Abstract
Predicting water temperature and streamflow across large-scale river networks is critical for environmental management but often suffers from spatial performance biases due to data heterogeneity and scarcity. We present Fair-Graph, a novel graph-based recurrent neural network that integrates node-specific graph masking with a constraint-aware priority adjustment strategy. By employing trainable aggregation and adaptation masks, Fair-Graph effectively models complex spatial interactions and leverages cross-node knowledge. Furthermore, a global priority adjustment mechanism dynamically reweights learning rates to minimize performance disparities while respecting inherent data quality constraints. Evaluations on the Delaware River Basin and Houston River Network demonstrate that Fair-Graph significantly reduces spatial bias and improves overall prediction accuracy compared to state-of-the-art baselines.
Introduction
Accurate predictions of water quality and quantity are essential for environmental policy and disaster prevention, such as flood and drought mitigation. While Graph Neural Networks (GNNs) are adept at modeling the flow of information through river segments, they often exhibit significant spatial bias. In large basins, well-monitored segments with high-quality sensors often dominate model optimization, leading to poor performance in remote or data-poor regions. This disparity creates a fairness issue, potentially resulting in inequitable resource distribution or environmental protection.
Problem Statement
The authors identify two primary hurdles in river network modeling:
- Data Heterogeneity: River segments possess unique physical characteristics (e.g., soil composition, geometry) that are frequently missing from standard input features.
- Data Scarcity: Many segments lack sufficient historical observations to train robust local models, leading to a reliance on global patterns that may not apply locally.
Methodology
Fair-Graph utilizes a base architecture of LSTMs for temporal patterns and GNNs for spatial interactions, enhanced by two innovative components designed to promote geographic equity.
Dual Graph Masking
To address heterogeneity and scarcity, the model introduces two trainable masks:
- Aggregation Mask ($M^a$): This mask learns the specific relevance of upstream and downstream neighbors for a given node, rather than assuming uniform influence. This allows the model to discover latent physical relationships, such as groundwater impacts.
- Adaptation Mask ($M^s$): Designed for data-scarce nodes, this mask enables segments with few observations to "borrow" relevant training samples from similar nodes across the network to fine-tune local parameters.
Constraint-Aware Priority Adjustment
A "global referee" monitors performance across the network during training. If a segment performs poorly relative to the average, its learning rate is increased to force the model's attention. Crucially, this mechanism recognizes "performance ceilings" caused by noisy data; it prevents the model from over-optimizing for "hopeless" segments at the expense of the broader network's accuracy.
Datasets
The model was evaluated on two distinct, large-scale systems:
- Delaware River Basin (DRB): 456 stream segments with 40 years of daily water temperature data.
- Houston River Network: 412 segments using simulated streamflow data from the U.S. National Water Model (NWM), characterized by frequent flash flooding.
Results and Discussion
Fair-Graph consistently outperformed standard GNNs and existing fairness-preserving algorithms like REG and ADL. In the Delaware River Basin, Fair-Graph achieved a Root Mean Squared Error (RMSE) of 1.786, a notable improvement over the base model's 1.833. In the Houston network, it maintained a high Nash–Sutcliffe efficiency (NSE) of 0.758 while significantly reducing fairness error.
Robustness tests involving "perturbed" data (dropping 90% of observations) showed that Fair-Graph maintained its performance edge, validating the effectiveness of the adaptation mask in leveraging cross-graph knowledge.
Conclusion
Fair-Graph reframes spatial fairness as a matter of geographic equity, ensuring that model reliability does not depend on a location's monitoring convenience. Future work will focus on predicting for segments with zero historical observations and developing more efficient, model-agnostic methods for estimating performance upper bounds.
Summary updated