Abstract
Smooth bromegrass (Bromus inermis) was adopted as experiment materials for identifying the seed maturity using a combination of multispectral imaging and machine learning. The trials were conducted to investigate the effects of three nitrogen application levels (0, 100 and 200 kg N ha⁻¹, defined as CK, N1 and N2 respectively) and two spikelet grain positions: superior grain (SG) at the basal position and inferior grain (IG) at the upper position, on smooth bromegrass seeds. The germination characteristics of the seeds revealed that the variations in nitrogen application and grain positions significantly influenced seeds vigor. The seed vigor increased gradually with their maturity, reaching a high level at 30 and 36 days after anthesis. A stacking ensemble learning approach was employed to identify the seed maturity based on multispectral imaging and autofluorescence imaging. The results demonstrated that the Ensemble model outperformed Support Vector Machine, Bayesian, XGBoost and Random Forest across all evaluated metrics in different scenarios. The model accuracy in CK, N1 and N2 were 89%, 87% and 93%, respectively. Furthermore, the SHapley Additive exPlanations method was selected to interpret the Ensemble model, identifying important features such as 405, 430, 540, 630, 645, 690, 850, 880 and 970 nm. These features exhibited a significant correlation with fresh weight, shoot length and vigor index. These findings showed the high accuracy and generalizability of the Ensemble model for identifying the maturity and quality of smooth bromegrass seeds. Therefore, a new strategy would be offered for evaluating seed maturity and vigor level.
Introduction
Smooth bromegrass (Bromus inermis) is a vital perennial forage grass, but it presents a major challenge for seed producers: inconsistent maturity. Harvesting too early results in seeds that lack vigor and yield, while harvesting too late leads to "shattering," where seeds fall off the plant, causing massive losses. Determining the optimal harvest window traditionally requires labor-intensive germination assays that take weeks and destroy the samples. Furthermore, seed quality is not uniform across a field, varying based on nitrogen application and the seed's physical position on the spikelet (superior vs. inferior grains). This research addresses these challenges by developing a robust, non-destructive way to predict seed quality despite environmental and biological variables.
Methods and Materials
The study utilized a dataset of 4,431 seed samples collected from the Yuershan Ranch in Hebei Province, China. The dataset explicitly incorporated agronomic variance, including four maturity stages (16, 23, 30, and 36 days after anthesis), three nitrogen levels (0, 100, and 200 kg N ha⁻¹), and two grain positions (superior grains at the basal position and inferior grains at the upper position).
The technical approach integrated high-dimensional imaging with advanced machine learning:
- Multispectral & Autofluorescence Imaging: Using a VideometerLab4, the team captured 19 multispectral wavelengths (from UV at 365 nm to NIR at 970 nm) and 8 excitation-emission autofluorescence combinations, providing a detailed chemical and morphological profile of each seed.
- Stacking Ensemble Learning: Instead of relying on a single algorithm, the researchers used a "stacking" approach. They trained four base learners—Random Forest (RF), Support Vector Machine (SVM), XGBoost, and Bayesian models—and then used a "meta-learner" to combine their outputs. This architecture is designed to handle complex, non-linear relationships in biological data better than individual models.
- Interpretability: To avoid the "black box" problem, the team used SHapley Additive exPlanations (SHAP) to identify exactly which wavelengths were driving the model’s decisions.
Key Results
The study confirmed that "superior" grains at the bottom of the spikelet are heavier and more vigorous than "inferior" grains at the top. High nitrogen application (N2 level) not only increased yield but also improved the model's ability to distinguish maturity, resulting in the highest prediction accuracy of 93%.
Seed vigor was found to peak at 30 and 36 days after anthesis, providing a clear target for harvest timing. The Ensemble model consistently outperformed all individual algorithms, achieving accuracy rates of 89% (Control), 87% (Low Nitrogen), and 93% (High Nitrogen). SHAP analysis identified 405, 430, 540, 630, 645, 690, 850, 880, and 970 nm as the most critical features. These spectral markers showed significant correlations with physical vigor indicators like shoot length and fresh weight.
Conclusion
This research demonstrates that combining multispectral imaging with a stacking ensemble model provides a high-accuracy, non-destructive "spectral fingerprinting" method for identifying seed maturity and vigor. While future work is needed to validate the model across different geographical locations and explore real-time field imaging, this approach offers a promising new strategy for precision agriculture in forage seed production.
Summary updated