A Secure Federated Transfer Learning Approach for Cotton Leaf Disease Classification Using Deep Learning Neural Networks

K. Rajeshwari¹   and S. Elamparuthi²
1. Department of Computer and Information Science, Annamalai University, Chennai, Tamil Nadu, India
2. PG Department of Computer Science, Govt Arts College, Chidambaram, Tamil Nadu, India
Correspondence to: K. Rajeshwari, rajeshwari.au2025@gmail.com

DOI: https://doi.org/10.70389/PJPB.100025

Additional information

• Ethical approval: N/a
• Consent: N/a
• Funding: No industry funding
• Conflicts of interest: N/a
• Author contribution: K. Rajeshwari – Conceptualization, Writing – original draft, review and editing
• Guarantor: K. Rajeshwari
• Provenance and peer-review: Unsolicited and externally peer-reviewed
• Data availability statement: N/a

Keywords: Cotton leaf disease, Deep learning, Distributed learning, Federated transfer learning, Privacy preserving AI, Smart agriculture, Xception.

Peer Review
Received: 13 February 2026
Last revised: 21 April 2026
Accepted: 11 May 2026
Version accepted: 6
Published: 18 May 2026

Plain Language Summary Infographic

Introduction

South Asia’s economy still depends heavily on agriculture, but the consequences of climate change are posing a growing danger to its viability. These environmental changes have accelerated the spread of crop diseases, leading to a noticeable decline in agricultural productivity. Among various crops, cotton—commonly known as ‘white gold’—holds significant economic importance as a major cash crop, with its global market value projected to approach 60 billion USD by 2030. In countries like Bangladesh, cotton plays a crucial role in supporting the textile industry. However, its production is frequently disrupted by factors such as pest attacks, plant diseases and adverse environmental conditions. Traditional methods of disease identification in cotton crops largely depend on manual inspection by farmers or agricultural experts.

Despite being widely used, these methods are frequently labour-intensive, slow and prone to human error. Inaccurate diagnosis might therefore result in incorrect pesticide use, higher production costs and significant yield losses. Automated plant disease identification has drawn a lot of attention with the development of artificial intelligence, especially deep learning. Deep neural network-based image-based analysis has shown encouraging results in accurately identifying crop diseases. However, traditional deep learning models usually depend on centralized data collection, where big data sets are processed and kept in one place. Particularly in distributed agricultural settings, this centralized approach presents serious questions about data ownership, privacy and secure information sharing.

Federated learning has evolved as a successful paradigm that allows collaborative model training without necessitating the exchange of raw data in order to overcome these issues. This method maintains data secrecy by having several decentralized participants—such as farms, research facilities or smart devices—train models locally and share only model updates. This article proposes a federated transfer learning system specifically designed for Bangladeshi agricultural circumstances for the diagnosis of cotton leaf disease.

The system makes use of pre-trained deep learning architectures, such as VGG-16, VGG-19, Inception-V3 and Xception, which are optimized at each client node using locally accessible data sets. The locally trained models are then combined into a global model using a federated aggregation technique, guaranteeing better generalization and data security. With an accuracy of 98.70%, the experimental evaluation shows that the federated model based on the Xception architecture performs better than alternative models. Furthermore, the optimized model is included in a web-based tool that enables quick, accurate and private disease diagnosis in real-time agricultural situations.

Literature Survey

Angamuthu and Arunachalam¹ proposed a hybrid approach combining k-means clustering with SVM for soybean leaf disease detection and achieved approximately 90% accuracy, demonstrating the effectiveness of machine learning in plant disease classification. Revathi and Hemalatha² applied a KNN-based model for cotton leaf disease identification using handcrafted image features and reported around 92% accuracy.

Sarangdhar and Pawar³ developed an SVM-based multiclass classifier for cotton disease detection, showing improved performance over traditional rule-based methods. Similarly, Balafas et al.⁴ presented a detailed survey on plant disease detection using machine learning and deep learning models, highlighting CNN-based architectures and their performance advantages. Junaidi et al.⁵ discussed the integration of deep learning and edge computing for real-time agricultural applications, highlighting low-latency disease detection and deployment challenges. Upadhyay et al.⁶ presented a comprehensive review of deep learning and computer vision techniques for plant disease detection, emphasizing modern architectures and challenges in precision agriculture.

