Querying remote sensing and GIS repositories with spatial association rules

The authors describe the fusion of multispectral image and GIS data mining functions in a module that implements spatial association rules. Spatial association rules may represent: 1) topological relationships between spatial objects; 2) spatial orientation or ordering; and 3) distance information. Their approach relies on a fast and unique segmentation algorithm that combines information with edge-based information in a variational framework. The indexing strategy distinguishes between three levels of features: 1) pixel level, 2) region level, and 3) scene level features. They use pixel level information for the extraction of higher level features, and in the process of query refinement. Region level features describe groups of contiguous pixels. Following segmentation, they describe each region with a boundary and a number of attributes, like spectral endmember types and percentages, textural classes, shape, size, fractal scale, etc. They quantify shape, orientation and other geometric properties of the segments by computing 32 moments. Scene level features describe global properties of whole scenes, and the spatial relationships of the largest regions in them. The second level of features can optionally support definition of semantic labels from multiple attributes, which include GIS attributes, if present. At a higher level, they can attempt to construct semantic labels from a tree of topological relationships. They store images and features in a database to enable fast access, data integrity and easy interfacing with other applications.