Hi @hongwei2017,
I believe your question is a duplicate of this thread. In brief, I would not recommend using read extraction when creating an ITS classifier. ITS classification performs much better with full-length ITS sequences. That thread gives other details on choosing optimal parameters for ITS classification (see the linked preprint).
If you are still getting kingdom and phylum-level classifications with full-length sequences, please let us know as that would indeed be unusual.
Hope that helps!