Flood forecasting with machine learning models in an operational framework

Abstract. Google's operational flood forecasting system was developed to provide accurate real-time warnings agencies and the public with a focus on riverine floods in large, gauged rivers. It became 2018 has since expanded geographically. This consists of four subsystems: data validation, stage forecasting, inundation modeling, alert distribution. Machine learning is used for two subsystems. Stage modeled long short-term memory (LSTM) networks linear models. Flood computed thresholding manifold models, where former computes extent latter both depth. The model, presented here first time, provides machine-learning alternative hydraulic modeling inundation. When evaluated historical data, all models achieve sufficiently high-performance metrics use. LSTM showed higher skills than while achieved similar performance extent. During 2021 monsoon season, warning India Bangladesh, covering flood-prone regions around rivers total area close 470 000 km2, home more 350 people. More 100 alerts were sent affected populations, relevant authorities, emergency organizations. Current future work includes extending coverage additional locations improving capabilities accuracy.