<![CDATA[The classification of Fresh Fruit Bunch (FFB) ripeness levels in oil palm is a crucial aspect of agricultural automation for maintaining production quality. This study evaluates the performance of three deep learning architectures through a comparative analysis of EfficientNet-B0, ConvNeXt-Tiny, and MobileNetV2 for distinguishing different fruit ripeness classes. The experimental results show that MobileNetV2 achieved the highest accuracy of 97%, followed by ConvNeXt-Tiny at 94%, whereas EfficientNet-B0 achieved only 32%, primarily due to systematic prediction bias on the complex dataset. The confusion matrix analysis indicates that although MobileNetV2 and ConvNeXt-Tiny performed effectively, visual ambiguity during transitional ripeness stages remains a challenge for purely computer vision-based approaches. These findings highlight the importance of selecting an appropriate backbone architecture and the potential need for hybrid approaches to improve system reliability. As a direction for future research, the application of Neuro-Symbolic AI is proposed to combine deep learning-based feature extraction with knowledge-based reasoning to resolve ambiguous predictions, together with model optimization to achieve greater computational efficiency on edge devices. This study provides a strategic contribution to the development of transparent and accurate automated sorting systems for the oil palm plantation sector.]]>
Copyrights © 2026
