Fecha de publicación: -- Fuente: PubMed "swarm" Talanta. 2026 Oct 26;314:130834. doi: 10.1016/j.talanta.2026.130834. Online ahead of print.ABSTRACTCoal is a host of critical metals, and rapid, non-destructive quantification of these metals is important for resource assessment. This study investigated lithium (Li) and rare earth elements (REE) in coal samples from the Sihe Coal Mine, Shanxi Province, and the Hebi Coal Mine, Henan Province, China. Hyperspectral data (350-2500 nm) were processed using First-order Derivative (FD) and Continuum Removal (CR), followed by feature selection with Competitive Adaptive Reweighted Sampling (CARS) and Boruta. Partial Least Squares Regression (PLSR), Back-Propagation Neural Networks (BPNN), and Particle Swarm Optimization-Support Vector Regression (PSO-SVR) were used to develop quantitative retrieval models. Lithium was relatively enriched, whereas REE concentration were generally within the normal range. The FD-Boruta-BPNN model achieved the highest prediction accuracy for both Li (R2=0.912) and REE (R2=0.968), using fewer input wavelengths. These findings demonstrate that spectral preprocessing, feature selection, and modeling algorithms jointly influence retrieval performance, highlighting the need to balance predictive accuracy and model complexity. Predictions likely rely on indirect spectral information associated with mineral carriers rather than direct elemental absorption features. However, the limited sample size and geographical coverage constrain generalizability. Future research should incorporate larger, independent datasets from diverse geological settings, verify spectral-response mechanisms through mineralogical analysis, and extend the framework to other coal-hosted critical metals and hyperspectral imaging.PMID:42987522 | DOI:10.1016/j.talanta.2026.130834