Z. N. Nemer, A. Mahdi Jiyad, M. Ali Abdulsamad, E. J. Harfash
Purpose: This research presents a framework based on convolutional neural networks (CNNs) for automatically classifying soybean seeds into five categories: intact, spotted, immature, broken, and damaged. Design/Methodology/Approach: The framework includes nine pre-trained CNN models, including DenseNet(169,121,201), InceptionV3, MobileNetV2, ResNet50V2, VGG (16,19), and Xception, which were evaluated and compared in terms of accuracy, recall, precision, and F1 parameter using the dataset. Data augmentation was used to improve the model's performance and prevent overlearning. An ensemble learning approach using majority voting (MV) was applied to improve the model's generalisation. Research Limitation: The controlled imaging setting limits its generalisability to the real world. But the presented method is definitely one of the best options for evaluating soybean seed quality. Findings: The DenseNet201 and DenseNet169 models showed superiority over the DenseNet201 and DenseNet169 models due to their superior ability to extract features and share parameters efficiently. These results confirm the effectiveness of deep modelling and group model integration techniques in improving classification reliability and reducing over-allocation. It outperforms all individual models, achieving 98.22% accuracy and 97.68% test accuracy. Practical Implication: Adopting artificial intelligence (AI) technologies offers an effective solution to enhance agricultural productivity and improve seed sorting and quality assessment. Social Implication: Reduced computational demand means lower energy consumption per classification event. At the scale of a full harvest season, this represents a meaningful reduction in the digital carbon footprint of precision agriculture. Originality / Value: These models, which feature a densely interconnected structure, not only facilitate more efficient feature reuse but also help prevent diminishing gradients.