Spatially Cost-sensitive Active Learning

In active learning, one attempts to maximize classifier performance for a given number of labeled training points by allowing the active learning algorithm to choose which points should be labeled. Typically, when the active learner requests labels for the selected points, it assumes that all points require the same amount of effort to label and that the cost of labeling a point is independent of other selected points. In spatially distributed data such as hyperspectral imagery for land-cover classification, the act of labeling a point (i.e., determining the land-type) may involve physically traveling to a location and determining ground truth. In this case, both assumptions about label acquisition costs made by traditional active learning are broken, since costs will depend on physical locations and accessibility of all the visited points. This paper formulates and analyzes the novel problem of performing active learning on spatial data where label acquisition costs are proportional to distance traveled.