CSTFSeg: A high-resolution Chinese tidal flat dataset and multi-scale attention semantic segmentation network

Tidal flats act as a critical ecotone connecting land and sea, maintaining unique biodiversity and global blue carbon functions. High-precision dynamic mapping of these areas is paramount for addressing climate change and coastal disaster mitigation. Although deep learning has been widely applied in remote sensing, fine-grained semantic segmentation of tidal flats remains constrained by two major bottlenecks: the lack of public, high-resolution pixel-level benchmark datasets, and the difficulty in precise segmentation due to spectral ambiguity at boundaries caused by hydrodynamic tides and sediment mixing. To overcome these limitations, this study presents a twofold contribution.First, we constructed and open-sourced the China Sentinel-2 Tidal Flat (CSTF) dataset, the first high-resolution semantic segmentation benchmark covering typical tidal flats along the Chinese coast. Based on Sentinel-2 imagery, the dataset spans 15 representative regions from the Liaohe Estuary to the Nanliujiang Estuary (comprising 1,136 samples). Through human-in-the-loop annotation and rigorous dual-verification, CSTF not only fills the data gap in this domain but also provides a high-quality evaluation benchmark (Overall Accuracy = 94.6%). Second, as a specialized solution tailored for this benchmark, we propose CSTFSeg, a multi-scale attention network integrating fuzzy logic and edge guidance. Driven by the physical characteristics of tidal flats, CSTFSeg incorporates three core modules: (1) An enhanced ResNet50 backbone, combined with Multi-Grid Dilated Convolutions and Channel Attention to bolster feature discrimination; (2) A Multi-scale Feature Transformer (MFT) to effectively aggregate local details with long-range global dependencies; and (3) An Edge-Guided Decoder, which innovatively incorporates a differentiable Fuzzy Layer to explicitly model the uncertainty of boundary pixels (i.e., mixed pixels), achieving deep interaction between edge and semantic features via a cross-scale fusion module.Experiments on the CSTF dataset demonstrate that CSTFSeg achieves a mean Intersection over Union (mIoU) of 93.78%, with the key tidal flat category reaching 90.55%, significantly outperforming eight mainstream models such as DeepLabv3+ and SegFormer. Furthermore, the model achieves an optimal balance between accuracy and efficiency, with only 29.27M parameters and an inference speed of 35ms per image. Spatial cross-validation across distinct regions further confirms the model's robust generalization capability. This study not only provides critical data resources for coastal remote sensing but also offers an effective methodological paradigm for solving segmentation challenges in coastal wetlands and other remote sensing targets characterized by weak boundaries.Data and Code Availability: The network code and CSTF dataset download links are available at: https://github.com/GGboywx/CSTFSeg