Assessment of Exploration Bias in Data-Driven Predictive Models and the Estimation of Undiscovered Resources

Abstract

The spatial distribution of discovered resources may not fully mimic the distribution of all such resources, discovered and undiscovered, because the process of discovery is biased by accessibility factors (e.g., outcrops, roads, and lakes) and by exploration criteria. In data-driven predictive models, the use of training sites (resource occurrences) biased by exploration criteria and accessibility does not necessarily translate to a biased predictive map. However, problems occur when evidence layers correlate with these same exploration factors. These biases then can produce a data-driven model that predicts known occurrences well, but poorly predicts undiscovered resources.

Statistical assessment of correlation between evidence layers and map-based exploration factors is difficult because it is difficult to quantify the “degree of exploration.” However, if such a degree-of-exploration map can be produced, the benefits can be enormous. Not only does it become possible to assess this correlation, but it becomes possible to predict undiscovered, instead of discovered, resources.

Using geothermal systems in Nevada, USA, as an example, a degree-of-exploration model is created, which then is resolved into purely explored and unexplored equivalents, each occurring within coextensive study areas. A weights-of-evidence (WofE) model is built first without regard to the degree of exploration, and then a revised WofE model is calculated for the “explored fraction” only. Differences in the weights between the two models provide a correlation measure between the evidence and the degree of exploration.

The data used to build the geothermal evidence layers are perceived to be independent of degree of exploration. Nevertheless, the evidence layers correlate with exploration because exploration has preferred the same favorable areas identified by the evidence patterns. In this circumstance, however, the weights for the “explored” WofE model minimize this bias. Using these revised weights, posterior probability is extrapolated into unexplored areas to estimate undiscovered deposits.