EarthNet2021x Scoring

Scoring your predictions

You can score your predictions using the EarthNet toolkit (pip install earthnet)

Save your predictions for one test set in one folder in the following way: {pred_dir/region/cubename.nc} Name your NDVI prediction variable as "ndvi_pred".

Then use the data_dir/dataset/split as the targets.

Then compute the normalized NSE over the full dataset:

import earthnet as entk
scores = entk.score_over_dataset(Path/to/targets, Path/to/predictions)
print(scores["veg_score"])

Alternatively you can score a single minicube:

import earthnet as entk
df = entk.normalized_NSE(Path/to/target_minicube, Path/to/prediction_minicube)
print(df.describe())

Vegetation Score

EarthNet2021x uses a vegetation score to benchmark different models.

It is the average Nash Sutcliffe Model Efficiency (sometimes equivalent to the Coefficient of Determination R^2) on cloud-free observations of Vegetation Pixels.

More specifically:

  1. For each pixel compute the Nash-Sutcliffe Model Efficiency (NSE) at cloud-free observations
  2. Rescale this with 1 / (2-nse) to the range 0-1 for robust averaging
  3. Averaging over all natural vegetation pixels (Landcover class Trees, Scrub or Grassland)
  4. Scaling back with 2 - 1/mean_nnse to the range -Inf,1

The Vegetation Score is 1 if the prediction is perfect. It is 0 if on average predictions are as good as the mean over the target period. It is negative if on average predictions are worse than the mean of the target period.

In Pseudo-Code it is computed as follows:

nse = NSE(targ_ndvi, pred_ndvi).where(targ_ndvi has no clouds)
nnse = 1 / (2-nse)
veg_score = 2 - 1/mean(nnse.where(landcover == Trees, Scrub or Grassland))

Models can use a context length for spin-up and are benchmarked over a target length, which is specified for the different test sets (tracks) as follows (same as EarthNet2021):

  • IID: 50 days context, 100 days target
  • OOD: 50 days context, 100 days target
  • Extreme: 100 days context, 200 days target
  • Seasonal: 350 days context, 700 days target

Here, five days equal one Sen2 observation.