PlantCLR: contrastive self-supervised pretraining for generalizable plant disease detection.

Abstract

Deep learning has improved automated plant disease detection by increasing recognition accuracy and robustness compared with traditional vision-based methods. Self-supervised learning (SSL) further reduces dependence on manual labels, but its transferability across heterogeneous agricultural datasets remains insufficiently characterized. Here, we evaluate a contrastive SSL pretraining and fine-tuning pipeline, termed PlantCLR, for plant disease classification under cross-dataset transfer with target-domain fine-tuning. PlantCLR combines SimCLR-style contrastive pretraining with a lightweight convolutional classifier to balance representation quality and deployment efficiency. Experiments on PlantVillage and Cassava Leaf Disease show strong performance, achieving 99.10% accuracy and 99.04% F1-score on PlantVillage, and 96.83% accuracy and 96.70% F1-score on Cassava. Feature embedding visualization using t-SNE and explanation maps using Grad-CAM indicate improved class separability and attention to disease-relevant regions. These results suggest that contrastive SSL can improve representation transfer while maintaining computational efficiency, supporting scalable plant disease diagnostics in practical agricultural settings. Code is available at GitHub.