Transformer-Based Semantic Segmentation for Flood Region Recognition in SAR Images
Lifan Zhou, Xuanyu Zhou, Huanghao Feng, Wei Liu, Hao Liu · IEEE Journal on Miniaturization for Air and Space Systems · 2025
Monitoring and evaluating floods is crucial for geographic information systems (GISs). The low backscattering coefficient of flood surfaces makes them appear darker in synthetic aperture radar (SAR) images, which is advantageous for flood segmentation. In recent years, with the advancement of deep learning, semantic segmentation of flood regions in SAR images using convolutional neural networks (CNNs) has become a focal point in earth observation tasks. However, challenges, such as the similarity between the texture and shape of flood regions and the background in SAR images, the segmentation discontinuity at flood edges, the loss of information on small water bodies, and the variability of flood regions in different scales and morphologies, remain inadequately addressed. To tackle these issues, we propose a transformer model based on an encoder–decoder architecture for precise segmentation of flooded areas in SAR images. First, we utilize the mix transformer as the model’s encoder to compensate for CNNs’ limitations in global modeling, enhancing the discrimination of similar features in the image. Second, we introduce a noise filtering module (NFM) to filter redundant semantic information within low-level feature maps during the feature fusion process, thereby mitigating segmentation discontinuities at flood edges and the loss of small water body information. Finally, we design a multiscale depth-wise convolution module (MDCM) to boost the network’s multiscale feature representation capability, addressing issues arising from flood scale variability. Experimental results demonstrate that our method surpasses other mainstream approaches on the Sen1Floods11 dataset.