Abstract:

Synthetic aperture radar (SAR) image water body segmentation is widely applied in critical fields such as disaster response, water resource management, and environmental monitoring, holding significant practical importance. To address the issue of low segmentation accuracy for water bodies in complex-background SAR images, a dual-encoder adaptive feature fusion network (DEAFFNet) is proposed to achieve accurate water body segmentation. First, the model employs a lightweight residual network and Swin Transformer to construct a dual-encoder architecture, collaboratively extracting local detail information and global contextual information to mitigate the problem of insufficient information representation capability in complex backgrounds. Second, a feature fusion module based on cross-attention and adaptive weight learning is designed. This module utilizes cross-attention for interaction between local and global information and achieves hybrid feature fusion through adaptive weight learning, thereby enhancing the model's perception of water body structures. Then, a multi-scale convolutional pooling module is integrated into the decoder to reinforce multi-scale contextual information, combined with a lightweight content-aware upsampling method to alleviate feature distortion caused by upsampling. Finally, a composite loss function combining focal loss and active contour loss is adopted to strengthen the model's constraints on sample balance and water body boundary information. Water body segmentation experiments conducted on the ALOS PALSAR and Sen-1SAR datasets demonstrate that DEAFFNet outperforms existing methods across multiple evaluation metrics, achieving more accurate water body segmentation.

Key words: SAR image water body segmentation, dual-encoder, cross attention, adaptive weight learning, composite loss