Hossen et al.⁷ reviewed transfer learning approaches in agriculture, emphasizing their effectiveness in handling limited datasets for plant disease detection. Dahiya et al.⁸ conducted a comparative analysis of deep learning architectures for plant disease detection, highlighting the impact of hyperparameters and optimizers on model performance. Bi et al.⁹ introduced a MobileNet-based model for apple leaf disease classification, achieving efficient and lightweight performance suitable for real-time applications. Qiu et al.¹⁰ developed a deep neural network combined with colour imaging for detecting wheat diseases, achieving improved accuracy in early-stage identification. Pandian et al.¹¹ developed a CNN-based model for plant disease detection, showing improved accuracy over conventional methods.

Angamuthu and Arunachalam¹² applied fine-tuned transfer learning models for agricultural disease detection and reported improved performance over baseline CNNs. Sandler et al.¹³ proposed the MobileNetV2 architecture, which enables efficient deep learning with reduced computational complexity. Sujatha et al.¹⁴ compared machine learning and deep learning methods for plant disease detection, showing the superior performance of CNN-based models. Ananthi¹⁵ proposed a fused segmentation algorithm using SAR images for detecting nutrient deficiencies in crops. Mohanty et al.¹⁶ demonstrated the effectiveness of deep convolutional neural networks for large-scale plant disease classification using the Plant Village data set, achieving high accuracy and establishing a benchmark for automated disease detection. Ferentinos¹⁷ evaluated multiple deep learning architectures for plant disease recognition and reported that CNN-based models significantly outperform traditional machine learning approaches in both accuracy and robustness.

Too et al.¹⁸ compared deep transfer learning models such as VGG, ResNet, Inception and DenseNet for plant disease classification and showed that fine-tuned pre-trained networks provide superior performance with limited training data. Angamuthu et al.,¹⁹ developed a deep learning-based system for real-time crop disease detection in field conditions and highlighted the challenges of model generalization under varying environmental factors. Majeed et al.²⁰ utilized deep learning-based segmentation for automating agricultural tasks, demonstrating accurate image segmentation in orchard environments. Kairouz and McMahan²¹ presented a comprehensive survey on federated learning, outlining its advantages for privacy-preserving distributed model training across multiple clients without sharing raw data.

Yang et al.²² introduced federated learning concepts for mobile and edge devices and demonstrated how collaborative training can improve global model performance while maintaining data privacy. Liu et al.²³ applied federated learning to image classification tasks and showed that federated optimization methods such as FedAvg achieve performance comparable to centralized training. Angamuthu and Arunachalam²⁴ implemented a privacy-preserving federated learning model for plant disease detection and demonstrated its effectiveness in distributed agricultural environments with reduced communication overhead.

Methodology

The architecture of the suggested federated transfer learning system for classifying cotton leaf diseases is shown in this section (Figure 1). By storing raw photos on local devices and only transmitting model changes with a central server, the method guarantees precise detection while protecting data privacy. For practical usage, the combined global model is thereafter made available via a web-based platform. Pre-trained models like VGG-16, VGG-19, InceptionV3 and Xception are used by the framework. To replicate real-world settings, the data set is divided across several clients, each of which carries out local preprocessing and model fine-tuning. A generalized global model is then constructed by aggregating updated model parameters using a federated approach. Standard criteria including accuracy, precision, recall and F1-score are used to assess the final model on unseen data.

Data Collection

2,310 photos of both healthy and sick cotton leaves, and plants from a publicly accessible Kaggle repository make up the data set utilized in this study (Dev3v D3V, 2022). Healthy cotton leaf, healthy cotton plant, diseased cotton leaf and diseased cotton plant are its four classifications. In order to replicate real-world decentralized data settings like farms or edge devices, the data set is divided into many subsets and distributed among multiple clients for the federated learning configuration. By storing data locally at each client, this method protects data privacy. Figures 2 and 3 display some photos from the data set.

