Bauxite Residues as a Source of Spectral Signatures for Fe and Al Exploration Using Machine Learning and Sentinel-2 Data

Abstract

The Dom Bosco Syncline (DBS), located in the southern portion of the Quadrilátero Ferrífero, Brazil, contains significant Fe and Al deposits. However, dense vegetation cover and geological heterogeneity impose substantial limitations on mineral prospecting using orbital remote sensing. In this context, the machine learning model MinMap-S2 (Mineral Mapping via Sentinel-2) is proposed to support mineral exploration in the DBS. MinMap-S2 is a hybrid model based on spectral signatures derived from bauxite residues deposited in the Marzagão tailings dam (MTD), located within the DBS, for the detection of Fe- and Al-rich mineralization. The model was developed using Python, and its methodology integrates the algorithms Self-Organizing Maps (SOM) and Extreme Gradient Boosting (XGBoost) to process Sentinel-2 multispectral data. SOM is employed to generate a spectral library from bauxite residues and surrounding surface materials, which is subsequently used by XGBoost for supervised classification and spatial prediction. The model was applied to a region located in the eastern portion of DBS, characterized by detrital-lateritic covers and metasedimentary rocks, including itabirites, phyllites, schists, quartzites, and sericitic metarenites. Validation of the results was conducted through geological mapping, field sampling, and petrographic analyses. The results demonstrate that MinMap-S2 effectively identifies areas with spectral responses associated with Fe- and Al-rich mineralization, supporting the discrimination of mineralized zones in heterogeneous terrains and under dense vegetation cover. Overall, the proposed methodology demonstrates potential as a supporting tool for mineral exploration using Sentinel-2 multispectral data.