Comparative Analysis of ResNet50, ConvNeXt, and Vision Transformer Architectures for Automated Plant Disease Recognition

This paper presents a controlled comparative study of three deep learning architectures, ResNet50, ConvNeXt-Tiny, and Vision Transformer ViT Base 16, for automated plant disease classification using the PlantVillage dataset, with the goal of identifying models that balance predictive accuracy and computational efficiency for precision agriculture applications. The authors fine tune all models from ImageNet pretrained weights under identical training conditions, including the same hyperparameters, preprocessing pipeline, and evaluation metrics, ensuring a fair benchmark across 38 plant disease classes derived from a stratified subset of the dataset. Performance is evaluated using accuracy, weighted and macro F1 scores, weighted precision, parameter count, inference latency, and throughput, providing both classification quality and efficiency perspectives. The results show that ConvNeXt-Tiny achieves the highest classification performance with 98.63 percent accuracy and strong consistency across classes, while ViT achieves slightly lower accuracy at 98.00 percent but demonstrates the fastest inference speed due to efficient parallelization on GPU hardware, and ResNet50, although less accurate at 96.00 percent, remains the most lightweight and computationally economical option. Analysis of challenging disease categories such as visually similar tomato diseases reveals that newer architectures like ConvNeXt and ViT handle fine grained distinctions better than ResNet50. Efficiency analysis further highlights trade offs, where ConvNeXt incurs significantly higher latency despite moderate parameter size, while ViT offers superior throughput despite having the largest parameter count. Training curves and confusion matrices shown in pages 4 to 7 indicate stable convergence and generally accurate class level predictions across models, with minimal misclassification outside visually ambiguous categories. The paper concludes that ConvNeXt is best suited for accuracy critical applications, ViT for high throughput systems, and ResNet50 for resource constrained deployments, and suggests future work on scaling to the full dataset, improving robustness to real world conditions, and applying model compression techniques for deployment feasibility.