Somoml Eu Eval

---
name: somoml-eu-eval
description: Evaluates the accuracy and spatial-temporal fidelity of a machine learning-derived daily soil moisture product for Europe. It probes the model's ability to generalize across diverse climates, capture drought dynamics, and outperform existing reanalysis and satellite-based soil moisture datasets. Use when the user wants to benchmark on SoMo.ml-EU, or asks about evaluating this task. Reports uRMSD.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2205.10753
  bibtex_key: o2022somoml
  confidence: high
---

# somoml-eu-eval

> High-resolution European daily soil moisture derived with machine learning (2003-2020) — Sungmin O et al. (2022) (arXiv:2205.10753, 2022)

## What this evaluates

Evaluates the accuracy and spatial-temporal fidelity of a machine learning-derived daily soil moisture product for Europe. It probes the model's ability to generalize across diverse climates, capture drought dynamics, and outperform existing reanalysis and satellite-based soil moisture datasets.

## Datasets

- **SoMo.ml-EU** — total ?; splits: train (-1), test (-1)

## Metrics

- `uRMSD` **(primary)** — range: other
  - Unbiased root mean square difference: RMSE computed after removing the mean bias between predicted and observed soil moisture values. Lower values indicate better agreement.
- `Pearson correlation (r)` — range: [0, 1]
  - Standard linear correlation coefficient between simulated and in-situ soil moisture time series.

## Input / output format

**Input**: Daily meteorological forcing (temperature, precipitation, net radiation, skin temperature) and static features (topography, vegetation, soil properties) over a 365-day lookback window, plus upper-layer soil moisture for deeper layers.

**Output**: Volumetric soil moisture (m³/m³) for three depth layers (0–10 cm, 10–30 cm, 30–50 cm) at 0.1° spatial resolution and daily temporal resolution for Europe (2003–2020).

## Scoring recipe

```python
pred = regrid_to_01deg(model_output)
obs = regrid_to_01deg(in_situ_data)
bias = np.mean(pred) - np.mean(obs)
pred_unbiased = pred - bias
uRMSD = np.sqrt(np.mean((pred_unbiased - obs)**2))
r = np.corrcoef(pred, obs)[0, 1]
anomalies = (pred - seasonal_mean) / seasonal_std
```

## Common pitfalls

- Training data is heavily biased toward arid/warm climates (US sites), limiting extrapolation accuracy in humid/cold European regions.
- Depth definitions vary across comparison datasets (e.g., SoMo.ml uses 0-10 cm vs CLM-DA's 3 cm), causing systematic dry biases that are not algorithmic failures.
- Freeze-thaw processes in Northern Europe introduce high uncertainty in both observational and model data, complicating temporal validation.

## Evidence (verbatim from paper)

> In terms of unbiased root mean square difference (uRMSD), SoMo.ml-EU shows smaller deviations from the in-situ data, while the median range of uRMSD across the datasets stays narrow between 0.04 to 0.05. ... The model shows a better performance for Layer 2 and Layer 3 (10-30 cm and 30 - 50 cm depths, respectively) ... (r = 0.8 for correlation between pixels in Layer 1) according to the five-fold cross-validation in which the training data from Europe are included (Fig. S2).

## Citation

```bibtex
@misc{o2022somoml,
  title={High-resolution European daily soil moisture derived with machine learning (2003-2020)},
  author={Sungmin O et al. (2022)},
  year={2022},
  note={arXiv:2205.10753}
}
```

- arXiv: 2205.10753