Semantic sensitive satellite image retrieval.
Date of Issue2008
School of Computer Engineering
In recent years, numerous remote sensing platforms for Earth observation have been developed and together acquire several terabytes of image data per day. However, due to this data volume, straightforward access to the data has become increasingly complex. Most common database systems retrieve the satellite images based on their world-oriented information. Queries not directly related to this type of information, for example, the search for a scene that possesses a similar ground cover characteristic as a query scene, cannot be processed. Therefore, a variety of content-based image retrieval techniques were developed to retrieve images based on abstract automatically extracted features. Applied to remote sensing databases, varying success was reported. The main difficulty is that users think in terms of high-level semantics, which cannot be directly mapped onto extracted low-level features / signals. The result is a gap between the signal classes on the one side, that is, pixels, regions or images that are clustered in the feature space, and semantic concepts on the other side. Although various relevance feedback approaches have been proposed to bridge this gap, most of them can only work well under the assumption that regions belonging to the same or similar signal classes are very likely to share the same semantic concepts as well. Often this assumption is invalid, since there are rarely one-to-one relations among signal classes and semantic concepts. On the one hand, a signal class may be related to multiple semantic concepts under different circumstances. On the other hand, multiple signal classes may be related to the same semantic concept. An example is the change in observed characteristics of the same scene due to changes in sun inclination and season.
DRNTU::Engineering::Computer science and engineering::Information systems::Information storage and retrieval