Performance of a deep learning generative surrogate model for flood inundation forecasting

Conventionally, models used for flood inundation forecasting are typically physically based and computationally intense. This limits their suitability for operational flood inundation forecasting where high-resolution data are critical. Deep Learning (DL) models have been proven to be able to reduce the computational burden while maintaining acceptable accuracies. However, some DL surrogate models often require complex model architectures that result in high computational costs to capture flood dynamics across the entire domain.

With the recent development of advanced DL models, generative models have the potential to overcome the need for computationally expensive model architecture and to be useful in flood inundation forecasting. Generative models can: generate synthetic data, capture complex relationships between different variables (e.g., hydrological, meteorological and topographical estimates) and allow for domain transferability. In this study, we developed a deep generative model as a surrogate model for flood inundation forecasting and investigated its performance under various spatial and temporal resolutions. The initial results indicate that increasing spatial resolution has a bigger impact on model training time compared to increasing temporal resolution; however, does not impact model prediction time. Additionally, the model accuracy tends to increase with the increase in resolution at the expense of computational costs. Enlarging the computation sub-domain can shorten the overall model run time and improve model accuracy but it's subject to hardware capacity. These findings indicate that the proposed generative surrogate model has the potential for operational flood forecasting.