Leveraging Inception-V3 and Transfer Learning for Early Detection and Classification of Cotton Crop Diseases

Cotton is a significant crop in India, vital for both its economy and agricultural sector. However, farmers often face challenges in diagnosing diseases in cotton leaves due to a lack of in-depth information. This leads to excessive use of fertilizers, impacting the environment and crop quality.
       
To address this, advancements in image processing, neural networks and machine learning are being applied. A mobile image processing system is being developed to help farmers identify cotton leaf diseases. This system processes leaf images, compares them with reference data and classifies them as diseased or healthy based on similarities. Machine learning enhances the system’s accuracy over time, improving with each iteration. By detecting infected leaves early, farmers can remove them to safeguard the harvest and reduce pest-related damage. This approach aims to protect both the cotton yield and the farming community, promoting sustainable agricultural practices.
 
Cotton diseases
 
Cotton plants are highly susceptible to fungal, bacterial and viral infections. These diseases typically manifest on the leaves, showing visible spots or lesions. However, the pathogens are not limited to foliage and can affect bolls and roots as well. Such infections can significantly reduce crop quality and yield if not managed properly. Early detection and precise classification are essential for effective disease control. To be familiar with the fertilizers that need to be given we have to make sure that the detection of the disease must be correct. The diseases that frequently occur on cotton leaves are: Aphids, Army Worm, Bacterial Blight, Powdery Mildew and Target spot.

Meta model in deep learning
 
The ‘cotton plant disease’ dataset contains 10673 images of various cotton leaves considered from ‘Kaggle’ is categorized into 6 classes, among them the data split training and testing is 80:20 respectively (Training: 7246 images and Testing: 3427 images). There are various models where we can predict that a leaf is either diseased leaf or an healthy leaf. Fig 1 shows the implementation flow of our model.


 
Model selection
 
There are multiple methodologies to predict whether a cotton leaf is a disease or not. Table 1 mentioned below depicts the review of different models:


                                                               
Architecture of inception-v3
 
The Inception-v3 is an architecture of a deep convolutional neural network, which combines inception modules for multi-scale feature extraction, as demonstrated in Fig 2. The inception modules use parallel convolution pathways including filters of different sizes to capture diverse spatial information. Due to its deep structure, Inception-v3 is, however, still computationally effective and can provide a good result in tasks of image classification.


 
Literature review
 
The health of cotton crops plays a critical role in global agricultural economies, especially in regions that rely on cotton production for economic growth and stability. With the advent of precision agriculture, leveraging machine learning (ML) and deep learning (DL) models has become essential in automating and improving disease detection and classification in cotton plants. One of the most promising approaches in this domain is the application of the Inception-v3 classifier, coupled with transfer learning (TL), to enhance the accuracy and efficiency of disease detection. This literature review synthesizes key research in cotton disease detection, focusing on Inception-V3 and TL methods to understand their contributions, challenges and future potential.
 
Deep learning in cotton disease detection
 
Deep learning models have significantly advanced the automation and optimization of cotton crop disease detection. Numerous studies highlight the role of convolutional neural networks (CNNs) and hybrid architectures in identifying and categorizing diseases in cotton plants. A landmark study by Tao et al. (2022) employed a CNN-based approach leveraging the ConvNeXt architecture, which combines transformers for enhanced feature extraction. This method achieved exceptional accuracy rates of 97.2%, 99.7% and 100% across various datasets while maintaining low inference times. However, the study mainly focuses on CNNs, with limited exploration of transfer learning (TL) and alternative algorithms like SVM, KNN, Random Forest and Decision Tree for classification. In contrast, Shaikh et al. (2023) implemented the Inception-V3 architecture as a classifier using TL to identify three cotton diseases: bacterial blight, curl virus and fusarium wilt. Their model, trained on a dataset of 950 images sourced from Kaggle and pre-processed with Gabor filters, achieved 97% accuracy. The use of TL improved feature extraction and the model effectively utilized supervised learning techniques. This study underscores the advantages of Inception-V3 and TL in accelerating training and achieving high accuracy for cotton disease detection.
 
