Spatiotemporal rainfall forecasting using deep learning under computational constraints: an IMERG-based study

Accurate short-term rainfall forecasting is essential for flood risk management, especially in tropical regions prone to localized convective storms. This study evaluates the performance of three deep learning architectures, ConvLSTM, 3D-CNN, and Temporal Convolutional Network (TCN), for 6-hour rainfall forecasting using satellite-based IMERG precipitation data. All models were trained, validated and tested on a consistent spatial-temporal dataset and compared based on standard metrics including RMSE, MAE, training stability, and spatial prediction quality. Quantitative results show that ConvLSTM achieved the lowest RMSE (0.3932 mm) and MAE (0.1148 mm), outperforming 3D-CNN and TCN across all evaluation criteria. Visual inspections of cumulative rainfall forecasts further confirm ConvLSTM’s ability to preserve convective structure and suppress background noise, while TCN struggled with spatial generalization and temporal consistency. Spatial RMSE maps averaged across the forecast horizon revealed that ConvLSTM maintained low error regions throughout the domain. These findings underscore the effectiveness of spatiotemporal modelling in rainfall forecasting and indicate that ConvLSTM shows promising performance for operational early warning systems, though further validation is required.