Data Analysis and Preprocessing

Real-world agricultural photos frequently have complicated backdrops, noise and fluctuating lighting, all of which might impair model performance. Data preprocessing is carried out locally at each federated client to address this, guaranteeing both enhanced data quality and privacy protection. For reliable and consistent training, all images are normalized to a range of 0–1 and scaled to 224 × 224 pixels. Data augmentation methods including rotation, zooming, shearing and horizontal flipping are used at the client level to improve model robustness and decrease overfitting. In order to track model performance and guarantee good generalization, 20% of each client’s local data is set aside for validation.

Federated Transfer Learning Model Training

ImageNet weights are used to initialize pre-trained models like VGG-16, VGG-19, InceptionV3 or Xception at each client. Custom completely connected layers and a softmax output for four-class categorization are used in place of the top layers. Lower levels are kept frozen to minimize computation, while only the upper layers are adjusted. The Adam optimizer is used for training, with a dropout rate of 0.02 and a learning rate of 0.001. Clients only exchange model parameters with the central server following local training. The Federated Averaging (FedAvg) technique is used by the server to combine these updates into a global model. Until convergence, this procedure is repeated across several rounds of communication.

Fine-Tuned Transfer Learning Models

Several transfer learning-based CNN architectures, such as VGG-16, VGG-19, Inception-V3 and Xception, which were chosen for their excellent performance in image classification tasks, are trained in this study using preprocessed cotton leaf pictures.

• VGG-16: A 16-layer CNN using small 3 × 3 filters, enabling effective hierarchical feature extraction with a simple and deep architecture.
• VGG-19: An extended version of VGG-16 with increased depth, improving feature representation while maintaining architectural consistency.
• Inception-V3: Utilizes parallel convolutional layers to capture multi-scale features efficiently, reducing computational cost through factorized convolutions.
• Xception: Uses depth-wise separable convolutions to enhance feature learning while significantly reducing computational complexity.

Experimental Study

This unit grants the experimental configuration, computational environment and evaluation methodology adopted to validate the proposed smart. Table 1 and Figure 4 summarize the primary hyperparameters utilized during the training and trying points of the federated transfer learning (FTL) model.

Table 1: Federated learning alignment slab.
Parameter	Value
Learning Rate	0.001
Dropout Rate	0.02
Optimizer	Adam
Shearing Range	−0.2 to +0.2
Horizontal Flipping	Enabled (True)
Rotation Range	−20° to +20°
Zoom Range	0.8 to 1.5
Validation Split	0.2

Experimental Set-Up

A single machine with an Intel Core i5 processor running at 1.80 GHz, 8 GB of RAM and an 8 GB dedicated graphics card was used for each experiment. It ran on a 64-bit version of Windows 11. A large-scale global data set is used to establish the pre-trained convolutional neural network in the federated transfer learning framework. The network is then locally fine-tuned on distributed client datasets without disclosing raw data. This approach improves the model’s capacity to generalize across many data sources while protecting data privacy. Each client’s locally updated model parameters are sent to a critical server, where a federated averaging technique is used to aggregate them and produce an optimal global model.

                       (1)

                 (2)

                        (3)

                           (4)

ReLU is an Activation Function, Not a Fitness Function

In the proposed deep learning framework, the Rectified Linear Unit (ReLU) is used as the activation function in the hidden layers. ReLU is chosen for its ability to model non-linear relationships and improve learning efficiency. It is widely used in deep neural networks due to its simplicity, fast convergence and computational effectiveness. The ReLU function is mathematically defined as categorical cross-entropy

              f (x) = max(0, x)        (5)

By mapping all negative input values to zero while keeping positive values unchanged, the Rectified linear unit (ReLU) activation function is used to add non-linearity to the model. This feature enables the network to extract intricate and significant patterns from picture data, which is crucial for accurate classification of cotton leaf diseases.