Transfer learning for enhanced accuracy
 
Transfer learning (TL) is a pivotal method for enhancing disease detection models, especially in agriculture, by utilizing pre-trained models to streamline training. Islam et al. (2023) demonstrated the effectiveness of TL using models like VGG-16, Xception and Inception-v3 for cotton disease classification. Among these, Xception excelled with an impressive accuracy of 98.70%. This approach not only boosts model efficiency but also addresses challenges posed by limited datasets in real-world agricultural scenarios. Shaikh et al. (2023) expanded on this by incorporating Inception-V3 alongside TL and techniques like Gabor filtering for data augmentation. This combination significantly improves edge detection and texture classification, enabling early disease identification. These methods highlight the practicality of TL in achieving high accuracy with smaller datasets. Other researchers, including Noon et al. (2022), explored CNN-based ConvNeXt architecture with transformers, yielding promising results. However, TL-centric models, such as Inception-v3, offer faster training and adaptability across diverse datasets, making them ideal for practical applications. In summary, TL emerges as an invaluable tool in agricultural disease detection, delivering remarkable accuracy and efficiency while addressing the challenges of limited labelled data availability. Its adaptability and scalability make it indispensable for modern agricultural practices.
 
Meta-learning and hybrid approaches
 
Hybrid architectures and meta-learning approaches are pivotal in advancing model performance for agricultural applications. Memon et al. (2022) introduced a meta-deep learning framework that integrated Custom CNN, VGG-16 and ResNet50, achieving an impressive 98.53% accuracy in identifying cotton diseases like leaf spot and bacterial blight. This innovative approach leveraged the combined strengths of multiple models, enhancing generalization capabilities and minimizing overfitting risks. Although the study did not involve Inception-V3, its findings highlight the effectiveness of hybrid models in improving disease detection for cotton crops. Similarly, Aggarwal et al. (2023) investigated pre-trained deep neural networks for rice disease classification, yielding encouraging results with potential applicability to cotton crop health assessment. Their research emphasized the cri,tical role of feature selection in transfer learning (TL) models, further advocating for the use of architectures like Inception-v3. This model excels in extracting high-level features, making it highly suitable for disease classification tasks. Together, these studies underscore the importance of exploring integrated architectures and transfer learning techniques to enhance agricultural disease detection systems. Similar hybrid approaches were also supported by Ahmed (2021), Moyazzoma et al. (2021) and Reddy et al. (2023), who demonstrated the advantages of CNNs and TL for leaf disease recognition. Additionally, Arathi and Dulhare (2023) employed DenseNet-121 for cotton leaf classification, while Patil and Burkpalli (2021) explored WEKA-based ML pipelines for cotton disease detection, further validating the value of hybrid frameworks in this domain.
 
Challenges and limitations in cotton disease detection
 
Despite the successes of these methods, challenges remain in optimizing DL models for real-time, large-scale deployment in agricultural settings. The need for high-quality, large-scale datasets continues to be a limitation for many DL-based models. Hassan et al. (2022) addressed this issue by developing a custom CNN model on a dataset of 500 images from local fields in Multan, Pakistan, achieving an accuracy of 85%. However, their results highlight the gap between model performance on small, local datasets and the more generalized, large-scale applications of models like Inception-v3. Another challenge is the computational cost and complexity of deploying these models in real-time field conditions. Alotaibi and Rassam (2023) discussed the importance of leveraging high-performance computing resources to optimize ML models in cloud environments. Applying such principles to cotton disease detection would require efficient resource management to ensure the models remain feasible for use in the field, where computational power may be limited. Furthermore, Alsarhan et al. (2021) discussed ML-driven optimization in vehicular networks, principles of which can be extended to resource-constrained agricultural deployments.
       
Looking forward, integrating Inception-v3 with TL and other advanced architectures holds significant potential for improving cotton disease detection. Mobile applications, as proposed by Rajasekar et al. (2021), could make these models more accessible to farmers, providing real-time disease detection tools. This would democratize access to precision agriculture, allowing farmers in remote areas to benefit from advanced DL techniques. The use of cloud-based platforms, as suggested by Alotaibi and Rassam (2023), could further enhance the scalability of these models. By offloading computationally intensive tasks to the cloud, these systems could handle larger datasets and offer real-time insights to farmers. Furthermore, meta-learning approaches, like those explored by Memon et al. (2022), could lead to more adaptive models capable of generalizing to new datasets and environmental conditions, further enhancing the robustness of these systems. The interpretability and scalability shown in IoT-driven approaches such as that by Kundu et al. (2021) also align with this direction, promising practical implementations for future agricultural decision-making tools.
       
