A metric for spatial data layers in favorability mapping for geological events

The authors present a new metric for quantifying the information content of a categorical data layer with respect to a selected prediction target. When the data layer is used in a favorability prediction, its metric value indicates the amount of contribution the data layer can make to the prediction capability of the model. A small metric value normally means the layer contributes little to the prediction task. Given a data layer in the form of a categorical map, the authors define the metric as an average ranking measure among all known target occurrence locations, where the ranking measure is defined for the classes in the categorical map based on their relative favorabilities. Two independent sets of past known occurrence data are used in defining the metric: one as a training set to define favorabilities for ranking the classes, and the other as an evaluation set to define the metric. The metric is tested on a real data set in a study of predicting landslides. The calculated metric values for the data layers agree with observations as well as with other theories