Soil Temp Ndvi Mlp Eval

soil-temp-ndvi-mlp-eval

Inferring the relationship between soil temperature and the normalized difference vegetation index with machine learning — Mortier et al. (2023) (arXiv:2312.12258, 2023)

What this evaluates

Evaluates the ability of multilayer perceptrons to predict vegetation phenology parameters (start of season, peak of season, peak NDVI value) from soil temperature and meteorological variables in subarctic grasslands. It probes how well non-linear models capture complex, non-linear interactions between climate drivers and vegetation dynamics compared to simple linear baselines.

Datasets

• Subarctic grassland phenology dataset (Iceland, 2014-2019) — total ?; splits: train (-1), test (-1)

Metrics

• MSE (primary) — range: other
  • Mean Squared Error: the average of the squared differences between predicted and actual values.
• MAE — range: other
  • Mean Absolute Error: the average of the absolute differences between predicted and actual values.
• R^2 — range: [0, 1]
  • Coefficient of determination: 1 minus the ratio of residual sum of squares to total sum of squares, representing the proportion of variance explained by the model.

Input / output format

Input: Features include average annual soil temperature and meteorological variables (air temperature, precipitation, radiation) aggregated over 26 weekly averages per year.

Output: Continuous regression targets: SOS (start of season in weeks), POS (peak of season in weeks), and PEAK (maximum annual NDVI value).

Scoring recipe

def compute_metrics(y_true, y_pred):
    mse = np.mean((y_true - y_pred) ** 2)
    mae = np.mean(np.abs(y_true - y_pred))
    ss_res = np.sum((y_true - y_pred) ** 2)
    ss_tot = np.sum((y_true - np.mean(y_true)) ** 2)
    r2 = 1 - (ss_res / ss_tot)
    return {'MSE': mse, 'MAE': mae, 'R^2': r2}

Common pitfalls

• Low R^2 values (0.19–0.32) indicate substantial unexplained variance despite statistically significant p-values, which can be misinterpreted as model failure rather than inherent environmental noise.
• Naive baselines (predicting the training mean) perform nearly as well as the MLP on PEAK, highlighting limited predictive gain for that specific target and requiring careful baseline comparison.

Evidence (verbatim from paper)

The performance of the MLPs (multilayer perceptron) can be found in Table 4. From Tables 3 and 4, it becomes evident that the inclusion of the meteorological variables and the utilization of MLPs (multilayer perceptron) enabled us to explain a significantly larger part of the variance compared to the linear models. The test set consists of 20% of the total data, and is split evenly across the years of data taking. The naive MSE (mean squared error) (MAE (mean average error)) is the MSE (mean squared error) (MAE (mean average error)) when the mean of all training samples is used as the prediction.

Citation

@misc{mortier2023soiltemp,
  title={Inferring the relationship between soil temperature and the normalized difference vegetation index with machine learning},
  author={Mortier et al. (2023)},
  year={2023},
  note={arXiv:2312.12258}
}

• arXiv: 2312.12258