Estimation of ground PM2.5 based on GOCI TOA reflectance using Deep Neural Network

Fine particulate matter (PM_2.5), which consists of solid and liquid particles and mixture of both suspended in the near surface atmosphere, is known to be one of the most threatening elements to human health by penetrating skin, lungs and bronchi. There have been diverse studies with regard to monitoring near-surface PM_2.5, particularly over East Asia, where recent rapid industrial development has produced serious air pollution. Some countries have already been operating ground-based monitoring networks and collecting relevant data. However, due to their poor spatial representativeness and inhomogeneous data quality, many of the previous studies were conducted based on space-borne observations. In this study, we tried to monitor concentrations of PM_2.5, particles with aerodynamic diameters less than 2.5 µm, based on GOCI top of atmosphere reflectance using a deep neural network (DNN) method. DNN is a kind of machine learning developed from artificial neural networks. In order to enhance the model performance, near-surface atmospheric information from Unified Model was also used as input variables such as surface temperature, dew point temperature, surface pressure, height of planetary boundary layer, relative humidity and wind fields. Sensitivity examinations were conducted to find optimal structures of training models and several techniques (e.g., regularization, early stopping, and normalization of input variables) were applied to prevent over-fitting training datasets. The retrieved data were characterized by comparing with estimates from the operational MCAQ model, which is used in air quality forecasting, and conventional linear regression results.