Strategic potential assessment of lanthanum and scandium through geochemical-lithological analysis with unsupervised machine learning in southern Ecuador

The preliminary identification of areas with strategic geochemical potential poses a major challenge in mineral exploration when only surface-level and unclassified data are available. This study aimed to integrate geostatistical techniques with unsupervised machine learning algorithms to classify zones of high, medium, and low potential for lanthanum (La) and scandium (Sc) in southern Ecuador. A database comprising 3998 geochemical samples was used, with concentrations estimated via Ordinary Kriging (OK), employing variogram structures tailored to each element. The K-means, Gaussian Mixture Models (GMM), Mini-Batch K-means (MBKM), and Spectral Clustering (SC) algorithms were applied to the interpolated values to achieve automated spatial segmentation. Validation against a traditional percentile-based classification yielded high accuracy, with SC (accuracy = 0.898) and KM (0.860) performing best for La, and GMM (0.899) for Sc. Additionally, total metal contents per zone were estimated, reaching up to 725.10 t of La (average grade: 11.98 mg/kg) and 103.23 t of Sc (average grade: 2.08 mg/kg) in medium-potential zones according to GMM and SC, respectively. Strong lithological associations were identified, particularly highlighting the JUB unit as key for scandium occurrence. Overall, the results confirm that the combination of kriging and unsupervised clustering enables effective classification of mineralogical domains with high spatial coherence, providing a robust tool for prioritizing target areas in early exploration stages.