Streamflow Hydrology Eval

streamflow-hydrology-eval

Hierarchical Conditional Multi-Task Learning for Streamflow Modeling — Xu et al. (2024) (arXiv:2410.14137, 2024)

What this evaluates

Evaluates hydrological forecasting models on predicting streamflow across 319 US basins. It tests the model's ability to capture long-term temporal dependencies and intermediate hydrological states (soil water, snowpack) under varying data segmentation, training sizes, and noise conditions.

Datasets

• 319 US Basins — total 319; splits: test (319)

Metrics

• RMSE (primary) — range: other
  • Root Mean Squared Error: sqrt(mean((y_true - y_pred)^2)). Measures the average magnitude of prediction errors in the same units as the target.
• NSE — range: other
  • Nash-Sutcliffe Efficiency: 1 - sum((y_true - y_pred)^2) / sum((y_true - mean(y_true))^2). Compares model performance to a naive baseline predicting the mean. Higher is better.

Input / output format

Input: Daily time series segments (14 to 365 days) of streamflow, soil water, and snowpack for each basin, processed via sliding windows.

Output: Predicted streamflow values for each time step in the segment.

Scoring recipe

def compute_metrics(y_true, y_pred):
    rmse = np.sqrt(np.mean((y_true - y_pred) ** 2))
    y_mean = np.mean(y_true)
    nse = 1 - np.sum((y_true - y_pred) ** 2) / np.sum((y_true - y_mean) ** 2)
    return rmse, nse

Common pitfalls

• NSE can be negative if the model performs worse than predicting the mean baseline, which is common in hydrological forecasting.
• RMSE is scale-dependent, making direct cross-basin comparisons difficult without normalization or careful interpretation of flow magnitudes.
• Sliding window segmentation must strictly avoid data leakage between training and test segments to ensure valid performance estimates.

Evidence (verbatim from paper)

Table [2] shows that the proposed HCMTL model delivers the best overall performance, with lower RMSEs and higher NSEs. HCMTL consistently achieves the lowest RMSE across all four regions, highlighting its superior accuracy. We conduct a detailed study of the United States in the following sections. Figure [4] shows HCMTL performing best in 123 out of 319 U.S. basins.

Citation

@misc{xu2024hcmtl,
  title={Hierarchical Conditional Multi-Task Learning for Streamflow Modeling},
  author={Xu et al. (2024)},
  year={2024},
  note={arXiv:2410.14137}
}

• arXiv: 2410.14137