An Ensemble of EfficientNetV2B3 and EfficientNetB4 for Crop Disease Detection Using the PlantVillage Dataset

Abstract

Crop diseases continue to threaten global food security, causing annual yield losses of 20–40% worldwide. Farmers in developing nations often lack timely access to ex-pert diagnosis, leading to delayed interventions and reduced harvests. This study presents a deep learning-based solution that automates crop disease identification using leaf images. We trained and evaluated two state-of-the-art convolutional neural networks—EfficientNetV2B3 and EfficientNetB4—on the publicly available PlantVillage dataset, which contains 54,303 images spanning 38 disease categories across 14 crop species. To improve classification robustness, we developed an ensemble model that combines the predictions of both architectures via weighted averaging. EfficientNetV2B3 achieved 98.0% accuracy individually, while EfficientNetB4 reached 94.0%. The proposed ensemble model attained an accuracy of 98.5% and an area under the curve (AUC) of 0.98, outperforming both parent models and several established baselines, including VGG16, ResNet50, InceptionV3, MobileNetV2, and DenseNet121. Beyond model development, we deployed the ensemble inside a Flask-based web application with user authentication, confidence scoring, and a searchable disease knowledge base. This end-to-end system bridges the gap between research and practice, offering farmers an accessible tool for rapid, reliable disease diagnosis.

Introduction

Agriculture suffers major losses due to plant diseases, especially in regions like South Asia, where smallholder farmers often lack timely access to expert diagnosis. Traditional methods are slow, costly, and ineffective for early detection. To address this, the study uses computer vision and convolutional neural networks (CNNs) to enable automatic disease identification from leaf images, even using smartphone cameras.

The proposed work improves upon existing research by focusing not only on model accuracy but also on real-world deployment. It highlights three main gaps in previous studies: reliance on single models, lack of practical applications, and poor reproducibility. To solve these, the authors build a weighted ensemble of two EfficientNet models (EfficientNetV2B3 and EfficientNetB4), trained on the PlantVillage dataset with over 54,000 images across 38 classes.

The system achieves 98.5% accuracy and an AUC of 0.98, outperforming individual models. It is also deployed as a Flask-based web application that allows users to upload leaf images, receive disease predictions with confidence scores, and view treatment recommendations.

Conclusion

A. Summary of Findings This paper presented a complete pipeline for crop disease detection, from data preprocessing and model training to web-based deployment. Using the PlantVillage dataset of 54,303 images spanning 38 disease classes, we trained Efficient-NetV2B3 and EfficientNetB4 under a two-phase transfer learn-ing schedule. EfficientNetV2B3 achieved 98.0% test accu-racy, outperforming EfficientNetB4 (94.0%) and several other baselines including VGG16, ResNet50, and InceptionV3. By combining both models into a weighted ensemble (Pensemble = 0.6 • PV2B3 +0.4 • PB4), we further improved accuracy to 98.5% with an AUC of 0.98. Crucially, we did not stop at reporting numbers. We built and deployed a fully functional Flask web application with user authentication, confidence scoring, and an integrated dis-ease knowledge base. This deployment transforms a research model into a practical tool that farmers can actually use. B. Limitations Several limitations merit acknowledgment. First, the PlantVillage dataset, while valuable, contains images captured under controlled conditions with uniform backgrounds. Real-world field images include varying illumination, occlusion, and background clutter. Performance on such images is likely lower. Second, the application currently requires an internet connection because inference runs on a remote server. Offline mobile deployment would be more practical for rural farmers. Third, the knowledge base, while useful, is static and does not incorporate regional treatment guidelines or real-time updates. C. Future Work We plan to address these limitations in the following ways: 1) Field Dataset Collection: We are collaborating with local agricultural universities to collect a realistic dataset of field-captured leaf images. This dataset will include variations in lighting, background, and disease severity. 2) Mobile Application: We intend to convert the Flask back-end into a lightweight TensorFlow Lite model that can run entirely on a smartphone without an internet connection. An accompanying Android app would make the system accessible to farmers in remote areas. 3) Explainability Module: Farmers and agricultural experts need to trust the model’s predictions. We plan to integrate Grad-CAM or similar saliency mapping techniques that high-light which regions of the leaf influenced the model’s decision. Visual explanations can build confidence and help identify model failures. 4) Multi-Lingual Support: The current interface is in En-glish. Future versions will support Hindi and other regional languages to improve accessibility for Indian farmers. 5) Continuous Learning: The ensemble model is static. We plan to implement a feedback loop where user corrections are periodically used to fine-tune the model, adapting to new disease variants or emerging outbreaks.

References

[1] U¨ . Atila, M. Uc¸ar, K. Akyol, and E. Uc¸ar, “Plant leaf disease classifi-cation using EfficientNet deep learning model,” Ecological Informatics, vol. 61, p. 101182, Mar. 2021. [2] S. Sadiq, K. R. Malik, W. Ali, and M. M. Iqbal, “Deep learning-based disease identification and classification in potato leaves using EfficientNet B4,” Journal of Computational and Biomedical Informatics, vol. 5, no. 1, pp. 45–58, 2023. [3] J. Padhi, L. Korada, A. Dash, and S. Mishra, “Paddy leaf disease classification using EfficientNet B4 with transfer learning,” IEEE Access, vol. 12, pp. 24500–24512, 2024. [4] M. Aggarwal, V. Khullar, N. Goyal, A. Singh, and R. K. Bansal, “Pre-trained deep neural network-based features selection supported machine learning for rice leaf disease classification,” Agriculture, vol. 13, no. 5, [5] p. 936, 2023. [6] F. Mashroor, I. Ishrak, S. Alvee, and M. R. Islam, “Rice paddy disease detection and segmentation using convolutional neural networks,” in Proc. IEEE TENCON, 2021, pp. 1–6. [7] B. Paul, R. Rozario, A. Roy, and M. M. Rahman, “Deep learning techniques for automatic detection and classification of rice diseases,” in Proc. International Conference on ICT for Competitive Strategies, 2022, [8] pp. 120–128. [9] S. Saleki and J. Tahmoresnezhad, “Agry: A framework for plant dis-eases classification via pretrained EfficientNet,” Multimedia Tools and Applications, vol. 83, pp. 11245–11270, 2024. [10] F. Adnan, M. J. Awan, S. O. Abdellatif, and H. Alshahrani, “EfficientNetB3-AADL for multi-class plant disease classification,” IEEE Access, vol. 11, pp. 78900–78915, 2023. [11] M. Ahdi, A. Kunaefi, B. Nugroho, and T. Susanto, “CNN EfficientNet-B0 for paddy disease classification with limited data,” in Proc. Inter-national Conference on Information Technology and Smart Computing, 2023, pp. 88–94. [12] X. Bi and H. Wang, “Double-branch CNN-based rice leaf disease recog-nition with attention mechanism,” Journal of Agricultural Engineering, vol. 55, no. 1, pp. 112–125, 2023.

Copyright

Copyright © 2026 Gautam Tyagi, Anshul Tyagi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.