Year: 2025
Type: article
Abstract: In the agricultural sector, pest detection is vital, and given the susceptibility of human analysis to errors, deep learning solutions such as Convolutional Neural Networks (CNNs) provide a promising alternative. Classifying insect pests is challenging due to the high variability among species across different regions and their various life stages. In this study, we evaluate several deep learning models and training strategies for automatic pest image classification. We analyze four CNN architectures—AlexNet, ResNet-50, EfficientNet, and Vision Transformer (ViT). Following a hyperparameter optimization step, the models were fine-tuned, and we examined the impact of four data augmentation strategies on classification performance using the Agricultural Pests dataset. ViT demonstrated superior performance, achieving an accuracy of 0.9574 without data augmentation. Although ViT did not benefit from data augmentation, these techniques proved essential for enhancing the performance of AlexNet, ResNet-50, and EfficientNet. Our findings underscore the potential of deep learning methods for pest classification, offering valuable tools to help professionals maintain crop quality and value.
Institution: Universidade Federal de Viçosa
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: diamond
Award: G29316538
OpenAlex ID: W4407772887