PNNL

The Science

Measuring sediment grain size in streams and rivers is important for understanding how surface water flows, how microorganisms respire and cycle nutrients, how extensively surface water/groundwater interact, and ultimately, the role of a given watershed as part of the land surface in regulating ecosystem function. Median grain size is known as D50 and has traditionally been measured manually—this is tedious and limits the number of data points collected. Other “top down” methods use watershed characteristics to predict grain size across entire river basins, but may over-generalize, provide less-accurate estimates, and are often of limited use for the large number of small or feeder streams in any given watershed. A multi-institutional team of researchers explored how machine learning (ML) techniques could complement standard approaches using estimates derived from a process using the You Only Look Once (YOLO) object detection model. This approach was found to be faster than manual methods and more site-specific than models based on watershed characteristics. The team compared ML methods for sites across the Yakima River Basin in Washington State, demonstrating both advantages and limitations of object detection-based methods for estimating D50. These findings showed that advanced technologies, like object detection algorithms, could improve the ability to understand and predict watershed function based on water flow and material transformation. 

The Impact

New research explored how ML-enabled methods could make faster and more accurate riverbed sediment grain size estimates. Traditionally, methods either focused on small areas measured manually or predicted grain sizes across broad scales that lacked local-scale precision. Researchers compared several conventional methods to a ML approach across the Yakima River Basin in Washington State to understand the strengths and limitations of each method. By combining conventional and ML approaches, this method was found to support subsequent research on how changes in the size distributions of sediments influence water flow, habitats, and nutrient cycles. Other fields, such as environmental modeling and hydrology, could benefit from these techniques to analyze sediment data and predict river system changes.

Summary

Sediment grain size for rivers and streams, often reported as D50, plays a critical role in understanding water flow, ecosystem health, and biogeochemical processes in flowing water bodies. Researchers explored an advanced approach using the YOLO ML algorithm to use artificial intelligence to analyze sediment images and provide precise grain size measurements. This method complements conventional techniques by providing high-throughput, non-destructive sampling, along with site-specific information, and the ability to estimate within-site D50 variability.

The study revealed differences in average D50 values, variability, and their relationships with the characteristics of rivers and streams within a given watershed. YOLO’s ability to capture localized sediment data on a larger scale, relative to manual methods, opens pathways for improved understanding of sediment dynamics and better integration of these findings into watershed and basin-wide models. By bridging the gap between small-scale precision and broad spatial coverage, ML-enabled object detection methods like YOLO are valuable tools for hydrology, environmental science, and other disciplines studying rivers and streams within watersheds. This approach provides an ability to quickly and easily assess spatiotemporal patterns of grain size and its impacts on river processes with greater efficiency and reproducibility.

The initial draft of the text above was created using ChatGPT (version 5.5 or lower, OpenAI). The language and content were subsequently edited by the author for grammar, clarity, and accuracy, and the final document was reviewed by the author.

Research Contacts

James Stegen, Pacific Northwest National Laboratory

Peter Regier, Pacific Northwest National Laboratory

Funding

This work was supported by the Department of Energy (DOE), Biological and Environmental Research program, Environmental System Science (ESS) program as part of the River Corridor Science Focus Area at Pacific Northwest National Laboratory (PNNL). PNNL is operated by Battelle Memorial Institute for the DOE. 

Related Links

• Regier P ; Chen Y ; Son K ; Bao J ; Forbes B ; Goldman A E ; Kaufman M H ; Rod K A ; Stegen J C (2023): Data associated with “Machine learning photogrammetric analysis of images provides a scalable approach to study riverbed grain size distributions.” River Corridor and Watershed Biogeochemistry Science Focus Area, ESS-DIVE repository. Dataset. DOI:10.15485/1972232
• River Corridor Science Focus Area Project Website