Scholarlink

This study investigates the potential of Deep Learning (DL) for automated monitoring of beach seagrass wracks (BW) and shoreline position using snapshots from beach imaging systems (SIRENA and CoastSnap, respectively) and specialized-labelled datasets. BW detection and segmentation is approached using U-Net, YOLOv8–9, and the Segment Anything Model (SAM). The shoreline extraction is tackled using U-Net and Bidirectional Long Short-Term Memory (Bi-LSTM) networks. For each challenge, experiments involving variable data processing and models’ parametrization are explored. Results for shoreline extraction from CoastSnap images indicate that the U-Net outperforms Bi-LSTM, while Bi-LSTM, trained beach-wise, offers a viable option for limited training data scenarios. For BW monitoring from SIRENA images, YOLO outperforms the U-Net and SAM, but detection and segmentation performance decreases with lower density BW. Overall, this research presents open-access datasets, insights into model performance, and findings transferable to different imaging systems, showcasing the potential of DL applied to natural image processing for enhancing coastal monitoring.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine Learning-Driven Shoreline Extraction and Beach Seagrass Wrack Detection from Beach Imaging Systems

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Jesús Soriano-González,
  • Jose David Pérez-Cañellas,
  • Josep Oliver-Sansó,
  • Francisco Fabián Criado-Sudau,
  • Àngels Fernández-Mora,
  • Elena Sánchez-García

摘要

This study investigates the potential of Deep Learning (DL) for automated monitoring of beach seagrass wracks (BW) and shoreline position using snapshots from beach imaging systems (SIRENA and CoastSnap, respectively) and specialized-labelled datasets. BW detection and segmentation is approached using U-Net, YOLOv8–9, and the Segment Anything Model (SAM). The shoreline extraction is tackled using U-Net and Bidirectional Long Short-Term Memory (Bi-LSTM) networks. For each challenge, experiments involving variable data processing and models’ parametrization are explored. Results for shoreline extraction from CoastSnap images indicate that the U-Net outperforms Bi-LSTM, while Bi-LSTM, trained beach-wise, offers a viable option for limited training data scenarios. For BW monitoring from SIRENA images, YOLO outperforms the U-Net and SAM, but detection and segmentation performance decreases with lower density BW. Overall, this research presents open-access datasets, insights into model performance, and findings transferable to different imaging systems, showcasing the potential of DL applied to natural image processing for enhancing coastal monitoring.