<p>This study investigates the inundation depths of urban floods induced by real storm events, focusing on the development and assessment of super-resolution model based on ensemble learning methods. Unlike traditional deep neural networks which require extensive training and high parameterization, this study utilizes ensemble learning model to reconstruct high-resolution flood predictions from low-resolution hydrodynamic simulations. Hydrodynamic modeling results of real pluvial flood event at various spatial resolution are used for constructing datasets and for training and testing the point-based super-resolution model. Influencing factors related to urban terrain, subsurface, rainfall inputs and the hydrodynamic modeling results at coarser resolutions are used as features in the super-resolution model on basis of Random Forest, in which hyperparameters are tuned with Bayesian optimization method. The trained super-resolution models effectively reconstruct high-resolution inundation conditions from 30 m to 5 m coarse resolution inputs, highlighting an increase in correlation coefficients and a decrease in root mean squared error (RMSE) as resolution improves. Dominant influencing factors in the super-resolution models are identified together with variances in their contributions to the model performance. Two optimization approaches are applied to enhance accuracy and mitigate overestimation at coarse resolutions for the super-resolution models. The first integrates outputs from various coarse resolution models as features, notably reducing overestimation, especially with finer 5 m resolutions. The second employs ensemble modeling with super-resolution models from different datasets, which improves the performance across all tested resolutions, demonstrating the robustness of combining multiple predictive models for better flood forecasting in urban environments.</p>