Extracting domain-specific knowledge from image databases is challenging and requires a deep understanding of the domain. For example, in the geospatial domain knowledge discovery computationally expensive due to the huge amount of generated imagery. Existing content-based image retrieval systems utilize models that are trained and optimized to experts' knowledge using expert-in-the-loop approaches. However, such approaches may lead to suboptimal models especially when the number of training images is small. In this paper, we propose incorporating existing domain knowledge resources into knowledge discovery. More specifically, we have developed methods for using ontological relationships between geospatial semantics to oversample under-represented semantics. Our experimental results show that our technique improves the knowledge discovery process, as evidenced by increased precision of semantic queries.