Appication of ground observations together with neural network technique to PM2.5 estimation from satellite aerosol optical depth product.

摘要

Fine particular matter with diameters less than 2.5microm (PM2.5) is a major criteria pollutant with concentration limits set by the U. S. Environmental Protection Agency (EPA). Direct monitoring of PM2.5 is difficult due to the high cost and manpower of existing surface samplers. This motivates the search for retrieval tools that can be used to provide extended coverage of PM2.5 in poorly monitored areas. In this thesis, we develop a Neural Network (NN) regional tool for the New York state area that combines satellite aerosol optical depth (integrated aerosol extinction) with appropriate meteorology inputs from the Weather Research and Forecast (WRF) model to better account for vertical distribution in order to obtain an optimal PM2.5 estimator. To test this approach, we first focused on New York City, where LIDAR derived PBL (Planetary Boundary Layer) heights were used together with AERONET optical depth, and demonstrated that the PBL height is the dominant meteorological factor in producing accurate surface PM2.5 estimations. Later, this approach was extended to the New York State. In this case, we found that temperature and geo-location were the most important factors and explained why WRF PBL was not as relevant as in the NYC analysis. Extensive numerical experiments were made to develop the most robust system with the least number of inputs as well as the simplest network architecture. The final results were compared against both the operational Infusing Satellite Data into Environmental Applications (IDEA) product and the Community Multi-scale Air Quality (CMAQ) outputs, resulting in higher correlation and less RMSE in contrast to these existing methods. Furthermore, this tool was applied with MODIS AOD data and WRF meteorology to create high spatial coverage imagery. In addition, we demonstrated that temperature is a strong factor leading to biases in the CMAQ PM2.5 outputs and developed a neural network bias correction scheme to reduce these errors. We also confirmed that the full spatial domain is needed for training purposes and applications outside this domain degrade significantly. More robust results are to be expected as the length of the data sets increase above the three-year data sets we had available.