Intelligent water level measurement based on visual foundation models

Abstract

Image-based water level measurement offers low-cost and visualization advantages, making it suitable for high-frequency, multi-location data collection. Although deep learning-based methods achieve high accuracy, their transferability and robustness are constrained by the need for site-specific large-scale training data. This study proposes an intelligent water-level measurement method based on visual foundation models (VFMs) to address this limitation under nonsufficient data. First, Language-guided Generative Data Augmentation (LGDA) is proposed to generate high-fidelity and diverse training images by simulating weather conditions or lighting variations. Then, the Segment Anything Model (SAM) is guided by representative pixel points prompts sampled from Deeplabv3+ probability maps to perform few-shot water surface segmentation. Finally, photogrammetry techniques are used to convert pixel coordinates into water level elevations. Compared with traditional data augmentation methods and other semantic segmentation models, the proposed method achieves higher accuracy in water surface segmentation, with an Intersection over Union (IoU) of 0.904, representing a 9.7 % improvement over the best-performing Deeplabv3+ baseline (IoU = 0.824). Transferability is validated across three hydrological stations using fewer than 50 training images per site. The proposed method achieves more accurate water level estimation, with a mean absolute error (MAE) below 0.026 m and a coefficient of determination (R²) exceeding 0.9 across all stations during one-month monitoring, significantly outperforming the baseline (MAE up to 0.055 m, R² as low as 0.521). The proposed model demonstrates strong transferability under limited data and enables more accurate and reliable water level measurement.