A Green Engineering Approach to Lightweight CNNs: Moringa Leaf Classification Using MobileNetV3-Small

Abstract

Agriculture plays a crucial role in maintaining global food security, particularly in developing countries, where Moringa oleifera is a high-value crop. However, leaf diseases can reduce productivity, while manual identification remains challenging due to limited expertise. This study aims to develop an accurate and energy-efficient image-based moringa leaf disease classification system. The method used was the MobileNetV3-Small architecture with three variants (small-050, small-075, and small-100) applied to the MoringaLeafNet dataset, accompanied by preprocessing and data imbalance handling. Experimental results show that increasing model complexity is directly proportional to performance improvement. The MobileNetV3-Small-075 model achieved 89.36% accuracy, 89.84% precision, 90.14% recall, and 89.83% F1-score with an average energy consumption of 8.84 mJ per image, while MobileNetV3-Small-100 achieved 93.62% accuracy, 94.50% precision, 93.91% recall, and 94.17% F1-score with an energy consumption of 10.28 mJ. These findings indicate a trade-off between classification performance and energy consumption. Thus, MobileNetV3-Small-075 is the most optimal configuration because it can provide the best balance between performance and efficiency for implementation on devices with limited resources.