Stormnet Bias Eval

---
name: stormnet-bias-eval
description: Evaluates a spatio-temporal graph neural network's ability to predict and correct systematic biases in storm surge water level forecasts. It probes the model's capacity to leverage spatial dependencies among coastal gauge stations and temporal patterns to improve long-horizon (up to 72h) hydrodynamic predictions. Use when the user wants to benchmark on Gulf Coast Gauge Network (NOAA/TCOON), or asks about evaluating this task. Reports RMSE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.20688
  bibtex_key: [REDACTED]
  confidence: high
---

# stormnet-bias-eval

> Storm Surge Modeling, Bias Correction, Graph Neural Networks, Graph Convolution Networks — Noujoud Nader et al. (2026) (arXiv:2604.20688, 2026)

## What this evaluates

Evaluates a spatio-temporal graph neural network's ability to predict and correct systematic biases in storm surge water level forecasts. It probes the model's capacity to leverage spatial dependencies among coastal gauge stations and temporal patterns to improve long-horizon (up to 72h) hydrodynamic predictions.

## Datasets

- **Gulf Coast Gauge Network (NOAA/TCOON)** — total ?; splits: train (-1), test (-1)

## Metrics

- `RMSE` **(primary)** — range: meters
  - Root Mean Square Error between observed and predicted (or bias-corrected) water levels. Calculated as the square root of the mean of squared differences between true and predicted values.

## Input / output format

**Input**: Time-series water level observations and raw forecasts at 16 coastal gauge stations, along with a static graph adjacency matrix constructed from Pearson correlation and geographic distance thresholds.

**Output**: Predicted bias offset (in meters) for each station and prediction window, which is added to the raw forecast to yield the corrected water level.

## Scoring recipe

```python
def compute_rmse(observed, predicted):
    errors = observed - predicted
    return (errors ** 2).mean() ** 0.5

# Evaluated per station and per prediction window (12h, 48h, 72h)
rmse_values = [compute_rmse(obs, pred) for obs, pred in zip(observed_levels, corrected_levels)]
```

## Common pitfalls

- The graph structure is constructed exclusively from training storms to prevent data leakage; the test storm (Idalia) is only used for inference.
- Evaluation is limited to a single test event (Hurricane Idalia), so generalization to other storm types or regions is not assessed.
- RMSE is reported as a distribution across stations and prediction windows, not just a single aggregate number.

## Evidence (verbatim from paper)

> The overall performance of StormNet to predict water level biases for different prediction window values in the range of 6-72 h, is presented in Figure[6]. Individual points represent the RMSE between real and predicted offsets for the bias predictions in each station, overlaid on a boxplot representing their distribution, for each value of the prediction window. Applying StormNet leads to a noticeable improvement in water level forecasts, as can be seen from Figure[7] it can be seen that applying StormNet with a 48 h-ahead prediction window leads to a reduction in RMSE in water level forecasts by 60%, 65% and 73% for the Cedar Key, Apalachicola and Galveston Bay Entrance North Jetty stations, respectively.

## Citation

```bibtex
@misc{nader2026stormnet,
  title={Storm Surge Modeling, Bias Correction, Graph Neural Networks, Graph Convolution Networks},
  author={Noujoud Nader et al. (2026)},
  year={2026},
  note={arXiv:2604.20688}
}
```

- arXiv: 2604.20688