Fast and Accurate Identification of Kiwifruit Diseases Using a Lightweight Convolutional Neural Network Architecture

Abstract

Kiwifruit (Actinidia chinensis Planch.) is highly valued for its nutritional benefits and unique flavor. However, diseases like bacterial canker and soft rot threaten its production, causing significant economic losses. Traditional disease identification methods, which rely on human expertise, are time-consuming and lack scalability. This study utilizes deep learning to enhance kiwifruit disease identification by evaluating eight advanced convolutional neural network (CNN) architectures on real-world field data. Among these, ShuffleNet_V2_x0_5 proved to be the most effective model. By incorporating advanced optimization strategies, including the AdamW optimizer and OneCycleLR scheduler, the model demonstrated rapid convergence and robust performance, achieving over 99% accuracy within five epochs, with only 1.37M parameters and 0.04G FLOPs. The lightweight architecture and computational efficiency make it particularly suitable for resource-limited settings, including mobile and embedded platforms. These findings underscore the utility of ShuffleNet_V2_x0_5 in supporting scalable and efficient disease management within precision kiwifruit agriculture. Our code and models are available at https://github.com/zhanglab-wbgcas/kiwifruit-diseases-classifier.

Keywords

• Kiwifruit diseases
• plant disease identification
• deep learning in agriculture
• convolutional neural networks
• lightweight IoT applications