Groundnut (Arachis hypogaea L.) is a significant oilseed and legume crop cultivated in many tropical and subtropical regions around the world, especially in India. It serves as a major source of edible oil and protein, playing a vital role in the food, nutrition and livelihood of millions of smallholder farmers. Developing countries in Asia, Africa and South America account for over 97% of the world’s peanut cultivation area and 95% of the total production. The majority of production is concentrated in Asia (50% of global area and 64% of global production) and Africa (46% of global area and 28% of global production), where the crop is mostly grown by smallholder farmers under rainfed conditions (ICRISAT, 2014). According to recent global estimates, peanut cultivation now spans approximately 32 to 33 million hectares worldwide, with total production reaching around 54.4 million metric tons in 2023 (FAOSTAT, 2024; USDA, 2024). China continues to lead in global peanut production, contributing nearly 42% of the total output, followed by India at approximately 12% and the United States at 8% (USDA, 2024). Other notable producers include Argentina, Nigeria, Myanmar, Indonesia and Vietnam. The rise in global peanut output is attributed to expanded cultivation areas, favorable climatic conditions and varietal improvements (USDA, 2024). In 2014, global production was about 40 million metric tons from 24 million hectares (ICRISAT, 2014). By 2024, both cultivated area and production efficiency have grown. Southeast Asian countries like Myanmar, Indonesia, Vietnam and Thailand remain key producers. Vietnam, in particular, has seen steady growth due to better seeds and favorable conditions (USDA, 2024).
       
However, groundnut cultivation is often threatened by various foliar diseases, which significantly reduce both yield and quality. Among the most prevalent diseases are leaf spot (early and late), alternaria leaf spot, rust and rosette. These diseases manifest as visible symptoms on the leaf surface and can spread rapidly if not identified and managed in time (Bernabe and Sugui, 2019; Diga and Libunao, 2022). Traditionally, the identification and classification of these leaf diseases rely on manual inspection by agricultural experts or farmers. This approach is time-consuming, subjective and prone to human error (Subrahmanyam et al., 1995; Qi et al., 2021; Mohammad et al., 2026). Moreover, in rural and under-resourced settings, timely access to expert diagnosis is often unavailable. The limitations of traditional methods necessitate the development of automated, reliable and scalable solutions for early disease detection (Ozçelik et al., 2026; Paek et al., 2026).
       
Recent advancements in artificial intelligence have supported the development of practical tools for plant disease detection. Many studies have applied machine learning and deep learning techniques to detect diseases in groundnut plants (Bama and Priyadharsini, 2022; Cho, 2024; Kim and AlZubi, 2024). CNN-based models are capable of learning hierarchical features from raw pixel data and can be trained to classify complex patterns with high accuracy (Sriram and Kumari, 2026; Souza et al., 2026). In the context of crop health monitoring, CNNs offer a promising tool for identifying plant diseases from digital images, enabling rapid and accurate diagnosis (Aishwarya and Reddy, 2023; Desfita et al., 2025; Wirawan and Mahendra, 2024).
       
Maheswaran et al., (2022) developed a CNN-based model for identifying groundnut leaf diseases such as leaf spot, rust, bud necrosis, root rot and web blotch. The model was trained on a large farm-collected dataset and achieved 96.50% accuracy. It outperformed traditional methods, offering an efficient, scalable and accessible solution for disease detection and crop management. Bowrishandar and Prabha (2020) proposed a method for diagnosing groundnut leaf diseases using threshold-based color image segmentation and an artificial neural network (ANN). The study focused on segmenting groundnut leaf images into meaningful regions using a non-contextual thresholding approach. The segmented images were then classified using ANN. The paper highlighted the effectiveness of thresholding in color image segmentation and demonstrated its application in disease detection.
 
Research gap
 
Despite the progress in AI-based plant disease detection, several limitations remain. Most existing studies focus on binary classification or limited disease categories and often rely on controlled datasets, which restricts their applicability under real-field conditions (Bharvey and Sharma, 2023; Manonmani et al., 2024). Furthermore, there is still a lack of robust and efficient deep learning models specifically designed for multi-class classification of groundnut leaf diseases.
 
