you see in the screenshot that she used training vectors (red squares).
You can only use rasters as training inputs if they represent the final classes (e.g. land cover maps).
Please have a look here: Supervised and unsupervised classification, Sentinel 2