Label output files are generated by overwriting the cloud_type channel of a reference file from the training set.
The meta data (file descriptors, date, author, etc) are not changed and are therefore not accurate.