A Feasibility Study on Synthetic RGB-NIR Image Generation for Oil Palm Fresh Fruit Bunch Grading

Accurate ripeness of grading oil palm fruit bunches (FFBs) is essential for optimizing oil quality and harvesting decisions. While near-infrared (NIR) imaging provides useful spectral cues for ripeness assessment, its adoption in field conditions is limited by sensor cost and system complexity. This study presents a low-cost alternative by generating synthetic NIR images from RGB inputs using a U-Net-based image translation model and integrating the generated NIR with RGB channels for ripeness classification. Five deep learning models, including a custom CNN, ResNet-50, EfficientNet-B0, DenseNet-201 and MobileNetV3, were evaluated under RGB-only and RGB + synthetic NIR configurations using identical training protocols. Experimental results demonstrate consistent performance improvements when synthetic NIR was incorporated. EfficientNet-B0 achieved the highest overall accuracy of 90.3%, while MobileNetV3 obtained the highest macro-averaged F1-score of 85.4%, indicating strong and balanced classification across ripeness classes. Confusion matrix analysis further revealed complementary strengths between the models, where EfficientNet-B0 showed stronger robustness in late-stage maturity detection, and MobileNetV3 provided improved discrimination of early-stage ripeness. The results demonstrate that synthetic NIR augmentation enhances classification performance and training stability without requiring specialized imaging hardware.