Accurate and early detection of plant leaf diseases is crucial for ensuring crop health and improving agricultural productivity. This work proposes a hybrid deep learning model that combines ResNet18, Inception blocks, and fully connected Capsule layers to classify leaf images of apple, grape, and corn plants into healthy or diseased categories. ResNet18 is used as the backbone for deep feature extraction, while Inception modules enhance the network’s ability to capture multi-scale patterns. Capsule layers are employed at the final stage to retain spatial relationships and pose information, improving the model\'s ability to recognize complex disease features. The model is trained and evaluated using images from the PlantVillage dataset, with separate configurations for each crop. The proposed model achieved validation accuracies of 99.84% for apple, 100% for grape, and 97.27% for corn. Performance is further assessed using precision, recall, and F1-score, and compared against a baseline ResNet18 model. The results demonstrate that the proposed architecture significantly improves classification accuracy and feature understanding, making it a strong candidate for real-world agricultural disease monitoring systems.
Introduction
Agriculture is crucial for the global economy, especially in countries like India. However, crop productivity is often threatened by plant diseases, particularly leaf diseases, which are common and visually identifiable. Traditional disease detection methods (manual inspection by experts) are not scalable or fast enough. This has led to the development of AI-based automated detection systems, especially using deep learning (DL) models like Convolutional Neural Networks (CNNs).
2. Motivation & Limitations of Traditional CNNs
CNNs (e.g., ResNet, VGG, GoogLeNet) are widely used for image classification but face limitations in:
Capturing spatial hierarchies
Generalizing across visually similar diseases
Preserving orientation and feature relationships
To address these issues, this work proposes a hybrid deep learning architecture that integrates:
ResNet18 – deep feature extraction via residual connections.
Inception blocks – for multi-scale feature extraction.
Capsule Networks – to retain spatial relationships and orientation of features.
3. Literature Review Highlights
Hosny et al.: Hybrid CNN + LBP features; 98–99% accuracy on Apple, Grape, Tomato.
Madhurya & Jubilson: Advanced pipeline using CLAHE, YOLOv7, and ShuffleNetV2 with optimization; 99.69% accuracy.
Moupojou et al.: FieldPlant dataset from real-world farms; MobileNet outperformed others.
Balafas et al.: Review of 79 studies; ResNet50 and YOLOv5 were top performers.
Benfenati et al.: Unsupervised anomaly detection via autoencoders; 90.35% accuracy.
Hama et al.: Improved ResNet50 with selective layer freezing; 99% accuracy.
Babu et al.: DenseNet-121 and ResNet101 best for rice disease detection.
Uses a fully connected capsule approach (simpler than dynamic routing).
D. Classification Layers
Final fully connected layer → Softmax → Output.
Outputs the probability distribution across disease classes.
5. Architecture Summary
Input leaf image.
Processed by ResNet18 for deep feature extraction.
Features fed to Inception block for multi-scale analysis.
Resulting features go into Capsule layer to preserve spatial structure.
Final classification based on vector outputs via Softmax.
6. Experimentation: Datasets Used
Used the PlantVillage dataset for training and testing:
Apple: 3,171 images across 4 classes (Scab, Black Rot, Rust, Healthy).
Grape and Corn datasets (not fully detailed in excerpt but implied).
Images are standardized (e.g., 256x256 pixels, plain backgrounds).
Diseases show distinct visual symptoms for classification training.
? Key Advantages of the Proposed Model
Generalizable across multiple crops and diseases.
Modular and adaptable (easy to expand for new crops/diseases).
Works well in real-world agricultural scenarios.
Can be deployed in mobile apps, drones, or edge devices.
Conclusion
The proposed ResNet18 model enhanced with Inception blocks and a fully connected Capsule Network achieved good results for plant leaf disease classification in apple, grape, and corn datasets. By integrating residual learning, multi-scale feature extraction, and spatially aware capsule encoding, the model effectively captured disease-specific patterns and obtained higher accuracy than the baseline ResNet18. Experimental evaluation showed that the proposed model achieved 99.84% accuracy for apple, 100% for grape, and 97.27% for corn, outperforming the standard ResNet18 across all metrics. In addition, the precision, recall, and F1-scores are consistently above 93%, confirming the robustness of the approach.
These outcomes highlight the effectiveness of combining deep learning architectures for reliable agricultural disease detection, supporting timely intervention to safeguard crop yield and minimize economic losses. For future work, the model can be extended to larger and more diverse datasets collected under real field conditions, with further improvements such as attention-based feature refinement, optimized capsule routing, and lightweight deployment on mobile or edge devices to enable real-time disease detection for precision agriculture.
References
[1] K. M. Hosny, W. M. El-Hady, F. M. Samy, E. Vrochidou, and G. A. Papakostas, “Multi-Class Classification of Plant Leaf Diseases Using Feature Fusion of Deep Convolutional Neural Network and Local Binary Pattern,” IEEE Access, vol. 11, pp. 62307–62317, 2023, doi: 10.1109/ACCESS.2023.3286730. [Online]. Available:
https://doi.org/10.1109/ACCESS.2023.3286730.
