A novel low-cost stereo camera system for river monitoring.

Most existing developments in camera gauges focus on single-camera configurations supported by ancillary information, such as detailed three-dimensional (3D) channel geometry and ground control points (GCPs). In these approaches, the water surface is delineated from images and reprojected onto a predefined 3D terrain model. However, acquiring accurate and up-to-date 3D models is often challenging or impractical, particularly in dynamic river environments where channel geometry evolves over time. As a result, frequent model updates are required to maintain measurement accuracy. Another key limitation of conventional camera gauges is the limited quantification of uncertainty in water level estimation. Surface velocity is typically derived using image velocimetry or particle tracking velocimetry. While these methods can provide accurate velocity measurements, they are contingent upon the selection of several parameters that must be meticulously chosen for each monitoring site. Furthermore, the performance of these methods can be degraded by challenging camera poses, varying illumination conditions, and flow regimes.

To address these limitations, we introduce a novel low-cost camera gauge system that integrates stereo photogrammetry with artificial intelligence (AI). The system comprises two low-cost cameras connected to a microcomputer capable of capturing, storing, and transmitting images and short video sequences to an online server. An AI-assisted multi-epoch stereo photogrammetry workflow is then applied to estimate camera pose and reconstruct dense 3D model. This process eliminates the need for predefined 3D data of the cross-section and allows us to compute a new and updated 3D model for each image pair. The updated 3D models are the key component of our methodology, from each water level and water surface velocity can measured in scaled values. Additionally, geomorphologic process can be also measured comparing subsequently 3D models. River water surface segmentation is performed using two foundation models, Grounding DINO and the Segment Anything Model (SAM). River waterlines from both images are then matched and projected into the 3D model, from which the water level is retrieved. This approach enables explicit assessment of errors in water level measurements. Particles tracked in video sequences, using a robust AI model, in both images are further projected into the 3D model, enabling scaled estimation of water surface velocity and, subsequently, river discharge.

The proposed methodology provides a robust and scalable remote sensing solution for river monitoring, enabling the observation of hydrological variables and geomorphological processes. Its low cost and reduced reliance on site-specific ancillary data make it well suited for addressing observational data gaps and for densifying hydrological monitoring networks. Moreover, with an appropriate setup, the system can be used for real-time monitoring, making it a valuable tool in scenarios such as flash floods.