A novel framework for automated water level estimation using CCTV imagery in Yongseong Agricultural Reservoir, South Korea

Study region: The study region is Yongseong Reservoir, located in Gyeongsangbuk-do, South Korea, a small agricultural reservoir primarily used for irrigation and is subject to pronounced hydrological seasonality. Study focus: This study proposes a novel framework for estimating water levels in ungauged agricultural reservoirs using images from CCTVs originally installed for security purposes. The method integrates a U-Net-based water-body segmentation model with four machine learning regression algorithms (support vector regression, SVR; random forest, RF; extreme gradient boosting, XGB; and light gradient boosting machine, LGBM) to predict reservoir water levels from segmented water pixel counts. Importantly, we assess the potential of region of interest (ROI) filtering to enhance prediction accuracy, demonstrating that surveillance camera imagery can be effectively repurposed for hydrological monitoring in data-scarce environments. New hydrological insights for the region: The results revealed that ROI filtering significantly improved prediction performance, increasing R² by 10–20 % and reducing root mean squared error by up to 0.197 (for RF). The RF model achieved the highest overall accuracy (R² = 0.964), while SVR performed best during no temporal variations. XGB and LGBM showed balanced residuals but slightly underestimated water levels during peak fluctuations. This study demonstrates the feasibility of image-based water-level estimation in ungauged agricultural reservoirs using security CCTVs. The results underscore the importance of spatial input refinement (ROI filtering) for reliable hydrological modeling.