Predicting River Sediment Without Sampling: Machine Learning in Minnesota's StreamStats

How much sediment a river carries matters for aquatic habitat, flooding, excess nutrients and river restoration. Physical samples answer the question — but collecting them on every river, at every time of interest, is rarely practical, and older analytical and numerical methods could not carry what was learned at sampled sites over to unsampled ones.

So the U.S. Geological Survey (USGS) built machine learning (ML) models to predict suspended-sediment concentration (SSC) and bedload transport (BL) on Minnesota rivers that have no sediment data.

Image from the USGS fact sheet on StreamStats sediment prediction

The models

• Training data: about 1,300 SSC samples from 56 sites and 600 BL samples from 43 sites across Minnesota, plus streamflow records and geospatial data on each watershed, catchment, near-channel area and channel.
• Key inputs: streamflow, watershed and catchment characteristics, and how fast streamflow is changing (its slope).
• Performance: the models explain about 70 percent of the variability in the samples.
• Why ML: it can learn complex nonlinear relations and transfer what it learns from sites with data to sites without.

Sediment sampling sites used to train the models. Image from the USGS fact sheet on StreamStats sediment prediction

Built into StreamStats

Because the models are complex to run, USGS worked with the Minnesota Pollution Control Agency to put them into the USGS StreamStats web application — the first time sediment-prediction ML models have been part of StreamStats. Users can now get SSC and BL predictions for Minnesota rivers with no physical samples, supplying streamflow in one of three ways:

1. Pick a USGS streamgage and use its streamflow record.
2. Upload streamflow data as a comma-separated values file of daily or 15-minute flows — for example from the Minnesota Department of Natural Resources Cooperative Stream Gaging website.
3. Estimate streamflow with the flow duration curve transfer method when no data exist.

The inputs used and the predictions come back as graphs and as comma-separated values files.

Example inputs and outputs. Image from the USGS fact sheet on StreamStats sediment prediction

Limits

• The models cover the whole state, but a site may still fail — for example, because its basin was delineated wrongly or its data are incomplete.
• The tool does not run on smaller streams.
• It will not predict beyond the training data: nothing above 7,040 milligrams per liter of SSC or 1,885 tons per day of bedload.

What it is good for

• Saving money — predictions where there is no budget for extensive field sampling.
• Water-quality monitoring — spotting streams that stray from water-quality standards.
• River restoration — informing what to restore and in what order.

The catch: sampling has stopped

Funding for sediment monitoring has come and gone. After a decline in the 1990s, it surged with the approval of the Clean Water, Land and Legacy Amendment in 2008, creating the Clean Water Fund and letting USGS expand its network with the Minnesota Pollution Control Agency and the Department of Natural Resources. As priorities later shifted toward restoration, sites were dropped — and by 2025, no active USGS sediment-monitoring sites remained in Minnesota.

The long-term USGS records remain essential for tracking sediment trends, setting water-quality standards and guiding restoration. Keeping the models useful, the authors stress, will take a continued commitment to sampling.

Sources

Based on Using Machine Learning in Minnesota's StreamStats to Predict Fluvial Sediment, by Joel Groten, J. William Lund, Erin N. Coenen, Andrea Medenblik, Harper Wavra, Mike Kennedy and Gregory Johnson, U.S. Geological Survey Fact Sheet 2025–3005; a work of the United States government in the public domain.