A Robust Dual-Backbone CNN Ensemble for Maize and Potato Leaf Disease Diagnosis Using Logit-Level Fusion and Calibration
Keywords:
Plant disease classificationx, dual-backbone ensemble, logit-level fusion, temperature scaling, bootstrapped confidence intervalsAbstract
Plant leaf diseases remain a major threat to global food security, particularly for staple crops such as maize and potato. Traditional diagnosis is slow, labour-intensive and subject to human variability, making it unsuitable for large-scale field monitoring. Although recent CNN advances have improved automated disease recognition, most studies rely on laboratory-style datasets with uniform lighting and clean backgrounds, limiting real-world applicability. This work proposes a dual-backbone deep-learning framework for leaf disease diagnosis, trained on a heterogeneous dataset of 7,005 maize and potato leaf images from PlantVillage, Kaggle, and a dew-stress field-photography subset (51 images). A pHash deduplication step removed six duplicates, yielding a final training set of 5,091 images (validation: 949; test: 959). Two CNN backbones-EfficientNet-B0 (~4.0 M parameters, 0.39 GFLOPs) and ConvNeXt-Tiny (~27.8 M parameters, 4.48 GFLOPs) are integrated via logit-level weighted fusion with alpha=0.05 selected by grid search. Experiments across three random seeds (42, 123, 7) demonstrate mean test accuracy of 99.24% +/- 0.13% and macro-F1 of 98.66% +/- 0.21%. Calibration yields ECE=0.0091 before and ECE=0.0070 after temperature scaling (T=1.356), with bootstrapped 95% CIs confirming robustness including the rare Healthy Potato class ([90.6%, 100.0%]). Mixup augmentation yields the highest per-class F1, and AUC reaches at least 0.9990 for all classes (macro-average 0.9998). All results are reported on an internal holdout split; no external validation on datasets such as PlantDoc has been conducted, and cross-domain generalization remains a key open research question.
References
[1] E. C. Too, L. Yujian, S. Njuki, and L. Yingchun, ”A comparative study of fine-tuning deep learning models for plant disease identification,” Computers and Electronics in Agriculture, vol. 161, pp. 272–279, 2019. doi: 10.1016/j.compag.2018.03.032.
[2] S. Albahli and M. Masood, ”Efficient attention-based CNN network (EANet) for multi-class maize crop disease classification,” Frontiers in Plant Science, vol. 13, 2022. doi: 10.3389/fpls.2022.1003152.
[3] M. A. Haque et al., ”Deep learning-based approach for identification of diseases of maize crop,” Scientific Reports, vol. 12, no. 1, p. 6334, 2022. doi: 10.1038/s41598-022-10140-z.
[4] L. Liu et al., ”A multi-scale feature fusion neural network for multi-class disease classification on the maize leaf images,” Heliyon, vol. 10, no. 7, p. e28264, 2024. doi: 10.1016/j.heliyon.2024.e28264.
[5] A. Kolesnikov et al., ”Big Transfer (BiT): General Visual Representation Learning,” arXiv preprint, 2020. (https://arxiv.org/abs/1912.11370).
[6] A. Dosovitskiy et al., ”An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv preprint, 2021. (https://arxiv.org/abs/2010.11929).
[7] M. Tan and Q. V. Le, ”EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” arXiv preprint, 2020. (https://arxiv.org/abs/1905.11946).
[8] Z. Liu et al., ”A ConvNet for the 2020s,” arXiv preprint, 2022. (https://arxiv.org/abs/2201.03545).
[9] S. Rahman, A. Indrawati, M. Zen, B. Zealtiel, and S. S. Tanjung, ”Enhanced RegNetY-400MF for Fruit Fly Species Classification: Fine-Tuning Strategies and Data Balancing for Improved Accuracy,” Jurnal RESTI (Rekayasa Sistem Dan Teknologi Informasi), vol. 9, no. 6, pp. 1463–1473, 2026. doi: 10.29207/resti.v9i6.6973.
[10] I. Loshchilov and F. Hutter, ”Decoupled Weight Decay Regularization,” arXiv preprint, 2019. (https://arxiv.org/abs/1711.05101).
[11] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, ”mixup: Beyond Empirical Risk Minimization,” arXiv preprint, 2018. (https://arxiv.org/abs/1710.09412).
[12] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Doll´ar, ”Focal Loss for Dense Object Detection,” arXiv preprint, 2018. (https://arxiv.org/abs/1708.02002).
[13] I. Loshchilov and F. Hutter, ”SGDR: Stochastic Gradient Descent with Warm Restarts,” arXiv preprint, 2017. (https://arxiv.org/abs/1608.03983).
