Abstract
One way of enforcing both speed and accuracy in feature subset selection is by attaching to each operation (e.g., computation of features, feature selection, labeling method), and each type of error, a cost. This paper exploits the concept of cost-based feature subset selection, and compares cost-based with classical implementations. The novelty of this paper is the use of explicit, user-defined, cost and accuracy limits, which affect the choice both of the feature subset selection mechanism and of the features themselves. Examples of GIS-based digitizing of multispectral high-resolution satellite imagery illustrate the usefulness of the proposed method in real-time, interactive image analysis.