Özet:

: Flood modelling is essential for addressing a range of scientific and engineering challenges. In recent years, the high computational demands of solving shallow water equations numerically have led researchers to explore machine-learning-based emulators for predicting floods and flood risk. Specifically, the proliferation of convolutional neural networks in solving different scientific problems has encouraged researchers to investigate their applicability in flood modelling. Most of these studies, however, have focused on specific locations or hydrological conditions, meaning that their findings may not be directly applicable to other situations without additional data and further training. We present here a U-Net model, a popular deep learning algorithm, which has the capacity to approximate maximum flood depths across multiple return periods while maintaining catchment generalizability. The model was trained using the outputs from a 2D hydraulic model (JFlow) to predict maximum water depths for a set of rainfall return periods (20, 100 and 1000 years). The pre-trained model was then applied to estimate depths in three unseen catchment areas. Our results demonstrate that U-Net can be used to approximate water depths in previously unseen catchments with significantly less computational time compared to the 2D model.

Anahtar Kelimeler: