Machine Learning-Based 3D Modeling of Mineral Prospectivity Mapping in the Anqing Orefield, Eastern China

• Published in: Natural resources research (New York, N.Y.) N.Y.), 2021-10, Vol.30 (5), p.3099-3120
• Main Authors: Qin, Yaozu, Liu, Liangming, Wu, Weicheng
• Format: Article
• Language: English
• Subjects: Algorithms; Chemistry and Earth Sciences; Computer Science; Earth and Environmental Science; Earth Sciences; Exploration; Fossil Fuels (incl. Carbon Capture); Geography; Geological mapping; Geology; Learning algorithms; Machine learning; Mapping; Mathematical Modeling and Industrial Mathematics; Mineral Resources; Modelling; Original Paper; Performance evaluation; Physics; Prediction models; Statistics for Engineering; Sustainable Development; Three dimensional models; Volumetric strain
• ISSN: 1520-7439; 1573-8981
• DOI: 10.1007/s11053-021-09893-7

Highlights Actual geological data, accurate models and precise samples are critical for ore targeting. The RF-based prediction model is more applicable for mapping mineral prospectivity than other algorithms in this study. The determination of sample set is more important than algorithm if there is not enough field data. Successful delineation of high potential targets for exploration in maturely-explored orefields is still a tough challenge. A reliable prediction model achieved by integration of various ore-related geological factors and exploration information in the 3D space is an effective approach to deal with this challenge. The Anqing orefield has been intensively exploited for decades, and thereby the possible potential left must be at depth. The accumulated abundant data of exploration and research provide us a possibility for carrying out machine learning-based 3D modeling. The 3D block models of the main geological bodies, resistivity and volumetric strain field in this orefield were used as multi-resource geological data to construct prediction models by using weight-of-evidence and machine learning methods. Through performance evaluation and comparison, the following conclusions were obtained: (1) it is more scientific and reasonable to use all the geological prospecting factors concurrently for mineral prospectivity mapping (MPM) rather than use one or a part of them; (2) random forest (RF) algorithm seems capable of MPM because of its high accuracy and reliability in prediction; and (3) rational selection of training and learning samples, especially, those from actual geological objects and exploration engineering, plays a more critical role in MPM than algorithm and methods themselves. Two different RF prediction models were obtained for MPM in the east and the surrounding part of this orefield based on the outcome of geophysical prospecting. The spaces with prediction probabilities higher than 0.508 in the east part and 0.501 in the surroundings take up only 3.71% volume of the whole orefield, but contain 95.92% of the mineralized blocks. The high potential targets are most likely parts of the above spaces with high prediction probabilities that have not been drilled yet up to now.