Deep learning based stacking ensembles for tropical sorghum classification

Subtle morphological differences among sorghum varieties make varietal identification difficult in breeding and seed production, increasing the risk of contamination, reduced grain quality, and challenges in maintaining seed purity for farmers. This research integrates deep learning with a stacking ensemble approach to develop a rapid and accurate method for sorghum seed classification, supporting breeding programs, seed production, and on-farm varietal identification. The research was conducted at two experimental sites: Bajeng in Gowa and ICERI in Maros Regency, South Sulawesi. Five popular sorghum varieties (Numbu, Kawali, Super 1, Suri 3 Agritan, and Super 4) were classified using CNN models (VGG16, SqueezeNet, DenseNet-121 and ResNet-50), while a stacking ensemble was constructed with support vector machine, and k-nearest neighbors, using logistic regression as the meta-model. Grid and Bayesian search optimizations were applied for parameter tuning, and model performance was assessed using an 80/20 train-validation split and tenfold cross-validation. The results show that the SqueezeNet-LR stacking model via Bayesian hyper parameters search achieved the highest accuracy of 0.978, highlighting its effectiveness, particularly with high-dimensional features. In comparison, the VGG16-KNN model exhibited the lowest accuracy at 0.924, while ResNet-50 and DenseNet-121 showed slightly lower but comparable performance. Multi-dimensional scaling (MDS) analysis of misclassifications indicated that errors were concentrated among the Kawali, Numbu, and Suri 3 Agritan varieties. Overall, the SqueezeNet-LR stacking model outperformed the others, demonstrating its superiority in seed classification. Furthermore, the practical application of these findings could extend to mobile platforms through integration with Android Studio, Flutter, and TensorFlow Lite, for in-field varietal identification and future research. Overall, this framework is essential for breeding programs and in-field varietal identification, especially to prevent varietal contamination during seed production. • The integration of VGG16, SqueezeNet, and DenseNet-121 CNN models with machine learning techniques led to significant improvements in sorghum panicle classification accuracy. • Grid search optimization, combined with the ensemble stacking approach using LR as the meta-model, achieved the highest classification accuracy of 0.972. • SqueezeNet-LR stacking model outperformed other models, effectively handling high-dimensional features, while VGG16 and DenseNet-121 also demonstrated strong performance. • Misclassifications were predominantly observed among Kawali, Numbu, and Suri 3 Agritan varieties, as revealed by multi-dimensional scaling (MDS) analysis. • The research provides a valuable framework for further development of intelligent sorghum variety classification systems for real-world field applications in Indonesia.