Advancing methods of rapid flood scenario assessment using hybrid approaches of hydrodynamic modelling and machine learning
❄
| Flooding is the most frequent natural hazard in New Zealand and the second-most costly after earthquakes. It is also expected to become more severe as climate change impacts are realized. Generally, two-dimensional hydrodynamics models are used to simulate flood inundations across a floodplain since they can represent the flow through the terrain, simulating the corresponding hydrological processes and flood dynamics. These models can provide reliable estimations of the flood extent, flood magnitude and other flood-related variables. However, these models are also computationally expensive and time-consuming. Current computing power and memory pose a limitation in the number, level of detail and modelling time of assessed flooding scenarios, making rapid real-time or forecast flooding risk assessment challenging. This study investigates the application of hybrid hydrodynamic and machine learning techniques to develop a rapid flood scenario assessment model using the Wairewa catchment (Banks Peninsula, Canterbury) as the study site. This approach aims to reduce the numerical modelling load and enable probabilistic modelling, allowing us to make rapid predictions of potential flooding events from an ensemble of previously assessed scenarios. The objective is to develop a hybrid model that is able to predict flood extent and maximum flood depth based on the characteristics of the main inundation driver in the area, heavy rainfall. The methodology consists of the following steps. Firstly, a sample of synthetic storms was created based on the temporal and spatial characteristics of heavy rainfall events in the catchment. Then, each event in the sample of storms was modelled using a 2D hydrodynamic model, BG-Flood, to obtain the corresponding flood extent and maximum water depth map. Following this, the sample of storms (with their corresponding inundation maps) was divided into training and testing datasets. The training dataset is the one used to train a machine learning model capable of predicting flood extent and maximum water depth from a rainfall event. To first explore which machine learning algorithm performs better for flood prediction, 3 critical locations within the Wairewa Catchment floodplain were selected. In these locations, the performance of 4 different machine learning algorithms, including Random Forest, XG-Boost, Radial Basis Function Networks, and Support Vector Machine, was assessed. To this end, the ability of these algorithms to predict if the location floods or does not flood (first objective) and to estimate the maximum water depth (second objective) was evaluated using various metrics. Precision, recall, and F1-score performance metrics were used for the first objective. It was found that all algorithms consistently achieve high accuracy rates (minimum 95%) and perform similarly in all the locations. For the second objective, root mean square error, mean squared error, mean absolute error, and coefficient of determination performance metrics were used. In the results, Random Forest and XG-Boost emerged as top performers, demonstrating superior accuracy (mean absolute error from 0.023 to 0.041 m) and computational efficiency. |