U-Net-based flood segmentation using Sentinel-1 SAR imagery in Zimbabwe

Abstract

Flooding is a devastating climate-induced hazard that has severe effects in Sub-Saharan Africa, resulting in significant human, economic, environmental, and losses. Climatic changes have increased the frequency of flood events in Zimbabwe, especially in regions with high rainfall, such as Chimanimani. This investigation introduces a semantic segmentation framework for the detection of flooding that is based on U-Net and employing Sentinel-1 Synthetic Aperture Radar (SAR) imagery. The proposed model enhances flood mapping under cloud-cover conditions through the use of all-weather and day-night imaging capabilities of SAR. Preprocessing procedures which were implemented to enhance model generalization and mitigate class imbalance include speckle noise reduction, image normalization as well as binary mask generation. The model was evaluated and trained using benchmark flood datasets that included Sentinel-1 SAR imagery. The experimental results showed that the segmentation performance was exceptionally good, with an overall precision of 95%, F1-score of 0.88 and recall of 0.82 for flooded regions. Further, the intersection over Union (IoU) analysis confirmed the accurate delineation of floods at pixel level. The results prove the practicality of convolutional neural network-based SAR flood segmentation techniques for real-time disaster monitoring in Zimbabwe and provide a reliable and flexible system for future implementation with real-time meteorological data as well as national early warning mechanisms.