PAEE-Net:A multi-scale attention water detection network for multi-frequency SAR

Water body detection from multi-frequency SAR images has broad practical value, offering scientific guidance for sustainable urban development. To address the challenges of multi-scale water detection and the uncertainty in the decision-making process of attention mechanisms due to the black-box nature of deep learning, this paper proposes the Pooling Attention and Edge Enhancement Networks (PAEENet) framework.It consists of the encoder and the decoder. In the encoder, the proposed the Attention Atrous Spatial Pyramid (AASP) module enables interaction of water body information across different levels, with an embedded the Channel and Spatial Pooling Attention (CSPA) attention mechanism that reallocates weights across dimensions to enhance feature extraction while suppressing irrelevant background information. In the decoder, the Edge Enhancement Model (EEM) module effectively improves segmentation at water-land boundaries. Additionally, the Pooling Attention Gradient-weighted Class Activation Mapping (PAG-CAM) module is introduced for interpretability, generating heatmaps to visualize and explain the decision logic of the attention mechanism. To validate the effectiveness of the proposed network, independent tests were conducted on multi-frequency SAR images and compared with three other detection networks. Experimental results show that PAEE-Net achieves an average water detection accuracy of 90.22%, outperforming the comparison models and demonstrating superior capability in handling multi-scale water body detection.