RiverScan stops averaging tributaries: area-weighted accumulation tops the evaluated baselines at 608 Danube gauges for daily runoff
Synopsis
Large-basin runoff prediction must carry heterogeneous subcatchment responses downstream along the river network, yet standard graph message passing does not explicitly account for the contributing area each tributary represents. RiverScan writes this accumulation as an area-fraction-weighted selective state-space recurrence, combining completed tributary states with local encodings in topological order and letting tributary state and reach attributes control retention. On 608 Danube gauges it gives the best overall predictive performance among the evaluated temporal and graph baselines without observed discharge inputs, and structural controls show that the benefit of correct topology grows along the gauge hierarchy.
Interpretation
RiverScan reframes runoff prediction on a river network as learning how to accumulate: each subcatchment keeps its own meteorological history, and complete upstream catchment states are composed step by step along river topology instead of being compressed into one averaged input. Shared LSTMs learn rainfall–runoff dynamics from catchment-averaged inputs, which compress spatial variability within a basin; one-hop synchronous message passing, degree normalization and learned attention also do not explicitly account for the unequal contributing areas represented by tributary states. On 608 gauges and 607 directed edges of the Danube LamaH-CE network, RiverScan achieves the best overall performance among the evaluated temporal and graph baselines without observed discharge inputs, with graph comparisons holding local inputs and the LSTM architecture fixed.
The accumulation rule comes from depth normalization under linear routing: local and tributary contributions are shared by area fractions, so the composed state corresponds to a complete contributing catchment. The recurrence builds on ordered graph recurrences and tree state-space models, but replaces the combination coefficients of tributary states with area fractions set by catchment geometry, and makes retention depend on the completed tributary state and reach attributes. The upstream scan starts at headwaters and proceeds in topological order, covering the full upstream subtree in one traversal; the state admits an exact decomposition into source-area fractions, local encodings and path retention products, and experiments cover all 608 gauges, 314 headwaters and 294 non-headwaters.
Removing area weighting is the largest loss among five matched component removals, indicating that combining tributary states by contributing area is a key part of how the operator works. All five component removals produce negative relative paired NSE changes in both the all-station and non-headwater groups, with the area-weighting removal declining most. Without area weighting, median KGE falls both overall and at non-headwaters; under the topology permutation control, RiverScan's median paired NSE gain over permutation grows from headwaters to gauges deeper in the upstream hierarchy, and all 44 gauges improve at the level whose longest upstream path contains four edges.
Perspective
The conclusions rest on 608 gauges of the Danube LamaH-CE network for daily rainfall–runoff prediction, with meteorological histories and static attributes as inputs and no observed discharge; graph comparisons hold local inputs and the LSTM architecture fixed. The authors propose finer-grained, fully distributed AI hydrological models as the next direction.
The authors note that whether river connectivity helps depends on the task: models tested on LamaH-CE for six-hour-ahead forecasting from hourly discharge and meteorological histories showed no advantage from true topology, while graph routing modules reported gains for daily prediction in the Upper Danube. When and how topology provides an effective inductive bias therefore remains open, and the conclusions here are limited to the daily setting without observed discharge inputs.
