AI-Driven Photogrammetric Workflow for Low-Cost Wadi Monitoring

Arid regions such as Oman are increasingly susceptible to severe flash floods driven by climate change and rapid urbanization. Accurate water level measurements are vital for flood preparedness and the development of early warning systems required to mitigate severe socio-economic impacts, including substantial property damage and the recurring loss of life. Beyond disaster mitigation, recording runoff is essential for sustainable water management and for enhancing the understanding of hydrological processes in small-scale, ephemeral catchments that remain largely ungauged. However, traditional water level monitoring via pressure gauges or radar sensors is often hindered by high infrastructure costs, physical vulnerability to high-flow events, and the changing morphology of wadi channels. To address these limitations, we present a robust photogrammetric workflow integrated into a low-cost, Raspberry Pi-based optical monitoring system for measuring water level, surface velocities and discharge assessment.

The workflow relies on the synergy between a single fixed low-cost camera and high-resolution Digital Terrain Models (DTMs) generated through UAV-based Structure-from-Motion (SfM-MVS). To convert 2D image measurements into 3D object space, both the camera and the DTM must be referenced in a shared coordinate system. Traditionally, this is established using permanent Ground Control Points (GCPs) measured with RTK GNSS; however, establishing and maintaining such markers in adverse wadi conditions is logistically challenging and the physical markers are prone to being lost during flood events. We address this by employing the GIRAFFE (Geospatial Image Registration And reFErencing) workflow. This approach replaces physical markers by performing an image-to-geometry registration that aligns the real 2D camera view with a synthetic image rendered from the UAV-based 3D pointcloud. Using the AI-based LightGlue matching algorithm, the system automatically identifies homologous points between the views to create 2D–3D correspondences. These correspondences function as pseudo-control points, allowing for the precise determination of the camera’s 3D pose and orientation via spatial resection.

For the hydrological monitoring, the workflow further employs two AI-driven stages:

Water Level Estimation: Convolutional Neural Networks (CNNs) segment the water area in time-lapse images. The resulting waterlines are projected into 3D space and intersected with the DTM to derive accurate water levels.

Discharge Assessment: Surface flow velocities are measured using the PIPs++ (Persistent Independent Particle tracker) technique. Unlike traditional frame-by-frame methods, PIPs++ tracks particles across multiple time steps jointly, providing enhanced temporal smoothness and robustness against illumination changes or partial occlusions. Based on these surface velocities, the mean velocity is determined and combined with the wetted cross-section from the DTM to estimate total discharge.

Initial results from deployments in Wadi Al-Hawasinah, Oman, demonstrate that this solar-powered, remote system successfully captures ephemeral flow events. By leveraging GIRAFFE for automated localization and PIPs++ for robust surface velocity estimation, this workflow provides a scalable and cost-effective solution for enhancing flood early warning systems in complex, ungauged terrains.