A Water Body Extraction Network Based on Fusing SAR Images and DEM Data

Synthetic Aperture Radar (SAR) plays a crucial role in flood monitoring and early warning systems through water body extraction techniques. However, unlike optical remote sensing images, SAR images are affected by unique imaging characteristics, including shadows, noise, and complex environments, which complicate interpretation. Most existing deep learning methods rely solely on single SAR images for model training, without incorporating additional information, leading to lower accuracy in water body interpretation, especially in challenging scenarios such as mountainous areas and densely vegetated regions. To address these challenges, we propose a water body extraction algorithm that fuses SAR images with Digital Elevation Model (DEM) data to improve the accuracy and robustness in complex environments. Four different fusion strategies are explored at both the image level and feature level, incorporating DEM and slope data into an existing encoder-decoder architecture. These enhancements effectively address the aforementioned issues. Experimental results demonstrate that the fusion approach significantly improves water body extraction precision and reduces false alarm rates compared to non-fusion methods. Through a comparative analysis of the different fusion strategies, the best approach was identified as image-level fusion using DEM data, achieving a mean Intersection over Union (mIoU) of 0.95 on the GF-Floodnet dataset.