FSDC: A Frequency-Selective Deformable Convolution for Water Body Extraction From Remote Sensing Images

With the growing demand for high-precision water body extraction in complex environments, traditional segmentation models continue to face challenges such as insufficient boundary detection and limited multi-scale feature representation. To address the limitations of fixed dilation rates in conventional convolutional networks, which constrain the adaptability of receptive fields across varying spatial scales, this paper proposes a novel segmentation module: the Frequency-Selective Deformable Convolution. The Frequency-Selective Deformable Convolution consists of two key components: a Frequency-Selection module and a deformable convolution unit. The frequency-selection module performs frequency-domain decomposition via Fourier transform and applies learnable masks to selectively enhance or suppress different frequency bands, enabling the network to focus on the frequency-domain features of key spatial structures. The deformable convolution component introduces content-aware sampling through learned offsets, dynamically adjusting the receptive field in the spatial domain based on local geometric variations. This collaborative mechanism of the frequency-spatial dual domain not only strengthens the feature representation of multi-scale spatial structures through frequency-domain filtering but also improves the capture of geometric details via adaptive sampling in the spatial domain, ultimately enhancing the model’s capability for multi-scale and structurally adaptive feature modeling. We integrate Frequency-Selective Deformable Convolution into three mainstream architectures, namely UNet, Attention-UNet, and TransUNet, to construct enhanced models. Experimental results on high-resolution, multi-source remote sensing imagery demonstrate that the FSDC-introduced models significantly improve water body extraction performance, especially in terms of boundary accuracy, omission suppression, and false positive reduction. These findings underscore the potential of frequency-aware, adaptive convolutional mechanisms for advancing remote sensing applications such as water resource monitoring, ecological management, and land cover analysis.