Discovery of High Confidence Association Rules in Remotely Sensed Images

Spatial data mining of Remotely Sensed Images (RSI) has become an important field of research as extremely large amounts of data are being collected from remote sources such as the Landsat satellite Thematic Mapper (TM) and other remote imaging systems. Association Rule Mining (ARM) has become an important method for mining large amounts of data in many areas beyond its originally proposed market-basket domain. The popularity of ARM comes from the well-known a-priori algorithm that exploits a user-specified minimum support (called minsup). Rules of interest are defined as only those lying within the set of rules that exceed this support level. To work efficiently, rules of interest need to be restricted to those that occur frequently. While this restriction enables a-priori based data mining to perform efficiently it rules out the discover of an entire class of rules of interest which are pruned for lack of support. Such a class of rules is of interest in applications such as those found in the agricultural domain where a rule of interest might address early insect infestation; a rule with extremely low support but of extremely high interest to a producer. In this paper, we develop a conceptual decision cube called a P-cube that is derived from a P-tree storage of remotely sensed images. This conceptual P-cube is then used to help develop an efficient algorithm for discovering high confidence rules using a precisionhierarchy approach. This approach discovers high confidence rules without concern for 1 Patents are pending on the bSQ and P-tree technology from which the P-cube is derived. 2 This work is partially supported by NSF Grant OSR-9553368, DARPA Grant DAAH04-96-1-0329 and GSA Grant ACT#: K96130308. Discovery of High Confidence Association Rules in Remotely Sensed Images Without Regard for Support Page 2 support. This algorithm does not suffer from the computational explosion inherent in the apriori approach when considering low support threshold. This makes it feasible to discover a wide class of rules normally lost to pruning techniques in the name of efficiency.