A Comparative Performance Analysis of Deep Learning Model for Pest Identification in Smart Agriculture

The adoption of smart agriculture has been rapid and
with it has come a need for an accurate and automated way of
identifying pests to reduce crop losses and promote sustainable farming. However, existing manual and semi-automated approaches are often limited in accuracy, have inconsistent performance in different environmental conditions, and have limited generalizations to different pest species. This study aimed at performing a holistic comparative test of top deep learning(DL) models such as Convolutional Neural Network (CNN), EfficientNet-B0, MobileNetV3, ResNet50, DenseNet121, and Vision Transformer (ViT) to determine the best architecture for pest identification in agricultural environments. Using the
IP102 benchmark dataset which has 75,222 images of 102 pest
categories which are preprocessed by data augmentation,
normalization and resizing before training models under
standardized settings. Quantitative results show that ViT gave the best top-1 accuracy of 91.3% followed by EfficientNet-B0 of 89.7% while MobileNetV3 offered the best computational efficiency of 24 ms inference time. These findings highlight the fact that transformer-based models deliver better classification accuracy, while lightweight CNN is the best choice when it comes to edge deployment. Overall, the study offers useful information to researchers and agritech developers for choosing model
architectures that compromise accuracy, speed, and resource
constraints.