ABSTRACT
Widespread use of spatial databases [42], an important subclass of multimedia databases, is leading to an increased interest in mining interesting and useful but implicit spatial patterns [23, 29, 18, 40]. Traditional data mining algorithms [1] often make assumptions (e.g., independent, identical distributions) which violate Tobler’s rst law of geography: everything is related to everything else but nearby things are more related than distant things [45]. In other words, the values of attributes of nearby spatial objects tend to systematically affect each other. In spatial statistics, an area within statistics devoted to the analysis of spatial data, this is called spatial autocorrelation [12]. Knowledge discovery techniques that ignore spatial autocorrelation typically perform poorly in the presence of spatial data. The simplest way to model spatial dependence is through spatial covariance. Often the spatial dependencies arise due to the inherent characteristics of the phenomena under study, but in particular they arise due to the fact that imaging sensors have better resolution than object size. For example, remote sensing satellites have resolutions ranging from 30 m (e.g., Enhanced Thematic Mapper of Landsat 7 satellite of NASA) to 1 m (e.g., IKONOS satellite from SpaceImaging), while the objects under study (e.g., urban, forest, water) are much bigger than 30 m. As a result, the per-pixel-based classi- ers, which do not take spatial context into account, often produce classied images with salt and pepper noise. These classiers also suffer in terms of classication accuracy.
