Research map / Time series / Causal and generative time series
The Merit of River Network Topology for Neural Flood Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Models river discharge at 358 Danube gauges with GNNs (GCNII, ResGCN, ResGAT) under different river-topology adjacency definitions and finds topology gives negligible benefit over isolated gauges or an MLP.
Problem
Data-driven river discharge forecasting usually treats gauging stations in the same river network as isolated problems, ignoring known network topology; it is unclear whether incorporating that topology helps prediction.
Approach
Use a single end-to-end GNN over the river network and systematically compare isolated, binary, physically weighted (stream length, elevation difference, slope) and learned adjacency, plus edge orientations, with a relevancy-weighted loss emphasising high-variability discharge windows.
Key results
- Isolated gauges (no adjacency) achieve 84.12% GCNII and 85.07% ResGCN mean NSE, with no topology variant clearly outperforming the isolated setting or the 85.37% wide MLP baseline.
- Learned edge weights show Pearson correlations with physical edge weights ranging roughly from -0.375 to 0.139 for ResGCN, with signs flipping across architectures.
- Worst gauge #169 shows sudden narrow discharge spikes that GNNs struggle to forecast.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add river-network topology rather than isolating gauges (ablation) | GCNII discharge forecaster · LamaH-CE Danube A | NSE: 84.12 ± 1.88 (isolated) -> 85.56 ± 1.41 (learned bidirected) | No clear effect |
| Use physical or learned edge weights instead of isolated gauges (ablation) | GCNII discharge forecaster · LamaH-CE Danube A | NSE: 84.12 ± 1.88 -> 85.56 ± 1.41 | No clear effect |
| Increase the input window size (ablation) | Bidirected-learned GCNII · LamaH-CE Danube A, lead time 6 h | NSE: 82.35 ± 1.94 -> 87.57 ± 1.21 | Helped |
| Reduce the forecast lead time (ablation) | Bidirected-learned GCNII · LamaH-CE Danube A, window size 24 h | NSE: 67.26 ± 5.12 -> 99.05 ± 0.06 | Helped |
| Use river-network topology in the prediction model (ablation) | GNN discharge forecasters · Four LamaH-CE subnetworks | NSE | No clear effect |
Surprising findings
- Removing all edges (isolated gauges) performs as well as any topology-aware adjacency, so the GNN effectively acts like an MLP.
- Learned edge weights correlate with neither static physical weights nor each other consistently across architectures, with correlation signs sometimes flipping.
Limitations
- Single dataset (LamaH-CE, central Europe Danube region) limits generality.
- No flood event annotations or inter-gauge propagation time metadata, so only continuous discharge is modelled.
- Only a subset of 358 gauges after filtering and no hyperparameter tuning per topology reported.
Experimental setup
Datasets: LamaH-CE
Benchmarks: LamaH-CE Danube A
Metrics: relevancy-weighted Nash-Sutcliffe Efficiency Pearson correlation
Baselines: isolated gauges (no adjacency) 19-layer MLP wide 2-layer MLP binary adjacency physically weighted adjacency
Builds on: GCN GCNII GAT CAMELS-x LSTM multi-gauge training DAGNN
Closest papers by meaning
- Physics-Guided Fair Graph Sampling for Water Temperature Prediction in River Networks (AAAI 2025 · similarity 0.89)
- RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting (NeurIPS 2025 · similarity 0.88)
- Fair Graph Learning Using Constraint-Aware Priority Adjustment and Graph Masking in River Networks (AAAI 2024 · similarity 0.87)
- Topology-aware Neural Flux Prediction Guided by Physics (ICML 2025 · similarity 0.85)
- Pluvial Flood Emulation with Hydraulics-informed Message Passing (ICML 2024 · similarity 0.85)
- Topology-Aware Conformal Prediction for Stream Networks (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.