Goodling et al. (2025) Technical note: A low-cost approach to monitoring relative streamflow dynamics in small headwater streams using time lapse imagery and a deep learning model

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-11-19
• Authors: Phillip Goodling, Jennifer H. Fair, Amrita Gupta, Jeffrey D. Walker, Todd L. Dubreuil, Michael Hayden, Benjamin H. Letcher
• DOI: 10.5194/hess-29-6445-2025

Research Groups

• Earth Surface Processes Division, US Geological Survey, Catonsville, MD, USA
• S. O. Conte Research Laboratory, Eastern Ecological Science Center, US Geological Survey, Turners Falls, MA, USA
• Microsoft Corporation AI For Good Lab, Redmond, WA, USA
• Walker Environmental Research LLC, Brunswick, ME, USA

Short Summary

This paper introduces a low-cost, non-contact method for monitoring relative streamflow dynamics in small headwater streams using time-lapse imagery and a deep learning model (SRE) trained on human-annotated image pairs. The method successfully characterized relative streamflow dynamics, with modeled hydrographs closely matching observed ones (median Kendall’s Tau of 0.75 during the annotation period).

Objective

• To describe and evaluate a low-cost, non-contact, and low-effort method for monitoring relative streamflow dynamics in small headwater streams using time-lapse imagery and a deep learning model.
• To assess the accuracy of human annotators in ranking images by streamflow.
• To determine the accuracy of image-derived relative hydrographs developed using person-generated annotations.
• To identify factors influencing ranking model accuracy and suitability for low-cost camera monitoring in unmonitored catchments.
• To quantify the number of person-generated annotations required to achieve stable ranking model performance.

Study Configuration

• Spatial Scale: 11 cameras at 8 streamflow sites in western Massachusetts, USA. Drainage areas ranged from 1.3 square kilometers to 107.2 square kilometers.
• Temporal Scale: Imagery data spanning at least 1.5 years (e.g., 2017-09-14 to 2024-04-09 for one site). Imagery was collected every 15 minutes.

Methodology and Data

• Models used: Streamflow Rank Estimation (SRE) deep learning framework, utilizing a fine-tuned convolutional neural network (ResNet-18 architecture pre-trained on ImageNet) and a learning-to-rank approach with the RankNet loss function.
• Data sources:
  • Time-lapse imagery from low-cost cameras (Reconyx Hyperfire 2, Bushnell Trophy, and Essential models).
  • Person-generated annotations, where individuals visually compared pairs of images to rank relative streamflow.
  • Co-located streamflow discharge data from USGS gages (USGS National Water Information System and Fair et al., 2025) for model evaluation.

Main Results

• Overall annotation accuracy for image pair ranking averaged 92.2% across the 11 camera sites, with near 100% accuracy for large flow differences and approaching 50% for small differences.
• Modeled relative hydrograph dynamics showed good correspondence with observed hydrograph dynamics.
• Model performance, measured by Kendall’s Tau rank correlation for "test-in" data (unseen images coincident with the training period), ranged from 0.60 to 0.83 (median 0.75).
• Performance for "test-out" data (unseen images following the training period) was lower, ranging from 0.34 to 0.74 (mean decrease of 0.20).
• Model performance was positively correlated with streamflow variability (coefficient of variation of log-transformed streamflow) and annotation accuracy.
• Camera stability showed a weak association with annotation accuracy and no significant relationship with model performance.
• A sensitivity analysis indicated that model performance generally plateaued around 1000 person-generated annotation pairs.

Contributions

• Introduces a novel, low-cost, non-contact, and low-effort methodology for monitoring relative streamflow dynamics in small headwater streams using time-lapse imagery and deep learning.
• Provides the first real-world evaluation of the Streamflow Rank Estimation (SRE) model using person-generated annotations across a diverse set of 11 camera sites at 8 streamflow locations.
• Characterizes human annotator accuracy patterns, showing high accuracy for large flow differences and lower accuracy for small differences, and proposes a function for simulating this performance.
• Identifies streamflow variability as a primary factor influencing model performance, offering guidance for selecting suitable monitoring sites.
• Establishes a practical guideline of approximately 1000 pairwise annotations for achieving stable model performance, optimizing annotation effort.
• Demonstrates the utility of relative flow estimates for various applications, including habitat characterization, ecological modeling, and extending stream monitoring networks, even without absolute discharge measurements.

Funding

• US Geological Survey (USGS) Next Generation Water Observing System (NGWOS) research and development program.
• US Environmental Protection Agency (US EPA) Professional Services Contract (68HE0B24P0246) awarded to Walker Environmental Research, LLC, with funding from the US EPA Regional-ORD Applied Research (ROAR) program.
• Collaborative agreement between Microsoft Corporation AI For Good Lab and the USGS.

Citation

@article{Goodling2025Technical,
  author = {Goodling, Phillip and Fair, Jennifer H. and Gupta, Amrita and Walker, Jeffrey D. and Dubreuil, Todd L. and Hayden, Michael and Letcher, Benjamin H.},
  title = {Technical note: A low-cost approach to monitoring relative streamflow dynamics in small headwater streams using time lapse imagery and a deep learning model},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-6445-2025},
  url = {https://doi.org/10.5194/hess-29-6445-2025}
}