A satellite reveals the rivers that global models struggle to follow

On the Connecticut River, a pumped hydropower reservoir can alter water depth by more than a metre in a day. Hydrologist Colin Gleason uses that example to explain why predicting a river from weather alone can fail: the water also responds to electricity prices and the operator’s pumping decisions.[2]

A model might reproduce rainfall and snowmelt convincingly while missing the decisions that move water on a particular day. Training on other rivers cannot guarantee that those local operating rules will be recovered. The difficulty concerns what the model knows about the river as well as the statistical method it uses.

Gleason’s team compared SWOT observations with models for 68,347 reaches, requiring at least 20 same-day pairs during July 2023–December 2025. They used a machine-learning ensemble where available and a traditional-model ensemble elsewhere.[1]

A surface measurement can test a flow prediction

SWOT, the Surface Water and Ocean Topography mission, measures water-surface height, width and slope. NASA describes its river measurements as changes in those dimensions over time; the satellite does not directly measure the absolute volume of water hidden in a river channel.[4]

The height measurement comes from radar. Two antennas, separated by a ten-metre boom, receive reflections from the same patch of water. The small difference in the phase of the returning signals, combined with the antenna separation and radar wavelength, allows researchers to calculate surface height. Each side of the spacecraft covers a swath about 50 kilometres wide. The instrument was designed to observe rivers wider than 100 metres, so this study concerns a selected network of substantial waterways.[3]

Flow, or discharge, asks how much water passes a point per second. Height asks where the surface is. Hydrologists connect them using a rating curve specific to a monitoring location; USGS publishes these curves alongside its streamflow records. The need for a local relationship is why a metre of additional height cannot be converted into one universal increase in flow.[5]

The study’s score, SES, is the Spearman rank correlation between observed height and modelled discharge. It tests whether higher-water observations coincide with higher-flow predictions.[1]

Ranking accommodates a curved relationship: a rise from one level to the next need not correspond to an equal increment in flow. It also leaves a blind spot. Two predictions can follow exactly the same ordering while assigning very different volumes to each observation. The following invented example isolates that distinction.

Calibration at 1,391 gauges yielded an SES threshold of 0.49: 75% of model results above it had positive Nash–Sutcliffe efficiency. The threshold suggests skill; some good models fall below it.[1]

Positive Nash–Sutcliffe efficiency means a model beats the benchmark of predicting the observed mean flow throughout the evaluation period.[6]

In the invented table, a discharge measurement on the ground would help decide whether A, B or neither is close. A high rank score narrows one class of error. It does not remove the need to measure or calibrate volumes.

The awkward rivers are widespread

The models were run for naturalised flow, without explicitly representing dam operations or withdrawals. Low agreement is therefore partly a test of that modelling assumption.[1]

In the university’s account, Gleason identifies dammed rivers, arid regions and the Arctic as particularly difficult. He links some arid-region problems to groundwater use, whose effect on rivers can be delayed, and Arctic problems to scarce training examples. These are proposed explanations for the geographical pattern, with different implications for improving the models.[2]

Excluding dam-affected, coastal, glacial and multichannel reaches left roughly 7,500 of the study’s 68,000 reaches, about 11%. That fraction describes the selected river categories; it is not the share of rivers modelled correctly.[1]

Tides, wetland hydraulics, wrongly matched reaches and small height changes relative to measurement error can also weaken the comparison. A low score requires interpretation before blame is assigned to the model.[1]

The useful outcome is a map of where different kinds of knowledge are missing. A pumping schedule calls for information about human decisions; an unfamiliar snowmelt regime calls for better observations of that environment. Satellite measurements help locate the disagreement, while gauges and local river knowledge help explain it. That makes the comparison valuable even where it cannot provide a single definitive answer about model accuracy.

Report an error or suggest a correction