This study proposes a hybrid model that integrates high-resolution physical flood inundation simulations with machine learning to improve flood prediction accuracy and real-time applicability. The physical model, based on two-dimensional shallow water equations, uses the finite volume method on a triangular unstructured grid to simulate flood inundation in Dresden, Germany, with detailed topographic data, including individual building layouts. The machine learning model utilizes features such as ground elevation, land use, and effective rainfall to predict flood depths. By training the machine learning model with the output of the physical model, the hybrid approach aims to provide rapid flood inundation predictions during extreme rainfall events. Numerical experiments are conducted using various rainfall scenarios with peak intensities ranging from 40 mm/h to 70 mm/h. The results demonstrate that the machine learning model accurately predicts flood depth distributions, with evaluation metrics showing good performance.