Monitoring and short-term forecasting for Ground-level Particulate Matter Using Satellite Observations and Numerical Model Output

Abstract

Air pollution seriousness is on the rise over the word. Particulate matter (PM) is one of the atmospheric aerosols which have influenced on the adverse human health. Especially, PM with aerodynamic diameter less than 10 micrometers and 2.5 micrometers (i.e. PM10 and PM2.5, respectively) is available to inhales into the body without filtering, and it leads to cause significant health problems such as cardiovascular and respiratory-related diseases. The accurate PM concentration monitoring and forecasting is a prerequisite for providing early warning. In this study, the several variables from multi satellite sensors (i.e. GOCI, GPM, SRTM, MODIS), numerical models (i.e. RDAPS), emission model (i.e. SMOKE), and in-situ measurements were used to establish short-term forecasting model of hourly PM10 and PM2.5 concentrations. one to three hours before in-situ PM measurements were interpolated using kriging method, and these values also used as an input variable. The proposed forecasting model was conducted based on two different machine learning approach, i.e. random forest (RF), support vector machine regression (SVR), over South Korea during 2015 to 2016 time period. The early results showed underestimation at high PM concentration part due to low concentrations of PM10 and PM2.5 relatively predominate in our study area. Thus, over- and sub- sampling technique was addressed to make the dataset balance. Oversampling approach is applied to high concentration samples with 3x3 or 5x5 windows for construction of training dataset. Log transformation on the target variables was additionally conducted to improve the model accuracy. The real-time training, which can be implemented on-line in real time, was conducted to make the short-term forecasting model for hourly PM10 and PM2.5 concentration, and this real-time forecasting model was evaluated by the leave-one-out validation based on the station. The results indicate that the RF model has better performances in compare with SVR model. In addition, when using only 1 hour ago PM in-situ observation, the model accuracy was higher than when using only 3 hours ago PM in-situ observation.