Deep Learning and Reflectance Spectroscopy for Mapping Rare Earths in Brazil

Abstract

Found in various rocks and soils, rare earth elements (REEs) are essential to the high-tech industry and the energy transition due to their electrical and magnetic properties. However, identifying and mapping them remains a challenge, often requiring time-consuming chemical analyses and costly techniques. Brazil stands out internationally for its large and diverse rare earth reserves, yet domestic geological knowledge regarding these elements remains limited. Analyzing absorption bands in the reflectance spectra of rocks and soils may enable the identification of constituent minerals and elements, thereby facilitating their characterization and classification. Aiming to develop a tool to assist in the identification and mapping of REEs, this article integrates deep learning techniques with reflectance spectroscopy and orbital hyperspectral imagery to seek a faster, lower-cost solution for mineral prospecting. Using laboratory-measured reflectance spectra from rock, soil, and regolith samples obtained from various Brazilian regions, a multilayer perceptron neural network model was trained to map potential occurrences, while a Hopfield network model was trained for mineral identification. Spectral analysis revealed a predominance of clay minerals, and the prediction algorithm successfully identified mining sites and areas of exposed soil potentially containing REEs concentrations in the order of 144 ppm.