A machine learning-based lithologic mapping method

Abstract

In this study, a gradient boosting decision tree (GBDT)-based lithologic mapping method constituted by field survey and machine learning is introduced. The Duolong mineral district, Tibet, China is currently chosen for model test. During the practical application, geochemical data at a 1:50000 scale is analyzed to identify lithologic units, while a geological map at the same scale currently provides lithologic units identified by field survey. Lithologic units within a small area are firstly collected from the geological map. Correspondence between geochemical data and lithologic units within the small area can consequently be marked, by which the GBDT method is applied to reclassify the geochemical data and further predict lithologic units in the Duolong district. Transforming the result to a probability distribution, areas with low probability can be identified, and further investigation will be implemented to update geological knowledge and correspondence between geochemical and lithologic units. Iteration of the process will lead a reasonable lithologic mapping result. It is shown that the model accuracy increases with iteration growing, and reaches 87% after 7 iterations. The currently proposed method highlights deep integration of field survey and machine learning algorithm, and emphasizes importance of field work in the whole modeling process. Useful geo-information can be deeply mined from existing data and further updates former geological understandings. Meanwhile, lithologic units within un-explored areas can be identified based on the knowledge in explored areas. The GBDT-based method which effectively reduces field work is a meaningful exploration in lithologic mapping and will provide a new reference and supplementary way to geological mapping.

Keywords

• data mining
• information fusion
• geologic unit
• decision tree
• geological mapping