Assisted predictive mapping with machine learning

Different machine learning techniques have been used to evaluate underground prospectivity in various fields, such as Oil and Gas industry, Mineral prospection or Geothermal Energy. However, the quantification of associated confidence maps is not always systematically performed, although they can significantly benefit any decision making based on the outcomes of these machine learning workflows. 

Our objective is to allow the computation of prospectivity maps and the quantification of the associated confidence maps.

Our approach uses in competition among the most efficient methods of Machine Learning, such as XGBoosting, two types of Random Forest (Distributed & Distributed Uplift), Generalized Linear Model, Generalized Additive Models, Gradient Boosting Machine, Support Vector Machine, Stacked Ensembles and Deep Learning (Neural Networks). The confidence maps are quantified by the standard deviation of the predictions of the ten best models and by the deviation P90 – P10 between the 90th percentile and the 10th percentile of the hundred best predictions.

The method is applied to public data inspired by a true case of mineral prospectivity on volcanogenic massive sulfide Cu–Zn deposit in northwestern China. A prospectivity map produced by experts was available and allowed to successfully test our method using the West part of the map for model Learning  and the East part for testing the model. Unsurprisingly, the confidence map in the western part of the map exhibits very favorable scores, while the one in the eastern part are less favorable... while still being quite acceptable (data from https://doi.org/10.1007/s11004-019-09844-2   ).

The early results are promising and pave the way for further adaptation. The proposed method, initially intended for mineral prospectivity, is sufficiently generic to be adapted to other domains even outside mining industry, such oil & gas industry or in geothermal energy.