Effective Damage Data Generation by Fusing Imagery with Human Knowledge Using Vision-Language Models