When I look at the images from the training set with my eyes, I cannot determine the target variable. Images have low resolution and sometimes blurry. So this is a case where machine learning does better. :) I wonder if the experts had access to the original images during the markup.
It's hard! Normally we're worried about people hand-labelling images but for this one, I don't know if I could :) From my understanding: The original images (high resolution) were labelled by agronomists who had access to additional data (date, a series of images etc). The 'lower quality' labels were selected by the farmers who took the pictures (I think) and so they had the crop in front of them, but might not all have the same definitions for the growth stages.
Thanks for your reply! It seems I understand the problem better