the classifier wants to extract all pixel values under the training geometries. If one of them is defined as NaN (not a number) defined in the band properties. These are displayed transparent in SNAP.
Please also see here: Polarimetric Classification NaN problem