Answers are grounded in this paper and cite their sources.
Deep learning models have achieved near-perfect performance on benchmark plant disease datasets; however, their reliability under deployment-relevant agricultural perturbations remains insufficiently understood. Most existing studies emphasize classification accuracy while overlooking robustness, vulnerability, and deployment-related performance characteristics. This study presents a deployment-oriented evaluation framework for assessing the operational reliability of plant disease diagnosis models beyond conventional benchmark accuracy. Four representative architectures, including ResNet50, EfficientNet-B0, ViT-B16, and CLIP ViT-B/32, were evaluated using the PlantVillage dataset comprising 54,305 images across 38 classes. Model performance was assessed through agriculturally relevant perturbations, including Gaussian blur, illumination variation, shadow effects, occlusion, and JPEG compression. Robustness, perturbation sensitivity, disease vulnerability, crop vulnerability, deployment readiness, and computational characteristics were analysed to quantify deployment-related risks. All architectures achieved high benchmark performance accuracy (97.86%-99.89%); however, substantial differences emerged under perturbation conditions. EfficientNet-B0 achieved the highest baseline accuracy (99.89%) and mean robustness (98.75%), whereas ViT-B16 exhibited the strongest worst-case performance (95.45%) and the lowest perturbation sensitivity (PSS = 1.01). EfficientNet-B0 also provided the most favourable balance between predictive performance, robustness, and computational efficiency. In contrast, CLIP ViT-B/32 showed pronounced vulnerability to environmental variability, with worst-case accuracy decreasing to 54.98%. Illumination variation was identified as the most influential perturbation (mean PSS = 8.44%), while Corn (maize) Gray Leaf Spot emerged as the most vulnerable disease category (DVS = 0.0801). Disease- and crop-level analyses further revealed that classification difficulty was concentrated within a limited subset of categories rather than being uniformly distributed across the dataset. The proposed framework demonstrates that deployment-oriented evaluation can lead to model selection decisions that differ substantially from those based solely on benchmark accuracy. Overall, EfficientNet-B0 and ViT-B16 emerged as the most deployment-oriented architectures representing a trade-off between computational efficiency and robustness stability under the evaluated perturbation framework. The findings highlight the importance of integrating robustness, vulnerability, and failure-pattern assessment into agricultural artificial intelligence evaluation and provide a practical framework for developing trustworthy plant disease diagnosis systems.
No takes yet. Share an insight, caveat, or question.
Malkani et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: