Chen et al. (2025) Robust water level measurement using adaptive prompt staff gauge image segmentation based on EdgeSAM

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-09
• Authors: Hongyu Chen, Zhen Zhang, Jie Su, S. P. Wen
• DOI: 10.1016/j.jhydrol.2025.134724

Research Groups

• College of Information Science and Engineering, Hohai University, Changzhou, Jiangsu 213200, PR China
• College of Computer Science and Software Engineering, Hohai University, Nanjing, Jiangsu 211100, PR China
• Key Laboratory of Hydrologic-Cycle and Hydrodynamic-System of Ministry of Water Resources, Nanjing, Jiangsu 210024, PR China

Short Summary

This study proposes a robust image-based water level measurement method for staff gauges using an adaptive prompt EdgeSAM model, achieving high accuracy and generalization with minimal training data in complex field environments. The method addresses limitations of traditional and existing deep learning approaches by integrating image features with prompt information for precise segmentation and measurement.

Objective

• To develop a robust, accurate, and adaptable image-based water level measurement method for staff gauges that overcomes limitations of traditional contact/non-contact methods and existing deep learning approaches, particularly regarding large data requirements, computational complexity, and generalization in complex field environments.

Study Configuration

• Spatial Scale: Multiple field sites.
• Temporal Scale: Designed for real-time water level monitoring.

Methodology and Data

• Models used: EdgeSAM (fine-tuned from the Segmentation Anything Model - SAM framework).
• Data sources: Staff gauge images (few-shot learning with dozens of images). Preprocessing includes image denoising, blur preprocessing, grayscale covariance matrix calculations, and texture feature extraction for prompt point screening.

Main Results

• EdgeSAM enables fine-tuned model training with few-shot learning (dozens of staff gauge images), significantly reducing dataset construction costs while maintaining lightweight properties for edge device deployment.
• A hybrid strategy combining point and box prompts facilitates adaptive coarse-to-fine segmentation, enhancing model generalization in complex scenes.
• A screening process using grayscale covariance matrix calculations and texture feature extraction effectively mitigates optical noise and floating object interference.
• Field tests demonstrate that adaptive prompting substantially improves EdgeSAM’s segmentation accuracy in challenging conditions (blurring, shadows, glare, floating object obstructions).
• The fine-tuned model exhibits strong generalization capabilities, with over 95% of measurement results showing an error margin within ±1 cm across various water levels and water conditions.

Contributions

• Proposes a novel staff gauge image segmentation and water level measurement method based on EdgeSAM, integrating image features with adaptive prompt information through the SAM framework.
• Introduces a few-shot learning approach for fine-tuning EdgeSAM, drastically reducing the required training dataset size and computational resources.
• Develops a hybrid point and box prompt strategy for adaptive coarse-to-fine segmentation, enhancing model robustness and generalization in diverse and complex field environments.
• Implements an effective screening process for water surface prompt points using grayscale covariance matrix and texture features to counter optical noise and floating object interference.
• Achieves high measurement accuracy (over 95% within ±1 cm error) and strong generalization capabilities, making it a practical and promising solution for automated water level monitoring.

Funding

Not specified in the provided text.

Citation

@article{Chen2025Robust,
  author = {Chen, Hongyu and Zhang, Zhen and Su, Jie and Wen, S. P.},
  title = {Robust water level measurement using adaptive prompt staff gauge image segmentation based on EdgeSAM},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134724},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134724}
}