Image-based and biochemical multimodal phenotyping for explainable classification of chia (Salvia hispanica L.) genotypes.
TL;DR
The findings indicate that multimodal phenotyping, coupled with explainable machine learning, offers a practical and biologically interpretable decision-support approach for chia genotype classification.
Abstract
Background
This study developed an explainable machine learning framework integrating morphological, color, and biochemical characteristics for classifying chia (Salvia hispanica L.) genotypes. A dataset was assembled from 1200 seed images spanning four genotypes, from which 17 morphological and color features were extracted. These were complemented by six sample-level biochemical traits - crude protein, fat, ash, fiber, carbohydrate, and total sugar - obtained from the corresponding experimental-unit seed sample, resulting in a total of 23 variables in the integrated dataset. The dataset was evaluated comparatively with 10 machine learning algorithms under repeated 10-fold cross-validation, with all preprocessing confined to each training fold to avoid data leakage.
Results
The highest performance was obtained with XGBoost, reaching 86.99% accuracy, a Matthews correlation coefficient of 0.820, a receiver operating characteristic (ROC) area of 0.975, and a precision-recall curve (PRC) area of 0.933; Simple Logistic followed closely at 86.85% accuracy, with comparable ROC and PRC areas (0.974 and 0.933). Significant differences among the algorithms were confirmed by the Friedman test (P = 2.47 × 10-120), with post hoc comparisons placing XGBoost and Simple Logistic within the same top-performing group. Protein, fiber, ash, and fat were the most influential biochemical traits, while hue and saturation among color parameters and shape index and geometric mean diameter among morphological features also contributed appreciably. The G1 genotype, which showed comparatively high protein (27.62%) and fiber (40.62%) contents, was the most consistently distinguished class, with XGBoost and Simple Logistic achieving F-measures of 0.954 and 0.955, respectively, whereas greater phenotypic overlap between G2 and G3 resulted in more frequent mutual misclassifications.
Conclusion
These findings indicate that multimodal phenotyping, coupled with explainable machine learning, offers a practical and biologically interpretable decision-support approach for chia genotype classification. © 2026 The Author(s). Journal of the Science of Food and Agriculture published by John Wiley & Sons Ltd on behalf of Society of Chemical Industry.
