CWBSNet: A Segmentation Network for Comple Water Bodies in Remote Sensing Images

To address the challenges posed by diverse water body morphologies, irregular boundaries, and significant reflectance variations in remote sensing images—factors often overlooked by general-purpose models that lead to local omissions, boundary mislocalizations, and false detections—we propose a novel complex water body segmentation network (CWBSNet). This network is designed for water body extraction tasks on both the Sentinel-2 dataset S1S2-Water and the aerial dataset FLAIR#1. CWBSNet employs a transformer-based backbone to generate four-level feature embeddings with long-range global dependencies. To enhance global water body representation and mitigate local omission issues, we introduce the spectrum transform module (STM), which exchanges spectral information between water and nonwater regions on small-scale features via frequency-domain transformations. Furthermore, we design a multilevel feature extraction and fusion module (MFEFM), comprising Block1, a spatial-channel information fusion block (SIFB), and Block3, which enhances local feature representations through attention mechanisms, multiscale convolutions, and spatial-channel transformations. Directional convolutions are incorporated into shallow features to capture orientation cues and refine boundary detail representation. In addition, based on the strong response of near-infrared (NIR) and normalized difference water index (NDWI) to water bodies, we propose the information augmentation module (IAM). This module integrates rotary position embedding (ROPE) attention into a linear attention transformer to support water body recognition and reduce false positives. Finally, a decoder generates the saliency map. Experiments on two public datasets demonstrate that CWBSNet achieves superior performance in both training efficiency and detection accuracy compared to the state-of-the-art methods. The code is available at: https://github.com/Lingxp/CWBSNet