Modular machine learning systems for continuous hydrologic and hydraulic modeling in flood warning systems.

Flooding is one of costliest natural hazards annually across the world, and great efforts have been made to deploy real-time flood warning and predictive inundation modeling in recent decades. One of the largest hindrances to this technology has been the computational cost of running hydrologic and hydraulic (H/H) models. AI and machine learning techniques have been shown to be useful for hybrid or pseudo-modeling H/H computations; however, the necessity for static datasets limits their application. Any changes to land use or topography from development would require a recreation of the entire synthetic dataset with current machine learning techniques. This limitation requires a trade-off of continued accuracy against the computational cost of retraining. To address the static dataset limitation of AI for pseudo-modeling, a modular network design was implemented. Modular networks have been commonly used after data clustering to improve overall accuracy with independent AI modules for each cluster of a dataset. The novel use of modularity for this work is the division of the physical system into discrete groups (e.g. watersheds) rather than the dataset. The resulting setup is an interconnected system of individual AI models that can independently be updated for future use or development. This framework eliminates the drawbacks that hybrid AI models hold with current implementation, and allows for further development towards real-time flood warning and inundation mapping.