Transfer Learning Approaches for Knowledge Discovery in Grid-based Geo-Spatiotemporal Data
Recent advances in remote sensing technologies led to an explosive growth of geo-spatiotemporal data in domains such as geology, ecology, hydrology, and astronomy, to name a few. To effectively process such a high amount of complex data, scientists and researchers rely primarily on deep learning models. However, two major challenges remain. Firstly, deep learning models being purely data-driven, lack the required domain knowledge of the system that generated the data in the first place and, therefore, often perform sub-optimally. Secondly, geo-spatiotemporal features significantly vary with location; hence, such models often need to be trained repetitively for almost every new region, increasing the computational resources required to train such complex models. To address these challenges, this work demonstrates a knowledge-guided deep neural network architecture - HydroDeep, that couples a domain (process-based) model with a combination of deep convolutional neural network (CNN) and long short-term memory (LSTM) network to effectively learn from grid-based geo-spatiotemporal data. HydroDeep outperforms a data-driven CNN's and LSTM's prediction performance by 1.6% and 10.5%, respectively. In addition, four transfer learning approaches are proposed to effectively transfer knowledge from one region to another. Results show an improvement of performance by 9% to up to 108% in new regions with a 95% reduction in time. The merit of transfer learning as a tool for discovering and analyzing spatiotemporal variance between regions is also investigated. Flood prediction, one of the hardest problems involving complex geo-spatiotemporal data, is used as a use case.
Committee: Ali Jannesari (major professor), Robyn Lutz, and Chaoqun Lu.