Hyperspectral drill-core scanning adds value to exploration campaigns by providing continuous, high-resolution mineralogical data over the length of entire boreholes. However, multivariate mineralogical data must be transformed into lithological domains such that it is compatible with interpolation techniques and be usable for geomodeling. Manual interpretation of multivariate drill-core data is a challenging, time-consuming and subjective task, and automated or semi-automated approaches are needed. However, naive machine-learning techniques that ignore the distinct spatial structure and multi-scale nature of geological systems tend to produce geologically unreasonable results. Automated geological logging and multi-scale hierarchical domaining of drill-cores has been previously addressed in several studies by means of scalograms from a wavelet transform and tessellation, albeit exploiting only univariate information. The methodology involves the extraction of the local first principal component at a neighborhood of each observation, and the segmentation of the resulting series of scores with a continuous wavelet transform for boundary detection. In this way, the correlation pattern between the variables is incorporated into the segmentation. The scalogram accurately locates the geological boundaries at depth and yields hierarchical geological domains with mineralogical composition characteristics. The performance of this approach is demonstrated on a synthetic as well as a real multivariate dataset. The real dataset consists of mineral abundances derived from drill-core hyperspectral imaging data acquired in Elvira, a shale-hosted volcanogenic massive sulfide deposit located in the Iberian Pyrite Belt, where 7000 m of drill-core were acquired along 80 boreholes. The extracted domains are sensible from a geological point of view and spatially coherent across the boreholes in cross-sections. The results at relevant scales were qualitatively validated by comparing against the lithological log. This method is fast, is appropriate for multivariate geological data along boreholes, and provides a choice of scales for hierarchical geological domains along boreholes with mineralogical composition characteristics that can be modeled in 3D. Our approach provides an automatic way to transform hyperspectral image-derived mineral maps into vertically coherent geological units that are appropriate inputs for 3D geological modeling workflows. Moreover, the method improves the boundary detection and geological domaining by making use of multivariate information.

To advance Europe's self-reliance in sustainably sourcing lithium, innovative exploration technologies that enhance the efficiency of both exploration and extraction are urgently required. Lithium's role in rechargeable batteries necessitates advanced methods to support domestic resource development and meet rising demand. This study evaluates artificial neural networks (ANN) combined with hyperspectral data for lithium prediction, comparing the HySpex Mjolnir S-620 sensor and the Technical Research Centre of Finland Oy's (VTT) novel active hyperspectral scanner (AHS). The AHS employs a broadband short-wave infrared supercontinuum light source, avoiding dependency on external illumination and overcoming limitations of traditional systems in underground settings. At Germany's Zinnwald/Cínovec greisen deposit, lithium prediction was achieved via ANN trained on portable X-ray fluorescence (pXRF)-derived rubidium data, a proxy for lithium in zinnwaldite-rich mineralization. Hyperspectral processing extracted diagnostic absorption features (e.g. band ratios, minimum wavelength mapping) and classified lithium-bearing minerals. The ANN model produced a high-resolution lithium concentration map, demonstrating the AHS's ability to enhance mine-face characterization for both detailed exploration and operational workflows, such as grade control. Integrating hyperspectral sensing and machine learning supports Europe's sustainable objectives by enabling more precise and efficient resource extraction, offering scalable solutions for modern mining. The AHS–ANN framework also holds potential for agriculture, environmental monitoring and defence, where rapid material characterization is paramount.

The Goldeneye project combines different sensing technologies with proximal sensing to produce reference calibrated mineralogical maps through data fusion. In order to develop mineral detection applications, rock specimens, taken from an outcrop in Bockau (Erzgebirge, Germany), are analyzed with active hyperspectral scanning (AHS) as well as portable X-ray fluorescence (pXRF) devices. The received data is analyzed by means of artificial intelligence in order to develop an approach to automatically map the minerals with the samples. The analysis is carried out in advangeo® 2D Prediction, developed by Beak Consultants GmbH. Tin concentrations derived from pXRF measurements and AHS data from 2/3 of the specimen surface serve as training and validation data of the artificial intelligence algorithm (artificial neural networks). As a result, we developed a prediction model for the distribution of tin and its associated mineral cassiterite throughout the rock specimen, which allows to detect the mineral potential of hand specimens and larger outcrops in a fast and reliable manner.The paper has been prepared in the frame of the Horizon 2020 co-funded project GOLDENEYE, which has received funds through the Grant Agreement 869398.

