Investigating advanced deep learning models for improving potato disease detection rates on imbalanced datasets

Potato diseases pose a significant threat to global food security, making accurate automated detection systems essential for sustainable agriculture. The study presents a rigorous comparative evaluation of four deep learning architectures, namely EfficientNet-B0, Vision Transformer (ViT-B/16), ResNet-50, and a U-Net encoder, for classifying seven potato diseases plus healthy specimens using a dataset of 2,529 images characterised by substantial class imbalance, with the two blight classes comprising 79% of samples.

Several critical methodological gaps were addressed in the experimental design. Dataset splitting was performed before augmentation to prevent data leakage, stratified five-fold cross-validation was used to produce robust performance estimates, and paired statistical testing was applied to validate claimed model superiority. The augmentation strategy deliberately excluded texture-destroying transformations such as Gaussian blur; ablation studies confirmed that including blur reduced accuracy by 4.65 percentage points by degrading lesion boundary detail and fungal surface markings essential for disease discrimination. EfficientNet-B0, ResNet-50, and ViTB/16 were initialised from ImageNet pre-trained weights, while U-Net was trained from scratch owing to the absence of standard pre-trained weights for that architecture.

EfficientNet-B0 achieved the highest overall performance, reaching 98.43% accuracy (±0.41%) with 92.37% macro-averaged precision. Despite requiring 16.2 times fewer parameters than ViT-B/16 (5.3 million versus 86.0 million), EfficientNet demonstrated substantially faster inference (12.3 ms versus 45.2 ms) and lower memory consumption (2.1 GB versus 7.8 GB). On the most challenging minority class, Blackleg, EfficientNet achieved 78% precision compared with ViT’s 67%, ResNet-50’s 55%, and U-Net’s 30%. ResNet-50 achieved 96.85% overall accuracy but struggled on underrepresented classes whilst U-Net encoder produced the weakest results (91.75% accuracy, 57.55% macro precision), with near-random performance on several minority classes, illustrating the critical role of transfer learning on small datasets. Statistical testing confirmed EfficientNet’s superiority over ResNet-50 and U-Net (p < 0.01), while its advantage over ViT approached but failed to reach significance at the corrected threshold (p =0.051).

In order to evaluate the robustness of these findings beyond controlled laboratory conditions, the four architectures were subsequently retrained and evaluated on a curated real-world dataset combining the PotatoCare benchmark with the Mendeley ”Novel Dataset of Potato Leaf Disease in Uncontrolled Environment”. EfficientNet-B0 maintained its position as the highest-performing architecture under these more demanding conditions, achieving 96.07% cross-validation accuracy and 91% test-set accuracy, confirming that the rankings established under laboratory conditions are preserved when variability in image quality and background content is introduced. The work establishes a rigorous algorithmic baseline and demonstrates that EfficientNet-B0 represents the optimal choice for resource-constrained agricultural deployment, balancing superior classification accuracy with exceptional computational efficiency.