Semantic-sensitive satellite image retrieval

Content-based image-retrieval techniques based on query scenes are a powerful means for exploration and mining of large remote sensing image databases. However, the gap between low-level unsupervised extracted features in content-based retrieval and the high-level semantic concepts of user queries limits the performance. Therefore, this paper proposes a specialized approach using a context-sensitive Bayesian network for semantic inference of segmented scenes. The regions' remote sensing related semantic concepts are inferred in a multistage process based on their spectral and textural characteristics as well as the semantics of adjacent regions. During the actual retrieval, the semantics are employed for the extraction of candidate scenes which are evaluated and ranked in a consecutive step. The approach was implemented and compared with a different strategy that utilizes the extracted features from the imagery directly to infer the semantics. In summary, the developed system achieved higher precision and recall rates using the same training data.