In recent years, there’s been a growing interest in using machine learning and deep learning techniques to detect and classify plant leaf diseases, especially in crops like cotton. For instance, Patil and Burkpalli (2021) took a practical approach by using traditional machine learning algorithms in the WEKA platform to classify cotton leaf images, showing how models like decision trees can still hold value in agricultural applications. On the other hand, researchers like Reddy et al. (2023) have shifted towards deep learning, specifically convolutional neural networks (CNNs), to predict cotton leaf diseases with impressive accuracy. Adding to this, Moyazzoma et al. (2021) used a transfer learning approach with MobileNetV2, showing that pre-trained models can be powerful tools for plant disease detection, especially when resources or data are limited. While these studies mainly focus on disease detection, Bansal et al. (2024) looked at the broader picture of crop productivity by examining how different pigeonpea varieties and spacing patterns impact yield. Together, these studies reflect how AI and data-driven methods are shaping the future of smart farming and plant health management.
               
The application of the Inception-v3 classifier, using TL, represents a significant advancement in cotton crop disease detection. By leveraging pre-trained models and fine-tuning them for specific agricultural tasks, researchers have demonstrated that high accuracy and efficiency can be achieved, even with relatively small datasets. Hybrid models and meta-learning approaches offer further opportunities for enhancing disease detection and classification, while cloud-based solutions and mobile applications could ensure these technologies reach the field, helping farmers protect their crops in real-time. Continued research in this area promises to refine these methods, making cotton crop disease detection more accessible, accurate and scalable.

Pre trained models such as Inception-v3 has the potential to identify diseases at the early stages while the pictorial symptoms might not be clear to the natural eye. This kind of detection in advance facilitates to the timely treatment. These Pre trained models can be deployed on varied platforms including drones and edge devices for remote monitoring and scalable of large fields of cotton. The methodology we proposed initiated with a step of pre-processing and data augmentation at University College of Engineering (A), Osmania University, Hyderabad, Telangana during the year 2023 as discussed below.
 
Data-augmentation
 
The process of data augmentation retained in this model enclosed a considerably wide shuffle of transformations in order to improve the robustness and diverseness of our dataset.The input image is taken to a size of 224*224 and processed. The practised transformations include shear range of 0.2, zoom range of 0.2, rotation range specified as 80, width_shift_range as 0.2, height_shift_range being 0.2, brightness range from 0.5 to 1.5 and horizontal vertical flips set to true. These measures help the model generalize better and prevent over-fitting. A pre-processing function is employed which is labelled “re-scale” set to ‘1.0/255’ to process the images into a desired format before being sent to be augmented. The validation split used in our model is a 10% split that is integrated into the process to access model performance during training. Fig 3 shows the Augmentation.


 
Feature extraction using inception-v3
 
Inception-v3 is a pre-trained CNN model with 17 layers, including convolution layers, Inception blocks and a fully connected layer. It is used for feature extraction by removing the last classification layer and using the activation inputs from this layer for describing features. Inception-v3 integrates techniques like factorized convolutions, regularization, dimension reduction and parallelized computations to optimize network performance. It is known for its computational efficiency, reduced complexity and versatility in feature extraction across various domains. The extracted features can be used to build a classification model with components like LSTM layers, dense layers and dropout layers, trained using the Adam optimization algorithm and specific hyper-parameter values. The Inception-v3 architecture is used for feature extraction in Model 1 of a study, which helps to build a NN model with greater stability and consistency in the year 2024.
 
Proposed model design
 
The features that have been extracted are used in the proposed method Inception V3 transfer learning model. Inception-v3 is from the Inception family that makes several improvements including using Label Smoothing, Factorized 7 × 7 convolutions and the use of an auxiliary classifer to propagate label information lower down the network.
 
Hyper tuning inception-v3
 
Hyper-parameter tuning for the Inception-v3 model in cotton plant disease detection involves optimizing hyper-parameters like learning rate, batch size, optimizer settings, activation functions and regularization techniques to enhance model performance. By systematically adjusting these hyper-parameters using techniques like grid search or random search, researchers can fine-tune the model for improved accuracy, convergence speed and generalization to unseen data. The learning rate controls how quickly the model learns, epochs determine training frequency, batch size specifies data samples used for parameter updates and dropout rate influences neuron dropout during training. This tuning process is crucial for achieving optimal model effectiveness in identifying and classifying diseases affecting cotton plants. Table 2 determines the metrics for tuning.


