Particulate matter with a mass concentration of particles with a diameter less than 2.5 μm (PM2.5) is a key air quality parameter. A real-time knowledge of PM2.5 is highly valuable for lowering the risk of detrimental impacts on human health. To achieve this goal, we developed a new deep learning model-EntityDenseNet to retrieve ground-level PM2.5 concentrations from Himawari-8, a geostationary satellite providing high temporal resolution data. In contrast to the traditional machine learning model, the new model has the capability to automatically extract PM2.5 spatio-temporal characteristics. Validation across mainland China demonstrates that hourly, daily and monthly PM2.5 retrievals contain the root-mean-square errors of 26.85, 25.3, and 15.34 μg/m3, respectively. In addition to a higher accuracy achievement when compared with various machine learning inversion methods (backpropagation neural network, extreme gradient boosting, light gradient boosting machine, and random forest), EntityDenseNet can "peek inside the black box" to extract the spatio-temporal features of PM2.5. This model can show, for example, that PM2.5 levels in the coastal city of Tianjin were more influenced by air from Hebei than Beijing. Further, EntityDenseNet can still extract the seasonal characteristics that demonstrate that PM2.5 is more closely related within three month groups over mainland China: (1) December, January and February, (2) March, April and May, (3) July, August and September, even without meteorological information. EntityDenseNet has the ability to obtain high temporal resolution satellite-based PM2.5 data over China in real-time. This could act as an important tool to improve our understanding of PM2.5 spatio-temporal features.
Keywords: Deep learning; Himawari-8; PM(2.5); Satellite.
Copyright © 2020 The Authors. Published by Elsevier Ltd.. All rights reserved.