A CNN-Based Deep Learning Architecture for Discriminating Botanical Adulteration and Complexities Among Commercial...

This study presents a CNN-based deep learning framework to automate the accurate authentication of mericarp seeds from 15 commercially important medicinal Apiaceae species, which exhibit subtle morphological overlaps and face risks of toxic adulteration (e.g., hemlock contamination), threatening consumer safety and trade integrity. Six CNN models were evaluated, with DenseNet121 demonstrating superior performance in terms of accuracy, precision, recall, F1-score, and convergence stability, followed by MobileNetV2, InceptionV3, and VGG16. Confusion matrix analysis revealed that higher-resolution inputs significantly improved discrimination, particularly for similar seeds like Anethum graveolens L. and Apium graveolens L. Batch size minimally influenced outcomes. Analysis of accuracy-loss dynamics further indicated that EfficientNetB0 and ResNet50 underperformed, while DenseNet121, which excelled in performance, convergence stability, strong generalization, and minimal overfitting, highlighted the critical role of architectural design in feature learning and optimization of key performance metrics. Additionally, t-SNE visualization confirmed DenseNet121's superior feature learning, achieving distinct separation of all 15 species and resolving intricate visual ambiguities that confound traditional methods and limit clearer clustering in other architectures. These findings underscore the potential of CNNs as scalable tools for botanical authentication-particularly for detecting adulteration and species complexities using seed digital morphometric characters-to safeguard public health and reinforce supply-chain safety.