Indeed, what we apply is a supervised classification, meaning that a part of the parcels are used to train the classification model, and then this classification model is applied to all the parcels (that are included in the classification). On top of that, the parcels that are classified but not used for the calibration are used for the classification validation. This comes from the assumption that in the CAP context, the big majority of the farmers’ declarations are correct. And the type of classification which is used, a random forest classification, is relatively low-sensitive to outliers; it means that a few not correct declarations will not impact much the classification model (and so the classification accuracy).
Concerning the process of selecting the parcels for the classification, and among them for the calibration, this process is explained in the “ATBD for L4A crop type mapping 1.2” doc, accessible here: http://esa-sen4cap.org/content/technical-documents. Depending on the number of parcels in each crop type, you will see that 3 different approaches are used. The aim of it is to improve the accuracy results in the crop types with relatively less parcels, which would generate otherwise less accurate results. Of course, the parameters defining these 3 approaches can be easily adapted.
Do not hesitate if you have further detailed questions. About a possible tutorial, what kind of tutorial have you in mind (support, etc.)?