Abstract

Ginkgo biloba is a dioecious species valued for landscaping and medicine, but rapid sex identification outside the flowering and fruiting stages remains challenging. Hyperspectral imaging offers a potential solution, though whether spectral sex markers are stable across phenological stages and can support a single year-round model is unclear. Here, leaf hyperspectral reflectance (400–1000 nm) was acquired from mature G. biloba at flowering (n = 360), green-leaf (n = 1395), and yellow-leaf (n = 243) stages. Sex identification models were built using multiple machine learning classifiers, with leaf flavonoid content as biochemical validation. Stage-specific models achieved optimal test accuracies of 96.67% (flowering, MSC + PLS-DA), 95.77% (green-leaf, raw spectra + LDA), and 97.12% (yellow-leaf, MSC + LDA). However, discriminative bands shifted from the visible (520–690 nm) at the green-leaf stage to the near-infrared (700–1000 nm) at the yellow-leaf stage, with almost no bands shared across all stages. Furthermore, cross-stage prediction accuracy dropped to near-chance levels (~50%). As t-SNE analysis revealed, the universal model had learned phenological rather than sex-specific information. In addition, flavonoid measurements revealed a highly significant sex × stage interaction (p < 0.001) and a reversal of the sex difference between green-leaf (male > female) and yellow-leaf (female > male) stages. Thus, the spectral and biochemical sex markers examined in this study varied substantially with phenology, and stage-specific models are required for practical sex identification in G. biloba.

IPC Classification

Keywords