Application of random-forest machine learning algorithm for mineral predictive mapping of Fe-Mn crusts in the World Ocean

Abstract

Mineral prospectivity mapping constitutes an efficient tool for delineating areas of highest interest to guide future exploration. Multiple knowledge-driven approaches have been applied for the creation of prospectivity maps for deep-sea ferromanganese (Fe-Mn) crusts over the last decades. The results of a data-driven approach making use of an extensive data collection exercise on occurrences of Fe-Mn crusts in the World Ocean and recent increase in global marine datasets are presented. A Random Forest machine learning algorithm is applied, and results compared with previously established expert-driven maps. Optimal predictive conditions for the algorithm are observed for (i) a forest size superior to a hundred trees, (ii) a training dataset larger than 10%, and (iii) a number of predictors to be used as nodes superior to two. The confusion matrix and out-of-bag errors on the remaining unused data highlight excellent predictive capabilities of the trained model with a prediction accuracy for Fe-Mn crusts of 87.2% and 98.2% for non-crusts locations, with a Kohen’s K index of 0.84, validating its application for prediction at the World scale. The slope of the seafloor, sediment thickness, sediment type, biological productivity, and abyssal mountain constitute the five strongest explanatory variables in predicting the occurrence of Fe-Mn crusts. Most ‘hand-drawn’ knowledge-driven prospective areas are also considered prospective by the random forest algorithm with notable exceptions along the coast of the American continent. However, poor correlation is observed with knowledge-driven GIS-based criterion mapping as the Random Forest considers un-prospective most target areas from the GIS approach. Overall, the Random Forest prediction performs better in predicting a high chance of Fe-Mn crust occurrence in ISA licensed area than the GIS approach, which constitutes an external validation of the predictive quality of the random forest model.