Drill core samples have been traditionally used by the mining industry to make resource estimations and to build geological models. The hyperspectral drill core scanning has become a popular tool in mineral exploration because it provides a non-destructive method to rapidly characterise structural features, alteration patterns and rock mineralogy in a cost effective way. Typically, the hyperspectral sensors cover a wide spectral range from visible and near-infrared (VNIR) to short and long wave infrared (SWIR and LWIR). The spectral features in this range will help to characterize a large number of mineral phases and complement the traditional core logging techniques. The hyperspectral core scanning provide mineralogical information in a millimetre scale for the entire borehole, which fills the gap between the microscopic scale of some of the laboratory analytical methods or the sparse chemical assays and the meter scale from the lithological descriptions. However, applying this technique to the core samples of an entire ore deposit results in big datasets. Therefore, there is the need of a workflow to build a 3D geological model conditioned by the data applying suitable data reduction methods and appropriate interpolation techniques. This contribution presents a case study in the combination of traditional core logging and hyperspectral core logging for geological modelling. To attain mineral and alteration maps from the hyperspectral data, unsupervised classification techniques were applied generating a categorical data set. The amount of data was reduced by the application of a domain generation algorithm based on the hyperspectral information. The domain generated by the algorithm is a compositional categorical data set that was then fed to condition the application of stochastic Plurigaussian simulations in the construction of 3D models of geological domains. This technique allows to simulate the spatial distribution of the hyperspectral derived categories, to make a resource estimation and to calculate its associated uncertainty.

To ensure Europe increases its domestic production of high quality and responsibly produced raw materials, the development of innovative technologies for 3D geological modeling in mineral exploration is paramount. The Erzgebirge in Germany provides an excellent framework to showcase the application of artificial intelligence and in particular Artificial Neural Networks (ANN) for 3D mineral prospectivity mapping. The Erzgebirge belongs to the Variscan Belt, withholding 800 years of mining history and it is also famous for Ag, Sn, W, Fe, Cu, Li mineralizations among others. The Bockau deposit is located at the western section of the Erzgebirge. The target area is a Paleozoic metasediment body that was formed during the Variscan orogeny. The metasediment body consists primarily of alternating micaschist, phyllite and quartzite and dips mostly 25° to 240° SW. The metasediment is surrounded by Late Variscan plutons which partly led to contact metamorphic zones. In addition there is a large Quartzite body which was mined near to the surface in the 17th century for Sn, following a stratiform tin anomaly which can reach up to 4000 ppm Sn. Thanks to the long mining history, the Bockau deposit condenses a large amount of geological, geochemical, geophysical and mineral data. To increase mineralogical knowledge of the deposit and to help identify drilling targets, a hybrid approach for 3D mineral predictivity mapping is implemented. Potentially mineralisation-controlling factors are identified in knowledge-driven genetic exploration models, taking into account the borehole data, major faults, electromagnetic data, intrusive bodies, contact metamorphic zones and lithological borders, followed by data-driven weighted ANN predictive modelling implemented in the in-house developed advangeo® 3D Prediction Software. The predictive model is guided by structural variables such as the euclidean distance to fault planes, lithological surfaces and to metamorphic contact zones. The model is also constrained by geophysical data by a magnetic susceptibility model obtained from an airborne magnetic data inversion. Finally, Sn anomaly data from boreholes is implemented as training data for the prediction. The results show the probability distribution of Sn mineralisation occurrence in 3D over a voxel model formed by blocks of approximately 684 m3 13(x), 13.5(y) and 4(z), increasing the mineralogical knowledge of the deposit and guiding exploration efforts complementing the decision making process for drilling new targets. The results are validated by iteratively implementing the jackknife method, splitting the training data into validation and training subsets. The first prediction iteration is performed with a subset containing 77 % of the Sn content data from boreholes as training data, followed by 50 and 30 % subsets. Thus, allowing at each iteration to perform a quantitative evaluation of the prediction by comparing the validation subset with the Sn content of the borehole that was not used for the prediction. The paper has been prepared in the frame of the Horizon 2020 co-funded project GOLDENEYE, which has received funds through the Grant Agreement 869398.