PM10 Estimation using Applying Time Variation and recurrent neural network

Air pollution such as aerosols particulate matters (PMs) is a serious environmental problem in Republic of Korea. Therefore, higher spatiotemporal observation of PM concentration is becoming more important. Ground-based observations could provide relatively accurate PM data but location of station is sparse. To overcome the limitation, a number of studies have been focused on developing empirical PM models by utilizing high resolution Satellite Aerosol Optical Depth (AOD) as well as various climate and other data. However, it is difficult to accurately reflect the time variation attribute only by simple regression methods.

In this study, we developed surface PM10 estimation model using AOD data from Geostationary Ocean Color Imager (GOCI), Lidar estimated Planetary Boundary Layer Height (PBLH), Humidity, Surface Pressure, Wind Speed, Dew Point Temperature, Temperature, Cloud Cover, Wind Direction data over Seoul metropolitan area for 6 years from 2011 to 2016. First, PM10 was estimated with regression model using AOD and climate data and investigated effect of applying time variation attributes such as Daily Summary, Biweekly Summary, Season analysis. In addition to that, we applied recurrent neural network - long short term memory (RNN-LSTM) method, one of the time-series deep learning techniques. We have found that models that reflect time variation shows better results and the RNN-LSTM model can effectively estimate the surface PM10 level.