Rough logic for building a landmine classifier

Landmines are significant barriers to financial, economic and social development in various parts of this world. Metal detectors currently used by teams engaged in the decontamination of mines cannot differentiate a mine from metallic debris where the soil contains large quantities of metal scraps and cartridge cases. So what is required is a sensor that will reliably confirm that the ground being tested does not contain an explosive device with almost perfect reliability. The various sensors provide different attributes about the nature of the soil. Even human experts are unable to give belief and plausibility to their rules. Thus the conventional Dempster-Shafer theory cannot be applied to build an expert system. Thus rough sets are applied to classify the landmine data because here, any prior knowledge of rules is not needed; these rules are automatically discovered from the database. For the application of rough set theory, first, approximation space is to be identified. This can be done by a human expert, otherwise, principal component analysis can be used. In fact the problem of identifying application space is similar to that of identifying redundant attributes (which carry no useful information for the purpose of classification) and throwing them away. It is observed that aspect ratio, blob size, and grayscale are important features for landmine classification. Now, all data tuples which agree on all task relevant attributes, form one elementary set. Thus the whole of the database is divided into mutually exclusive elementary sets. As no crisp rules exist for classification, this essentially implies the decision sets (the sets formed by the presence or absence of landmines) are not definable (cannot be expressed as union or intersection of elementary sets). Thus lower and upper approximations of the decision sets are taken and the boundary set is found.