Strategies for Predicting Flood Inundation in a Large and Complex Floodplain Based on Low‐Fidelity Hydrodynamic Models

Abstract Accurate and computationally efficient flood inundation prediction is critical for effective flood risk management. While high‐fidelity hydrodynamic models provide detailed representations of flood processes, their computational demands limit real‐time flood forecasting or ensemble applications. To address this, surrogate models have been developed to deliver faster predictions with acceptable accuracy. Among them, the Low‐Fidelity, Spatial Analysis and Gaussian Process Learning (LSG) model has demonstrated superior performance with reasonable computational cost, effectively simulating flood dynamics in complex floodplains. However, its ability to handle large‐scale floodplains with intricate flow interactions and upstream dam releases remains untested. This study investigates the feasibility of applying the LSG model to the Lower Brisbane River floodplain, which presents additional challenges due to extensive urbanization, dam regulation, tidal influences, and many interacting tributaries. These conditions lead to backwater effects from the main river channel into the tributaries, further complicating flood behavior. Initial implementation in this study exhibited limited accuracy, prompting further investigation. Higher low‐fidelity resolution was found to significantly improve model performance. To predict maximum flood surfaces, we developed and investigated two LSG model variants: (a) LSG‐Max trained directly on maximum flood surfaces; (b) LSG‐TS trained on time series data and deriving maximum flood surfaces from the predicted flood evolution. Evaluation against historical and synthetic flood events showed that LSG‐TS consistently outperforms LSG‐Max due to its richer training information. These findings demonstrate that the LSG modeling approach can offer effective strategies for predicting flood inundation in large and complex floodplains.