I have a vector polygon which most of the features fall outside of my tiles (3 out 4) since I’m just only selecting one from the available input tiles in the custom jobs.
How will these affect may accuracy and classification results for an L4B product?
The features are clipped against the tiles, so the extra parts will be ignored.
Also, note that the processor splits the polygons into a training and validation set. It goes for a 75-25% split, but duplicates them if 25% is 0. In that case, the validation will use training data, so the accuracy will be over-estimated. So it might be better to have more polygons.
So I really need to have much more training vector polygons than I usually have.
We usually used a number in the hundreds or thousands for each crop class.