Application of Deep Learning for Air Quality Predictions, Remote Sensing Processing and Long-term Climate Index Forecasting

The advancement and development of new technology provide atmospheric scientists and modelers to acquire an overwhelming amount of data on meteorology and air quality from space, numerical simulations, and in-situ monitoring sites. Integrating these data sources provides unique opportunities to enhance understanding of atmospheric processes to better simulate and forecast these processes. While Global Climate Models and Chemical Transport Models have undergone significant optimizations and improvements over the past decades, they are still unable to provide fully reliable biogenic air quality predictions or long-term climate forecasting. These limitations can be alleviated and addressed by incorporating in-situ measurements and remote sensing products into data assimilation or reanalysis techniques. While ground-based remote sensing measurements provide detailed point observations, they lack the spatial coverage of remote sensing-derived measurements. Unfortunately, these remote sensing measurements experience issues caused by outside factors such as cloud cover contamination and false reflectance. Internal issues involve sensor errors that corrupt or lead to failed measurements of the data. This study utilizes the advanced capability of several deep learning models for the forecasting of pollen concentrations by up to 7 days; the imputation of remote sensing measurements spatially with partial convolutional neural networks and subsequent revision to incorporate spatio-temporal imputation; and long-term forecasting system of the climate index Nino3.4 by up to 36 months.