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
The original training/validation/test split were generated in the following way:
Split the original 1200km x 1200km tile into 1.2km x 1.2km patches (as shown in the Tile to patches figure)
Group the patches into 2 x 2 grids
Select two patches for the train, one for the validation and one for the test split (as shown in the split figure)
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.