Fully convolutional neural networks on semantic segmentation of flooded areas

Floods cause extensive economic damage and loss of life worldwide. Thus, automatic image detection is valuable for effectively minimizing response time to these impacts. Synthetic Aperture Radar (SAR) imaging has proven to be an important resource in flood management, as this remote sensing technology is highly sensitive to water. This study applies Fully Convolutional Neural Networks (FCNN), particularly U-Net and U-Net++ topologies, to semantic segmentation of flood-affected regions in Sentinel-1 satellite images from Cloud to Street Microsoft floods dataset. The U-Net++ architecture demonstrates a high capability in identifying flooded areas, achieving an Intersection over Union (IoU) metric of 0.8280, F1 score of 0.9053, and sensitivity of 0.9001.