Flood Detection in Data-Limited Regions Using Sentinel-1 SAR and U-Net: A Case Study in Zimbabwe
Abstract: Flood disasters pose major risks in Sub-Saharan Africa, where limited infrastructure and scarce annotated datasets constrain effective monitoring. This study proposes a flood detection framework using Sentinel-1 Synthetic Aperture Radar imagery and a U-Net convolutional neural network for flood segmentation in Zimbabwe. To address data scarcity, benchmark flood datasets were incorporated alongside Sentinel-1 imagery, with preprocessing steps including normalization, speckle reduction, augmentation, and class imbalance mitigation. Model performance was evaluated using accuracy, precision, recall, F1-score, Intersection over Union, and Dice coefficient. The model achieved strong segmentation performance, demonstrating the feasibility of SAR-based deep learning for flood detection in cloud-prone and data-limited environments. The findings support the potential application of deep learning-assisted flood mapping for disaster preparedness in Zimbabwe and similar regions. Future work will investigate integration with meteorological and multimodal data for predictive flood early warning systems.
Keywords: Flood mapping, Sentinel-1 SAR, U-Net, Deep learning segmentation, Disaster risk management, Remote sensing, Zimbabwe, Early warning systems
How to Cite?: Tambirai Gahadza, Simbarashe Mugova, "Flood Detection in Data-Limited Regions Using Sentinel-1 SAR and U-Net: A Case Study in Zimbabwe", Volume 15 Issue 5, May 2026, International Journal of Science and Research (IJSR), Pages: 1433-1439, https://www.ijsr.net/getabstract.php?paperid=SR26517164625, DOI: https://dx.doi.org/10.21275/SR26517164625