Incorporating Domain High-Level Concepts into Heuristic Searches: A Case Study on Identifying Plant Species in Remote Sensing Images

In order to take domain expert knowledge as a supplement of the machine "intelligence", a High-Level-Concept-based framework (HLC) for heuristic search is proposed. HLC has domain experts actively generalize their own psychological feelings about image features into high-level concepts, then write down them in descriptors, and make the domain high-level concepts enter a heuristic search with a multi-descriptor combination. These so-called "descriptors" are a kind of non-threshold models so that they have reusability and generalization. To demonstrate the idea, 18 new descriptor have been designed. The examinations of discrimination accuracy indicate that in a multi-descriptor space, the error rates of classification is 67.91% lower than that in a spectral-brightness space.