qhjqhj00 super-resolution-weather-eval
Super Resolution On Global Weather Forecasts — Zhang et al. (2024) (arXiv:2409.11502, 2024)
What this evaluates
Evaluates a model's ability to perform spatial super-resolution on global weather forecast data, specifically upscaling temperature and cloud coverage maps from 1° to 0.5° resolution. It measures pixel-wise reconstruction accuracy against high-resolution ground truth.
Datasets
- GraphCast-ERA5 Paired Dataset — total 400; splits: train (-1), test (-1)
Metrics
MSE(primary) — range: other- Mean Squared Error: the average of the squared differences between predicted and ground truth pixel values. Formula: (1/N) * Σ(y_true - y_pred)^2.
Input / output format
Input: Single-channel low-resolution (1°) weather map images representing temperature or cloud coverage.
Output: Single-channel high-resolution (0.5°) weather map images.
Scoring recipe
def compute_mse(predictions, ground_truth):
# predictions and ground_truth are numpy arrays of identical shape
diff = predictions - ground_truth
mse = np.mean(diff ** 2)
return mse
Common pitfalls
- MSE heavily penalizes large pixel-wise errors but does not capture perceptual quality or meteorological utility (e.g., storm structure preservation).
- The evaluation uses a small dataset of 400 paired images without explicit train/val/test splits, making it difficult to assess generalization or overfitting.
- Qualitative visual inspection is used alongside MSE, but no standardized meteorological accuracy metrics (e.g., anomaly correlation, bias, or RMSE per variable) are reported.
Evidence (verbatim from paper)
To train our models, we used Mean Squared Error (MSE) loss to minimize the difference between outputs and ground truth and as an evaluation metric.
Citation
@misc{zhang2024superresolution,
title={Super Resolution On Global Weather Forecasts},
author={Zhang et al. (2024)},
year={2024},
note={arXiv:2409.11502}
}
- arXiv: 2409.11502