Augmentation and extrapolation of streamflow and rainfall datasets using deep learning

Earth systems are inherently chaotic and hard to model in a linear way. Because of this, scientists have been using black-box modeling techniques like deep neural networks. In order to pave the way for better Earth modeling using deep neural networks, this study proposes a cumulative framework that includes data collection, benchmark dataset generation, data augmentation, forecasting and communication. Deep neural networks are used to augment the data either temporally or spatially, and finally, using the augmentation model and the original collected data, forecast models employing deep neural networks are built. To demonstrate how this framework applies to various Earth data types, this study focuses on forecasting streamflow and rainfall which is a significant factor on streamflow prediction. At the final step of the study, a web platform is presented that enables domain scientists, educators, and public to experiment with state-of-the-art deep learning models on the web without the know-how of numeric computation libraries.