Finding gold and copper in Catanduanes with random forests
- Article
- Geoscience
Mineral prospectivity mapping asks a practical question: given what is known about an area’s geology, where are undiscovered deposits most likely to be? In a 2015 paper in Natural Resources Research, Emmanuel John M. Carranza and Alice G. Laborte tested whether the random forest algorithm could answer that question for gold-copper deposits on Catanduanes Island, in a setting with very little training data.
The setup
The study area had 17 known gold-copper (Au-Cu) deposits. The authors used those as positive examples and paired them with 17 locations judged non-prospective as negatives, giving a training set of 34 points. Each location was described by a set of predictors drawn from geological maps: distance to particular faults, how favourable the rock at the site is for mineralisation, and geochemical anomalies. Some predictors had missing values at some locations, which is normal for field data and a problem for many methods.
The random forest model was trained to estimate the probability of a deposit at every cell across the island, producing a prospectivity map. That map was then compared with one produced earlier for the same area using an evidential belief (EB) model, a more traditional knowledge-driven approach.
What they found
The spatial relationships the random forest learned between deposits and geological features matched those already established in the evidential belief model, which is reassuring: the algorithm was not finding patterns that a geologist would dispute. On performance, the random forest had a success rate about 5 percent better than the EB model, and a prediction rate about 4 percent worse. In other words, it fit the known deposits slightly better and generalised to held-out ones slightly less well. The two methods are close.
The clearer advantage was on missing data. Random forests can impute missing predictor values using the same tree-based machinery, so the model could use every location in the training set. The EB model has no equivalent feature.
The authors’ conclusion
Carranza and Laborte concluded that random forests have potential for mineral prospectivity modelling in areas with few known occurrences, and are particularly useful where the predictors have gaps. For a country whose mineral resources are still incompletely mapped, being able to model where deposits are likely, from a handful of known ones and imperfect data, is a useful capability.