Introduction

Bread wheat (Triticum aestivum L.) is a strategic crop for ensuring global and national food security. Seed quality is a critical factor for successful seedling establishment and achieving optimal yield. However, seed ageing during storage remains a major challenge in this field, as it significantly reduces seed vigor and viability. To assess seed quality accurately and quickly, traditional methods are time-consuming and have limitations due to their high demand for time and resources; therefore, machine learning (ML) offers faster alternatives for predicting germination and seedling growth traits. This study evaluates six ML models, including Linear Regression, XGBoost, Support Vector Regression (SVR), Gaussian Process, Lasso, and Ridge, for predicting 14 germination and seedling growth traits under accelerated ageing.

Materials and Methods

A total of 228 bread wheat genotypes were tested during the 2022–2023 season. The experiment was conducted as a factorial trial in a completely randomized design with three replications at the Seed Laboratory, Faculty of Agriculture, Urmia University. Accelerated aging treatments (0, 48, 72, and 96 hours) were applied at 45°C and near-saturated relative humidity, following ISTA standards. Germination tests followed ISTA standards under controlled laboratory conditions. Traits measured included germination percentage (GP), germination index (GI), vitality index (VI), simplified vitality index (SVI), mean germination time (MGT), mean daily germination (MDG), seedling length (SL), shoot length (SHL), root length (RL), shoot wet weight (SHWW), shoot dry weight (SHDW), root wet weight (RWW), root dry weight (RDW), and allometric coefficient (AL). Machine learning models were implemented in Python 3.9 using the StandardScaler from the scikit-learn library, along with XGBoost and NumPy packages. Data were randomly split into training (80%) and testing (20%) subsets, with fixed random seed for reproducibility. Continuous variables were normalized, and model performance was evaluated using R² and MAE. Trait correlations were visualized via heatmap.

Results and Discussion

The results revealed strong positive correlations among most germination and seedling growth indices. In addition, a significant declining trend (P<0.05) was observed in most traits with increasing aging duration. For instance, germination percentage (GP) decreased from an average of 91.96% under control conditions to 11.61% at 96 h of aging. Furthermore, cluster analysis identified KOOHDASHT, FONG, ALVAND, 627849, 626978, 627787, 624864, 627414, 626215, and 627410 as tolerant genotypes, exhibiting the highest seed performance under aging conditions. Evaluation of the models based on MAE and R2 metrics demonstrated that the Gaussian Process (GP) model outperformed other algorithms for key traits including germination percentage, germination index, vitality index, simplified vitality index, and mean daily germination achieving higher R2 values (ranging from 0.90 to 0.94) and the lowest MAE. Conversely, for seedling traits such as seedling length, shoot length, root length, and shoot/root fresh and dry weights, the XGBoost model was identified as the optimal algorithm, exhibiting superior predictive power (R2: 0.61–0.80) and reduced prediction error (MAE: 0.0011–3.12).

Conclusion

ML models, especially XGBoost and Gaussian Process, can effectively predict germination and seedling growth traits in bread wheat under accelerated ageing. By identifying tolerant genotypes and providing an intelligent framework for trait prediction, the findings of this study offer an efficient integration of statistical methods and machine learning to accelerate breeding programs and facilitate precise seed quality management.