In July 2021, Central Europe was hit by a large rainfall event that resulted in 220 people losing their lives and the destruction of over US$30 billion worth of assets. In February 2022, the north-eastern parts of Australia were impacted by a series of floods, killing 23 people, and causing damages worth over US$1.8 billion. Recently, in August 2023, large parts of China experienced severe flooding, which displaced over 1 million people and killed over 30. These are just a few examples of how severe and dangerous floods can be. One of the main tasks for hydraulic engineers is to mitigate the impact of floods, by providing accurate information of incoming floods to help with evacuation planning and/or designing infrastructure to reduce the impact of floods. This requires accurate predictions of floods before they occur; thus scientists and engineers have developed and improved the use of hydrodynamic models for this purpose over the last century.  Hydrodynamic models are numerical models that simulate flooding by dividing an area of interest into small subareas (i.e. grid cells) and then calculating how water moves between grid cells. The water movement is described by solving complex differential equations based on the physical principles of water flow. Hydrodynamic models are well-documented and can accurately simulate flood events. However, when using hydrodynamic models to simulate flooding over large areas in high resolution, millions of grid cells are needed. Solving complex equations for millions of interconnected grid cells is intractable and results in high computational costs for running hydrodynamic models. This means many flood events happen faster than we can predict them using these high-resolution hydrodynamic models, resulting in no or limited emergency response time for evacuation and limited effectiveness of mitigation strategies. For this reason, an efficient and accurate approach to predicting flood events is urgently needed to provide valuable information as floods unfold, as well as to design robust infrastructure that can mitigate the impact of floods.

The Low-fidelity, Spatial analysis, and Gaussian Process learning (LSG) model recently developed is an approach that can be used to predict flood extent and depth much faster than flood events unfold. The idea of the LSG model is to use a low-resolution hydrodynamic model to provide an initial flood estimate. The low-resolution hydrodynamic (i.e. low-fidelity) model has a much lower computational demand than the traditional high-resolution model but at the cost of accuracy. To improve the accuracy, the LSG model upskills the initial flood estimate to high resolution and accuracy, similar to what a high-resolution hydrodynamic model would predict. The upskilling process consists of matching flood inundation patterns in time and space and is performed using mathematical methods that transform the low-fidelity estimates to what a high-resolution hydrodynamic model would have produced.

Previously, it was believed that only moderate speed-ups (e.g., 10 times) compared to the high-resolution hydrodynamic model could be achieved via this approach. For this reason, researchers have mainly focused on using fast machine-learning approaches as surrogates for high-resolution hydrodynamic models. This study overturns this belief and shows speed-ups over 1000 times faster than high-resolution models can be achieved using the LSG model while maintaining high accuracy of the flood predictions. The key to achieving this massive speed-up is the development and use of an extremely coarse and simplified low-fidelity model. The low-fidelity model has grid cells covering over 1000 x 1000 m, but using the LSG model methodology, the estimate from using such a coarse model can be upskilled to make accurate predictions for grid cells that are more than 50 times smaller.

We demonstrated the LSG model for two large river systems in Australia. The first is the flat and complex Chowilla floodplain in southern Australia (740 km²), and the second is the steep and fast-flowing Burnett River in northeast Australia (1479 km²). The distinct differences between these case studies make them a challenging test of the LSG model’s ability to provide fast and accurate flood predictions. The results show the LSG model can simulate the dynamic evolution of flood inundation in both case studies. The LSG model provides information on arrival time, flood extent, peak water depth, and full water depth hydrographs with similar accuracy to a traditional high-resolution hydrodynamic model.

The result of this study is a big leap forward toward providing useful flood predictions; both during emergencies to help make informed decisions to save lives and protect valuable infrastructure, as well as in the preparation before flood events in the design of robust infrastructure. Currently, flood inundation predictions are mainly based on deterministic approaches, where the most likely scenario is simulated, because of the high computational costs of running a high-resolution hydrodynamic model. However, the LSG model can make it possible to simulate all scenarios of flood events, both before the events and as they unfold. This can shift the current practice from using deterministic predictions to risk-based probabilistic forecasts. In addition, this methodology can be used to design a more robust infrastructure by enabling the use of Monte Carlo methods to simulate how different combinations of flood drivers affect the severity of flood events. Future studies should focus on implementing this technology in operation and design frameworks to maximise the capabilities and benefits of the LSG model.