 
Proposed model customized parameters
 
In the context of the Inception-v3 model, the hyper-parameter tuning process using Grid Search aims to find the best combination of learning rate, dropout rate and optimizer settings that maximize the model’s disease detection accuracy. By exploring various hyper-parameter configurations and evaluating their impact on the model’s performance, researchers can identify the most effective settings for these parameters. This systematic approach ensures that the Inception-v3 model is fine-tuned to achieve optimal results in classifying cotton plant diseases based on extracted features. Table 3 discover the metrics that are customizing  the model proposed is Inception-v3.

A hyper-parameteris known as “epochs” describes that the number of iterations will be performed the learning method over the totaldataset given to training. For every epoch completion, the parameters of the model have been modified for the data collection in training dataset. A portion of the dataset is considered to train the neural network in a couple of lot that makesevery epoch. We can say the epoch is “iteration” as the process of transfering a bunch of training data. The accuracy is very low before tweaking settings Fig 4 depict the loss and accuracy on traditional approaches respectively.


       
The built-in dataset determines the indecent number in epochs. By using the entire training data, a neural network taking care of training for every cycle throughout the epoch. All the way each portion of knowledge is precisely used. If anepoch  moves both the directions backward and forward, it measures as a pass. The proposed methodlogy Inception-v3 adjusts the parameters, so accuracy has increesed. It rises to 98.85%. Minute values also considered and workable to the loss. Completed the training by consideration of 20 epochs. Fig 5 show the loss at 4.27% and accuracy at 98.85% attained by our proposed model after the tuning process respectively.


       
The curve of learning derived from the dataset given to training provides how effectively the model is able to picking up the new data. The curve of learning validation derived from a holds out validation dataset, it shows how well the model applies what it has learned to new, unseen data. The training loss drops sharply and it continues to rise till the last epoch. The curve in the proposed model of Fig 6 shows in the form of confusion matrix for the labels as diseases  and healthy explains that the loss is low.


       
In the above Table 4 mentioned the standard evaluation metrics for each disease class including healthy class and ensuring with Macro F1-score ≈ 97.17% and overall accuracy ≈ 98.03%.


       
The results and findings of the proposed model Inception-v3 are shown in Fig 8 and Fig 7(a). The model is successfully classify the class of diseased leaf with better accuracy and the the healthy leaf. Diseased leaves shown in  Fig 7(a) to 7(e) and healthy leaves are classified as healthy shown in the Fig 7(f). Here the approach is able to analyse the image disease characteristics and identifies the leaf and its particular disease.The designation of disease class is exact. Summary of the classification accuracy of each model is shown in Table 5.






 
Comparison with VGG-16
       
We observed the comparison between VGG-16 and showed them in the Fig 8(a) and 8(b) having accuracy and loss 96.69% and 35.47% to the proposed model, Inception-v3 shown in Fig 8(c) and 8(d) having accuracy and loss as 98.85% and 4.27% respectively showing that our model has achieved better results.

In conclusion, the application of Inception-v3 for detecting cotton leaf diseases has proven to be a promising advancement in agricultural practices. By utilizing image-based illness classification methods, meta-learning strategies and pre-trained models, the system achieved a remarkable accuracy of 98.85% by the 20^th epoch. This precision not only aids in identifying a wide range of diseases, such as bacterial blight, army worms and aphids, but also contributes to more sustainable farming by reducing the overuse of chemicals. The model’s ability to maintain crop yield and fibre quality, while minimizing environmental impact, highlights its potential in revolutionizing crop management.
       
Future efforts should focus on enhancing the robustness of the model through more diverse datasets and refining transfer learning strategies. By tailoring the model to detect diseases in their early stages, farmers could act proactively to prevent outbreaks, leading to higher crop yields and improved plant health. Ultimately, this research lays the groundwork for more efficient, eco-friendly agricultural practices that can drive significant improvements in cotton farming.

Disclaimers
 
The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided, but do not acceptany liability for any direct or indirect losses resulting from the use of this content.

The authors declare that there are no conflicts of interest regarding the publication of this article. No funding or sponsorship influenced the design of the study, datacollection, analysis, decision to publish, or preparation of the manuscript.