A machine learning approach to geochemical mapping

Geochemical maps provide invaluable evidence to guide decisions on issues of mineral exploration, agriculture, and environmental health. However, the high cost of chemical analysis means that geochemical survey sampling... [ view full abstract ]

Geochemical maps provide invaluable evidence to guide decisions on issues of mineral exploration, agriculture, and environmental health. However, the high cost of chemical analysis means that geochemical survey sampling density will always be limited. Traditional geochemical maps, produced by spatial interpolation alone, are therefore limited in resolution and lack useful fine-scale accuracy. A possible solution to this problem comes from implementing models that form predictions on the basis of high resolution auxiliary information instead. This study uses quantile regression forests (an elaboration of random forest) to investigate the potential of high resolution remote sensing and geophysical survey data to support the generation of accurate and interpretable geochemical maps in south west England. A summary of their performance is presented.
Through stratified 10-fold cross validation we find the accuracy of quantile regression forests in predicting soil geochemistry in south west England to be a general (but not quite unanimous) improvement over that offered by ordinary kriging. The concentrations of immobile elements whose distributions are most tightly controlled by bedrock lithology are predicted with the greatest accuracy (e.g. Al with a cross validated R2 of 0.79), while the concentrations of more mobile elements prove harder to predict. In addition to providing a high level of prediction accuracy, models built on high resolution auxiliary variables allow for informative, process based, interpretations to be made. In conclusion, this study has highlighted the ability to map and understand the surface environment with greater accuracy and detail than previously possible by extracting information from multiple datasets.