A Comparative Study of ConvNeXt-Tiny and Swin-Tiny for Multi-Label Plant Disease Classification on Real-World Leaf Images

Abstract

The current deep learning systems which detect plant diseases in farms require substantial financial resources because they operate with lab-based data and single-label classification methods which fail to function in actual field conditions where multiple diseases can affect one leaf. This study uses the Plant Pathology 2021 dataset, which has field images, to test the ConvNeXt-Tiny and Swin-Tiny architectures for classifying apple diseases. Both models started with pretrained weights from ImageNet 22k. The model used class- weighted loss to address the significant class imbalance problem. The evaluation process measured how long it took for computers to generate predictions and also checked the accuracy of these predictions. The results indicate that both models produce successful results. Swin-Tiny is better at remembering things and making guesses, while ConvNeXt- Tiny is better at both. The two models successfully detect diseases which present different symptoms but they fail to identify multiple diseases that exist simultaneously. The results show that the architecture you choose should depend on what you need to do with it. The two models offer different benefits for application development because ConvNeXt achieves better accuracy through its quick training time but Swin provides better real-time decision performance and memory efficiency. This study provides real-world evidence to help make architectural decisions in agricultural AI by showing that multi-label frameworks better show how diseases happen together in real life.