Authors: Achraf Tounsi, Marouane Temimi – Stevens Institute of Technology
Title: Towards an operational early warning system for precipitation and flood nowcasting
Abstract: The National Weather Service has continuously developed accurate rain and flooding forecasting models using a merger between machine learning and state-of-the-art NextGen National Water Model (NWM) frameworks. The ability to accurately forecast especially in the very short term the likelihood of rain and flood events can reduce the amount of associated property damage and loss of life.
This study aims to enhance NWM Flood Inundation Mapping (FIM) product by generating a new sub-hour-scale, very short-range forecast of FIM in line with NWS objectives. To that extent, this study developed and validated an operational framework called FloodCaster. This early warning system involves a Machine Learning (ML)-based operational framework that couples rain nowcasting and streamflow discharge prediction to generate short-term flood inundation maps over one of the Hackensack sub-watersheds, In New York City Metropolitan area. The precipitation nowcasting is first computed using the Lagrangian INtegro-Difference equation model with Autoregression (LINDA) method (Pulkkinen et al., 2021) to generate the next 3-hour rain prediction using MRMS data at 10-minute time step. Then, the streamflow forecast module, which integrates a sub-hourly LSTM framework, generates discharge predictions. The manifold method similar to the one proposed by Google Research (Nevo et al., 2022) for operational flood forecasting will be used for the flood inundation mapping module to generate real-time flood maps. The proposed system is integrated in the NextGen National Water Model via the Basic Model Interface (BMI) to generate flood extent maps over the entire Hackensack watershed. The effectiveness of the proposed approach is evaluated using case studies of extreme weather events such as Hurricane Ida that impacted the northeastern US. The results demonstrate the potential of the approach to predict floods accurately, generate flood inundation maps, and issue timely alerts to communities and decision-makers. Tounsi et al. (2023) showed that the LINDA model performed well over the continental United States for moderate to heavy rainfall. In addition, Tounsi et al., 2022 successfully predicted streamflow values at the USGS station at New Milford, right downstream of the Hackensack watershed using a trained Long Short-Term Memory (LSTM) model that led to a Nash-Sutcliffe Efficiency (NSE) value of 0.95. Both methods were coupled with the manifold method, which generated flood extent maps over the Hackensack upstream sub-watershed. Work is in progress for the assessment of the full integration of into the NextGen National Water Model to substitute the prediction of the selected sub-watershed using the BMI framework. The study has many novel components: it (1) contributes to the development of a very-short term operational flood prediction and management tools that can be used to mitigate the impact of floods on communities and infrastructure. It also (2) proposes an approach that has the potential to inform and guide decision-making processes related to flood management and emergency response, ultimately saving lives and minimizing the damage caused by floods, and it (3) finally uses and validates the effectiveness of machine learning models in supporting rain and flooding decision-making. The scope and the outcome of this work are aligned with CIROHs research priorities and the mission of the National Water Center.