Answers are grounded in this paper and cite their sources.
The low germination percentage of Bupleurum chinense DC. seeds represents a major bottleneck for its standardized, large-scale cultivation. To address this, we integrated machine vision, hyperspectral imaging, and single-seed germination test, proposing a strategy to improve seed lot germination based on maturity discrimination. Seeds were classified into four maturity grades: M0 (immature), M1, M2, and M3 (increasing maturity). M1∼M3 seeds exhibited significantly higher viability (70∼80%) than M0 seeds after four months of storage. A Stacking model (Model A) built from single-seed SG-SNV preprocessed hyperspectral data achieved 74.4% accuracy in viability discrimination. A Voting model (Model B) based on 54 machine vision features achieved 93.9% accuracy. An FD (First derivative)-Stacking model using hyperspectral reflectance achieved 97.2% accuracy. Further refinement using the CARS algorithm to select 112 key spectral bands yielded an FD-CARS (Competitive Adaptive Reweighted Sampling)-Stacking model (Model C) with 98.9% accuracy and 99.3% recall. SHAP analysis identified 503.1 nm as the most critical spectral band for maturity discrimination. Applied to an external seed batch with a baseline germination of 56.7%, Models A, B, and C increased the predicted germination rates to 65.1%, 80.0%, and 75.4%, respectively. Unlike direct prediction of seed viability, this study proposes a two‑step strategy: first, establishing seed maturity grade models using hyperspectral and machine vision data, and then linking maturity to germination performance. This approach significantly outperforms direct viability classification. This study establishes that maturity, corresponding to stable physico-chemical changes during seed development, is a more robust and optically tractable trait for non-destructive sorting than direct viability assessment, which is confounded by variable factors and noisier spectral signals.
No takes yet. Share an insight, caveat, or question.
Junjiao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: