Geographical Patch distribution#

The following map shows the geographical label distribution of BigEarthNet.


The polygons have been merged together to minimize page-load time and storage requirements

Make this Notebook Trusted to load map: File -> Trust Notebook
Make this Notebook Trusted to load map: File -> Trust Notebook

Split Distribution#

The original training/validation/test split were generated in the following way:

  1. Split the original 1200km x 1200km tile into 1.2km x 1.2km patches (as shown in the Tile to patches figure)

  2. Group the patches into 2 x 2 grids

  3. Select two patches for the train, one for the validation and one for the test split (as shown in the split figure)

1.2 km1.2 km1200 km1200 km

Each 1200km x 1200km tile was split into 1.2km x 1.2km patches #


The patches of each tile were grouped into 2 x 2 patches. Of those, 2 patches were added to the train split, 1 to the validation and the last one to the test split. #

As a result of this splitting strategy, the splits have a high geographical correlation. One could argue that this also relates to a high domain correlation, i.e. that the performance on the test split might not provide deep insights into generalizability of the tested model.