Abstract: This study focused on 77 powdered cobalt-rich crust samples from the Magellan Seamounts in the western Pacific. Reflectance spectra over the range of 350-2500 nm were acquired using a LabSpec^® 4 Hi-Res field spectrometer. Combined with chemical elemental data, a hyperspectral–element dataset was established to investigate spectral responses, evaluate retrievability, and analyse the underlying mechanisms. This study aimed to reveal the relationships between elemental contents and hyperspectral reflectance in cobalt-rich crusts, clarify their mineralogical mechanisms, and assess the hyperspectral retrievability of elemental contents. The study first used non-negative matrix factorization (NMF) to analyse the main variation directions of the mixed spectra of the samples and their possible mineralogical sources. Then, bandwise Pearson correlation analysis and continuum removal (CR) were applied to extract elemental absorption features and identify element-sensitive bands. Finally, partial least squares regression (PLSR) was used to construct retrieval models, and the spectral retrievability of elements was classified based on cross-validation metrics and RPD. The results show that: ① The selection of the number of NMF endmembers indicates that k=5 achieves a favourable balance between reconstruction error and endmember stability. The five mathematical endmembers exhibit different spectral characteristics, indicating the main spectral variation directions controlled by multi-component mixing. These responses are jointly controlled by the Fe-Mn oxide absorption background, phosphate/carbonate components, aluminosilicate/detrital components, and elemental covariation. ② The correlations between elemental concentrations and hyperspectral reflectance are mainly concentrated in the visible to early near-infrared region, with the principal response interval occurring at 463-1058 nm and the strongest correlations around 450-700 nm. P and Ca show positive correlations within this interval, whereas Fe, Mn, Co, Ni, Ti, and V show negative correlations. ③ PLSR cross-validation results reveal clear differences in spectral retrievability among elements. Ca and P show relatively better modeling performance, with the optimal models yielding cross-validated coefficients of determination ( R_\textcv^2 ) of 0.553 and 0.519 and RPD values of 1.506 and 1.451, respectively. Several major metallogenic elements, including Fe, Mn, and Co, also achieved R_\textcv^2 ＞0.4 and RPD＞1. Overall, the hyperspectral signatures of cobalt-rich crusts are primarily controlled by the mixed spectral effects of the Fe-Mn oxide background, phosphate/carbonate components, aluminosilicate/detrital components, and H₂O/OH-related shortwave infrared absorption features. The mixed spectra also contain weak spectral signals associated with major elements occurring through ion adsorption or lattice substitution mechanisms. These signals can be utilized for the quantitative or semi-quantitative prediction of certain elements, as well as for the preliminary screening of others.