DWFAF-Net: A water detection network for multi-frequency SAR systems

Water detection using SAR images has significant application value, providing guidance for natural disaster prevention. Although considerable progress has been made in water body detection for single-system SAR images, the detection results are poor when applying SAR images from different systems and different bands, which seriously affects its practical application value. To address this issue, this paper proposes a new neural network model, namely the Dualencoder Weighted Feature Attention Fusion Network (DWFAFNet), aimed at efficiently detecting water body information from multi-frequency SAR images. It consists of the encoder and the decoder. In the encoder, a dual backbone network is constructed, namely the ResNet-34 and the Asymmetric Convolution Encoder (ACE) backbone network, which collaboratively extract water body features and detailed information. In the decoder, feature maps of different sizes generated by the encoders were transmitted to the Channel and Spatial Attention Fusion (CSAF) module through skip connections. This module allocates different weights in channel and spatial dimensions, enhancing the extraction ability of water body features while suppressing background interference, thereby improving water body detection accuracy. To validate the effectiveness of the proposed network model, independent tests were conducted on multi-frequency SAR images (Gaofen-3 system in C-band, and COSMO-SkyMed system in X-band), and the results were compared with three commonly used detection networks. The experimental results show that DWFAF-Net significantly outperforms the other networks, with an average water detection accuracy of up to 96%. This indicates that DWFAF-Net demonstrates excellent performance in multi-frequency SAR image water detection, which may promote the wide application of deep learning in SAR image water detection.