A new study using data from NASA’s Surface Water and Ocean Topography (SWOT) satellite has revealed significant weaknesses in advanced computer models used to estimate river conditions worldwide. The findings highlight potential risks for climate change predictions irrigation planning, hydropower operations and water resource management.
The research led by Colin Gleason a professor of civil and environmental engineering at the University of Massachusetts Amherst used satellite observations to identify where machine learning based river models provide reliable estimates and where they produce inaccurate results.
According to the study published in the journal Geophysical Research Letters rivers that are affected by dams those flowing through arid regions and Arctic rivers are among the most difficult to model accurately. Although fewer than 10% of river reaches fall into the category of serious modelling errors, these areas are often particularly important for water resources and climate sensitive planning.
The findings raise concerns because many regions depend on river models to forecast water availability, plan irrigation, develop hydropower projects and prepare for climate-related risks, particularly where ground-based measurements are limited.
Dams and dry regions pose major challenges
Researchers found that river models perform better on relatively straightforward rivers that are not glacially fed have a single channel, lack estuaries and are unaffected by dams. However, only about 11% of the world’s rivers fit this description.
Gleason noted that complex rivers, including those with multiple channels and dams are more common than conventional modelling assumptions might suggest.
Dammed rivers present a particular challenge because water levels can change rapidly due to human operations rather than natural hydrological cycles alone. Gleason cited the Connecticut River in the United States, where a pumped storage hydropower facility can alter river depth by more than a metre in a single day.
Accurately predicting such changes may require information about the facility’s operations, electricity prices and the economic thresholds governing when water is pumped or released. Without this information, models may struggle to estimate river conditions correctly.
Arid regions, including Australia, Central Asia, the southwestern United States and Mexico, also pose difficulties. Researchers suggest that groundwater use may influence river flows in complex ways with the effects of pumping sometimes appearing after a delay.
Arctic rivers remain difficult to predict
The study also identified Arctic rivers as a major modelling challenge partly because machine learning systems lack sufficient data from these environments.
Iceland illustrates the problem. Its glaciers, snow and ice, active geology and distinctive geographical conditions make it difficult to find comparable locations from which machine learning models can learn reliable patterns.
According to Gleason, machine learning performs well when it can identify recurring patterns in available datasets but its ability to make reliable estimates is limited when comparable observations are scarce.
These data gaps can undermine river predictions in regions where reliable information is especially important for understanding changing water availability and preparing for climate impacts.
SWOT satellite data offers an alternative
The SWOT satellite mission is providing researchers with a new way to assess river conditions on a global scale. By comparing satellite observations with model estimates, scientists can identify where existing models are reliable and where their predictions need improvement.
The study’s findings could help researchers focus on river systems that require better observations, more representative datasets and improved modelling methods.
Gleason also suggested that satellite measurements could sometimes be used directly instead of relying on inaccurate model estimates. He argued that researchers should prioritise measuring rivers and trusting observed data where it provides a more reliable picture than computer predictions.
The research does not suggest that river models are universally unreliable. Rather, it demonstrates that their accuracy varies considerably depending on river characteristics, human interventions and the availability of observational data.
The findings underscore the importance of combining satellite measurements with improved modelling approaches to strengthen water management hydropower planning, irrigation forecasts and climate change preparedness.