Tile-Based CNN with Combined Optimizer for Urban Flood Prediction Under Various Deterministic Rainfall Scenarios

Urban flooding poses increasing risks globally due to climate change and urbanization, yet physics-based hydraulic models suffer from high computational costs that limit their application to flood analysis under various rainfall conditions. This study proposes an optimizer-improved tile-based convolutional neural network framework for efficient prediction of two-dimensional peak flooding maps under various rainfall scenarios. The framework integrates tile-based spatial learning for efficient localized flood-response learning and a combined Adam-VCA optimizer to mitigate local optimum convergence. The framework was applied to the Dorim basin in Seoul, South Korea, using 22 rainfall scenarios ranging from 10 to 800 mm, of which 20 scenarios were used for model training and 2 scenarios (500 mm and 700 mm) were reserved as independent test scenarios for performance evaluation. Prediction accuracy was evaluated using F1-score and critical success index (CSI), and Verification period based accuracy (VAC) and Peak flooding based accuracy (PAC) as study-specific depth-incorporated metrics. The proposed optimizer-improved CNN substantially outperformed the conventional CNN, achieving F1-score of 83.84%, CSI of 72.24%, VAC of 90.39%, and PAC of 70.93%, compared to 51.03%, 34.25%, 63.39%, and 47.29%, respectively. The results confirm the framework’s potential for efficient flooding assessment and urban flood risk management under diverse rainfall conditions.