Hola Amigos
While looking through the data I found some potentially mis-labeled data in the training set. Check out the IDs 451, 464, 473 and 477 from Orenburg_training_samples. The file labels them as 0 but from my view they are cropland. I crosschecked using ESAs WorldCover 10m v200 maps (for 2021).
I just checked a handfull of samples but this makes me wonder if there are also mislabeled datapoints in the test set.
Here an example:
ID: 451
Cropland: 0
POINT: (54.83862 52.03939)
Google Maps:
WorldCover cropland mask:
Doesn't really look like a cropland to me😅. How are you able to identify a cropland?
In ChatGPT we trust:
What’s Visible in the Image
I haven't done any form eda yet but are there several images that look like the mislabelled image?
Which IDs then if that's the case?
In ChatGPT we trust🤣