Floods are a significant threat to both lives and property, and pose environmental and public health hazards. Anticipating the timing and location of floods could empower hydraulic structure operators to implement timely flood mitigation strategies and enable citizens and local governments to enhance preparedness for potential emergencies. Thus, predicting floods in advance and solving optimization problems related to flood mitigation are therefore vitally significant endeavors. Classical flood forecasting is typically achieved by computing on detailed grid representations of terrain elevation maps of the entire watershed and solving partial differential equations (PDEs). Consequently, the simulations are often accurate but computationally prohibitive, especially on large watersheds. Furthermore, such physics-based simulations lack the model explainability. Flood management is usually achieved by pre-releasing a sufficient volume of water in advance of a big storm or tide so that water levels can stay within flood thresholds. However, determining “optimum” pre-release schedules is a greater challenge than flood forecasting since it needs thousands of such simulations. The corresponding optimization is NP-hard, and even the heuristic algorithms such as those using genetic algorithm are computationally intensive. Machine learning (ML) and Deep learning (DL) methods are known to be effective approaches to tackle such optimization problems. This proposal presents three specific aims as follows: (1) to develop robust, explainable, and real-time DL models for flood prediction and mitigation in coastal river systems; (2) extend the above DL models to handle extreme weather conditions, large watersheds, and non-coastal river systems; and (3) explore DL models to study the large-scale weather patterns, since high precipitation is a primary cause of river floods.

Jimeng Shi is a Ph.D. candidate in the group of Algorithms for Machine Learning and Data Analytics (AMaDAys), at the Knight Foundation School of Computing and Information Sciences, Florida International University (FIU), under the supervision of Prof. Giri Narasimhan. His primary research areas include Deep Learning, AI for Science, and model explainability. He holds a M.S. degree from FIU and a B.S. degree from Tianjin University of Science and Technology in China. This summer, he is working on AI for Climate Science as a Research Fellow at Columbia University. He has published papers in NeurIPS and ECML workshops, ICML, and Nature Machine Intelligence. He has also served as a reviewer for a ICLR workshop, SIGKDD, CIKM, and ECML conferences.