Results

Multiple transfer learning models were adapted and fine-tuned to perform cotton leaf disease classification, and their performance was assessed using standard evaluation metrics, including accuracy, loss, precision, recall, F1-score and confusion matrices. The training and validation curves demonstrate consistent learning behaviour, with gradual convergence and only minor signs of overfitting across the models (Figure 5). Among the evaluated architectures, the Xception model exhibits comparatively more stable performance, achieving a higher validation accuracy of 97.26%, whereas Inception-V3 attains 96.15%. In addition to improved accuracy, Xception shows better generalization and smoother convergence patterns during training. Overall, the Xception-based model offers an effective trade-off between predictive accuracy and model stability, making it a reliable choice for real-world deployment in cotton disease detection systems.

The effectiveness of the proposed method is summarized in Table 2, which presents a comparative analysis of key performance metrics, including accuracy, precision, recall and F1-score. The classification results indicate that the VGG-16, VGG-19, Inception-V3 and Xception models achieve accuracies of 90.22%, 96.74%, 97.83% and 98.70%, respectively. Among these models, Xception demonstrates the highest accuracy, highlighting its superior capability in cotton leaf disease classification. In terms of precision, VGG-16 records the lowest value at 92.59%, whereas the Xception model achieves the highest precision of 98.80%, as depicted in Figures 6–9. This further confirms the effectiveness of the Xception architecture in producing more reliable and accurate predictions compared to the other evaluated transfer learning models.

Table 2: Results of deep CNN models.
Algorithm	Accuracy (Std)	Accuracy (FL)	Precision (Std)	Precision (FL)	Recall (Std)	Recall (FL)	F1-score (Std)	F1-score (FL)
VGG-16	90.22	93.15	92.59	93.40	94.15	94.80	92.42	93.05
VGG-19	96.74	97.60	96.88	97.45	96.38	97.12	96.52	97.28
Inception-V3	97.83	98.32	97.88	98.10	98.29	98.58	98.08	98.44
Xception	98.70	99.15	98.80	99.05	98.90	99.20	98.85	99.12

Binary cross-entropy loss is commonly used for binary classification problems where the target output belongs to one of two classes. It measures the dissimilarity between the predicted probability and the actual class label by penalizing incorrect predictions. Minimizing this loss function during training helps the model achieve better classification performance and faster convergence.

              L = -[ ylog(p) + (1 – y)log(1 – p)]                (6)

The recall values for VGG-16, VGG-19, Inception-V3 and Xception are 94.15%, 96.38%, 98.29% and 98.90%, respectively, while the corresponding F1-scores are 92.42%, 96.52%, 98.08% and 98.85%. Xception consistently achieves the highest performance among all models. The confusion matrices presented confirm this trend. Compared to VGG-16, VGG-19 and Inception-V3, the Xception model achieves higher true-positive and true-negative rates, with fewer misclassifications across fresh and diseased cotton plant and leaf classes. Its superior performance is attributed to depth-wise separable convolutions and residual connections, which enable more effective feature extraction and accurate disease classification. EfficientNet-B0, MobileNetV2 (important for edge) and ViT (since your research already uses it!).

Discussion

The comparative results outlined in Table 3 provide an overview of the performance achieved by different machine learning and deep learning techniques for cotton disease detection. The proposed framework attains classification accuracies of 90.22%, 96.74%, 97.83% and 98.70% using VGG-16, VGG-19, Inception-V3 and Xception models, respectively. Across all evaluated metrics, the fine-tuned models in the proposed approach consistently outperform previously reported methods, demonstrating improved accuracy, precision, recall and F1-score.

Table 3: Comparative performance analysis with existing research works.
SL	Author	Method	Precision (%)	Recall/Sensitivity (%)	F-Score (%)	Accuracy (%)
1	Caldeira et al. (2021)	SVM	–	–	–	80.3
		NFC	–	–	–	71.1
		RNA	–	–	–	76.6
		KNN	–	–	–	78.8
		GoogleNet	96.1	90.6	93.3	95.2
		ResNet-50	92.2	94.1	93.2	95.1
2	Chen et al. (2020)	KNN	87.84	78.09	–	–
		RF	–	–	84.78	82.58
		LSTM	–	–	86.98	87.01
3	Proposed Model	Bi-LSTM	–	–	87.87	87.84
		VGG-16	92.59	94.15	92.42	90.22
		VGG-19	96.88	96.38	96.52	96.74
		Inception-V3	97.88	98.29	95.08	97.83
		Xception	98.80	98.90	98.85	98.70

