[Paper Review] HydroNets: Leveraging River Structure for Hydrologic Modeling

Zach Moshe, Asher Metzger+4 more|arXiv (Cornell University)|Jul 1, 2020

Hydrology and Watershed Management Studies8 references39 citations

TL;DR

HydroNets uses river network structure to jointly forecast multiple basins with a shared, basin-aware neural architecture, improving hydrologic predictions when data are scarce by leveraging upstream basin information.

ABSTRACT

Accurate and scalable hydrologic models are essential building blocks of several important applications, from water resource management to timely flood warnings. However, as the climate changes, precipitation and rainfall-runoff pattern variations become more extreme, and accurate training data that can account for the resulting distributional shifts become more scarce. In this work we present a novel family of hydrologic models, called HydroNets, which leverages river network structure. HydroNets are deep neural network models designed to exploit both basin specific rainfall-runoff signals, and upstream network dynamics, which can lead to improved predictions at longer horizons. The injection of the river structure prior knowledge reduces sample complexity and allows for scalable and more accurate hydrologic modeling even with only a few years of data. We present an empirical study over two large basins in India that convincingly support the proposed model and its advantages.

Motivation & Objective

• Incorporate prior river-network structure into hydrologic forecasting models.
• Reduce sample complexity so forecasts improve with limited training data.
• Enable both shared (physical) and basin-specific modeling within a modular architecture.

Proposed method

• Represent the hydrologic region as a directed river graph with basins as nodes and flow as edges.
• Use a three-part node design: a basin-specific Combiner, a Shared Hydrologic Model, and a Basin-Specific Prediction Model.
• Share weights in the upstream-embedding process to capture common hydrologic dynamics while allowing basin-specific outputs.
• Propagate temporal embeddings from sources to downstream basins following the region graph, forming a tree-like computation graph.
• Train the model with a weighted MSE loss across all basins to balance regional and local forecasting accuracy.

Experimental results

Research questions

• RQ1 Can river-network structure be effectively encoded into a neural forecasting model for multiple basins?
• RQ2 Does enforcing shared upstream dynamics reduce data requirements and improve long-horizon predictions?
• RQ3 How does HydroNets compare to a flat baseline across basins, especially with limited training data?
• RQ4 What is the impact of region depth (tree-like structure) on predictive performance?
• RQ5 Can HydroNets maintain performance across diverse basins with varying upstream influence?

Key findings

• HydroNets outperforms a flat linear baseline across representative basins in both regions.
• Deeper regional trees yield continued gains for HydroNets but not for the baseline.
• HydroNets show stronger relative gains as training data decreases, indicating reduced sample complexity.
• Across basins, HydroNets delivers better persist R-squared scores than the flat model, especially for upstream regions.
• The model consistently benefits from leveraging upstream sub-basin information through the shared upstream-embedding mechanism.