Abstract Hyperspectral remote sensing is transforming precision agriculture by capturing detailed spectral information that supports improved crop management. This study evaluated forage quality in soybean ( Glycine max ), tepary bean ( Phaseolus acutifolius ), and moth bean ( Vigna aconitifolia ) using in situ hyperspectral measurements and convolved wavebands to predict neutral detergent fiber, acid detergent fiber, in vitro true digestibility, and crude protein with machine learning models including k‐nearest neighbors (k‐NN), partial least squares regression (PLSR), support vector networks (SVN), and random decision forests (RDFs). To assess the potential of future hyperspectral satellites, hyperspectral wavelengths were convolved to represent the proposed spectral bands of CHIME. Additional analyses compared models incorporating legume type as a variable with those developed separately for each species. Results indicate that both in situ hyperspectral data and CHIME‐simulated bands effectively predicted key forage traits, with RDF and SVN consistently outperforming k‐NN and PLSR. SVN achieved the highest predictive accuracy while requiring fewer computational resources, whereas RDF predictions were slightly lower by 1%–6%. In contrast, k‐NN performance declined by 18%–25%, and PLSR showed the weakest predictive ability. Among additional model comparisons, it was found that model building of individual legume types generally outperformed models incorporating legume type as a variable. The results also provide insights into the nutritional composition of legume forages that can aid in their selection and utilization in various agronomic applications. Furthermore, these findings provide a foundation for advancements in hyperspectral remote sensing, providing crucial insights for improving precision agriculture practices.
K. et al. (Sun,) studied this question.