i am currently running the Random Forest Classifier on a Sentinel 2 Mosaic, however, the RF classifier seems to have problems with the overlapping part of the two original S2 images… one can clearly see the S2 overlapping part in the classification result…
the RF classifier is quite sensitive to small variations because it is based on pixel thresholds.
Smaller inconsistencies between both tiles are possibly causing this effect.
Some ideas
use L2A products instead of L1C (if you are not already doing this)
create training polygons in both sides of the image to include these variations
test the KD Tree Classifier which could be a bit more robust towards this.
Increase the number of trees in the RF classifier (not sure if it solves the problem but also makes the classifier more effective)
I see your point. Sentinel-2 has many additional bands which contain quality flags and geometric information. If they are used to train the classifier, you will inevitable introduce weird patterns.
three bands do not make much sense to me, because the strength of the RF classifier is to shuffle the input bands repeatedly to make use of the most relevant ones. Why are you not using all 12 bands?
I used all bands for mosacing, and then der RF classifier worked just fine, so i think the problem in the first place was, as you said, only using 3 bands.