[14] R. Müller, S. Kornblith, and G. Hinton, ”When Does Label Smoothing Help?” arXiv preprint, 2020. (https://arxiv.org/abs/1906.02629).
[15] E. D. Cubuk et al., ”AutoAugment: Learning Augmentation Policies from Data,” arXiv preprint, 2019. (https://arxiv.org/abs/1805.09501).
[16] J. Chen, J. Chen, D. Zhang, Y. Sun, and Y. Nanehkaran, ”Using deep transfer learning for image-based plant disease identification,” Computers and Electronics in Agriculture, vol. 173, p. 105393, 2020. doi: 10.1016/j.compag.2020.105393.
[17] K. P. Ferentinos, ”Deep learning models for plant disease detection and diagnosis,” Computers and Electronics in Agriculture, vol. 145, pp. 311–318, 2018. doi: 10.1016/j.compag.2018.01.009.
[18] J. G. A. Barbedo, ”Plant disease identification from individual lesions and spots using deep learning,” Biosystems Engineering, vol. 180, pp. 96–107, 2019. doi: 10.1016/j.biosystemseng.2019.02.002.
[19] M. Shoaib et al., ”An advanced deep learning models-based plant disease detection: A review of recent research,” Frontiers in Plant Science, vol. 14, 2023. doi: 10.3389/fpls.2023.1158933.
[20] I. Pacal et al., ”A systematic review of deep learning techniques for plant diseases,” Artificial Intelligence Review, vol. 57, no. 11, p. 304, 2024. doi: 10.1007/s10462-024-10944-7.
[21] M. Shoaib, A. Sadeghi-Niaraki, F. Ali, I. Hussain, and S. Khalid, ”Leveraging deep learning for plant disease and pest detection: a comprehensive review and future directions,” Frontiers in Plant Science, vol. 16, 2025. doi: 10.3389/fpls.2025.1538163.
[22] S. Yang et al., ”Classification and localization of maize leaf spot disease based on weakly supervised learning,” Frontiers in Plant Science, vol. 14, 2023. doi: 10.3389/fpls.2023.1128399.
[23] J. Hu, X. Jiang, J. Gao, and X. Yu, ”LFMNet: a lightweight model for identifying leaf diseases of maize with high similarity,” Frontiers in Plant Science, vol. 15, 2024. doi: 10.3389/fpls.2024.1368697.
[24] K. Ahadian et al., ”Maize disease classification using transfer learning and convolutional neural network with weighted loss,” Heliyon, vol. 10, no. 21, p. e39569, 2024. doi: 10.1016/j.heliyon.2024.e39569.
[25] S. Roy et al., ”SPUDNET5-R3: A Lightweight Hybrid and Explainable CNN Model for Potato Leaf Blight Detection,” 2025. doi: 10.21203/rs.3.rs-7377242/v1.
[26] T. Sultana and M. Reza, ”Identification of Potato Leaf Diseases Using Hybrid Convolution Neural Network with Support Vector Machine,” in Communications in Computer and Information Science, Springer, pp. 350–361, 2023. doi: 10.1007/978-3-031-35641-4_28.
[27] G. Sangar and V. Rajasekar, ”Optimized classification of potato leaf disease using EfficientNet-LITE and KE-SVM in diverse environments,” Frontiers in Plant Science, vol. 16, 2025. doi: 10.3389/fpls.2025.1499909.
[28] H. H. Luong, ”Improving Potato Diseases Classification Based on Custom ConvNeXt-Small and Combine with the Explanation Model,” International Journal of Advanced Computer Science and Applications, vol. 15, no. 4, 2024. doi: 10.14569/IJACSA.2024.01504121.
[29] M. M. Jawed et al., ”A hybrid deep learning framework using convolutional and transformer models for robust plant disease classification,” Scientific Reports, vol. 16, no. 1, p. 9704, 2026. doi: 10.1038/s41598-026-38209-z.
[30] D. Singh et al., ”PlantDoc,” in Proceedings of the 7th ACM IKDD CoDS and 25th COMAD, New York, NY, USA, pp. 249–253, 2020. doi: 10.1145/3371158.3371196.
[31] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, ”On Calibration of Modern Neural Networks,” in Proceedings of the 34th International Conference on Machine Learning, vol. 70, pp. 1321–1330, 2017. (https://arxiv.org/abs/1706.04599).
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Yousif Al-haj, AMMAR ALHATAMI, Mohammed Al-sharfi, Aidroos Al-mehdhar, Yassin Al-eaqariu, Ameen Ali, Mijahed Aljober, Somaia Y. A. Al-gwaini, Abdulrahman M. H. Obaid, Ammar Al-nami (Author)
This work is licensed under a Creative Commons Attribution 4.0 International License.
