# Evaluation

## Metric
Submissions are evaluated using **unbiased Root Mean Square Difference (ubRMSD)** as the primary metric:

$$
\text{ubRMSD} = \sqrt{\frac{1}{n} \sum_{i=1}^{n} ((\hat{y}_i - \mu_{\hat{y}}) - (y_i - \mu_y))^2}
$$

where:
- $\hat{y}_i$: predicted soil moisture
- $y_i$: ground truth soil moisture
- $\mu_{\hat{y}}$ and $\mu_y$: means of predictions and truth (removed to eliminate bias)

Lower ubRMSD is better. Typical targets: < 0.04 cm³/cm³ for state-of-the-art.

## Secondary Metrics (for reference)
- RMSE: $\sqrt{\frac{1}{n} \sum_{i=1}^{n} (\hat{y}_i - y_i)^2}$
- Bias: $\frac{1}{n} \sum_{i=1}^{n} (\hat{y}_i - y_i)$
- Pearson correlation coefficient (R)

## Leaderboard
- **Public Leaderboard**: Scored on ~30% of the private test set (visible during competition).
- **Private Leaderboard**: Scored on the remaining ~70% (final rankings, revealed at end to prevent overfitting).
- Final ranking based solely on private ubRMSD.
- Teams may select up to 2 submissions for final scoring.

## Additional Notes
- The test set covers similar temporal/spatial distribution as training but includes unseen regions/seasons.
- NaN predictions are not allowed; invalid submissions score infinitely.