Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Merit of River Network Topology for Neural Flood Forecasting

Nikolas Kirschstein, Yixuan Sun

TL;DR

This paper examines whether river network topology improves neural flood forecasting. Using an end-to-end GNN on the LamaH-CE dataset, it systematically varies adjacency definitions (including isolated, binary, physical, and learned weights) and evaluates performance across multiple splits. The key finding is that topology provides negligible benefits over isolated gauges, and learned edge weights do not meaningfully correlate with static physical relationships, even under deep or subnetworks. The work suggests focusing on spike forecasting and richer metadata (e.g., inter-gauge propagation times) or DAG-specific models to realize any potential gains from graph structure.

Abstract

Climate change exacerbates riverine floods, which occur with higher frequency and intensity than ever. The much-needed forecasting systems typically rely on accurate river discharge predictions. To this end, the SOTA data-driven approaches treat forecasting at spatially distributed gauge stations as isolated problems, even within the same river network. However, incorporating the known topology of the river network into the prediction model has the potential to leverage the adjacency relationship between gauges. Thus, we model river discharge for a network of gauging stations with GNNs and compare the forecasting performance achieved by different adjacency definitions. Our results show that the model fails to benefit from the river network topology information, both on the entire network and small subgraphs. The learned edge weights correlate with neither of the static definitions and exhibit no regular pattern. Furthermore, the GNNs struggle to predict sudden, narrow discharge spikes. Our work hints at a more general underlying phenomenon of neural prediction not always benefitting from graphical structure and may inspire a systematic study of the conditions under which this happens.

The Merit of River Network Topology for Neural Flood Forecasting

TL;DR

This paper examines whether river network topology improves neural flood forecasting. Using an end-to-end GNN on the LamaH-CE dataset, it systematically varies adjacency definitions (including isolated, binary, physical, and learned weights) and evaluates performance across multiple splits. The key finding is that topology provides negligible benefits over isolated gauges, and learned edge weights do not meaningfully correlate with static physical relationships, even under deep or subnetworks. The work suggests focusing on spike forecasting and richer metadata (e.g., inter-gauge propagation times) or DAG-specific models to realize any potential gains from graph structure.

Abstract

Climate change exacerbates riverine floods, which occur with higher frequency and intensity than ever. The much-needed forecasting systems typically rely on accurate river discharge predictions. To this end, the SOTA data-driven approaches treat forecasting at spatially distributed gauge stations as isolated problems, even within the same river network. However, incorporating the known topology of the river network into the prediction model has the potential to leverage the adjacency relationship between gauges. Thus, we model river discharge for a network of gauging stations with GNNs and compare the forecasting performance achieved by different adjacency definitions. Our results show that the model fails to benefit from the river network topology information, both on the entire network and small subgraphs. The learned edge weights correlate with neither of the static definitions and exhibit no regular pattern. Furthermore, the GNNs struggle to predict sudden, narrow discharge spikes. Our work hints at a more general underlying phenomenon of neural prediction not always benefitting from graphical structure and may inspire a systematic study of the conditions under which this happens.
Paper Structure (17 sections, 10 equations, 5 figures, 9 tables, 2 algorithms)

This paper contains 17 sections, 10 equations, 5 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Data Preprocessing
  5. The Forecasting Task
  6. Experiments
  7. Experimental Setup
  8. River Topology Comparison
  9. Learning the Weights
  10. Small-scale Subnetworks
  11. Worst Gauge Investigation
  12. Conclusion
  13. Appendix
  14. Preprocessing Algorithms
  15. Learned Edge Weights Statistics
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Historical occurrence of natural disasters by disaster type. The number of events increased over time, with floods being the most common. ritchie_natural_2022.
  • Figure 2: Geographical contextualisation of the LamaH-CE dataset. Circle colour indicates gauge elevation, while circle size indicates catchment size. klingler_lamah-ce_2021
  • Figure 3: Year-wise maximisers of the relevancy score $\mathop{\mathrm{\varrho}}\nolimits$. Each maximiser's discharge window (blue) exhibits both high variability as well as an excessive overall discharge level, in relation to the mean discharge of its gauge (gray).
  • Figure 4: Four subgraphs of the river network with different depth and sink in-degree. The node labels refer to the original gauge IDs from the dataset.
  • Figure 5: Worst predictions of the bidirected-learned GCNII (third fold) on its overall worst gauge #169. Negative time indicates past, and positive time indicates future discharge.