Residual wave vision U-Net for flood mapping using dual polarization Sentinel-1 SAR imagery

Abstract

The increasing severity, duration, and frequency of destructive floods can be attributed to shifts in climate, infrastructure, land use, and population demographics. Obtaining precise and timely data about the extent of floodwaters is crucial for effective emergency preparedness and mitigation efforts. Deep convolutional neural networks (CNNs) have shown astonishing effectiveness in various remote sensing applications, including flood mapping. One of the key limitations of CNNs is that they can only predict whether a desired feature will appear in an image, not where it can be recognized. To address this limitation, the incorporation of self-attention mechanisms deployed in vision transformers (ViTs) can be particularly effective. However, the self-attention modules in the ViTs are complex and computationally expensive, and they require a wealth of ground data to attain their full capability in image classification/segmentation. Thus, in this paper, we develop the Residual Wave Vision U-Net (WVResU-Net), a deep learning segmentation architecture that utilizes advanced Vision Multi-Layer Perceptrons (MLPs) and ResU-Net for accurate and reliable flood mapping using Sentinel-1 SAR’s dual polarization data. Results showed the significant superiority of the developed WVResU-Net algorithms over several well-known CNN and ViT deep learning models, including Swin U-Net, U-Net+++, Attention U-Net, R2U-Net, ResU-Net, TransU-Net and TransU-Net++. For example, the segmentation accuracy of TransU-Net++, SwinU-Net, ResU-Net, R2U-Net, Attention U-Net, TransU-Net, and U-Net+++, was significantly improved by approximately 5, 12, 13, 13, 16, 19, and 23 percentage points, respectively in terms of recall obtained by the WVResU-Net with a recall value of about 69.67%. The code will be made publicly available at https://github.com/aj1365/RWVUNet.

Keywords

• Flood mapping
• Deep learning
• Semantic segmentation
• Vision MLP
• ResU-Net