Mitigating Site‐Specific Data Dependency in Image‐Based Water Level Estimation Using Vision Foundation Models

Abstract The increasing frequency and severity of fluvial flooding underscores the urgent need for accurate and timely river water level monitoring. Visual gauges based on river cameras offer a cost‐effective means for water level observation. However, conventional end‐to‐end deep learning regression models that infer water levels from images typically require long‐term co‐operation with physical gauges to accumulate sufficient image‐water level data pairs for training. To reduce the calibration burden associated with such training data requirements, this study develops a data‐driven framework that embeds water level‐related priors into regression, and incorporates a vision foundation model to enable consistent and robust prior feature extraction. Specifically, the framework adopts a one‐shot foundation segmentation model to extract water masks and compute the Static Observer Flooding Index (SOFI), an indicator of water extent. A lightweight regression model then maps SOFI to river water levels. Tested on four representative rivers, the proposed framework demonstrates strong agreement with ground truth water levels, achieving an average Nash‐Sutcliffe efficiency exceeding 0.8. Moreover, it exhibits improved extrapolation capabilities at high water levels, with the average mean absolute error reduced by approximately 50% compared with end‐to‐end deep learning‐based regression models. The framework mitigates the high site‐specific data dependency in calibrating visual gauges, providing a scalable paradigm for river monitoring.