Heuristic Comparison

In addition, experiments were conducted using well-known meta-heuristic optimization techniques, particularly the particle swarm optimization (PSO) algorithm. In this experimental set-up, pre-trained deep learning models such as VGG-16, VGG-19 and Xception were used as feature extractors, and PSO was applied to select the optimal feature subsets for classification. The findings clearly indicate that the proposed fine-tuned models outperform both the original deep learning architectures and the PSO-based feature selection approach, thereby validating the effectiveness and robustness of the proposed methodology.

A comparison of model accuracies demonstrates that deep learning techniques perform noticeably better than conventional classifiers. Out of all the investigated methods, the suggested Xception model achieves the highest accuracy. Model weights were optimized using gradient descent with backpropagation. Validation performance was used to adjust hyperparameters including learning rate, batch size, regularization and network design. Crop disease detection was assessed using a number of deep learning models, including CNN, VGG-16, VGG-19, ResNet and Xception. For the categorization of cotton diseases, refined CNN-based models demonstrated better performance and generalization.

Conclusions

A comparison of the federated learning (FL) technique and traditional centralized training (Std) for the VGG-16, VGG-19, Inception-V3 and Xception models shows that the federated set-up consistently improves performance. When compared to conventional training techniques, the FL-based strategy produces better outcomes in terms of accuracy, precision, recall and F1-score for all assessed architectures. In particular, VGG-16 shows improvements in all major evaluation parameters, along with a significant increase in accuracy from 90.22% to 93.15%.

In a similar vein, VGG-19 shows improved performance, with accuracy increasing from 96.74% with normal training to 97.60% with federated learning. This method is also useful in more complicated designs, as Xception achieves the best overall performance, increasing from 98.70% to 99.15%, while Inception-V3 improves from 97.83% to 98.32% in terms of accuracy. These findings imply that by leveraging a variety of decentralized data sources, federated learning improves generalization while preserving model efficacy.

With peak values across all evaluation parameters, including accuracy (99.15%), precision (99.05%), recall (99.20%) and F1-score (99.12%), the combination of Xception with federated learning produces the most dependable and consistent results of all the models. Overall, the results show that combining deep convolutional architectures with federated learning is a viable strategy for enhancing classification performance while guaranteeing privacy protection and permitting distributed model training in practical settings.

