An enhanced tidal flat waterline extraction algorithm based on curated Sentinel-2 training data and its applications

The accurate mapping of waterlines in tidal flats is essential for coastal management, environmental monitoring, and disaster response. Coastline delineation from optical images is relatively straightforward due to the distinct spectral differences between land and sea. The Landsat series and Sentinel-2 A/B satellites provide extensive open-access optical data, facilitating the extraction and analysis of coastline changes over time. However, tidal processes, loose mud, the presence of aquaculture areas, and the varying widths of tidal channels introduce significant challenges to waterline extraction in tidal flats, regardless of whether thresholding methods or deep learning models are applied. To mitigate these limitations, this study proposes a novel deep learning algorithm developed using curated Sentinel-2 training data. The training dataset was generated by experts with field experience in intertidal waterline identification, employing thresholding and visual interpretation across 32 sites in the Yellow Sea. These sites represented diverse tidal conditions—low, medium, and high tides—to ensure data diversity and representativeness. The algorithm was evaluated at 10 sites, including locations in the Yellow Sea and other coastal areas globally. It achieved an average land-sea segmentation F1-score of 0.9869, and an IoU of 0.9747; a waterline distance error of 1.07 pixels, and a waterline Pratt's Figure of Merit of 0.73. The robustness of the proposed algorithm was further validated through comparisons with previous methods and assessments using drone-based tidal flat field data. Its practical applicability was demonstrated by generating land-masking vector data, intertidal mapping, and tidal flat DEMs. This study offers a robust and reliable deep learning solution for accurately extracting coastal waterlines, contributing to improved coastal management and tidal flat environmental monitoring. Furthermore, the algorithm holds potential for global applications in coastal waterline extraction and related studies in the future.