Answers are grounded in this paper and cite their sources.
Convolutional Neural Networks (CNNs) have recently shown strong potential for automating early plant disease diagnosis, an important need in modern agriculture where plant disease-driven yield loss is threatening global food security. In this study, we fine-tuned and evaluated deep learning models using RGB images for multi-class classification of soybean leaf diseases. The Auburn Soybean Disease Image Dataset (ASDID), which contains 9,648 field background images across eight diseases, including bacterial blight, frogeye leaf spot, soybean rust, and potassium deficiency, was used for this purpose. We fine-tuned three CNN architectures: Residual Network with 50 layers (ResNet50), Residual Network with 18 layers (ResNet18), and Mobile Network Version 2 (MobileNetV2), using transfer learning from ImageNet and adapting them with customized classification heads. To improve generalization under realistic field variation and to address class imbalance, we apply targeted preprocessing and training strategies, including image augmentation (rotation, flips, zoom, brightness/contrast variation, mild blur), class-weighted loss, weighted sampling, label smoothing, and learning rate scheduling. All three models achieved strong performance on held-out test data, with mean accuracies between 96.9% and 97.7% across three random seeds and weighted F1 scores between 0.969 and 0.977 across eight classes. The tuned ResNet50 variant delivered the highest accuracy of the three models (97.70), and the repeated runs verified the results using McNemar’s exact test and five-fold cross-validation. These results indicate that accurate soybean disease classification is achievable from standard RGB field imagery without manual feature engineering or segmentation, within the single dataset evaluated here. Although ResNet18 and MobileNetV2 achieved slightly lower accuracy, MobileNetV2 required 8.5 MB of storage and 36.5 MB of peak inference memory against of peak inference memory and ResNet18 47.2 MB and 76.1 MB, against 100.2 MB and 130.8 MB for ResNet50, which supports the candidacy of the lighter architectures for resource-constrained platforms such as mobile devices, drones, and edge cameras, though on-device performance was not measured.
No takes yet. Share an insight, caveat, or question.
Sengar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: