Learning River Flow Taught One Model to Track Water Chemistry Too

Hydrologists in Germany trained one neural network to predict river flow and six measures of water chemistry at once across 408 catchments, and report it held its flow accuracy while taking on the chemistry too.
Jean-Paul Brede and Ralf Loritz of the Karlsruhe Institute of Technology, with colleagues at the Helmholtz Centre for Environmental Research (UFZ) and at Stone Environmental in Vermont, published the work Nov. 18, 2026, in Hydrology and Earth System Sciences. The authors describe a data problem: river flow is gauged automatically, nearly every day, while river chemistry is collected by hand (in most of these catchments, every 20 to 50 days), and grab samples can miss short concentration peaks. They argue that the chemical measures move with flow and with one another, so learning them jointly partly compensates for a thin record.
The models were scored on a combined measure of how closely predictions match observations in timing, average level and spread, with a perfect match at the top of the scale. The authors report that flow predictions held a median of 0.84 to 0.87 across every configuration tested, while the six chemistry targets spread much more widely — from 0.40 for phosphate, the hardest of the set to reproduce, to 0.95 for water temperature.
The authors also report that training the network to predict flow alongside a chemistry variable matched or beat feeding it measured flow as an input for five of the six variables, the clearest gain being 0.05 for dissolved organic carbon; only dissolved oxygen came out slightly better with flow fed in. They recommend making streamflow an extra target in water-quality models of this kind, and note that such a model needs no flow measurement to run.
The team also tested a second training strategy, hiding variables at random and having the network rebuild them from the rest. The authors report no systematic accuracy gain over the supervised version and describe its advantage as flexibility: one trained model can be run on whatever inputs a catchment has, without retraining. Both were trained on two published German datasets, CAMELS-DE and QUADICA.
