Ogle et al. (2026) Image-based classification of stream stage to support ephemeral stream monitoring

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-02-06
• Authors: Sarah E. Ogle, Garrett McGurk, Anahita Jensen, F. Martin Ralph, Morgan Levy
• DOI: 10.5194/hess-30-709-2026

Research Groups

• Center for Western Weather and Water Extremes, Scripps Institution of Oceanography, University of California San Diego, San Diego, California, USA
• School of Global Policy and Strategy, University of California San Diego, San Diego, California, USA

Short Summary

This study develops a low-cost, image-based machine learning method to classify relative stream stage (no, low, or high water levels) in ephemeral streams using field camera imagery from 2017-2023 in the upper Russian River watershed, California, demonstrating its utility for monitoring and quality control in data-scarce intermittent river systems.

Objective

• To develop and evaluate a transferable, low-cost method using machine learning and field camera imagery to classify relative stream stage (no water, low water, high water) in intermittent rivers and ephemeral streams (IRES) to support monitoring and water management, particularly in data-scarce regions.

Study Configuration

• Spatial Scale: An ephemeral stream (Perry Creek, PEC) in the upper Russian River watershed, California, USA, with a drainage area of 7.05 km². Comparison with NOAA National Water Model (NWM) for a stream segment overlapping PEC. Transferability tested on two additional US sites (Dry Brook Upper, Massachusetts; Troy Creek, Nevada).
• Temporal Scale: Hourly field camera images (with gaps) from 2017–2023. Continuous 15-minute stage observations from August 2017 – October 2023. Hourly NWM discharge from August 2017 – November 2020. Manual discharge measurements from late 2017 – March 2024. Meteorological and soil moisture observations from August 2017 – November 2023.

Methodology and Data

• Models used: Supervised machine learning: Logistic Regression model (multinomial log-linear regression with L2 regularization). Image processing included grayscale conversion, Histogram of Oriented Gradients (HOG) transformation, and pixel value scaling. Hydrologic model for comparison: NOAA National Water Model (NWM) Retrospective Version 2.1 (based on NOAH-MP Land Surface Model).
• Data sources:
  • Field Camera Imagery: 15,821 hourly (or 30-minute) images from various trail cameras (Wingscapes TimelapseCam Pro, Spypoint Link S Dark, Force-S-Pro) at Perry Creek (PEC), with 537 manually labeled images (09:00-16:00 PST).
  • In-situ Stage Observations: Continuous 15-minute (or 5-minute) water level measurements from Solinst Levelogger or HOBO MX2001-04-SS-S pressure transducers at PEC, barometrically compensated.
  • Manual Discharge Measurements: 12 periodic measurements at PEC using handheld flow meters and a wading rod.
  • Meteorological and Soil Moisture Observations: 2-minute observations of precipitation, relative humidity, temperature, and soil moisture (at 5, 10, 15, 20, 50, and 100 cm depths) from the Deerwood (DRW) site.
  • Modeled Discharge: Hourly discharge from NOAA National Water Model (NWM) Retrospective Version 2.1 for a stream segment overlapping PEC.

Main Results

• The Logistic Regression model achieved a mean prediction accuracy of 91% and a mean balanced accuracy of 78% for classifying stream stage into "no water", "low water", "high water", and "obstructed" categories.
• High and medium confidence image classifications strongly agreed with observed stage: 94.6% of "any water" classifications corresponded to stage greater than 0 cm, and 99.9% of "no water" classifications corresponded to stage of 0 cm.
• The median stage for "high water" classifications was 27.1 cm, significantly higher than for "low water" (19.1 cm), with "high water" classifications capturing nearly all stage events above the 99th percentile (46.1 cm).
• Image classifications were effectively used for quality control of continuous stage data, identifying sensor malfunctions, noise, and periods of true dryness.
• The NOAA National Water Model (NWM) discharge often missed or underestimated short-duration, moderately-sized quickflow events, particularly during the early wet season, and showed discrepancies with manual discharge measurements (e.g., NWM predicted 0 m³/s while observed stage was greater than 10 cm in 1.8% of overlapping times).
• Environmental conditions analysis showed that "high water" stage positively correlated with daily and 30-day rolling sum precipitation (R² = 0.41 for both) and 5 cm daily rolling mean soil volumetric water content (R² = 0.37). Shallow soil moisture was a stronger predictor of water presence than deep soil moisture, with stage being 0 cm when 5 cm VWC was less than 0.15.

Contributions

• Provides a low-cost, transferable, and accurate method for monitoring categorical stream stage in sparsely observed intermittent rivers and ephemeral streams (IRES) using readily available field camera imagery and a simple machine learning model.
• Demonstrates the utility of image classification for quality control and validation of in-situ stage observations, particularly in IRES prone to sensor errors and data gaps.
• Highlights the limitations of generalized hydrologic models (like NWM) in representing IRES dynamics and shows how image classifications can augment and improve understanding of modeled discharge in these systems.
• Offers a practical approach to observe ecologically important "pooling" phases of IRES, which are not captured by traditional stage or discharge measurements.
• The method is adaptable for integration with existing platforms (e.g., USGS Flow Photo Explorer, CrowdWater) and can inform water management decisions (e.g., reservoir operations, habitat restoration) in climate-impacted freshwater systems.
• The model prioritizes simplicity and requires minimal manually-labeled training data (approximately 100 images per site).

Funding

• Hellman Fellows Program (faculty) award
• NSF Award (grant no. 2205239)
• U.S. Army Corps of Engineers Engineer Research and Development Center FIRO program (grant no. Award USACE W912HZ-24-2-0001)

Citation

@article{Ogle2026Imagebased,
  author = {Ogle, Sarah E. and McGurk, Garrett and Jensen, Anahita and Ralph, F. Martin and Levy, Morgan},
  title = {Image-based classification of stream stage to support ephemeral stream monitoring},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-709-2026},
  url = {https://doi.org/10.5194/hess-30-709-2026}
}