<p>Urban flood damage has intensified due to localized heavy rainfall and irregular precipitation. Effective management of urban stormwater systems would benefit from real-time discharge prediction that accurately represents the rainfall–runoff relationship. With the expansion of smart sensor networks, the demand for models using real-time data is increasing. Therefore, this study enhanced prediction performance by incorporating two-dimensional rainfall distribution data (radar data) into a previously developed convolutional neural network (CNN)–long short-term memory (LSTM) spatiotemporal deep learning model and compared its performance with the conventional model. The results showed that incorporating rainfall distribution significantly improved prediction accuracy, particularly that of the CNN–LSTM model, compared with those of models using only water level and inflow data. Among the single models, the CNN performed well under specific rainfall conditions, whereas the LSTM did not exhibit consistent improvement. Model performance depended on the rainfall characteristics, emphasizing the need for further analysis of how rainfall distribution affects prediction accuracy. Analysis of various rainfall events revealed that the CNN–LSTM model achieved superior performance under consistent spatial distributions, such as when the rainfall center shifted along the upstream–downstream axis. In contrast, the CNN model was more effective under random or irregular rainfall distributions. These findings demonstrate that rainfall spatial characteristics can guide model selection and highlight the importance of incorporating spatial variability in predictive modeling. This study provides a foundation for optimizing input data composition and model combination in spatiotemporal deep learning-based discharge prediction models using diverse rainfall events and measurement data.</p>