 µ g/m 3 to 133.7 µ g/m 3 [9]. Urban areas with high PM2.5 concentrations, such as Beijing, Shijiazhuang, Xingtai, and Handan, were effectively captured by the improved LME model.…”

A Generalized Additive Model Combining Principal Component Analysis for PM2.5 Concentration Estimation

“…However, most of them depend on presumed linear relationships between the ground-level measured PM 2.5 concentrations and the independent variables, despite the fact that the linear influencing mechanism on PM 2.5 concentration is not always suitable for all independent variables. Focusing on this issue, the generalized additive model (GAM) was introduced to capture the non-linear and non-monotonic relationships between variables in a few studies [21][22][23][24]. Results of those studies proved that the GAM is effective at identifying the effect of different factors on regional PM 2.5 concentrations, meaning GAM modelling is a robust method for estimating PM 2.5 concentrations.…”

“…The finalized regression model presented in this article was determined such that the model AIC value is among the lowest of all the models [40]. Additionally, a significant test was also employed using the 0.05 level to check whether each term remaining in the finalized model was statistically significant [22].…”

Two decades of satellite observations of AOD over mainland China using ATSR-2, AATSR and MODIS/Terra: data set evaluation and large-scale patterns

“…Many studies have been published on aerosols in relation to air quality in the eastern part of China, including satellite remote sensing, ground-based measurements, modeling and combinations thereof, which often focus on local or regional aspects (e.g., Song et al, 2009;Ma et al, 2016;Zou et al, 2017;Xue et al, 2017;Miao et al, 2017;Guo et al, 2017). Satellites offer the opportunity to obtain information, using the same instruments and methods, over a large area during a longer period of time.…”