References

1. Angamuthu T, Arunachalam AS. A comparative analysis of CNN, GA, RF & RNN for image classification: insights on performance and optimisation using hybrid approaches. J Theor Appl Inf Technol. 2025;103(9):3890–3897.
2. Revathi P, Hemalatha M Classification of cotton leaf spot diseases using image processing edge detection techniques. International Conference on Emerging Trends in Science, Engineering and Technology, Tiruchirappalli, India. IEEE; 2012:169–173.
   https://doi.org/10.1109/INCOSET.2012.6513900
3. Sarangdhar AA, Pawar VR Machine learning regression technique for cotton leaf disease detection and controlling using IoT. International Conference on Electronics, Communication and Aerospace Technology, Coimbatore, India. IEEE; 2017:449–454. https://doi.org/10.1109/ICECA.2017.8212855
4. Balafas V, Karantoumanis E, Louta M, Ploskas N. Machine learning and deep learning for plant disease classification and detection. IEEE Access. 2023;11:114352–114377. https://doi.org/10.1109/ACCESS.2023.3324722
5. Junaidi A, Hashim SZ, Othman MS, et al. Deep learning and edge computing in agriculture: a comprehensive review of recent trends and innovations. IEEE Access. 2025;13:137464–137490. https://doi.org/10.1109/ACCESS.2025.3595621
6. Upadhyay A, Chandel NS, Singh KP, et al. Deep learning and computer vision in plant disease detection: a comprehensive review of techniques, models, and trends in precision agriculture. Artif Intell Rev. 2025;58(3):92. https://doi.org/10.1007/s10462-024-11100-x
7. Hossen MI, Awrangjeb M, Pan S, Mamun AA. Transfer learning in agriculture: a review. Artif Intell Rev. 2025;58(4):97. https://doi.org/10.1007/s10462-024-11081-x
8. Dahiya S, Gulati T, Gupta D. Performance analysis of deep learning architectures for plant leaves disease detection. Meas Sensors. 2022;24:100581. https://doi.org/10.1016/j.measen.2022.100581
9. Bi C, Wang J, Duan Y, Fu B, Kang JR, Shi Y. MobileNet based apple leaf diseases identification. Mob Netw Appl. 2022;27(1):172–180. https://doi.org/10.1007/s11036-020-01640-1
10. Qiu Ruicheng, Yang C, Moghimi A, Zhang M, Steffenson BJ, Hirsch CD. Detection of fusarium head blight in wheat using a deep neural network and color imaging. Remote Sens. 2019;11(22):2658. https://doi.org/10.3390/rs11222658
11. Pandian JArun, Kumar V, Geman O, Hnatiuc M, Arif M, Kanchanadevi K. Plant disease detection using deep convolutional neural network. Appl Sci. 2022;12(14):6982. https://doi.org/10.3390/app12146982
12. Angamuthu T, Arunachalam AS. Employing vision transformers for high-precision sugarcane disease classification: a deep learning perspective. Int J Basic Appl Sci. 2025;14(1):222–227.
13. Sandler M, Howard A, Zhu M, Zhmoginov A, Chen LC. Mobilenetv2: inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE; 2018; 4510–4520.
14. Sujatha R, Chatterjee JM, Jhanjhi NZ, Brohi SN. Performance of deep learning vs machine learning in plant leaf disease detection. Microprocess Microsyst. 2021;80:103615. https://doi.org/10.1016/j.micpro.2020.103615
15. Ananthi VP. Fused segmentation algorithm for the detection of nutrient deficiency in crops using SAR Images. In: Artificial Intelligence Techniques for Satellite Image Analysis. Springer International Publishing; 2019:137–159.
16. Mohanty SP, Hughes DP, Salathé M. Using deep learning for image-based plant disease detection. Front Plant Sci. 2016;7:1419. https://doi.org/10.3389/fpls.2016.01419
17. Ferentinos KP. Deep learning models for plant disease detection and diagnosis. Comput Electron Agric. 2018;145:311–318. https://doi.org/10.1016/j.compag.2018.01.009
18. Too EC, Yujian L, Njuki S, Yingchun L. A comparative study of fine-tuning deep learning models for plant disease identification. Comput Electron Agr. 2019;161:272–279. https://doi.org/10.1016/j.compag.2018.03.032
19. Angamuthu T, Arunachalam AS. Comparative analysis of deep learning and optimization techniques for sugarcane disease classification. Science. 2025;3:100011.
20. Majeed Y, Zhang J, Zhang X, et al. Deep learning based segmentation for automated training of apple trees on trellis wires. Comput Electron Agri. 2020;170:105277. https://doi.org/10.1016/j.compag.2020.105277
21. Kairouz P, McMahan HB. Advances and open problems in federated learning. Found Trends Mach Learn. 2021;14(1–2):1–210. https://doi.org/10.1561/2200000083
22. Yang Q, Liu Y, Chen T, Tong Y. Federated machine learning: concept and applications. ACM Trans Intell Syst Technol. 2019;10(2):
    1–19. https://doi.org/10.1145/3298981
23. Liu Y, et al. Federated learning for image classification: a survey. Pattern Recognit Lett. 2020;135:235–242.
24. Angamuthu T, Arunachalam AS A Novel hybrid framework for image classification: merging convolutional and evolutionary strategies. Int J Adv Netw Appl. 2024;16(3):6409–6412.

Cite this article as:
Rajeshwari K, Elamparuthi S. A Secure Federated Transfer Learning Approach for Cotton Leaf Disease Classification Using Deep Learning Neural Networks. Premier Journal of Plant Biology 2026;8:100025

Export to EndNote Export to Mendeley