March 30, 2026 Dataset Open Access

Data for Upscaling mineralogy with hyperspectral data: a benchmark dataset and machine learning framework to enable hyperspectral geometallurgy

Thiele, Samuel Thomas; Kirsch, Moritz; Frenzel, Max; Tolosana Delgado, Raimon; Kamath, Akshay Vijay; Guy, Bradley Martin; Kim, Yongwhi; Laura, Tusa; Járóka, Tom; Gloaguen, Richard

Mineral liberation analysis (MLA) dataset accompanying the paper: Upscaling mineralogy with hyperspectral data: a benchmark dataset and machine learning framework to enable hyperspectral geometallurgy. This describes the mineralogy of 204 thick-sections prepared from 49 drillholes sampled across 7 different locations and coregistered with VNIR-SWIR-MWIR-LWIR hyperspectral data. It is intended to help develop, test and benchmark methods for predicting mineralogy from hyperspectral data. 

The data are stored as hycore (https://github.com/samthiele/hycore) Shed directories for easy loading, although individual MLA sections and corresponding hyperspectral images are all in ENVI format (so can be loaded by any hyperspectral analysis code or software). MLA outputs are also stored in their original (high-resolution) form as indexed bitmaps. The AbundanceMapping.xlsx file can be used to translate these MLA class indices into modal mineral abundances.

Finally, jupyter notebooks used to derive the benchmarks presented in the paper are also included, in the Code folder. These illustrate how the data can be loaded and manipulated using hycore and hklearn (https://github.com/samthiele/hklearn), and used to train machine learning models that predict modal mineralogy given hyperspectral data.