Low-cost stereo vision and deep learning for river water level measurement

Abstract. The increasing frequency of extreme hydrological events, driven by climate change, necessitates dense and scalable water level monitoring networks. We evaluate a low-cost, non-contact, stereo-vision camera system for automated water level estimation. We compare two distinct image-processing pipelines — with and without semantic masking — to determine their camera pose stability and measurement accuracy. Our results show that raw stereo-vision estimates are highly correlated with reference sensor measurements (correlations ranging from 0.70 to 0.77), capturing the overall hydrologic behavior. By implementing a masking technique to isolate static environmental features, we successfully corrected a baseline error (i.e., offset), aligning the system with the true physical geometry. Although masking improves absolute accuracy, it introduces transient instability (i.e., spikes in pose estimation). This study serves as a proof of concept for the deployment of low-cost, edge-based stereo-vision systems in hydrological monitoring.