Predicting spatiotemporal variations of PM2.5 concentrations during spring festival for county-level cities in China using VIIRS-DNB data

Abstract

Short-term PM2.5 increment is one of the severe air quality issues in the urban areas of China during the Spring Festival (SF) owing to the large amount of firework (FW) displays. This study proposed a hybrid self-adaptive deep learning algorithm, named Self-adaptive deep belief network (SADBN), which calibrates PM2.5-AOD relationship self-adaptively by the spatial relation between each predictor and the others around it, to estimate the short-term rapid changes of PM2.5 concentration during SF of China urban areas. Firstly, all the urban areas were extracted for 4 years (2015–2018) by using the optimal threshold method based on VIIRS-DNB data. Secondly, the SADBN model was constructed through integration of VIIRS-DNB data, PM2.5 ground measurements and meteorological fields. Then the PM2.5 concentrations were mapped for the urban areas of China during SF used SADBN model. Finally, the spatiotemporal changes of PM2.5 were analyzed based on three periods (Previous, During and Post) of SF from 2015 to 2018. The results demonstrated that the VIIRS-DNB product could be able to capture the short-term variability of PM2.5 concentration at high spatiotemporal resolution with acceptable accuracy, which is affected by human activity. The SADBN could be improved by the consideration of VIIRS-DNB data with increments R2 of 0.06 and decrements of RMSE of 4.98 μg/m3. The results indicated that the rapid increase of PM2.5 concentration caused by FW can affect the monthly average PM2.5 of most of the urban areas during SF. The mapping results showed that there were obvious temporal trends during and after SF, the PM2.5 concentration increased rapidly During period and decreased significantly Post period of SF in the most of the urban areas.