[2] C. Madhurya and E. A. Jubilson, \"YR2S: Efficient Deep Learning Technique for Detecting and Classifying Plant Leaf Diseases,\" IEEE Access, vol. 12, pp. 3789–3804, 2024, doi: 10.1109/ACCESS.2023.3343450.
[3] E. Moupojou, A. Tagne, F. Retraint, A. Tadonkemwa, D. Wilfried, H. Talamo, and M. Nkenlifack, \"FieldPlant: A Dataset of Field Plant Images for Plant Disease Detection and Classification With Deep Learning,\" IEEE Access, vol. 11, pp. 35398–35410, 2023, doi: 10.1109/ACCESS.2023.3263042 Paper Link: https://doi.org/10.1109/ACCESS.2023.3263042.
[4] V. Balafas, E. Karantoumanis, M. Louta, and N. Ploskas, “Machine Learning and Deep Learning for Plant Disease Classification and Detection,” IEEE Access, vol. 11, pp. 114352–114376, Oct. 2023, doi: 10.1109/ACCESS.2023.3324722. [Online]. Available: https://doi.org/10.1109/ACCESS.2023.3324722.
[5] A. Benfenati, P. Causin, R. Oberti, and G. Stefanello, “Unsupervised deep learning techniques for automatic detection of plant diseases: reducing the need of manual labelling of plant images,” Journal of Mathematics in Industry, vol. 13, no. 5, 2023. doi: 10.1186/s13362-023-00133-6. [Online]. Available: https://doi.org/10.1186/s13362-023-00133-6.
[6] H. M. Hama, T. Sh. Abdulsamad, and S. M. Omer, “Houseplant leaf classification system based on deep learning algorithms,” Journal of Electrical Systems and Information Technology, vol. 11, no. 18, pp. 1–15, Feb. 2024, doi: 10.1186/s43067-024-00141-5. [Online]. Available: https://doi.org/10.1186/s43067-024-00141-5.
[7] S. Babu, M. Maravarman, and R. Pitchai, “Detection of Rice Plant Disease Using Deep Learning Techniques,” Journal of Mobile Multimedia, vol. 18, no. 3, pp. 757–770, Jan. 2022, doi: 10.13052/jmm1550-4646.18314. [Online]. Available: https://doi.org/10.13052/jmm1550-4646.18314.
[8] R. Maurya, L. Rajput, and S. Mahapatra, “RAI-Net: Tomato Plant Disease Classification Using Residual-Attention-Inception Network,” IEEE Access, vol. 13, pp. 64832–64840, Apr. 2025, doi: 10.1109/ACCESS.2025.3559804. [Online]. Available: https://doi.org/10.1109/ACCESS.2025.3559804.
[9] S. K. Noon, M. Amjad, M. A. Qureshi, and A. Mannan, \"Handling Severity Levels of Multiple Co-Occurring Cotton Plant Diseases Using Improved YOLOX Model,\" IEEE Access, vol. 10, pp. 134811–134825, Dec. 2022, doi: 10.1109/ACCESS.2022.3232751.
[10] P. Theerthagiri, A. U. Ruby, J. G. C. Chandran, T. H. Sardar, and A. B. M. Shafeeq, \"Deep SqueezeNet learning model for diagnosis and prediction of maize leaf diseases,\" Journal of Big Data, vol. 11, no. 112, 2024. doi: 10.1186/s40537-024-00972-z.
[11] S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” in Advances in Neural Information Processing Systems (NeurIPS), pp. 3856–3866, 2017. [Online]. Available: https://arxiv.org/abs/1710.09829.
[12] T. V. Le, L. T. Ngo, and T. D. Pham, “Detection of Co-Occurring Tomato Leaf Diseases Based on Improved YOLOX,” Sensors, vol. 23, no. 15, p. 6841, Aug. 2023, doi: 10.3390/s23156841. [Online]. Available: https://doi.org/10.3390/s23156841.
[13] Y. Liu, Q. Wang, Y. Liu, and M. Ma, “A plant leaf disease image classification method integrating capsule network and residual network,” Computational Intelligence and Neuroscience, vol. 2022, Art. no. 1759507, pp. 1–12, 2022, doi: 10.1155/2022/1759507.
[14] K. P. A. Rani and S. Gowrishankar, “Pathogen-based classification of plant diseases: A deep transfer learning approach for intelligent support systems,” IEEE Access, vol. 11, pp. 64476–64493, Jun. 2023, doi: 10.1109/ACCESS.2023.3284680.
[15] J. Chaki and D. Ghosh, “Deep learning in leaf disease detection 2014–2024: A visualization-based bibliometric analysis,” IEEE Access, vol. 12, pp. 95291–95309, Jul. 2024, doi: 10.1109/ACCESS.2024.3425897.