On Blind Tests and Spatial Prediction Models

This contribution discusses the usage of blind tests, BT, to cross-validate and interpret the results of predictions by statistical models applied to spatial databases. Models such as Bayesian probability, empirical likelihood ratio, fuzzy sets, or neural networks were and are being applied to identify areas likely to contain events such as undiscovered mineral resources, zones of high natural hazard, or sites with high potential environmental impact. By processing the information in a spatial database, the models establish the relationships between the distribution of known events and their contextual settings, described by both thematic and continuous data layers. The relationships are to locate situations where similar events are likely to occur. Maps of predicted relative resource potential or of relative hazard/impact levels are generated. They consist of relative values that need careful quantitative scrutiny to be interpreted for taking decisions on further action in exploration or on hazard/impact mitigation and avoidance. The only meaning of such relative values is their rank. Obviously, to assess the reliability of the predicted ranks, tests are indispensable. This is also a consequence of the impracticality of waiting for the future to reveal the goodness of our prediction. During the past decade only a few attempts have been made by some researchers to cross-validate the results of spatial predictions. Furthermore, assumptions and applications of cross-validations differ considerably in a number of recent case studies. A perspective for all such experiments is provided using two specific examples, one in mineral exploration and the other in landslide hazard, to answer the fundamental question: how good is my prediction?