Abstract
Flooding is a destructive natural hazard that is exacerbating due to climate change, causing significant socio-economic losses. Flood inundation models are critical tools in the mitigation and management of such hazards. Conventional standalone process-based models can be limited in real-time applications due to their computational intensity and the trade-off between accuracy, resolution and speed. Data-driven (machine learning) models offer an alternative route as surrogate models, promising computational efficiency, however, are noted to have limitations of high data-dependency and a black-box nature. Hybridization, implemented in machine learning models to enhance strengths and limit tradeoffs has been shown to increase physics awareness, real-time applicability, and adaptability. This paper conducts a systematic review of state-of-the-art research on hybrid flood models, and presents a definition of hybridization. A classification of hybridization techniques into enhancing inputs, structure and processing is proposed. Sequentially, common metrics used to evaluate flood models are classified into accuracy and speed-based metrics, and the usefulness of hybridization is discussed. An extensive benchmarking framework is proposed for the comparison of models before application, based on quantitative classification of the model and standardized datasets and tests. Finally, a way forward is proposed in the context of physics informed machine learning for flood inundation modeling.