Advancing flood early warning systems: ensemble learning-based classifiers for urban flood forecasting

ABSTRACT Flood early warning systems (FEWS), which rely primarily on flood forecasting models, are becoming increasingly important for mitigating damage to natural and built infrastructure. Machine learning (ML) provides flexible modelling solutions to modelling challenges such as FEWS. In this work, a flood forecasting approach is developed as a pure classification problem to directly predict flood events. The advantages of the proposed approach are demonstrated by appropriately discretising the streamflow or stage into binary states, instead treating them as continuous variables. This distinction is highlighted through a juxtaposition of regression and classification approaches, each used to generate flood alerts. The research also features a systematic cross-comparison of five ML models and three ensemble learning frameworks to provide additional insights on modelling applicability. Classification performance is shown to be primarily dependent on the base learner; extreme learning machines and support vector machines (SVMs) exhibit the best performance. SVMs are the only type of model to outperform the mean model performance across all four metrics considered, with improvements ranging from 2.9 to 24.9% above the mean. Boosted models consistently outperform the other ensembles, by 1.4 to 14.2% above the mean.