Uranium anomalies detection through Random Forest regression

Abstract

Airborne geophysics provides one of the most relevant data for uranium exploration. However, the application of radiometric surveys is decreasing considerably as the depth of exploration is increasing. Notwithstanding, there is still a large potential for radiometric data, especially using recent data processing techniques such as machine learning methods. In this work, we propose a new method to detect uranium anomalies through regression using the Random Forest machine learning algorithm (RF). The RF regression allows combining airborne geophysical data to predict the expected uranium content, which represents the uranium content generated by environmental effects such as lithology and pedogenesis. Therefore, the deviation (Ud) between the measured uranium and the expected uranium represents the secondary effects such as weathering, soil alteration, hydrothermal alteration or mineralisation process. We evaluated the relevance of the geophysical parameters proposed by previous authors in the prediction of the expected uranium (thorium, thorium potassium ratio, uranium potassium ratio, and Total Gradient Amplitude). Randomly selecting only 10% of the database as training data, we estimate the expected uranium with an R 2 = 0.99 concerning the measured uranium. To assess the reliability of the Ud anomalies, we employed the proposed methodology in the Carajás Mineral Province (CMP), Brazil. In the CPM, the Ud anomalies showed a clear correlation with the several Iron Oxide Copper–Gold deposits (IOCG) and some IOCG-related and granite-related prospects. In situ measurements with a portable gamma-ray spectrometer in the Salobo mine supported the uranium anomalies. The Ud map also highlighted contrasts within granites that correlate with previously reported granitic facies. Therefore, the Ud map generated by RF regression is useful in setting exploration targets for conventional and unconventional uranium resources, as well as in high-detail granitic facies mapping.