Lightweight Lepidopteran Pest Classification Model Using Knowledge Distillation

Lepidopteran pests cause significant damage to plant foliage, stems, and roots, leading to substantial agricultural losses. Effective pest management requires timely and appropriate pesticide application. However, accurately classifying Lepidopteran pest species remains a challenge, even for experienced farmers. Convolutional neural networks (CNNs) have demonstrated outstanding performance in classification tasks, yet they typically demand substantial computational resources. In resource-constrained environments such as smartphones, drones, and IoT devices, deploying conventional CNN models is impractical. To address this limitation, we propose a lightweight Lepidopteran pest classification model optimized through knowledge distillation from a high-capacity teacher model. Experimental results demonstrate that the distilled MobileNetV3 model achieves the highest classification accuracy while maintaining the lowest inference latency and the smallest parameter size.