Blanch et al. (2026) AI image-based method for a robust automatic real-time water level monitoring: a long-term application case

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-02-12
• Authors: Xavier Blanch, Jens Grundmann, Ralf Hedel, Anette Eltner
• DOI: 10.5194/hess-30-797-2026

Research Groups

• Institute of Photogrammetry and Remote Sensing, Dresden University of Technology, Dresden, Germany
• Department of Civil and Environmental Engineering, Universitat Politècnica de Catalunya-BarcelonaTech, Barcelona, Spain
• Institute of Hydrology and Meteorology, Dresden University of Technology, Dresden, Germany
• Fraunhofer-Institut für Verkehrs- und Infrastruktursysteme IVI, Dresden, Germany

Short Summary

This study presents a robust, automated camera gauge for long-term, near real-time river water level monitoring, employing artificial intelligence for image-based segmentation and ground control point identification combined with photogrammetric techniques. Tested over 2.5 years at four sites, the system achieved high performance with mean absolute errors ranging from 0.96 to 2.66 cm, demonstrating resilience to adverse conditions and enabling continuous 24/7 monitoring.

Objective

• To develop and demonstrate a robust, fully automated, contactless, and low-cost AI image-based method for accurately quantifying river water levels with centimetre precision in near real-time, under various environmental conditions including nighttime and adverse weather, over extended observation periods.

Study Configuration

• Spatial Scale: Four river monitoring sites in Saxony, Germany: Elbersdorf (Wesenitz River), Großschönau 2 (Mandau River), Lauenstein 4 (Müglitz River), and Neukirch (Wesenitz River).
• Temporal Scale: Over 2.5 years (from late 2021 to June 2024), with images acquired at 15-minute intervals (96 images per day).

Methodology and Data

• Models used:
  • AI Segmentation: UPerNet architecture with a ResNeXt50 backbone, trained on the RIWA dataset (River Water Segmentation Dataset) and refined using Deep Active Learning. A separate model was trained for infrared (IR) nighttime images.
  • AI GCP Identification: Adapted R-CNN Keypoint detector neural network, retrained on KIWA imagery.
  • Photogrammetry: Structure-from-Motion Multi-View Stereo (SfM-MVS) for 3D model generation, PyBathySfM for refraction correction, collinearity equations for exterior orientation.
  • Data Processing: Sobel edge-detection filter for image quality, KNNImputer for missing GCP coordinates, modified Tukey filter (2 × IQR) for outlier rejection.
• Data sources:
  • Surveillance cameras (Axis Q1645 LE and Q1615 Mk III) capturing 1920 × 1080 pixel images.
  • Remotely controlled infrared (IR) illumination for nighttime monitoring.
  • Ground Control Points (GCPs) surveyed with centimetre accuracy using RTK-GNSS.
  • 3D site models generated from terrestrial and UAV imagery, supplemented by RTK-GNSS cross-sections.
  • Reference water level data from official gauging stations (float-operated and bubble gauges) providing 15-minute averaged values.
  • RIWA dataset, WaterNet, and ADE20K images for AI model training.

Main Results

• The system processed approximately 218,000 images over more than 2.5 years across four sites.
• High performance was demonstrated with mean absolute errors (MAE) for individual measurements ranging from 1.26 cm (LAU) to 2.66 cm (ELB). Daily averaged MAE values were 0.94 cm (LAU), 2.25 cm (ELB), and 1.56 cm (GRO).
• The system showed resilience to adverse weather and lighting conditions, achieving an image utilization rate above 95% (ranging from 86.9% to 98.1% of valid water level estimates relative to all acquired images).
• Integration of infrared illumination enabled continuous 24/7 monitoring capabilities.
• The 95th percentile of absolute differences for individual measurements ranged from 3.0 cm (LAU) to 7.0 cm (ELB).
• Pearson correlations ranged from 0.97 to 0.99 and Spearman correlations from 0.94 to 0.97, indicating strong agreement with reference gauges.
• The system met the German guideline of <2.0 cm systematic error for operational water-level monitoring at LAU (91.9% of daily means within ±2 cm) and GRO (71.0% of daily means within ±2 cm).
• Key factors influencing absolute error were identified as camera calibration, GCP stability, and vegetation changes.

Contributions

• First systematic documentation of an image-based contactless method's operational reliability over extended periods (>1 year) across multiple sites, including full 24/7 monitoring under adverse weather and nighttime conditions.
• Development of a robust, low-cost, and non-invasive camera gauge system that combines advanced AI techniques (UPerNet for water segmentation, R-CNN Keypoint detector for GCPs) with established photogrammetric methods for centimetre-precision water level measurements.
• Significant advancement in hydrological monitoring capabilities, particularly for flood detection and mitigation in ungauged or remote areas, by providing frequent, near real-time updates and visual context.
• Establishes a foundation for future multi-stage river monitoring, enabling the derivation of surface velocity and discharge from a single camera sensor without additional in-stream instrumentation.

Funding

• Bundesministerium für Forschung, Technologie und Raumfahrt (Federal Ministry of Education and Research, Germany)
  • Grant numbers: 13N15542 and 13N15543
  • Program: "Artificial Intelligence in Civil Security Research" (part of the Federal Government's "Research for Civil Security" programme)
• KIWA project (Künstliche Intelligenz für die Hochwasserwarnung – Artificial Intelligence for Flood Warning)

Citation

@article{Blanch2026AI,
  author = {Blanch, Xavier and Grundmann, Jens and Hedel, Ralf and Eltner, Anette},
  title = {AI image-based method for a robust automatic real-time water level monitoring: a long-term application case},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-797-2026},
  url = {https://doi.org/10.5194/hess-30-797-2026}
}