Objective of work
 
To address this gap, this study proposed a CNN-based model using the InceptionV3 architecture to classify five types of groundnut leaf conditions. The dataset, obtained from the Mendeley database, consisted of field images collected from Ramchandrapur village, Purba Medinipur district, West Bengal, India. Transfer learning was applied using a pretrained InceptionV3 model, along with custom dense layers, batch normalization and dropout to improve generalization. The novelty of this study lies in the use of a fine-tuned deep learning model combined with data augmentation techniques to enhance performance on diverse field images. The proposed approach aims to provide an accurate, efficient and scalable solution for groundnut disease detection to support precision agriculture.

Dataset collection and annotation
 
The dataset is taken from the publicly available mendeley data repository (Sasmal et al., 2024). Groundnut leaf images were collected from actual field conditions to ensure the dataset represents a realistic variation in appearance. The image acquisition was carried out in Ramchandrapur village, located in the Purba Medinipur district of West Bengal, India. A total of 1,720 high-resolution images, each with a pixel resolution of 4624 × 3472, were captured using a DSLR camera under natural lighting conditions. Images were taken from multiple plants at different times of day to incorporate lighting variability and leaf orientation diversity. The dataset encompasses five classes (Fig 1).
•  Healthy: Leaves with no visible signs of infection.
•  Leaf spot (Early and late): Includes symptoms ranging from small brown lesions to larger necrotic patches.
•  Alternaria leaf spot: Characterized by concentric rings of dark lesions caused by Alternaria species.
•  Rust: Exhibits reddish-brown pustules, typically on the underside of leaves.
•  Rosette: Marked by leaf stunting, yellow mottling and bunching at the top.


       
Each image was categorized and stored in one of five distinct directories, each named according to the corresponding class. This folder-based labeling approach ensured consistency during data loading and helped avoid annotation errors. Manual verification was performed to confirm the accuracy of each classification.
 
Preprocessing and data augmentation
 
Prior to training, all images were resized to 224 × 224 pixels to match the input requirements of the InceptionV3 architecture. The images were also converted to the RGB color space and normalization was performed using the preprocess_input() function from Keras’s InceptionV3 module. This function adjusts pixel intensity values to the range expected by the pretrained model, improving convergence during training.
def preprocess_image (image):
resized_image= cv2.resize [image, (224, 224)]
return preprocess_input [resized_image. astype(np. float32]
       
In addition to resizing and normalization, data augmentation techniques such as horizontal flipping, rotation and zoom were applied during model training using Keras’s Image Data Generator to enhance generalization and prevent overfitting. Specifically, augmentation included random rotations (up to ±20°), horizontal flipping, zoom range (0.2) and brightness adjustment to simulate real-field variability in lighting and orientation.
 
Label encoding and dataset split
 
To prepare the image labels for training, the class names were converted into a numerical format using scikit-learn’s label encoder. The dataset was then split into training and testing subsets using a stratified 70:30 ratio to ensure equal representation of all classes in both subsets. This stratification helped maintain class balance and ensured that minority classes such as rust and rosette were not underrepresented in the evaluation phase.
X_train, X_test, y_train, y_test = train_test_split (dataset, encoded_labels, test_size=0.3, stratify=encoded_labels, random_state=42).
       
To reduce the effects of class imbalance, particularly for the Rust and Rosette categories, class weights were computed using compute_class_weight from Scikit-learn. These weights were passed to the model during training to penalize misclassifications of underrepresented classes more heavily.
 
Deep learning model architecture
 
This study employed a transfer learning approach based on the InceptionV3 architecture, which was pretrained on the ImageNet dataset (Fig 2). The top classification layers of the original InceptionV3 model were removed and new custom layers were appended to adapt the network for five-class classification of groundnut leaf conditions. The base model used InceptionV3 with the top layers excluded (include_top=False) and global max pooling enabled to flatten the output feature maps.


       
The custom classification head consisted of three fully connected (dense) layers. The first dense layer had 128 units followed by batch normalization and a dropout rate of 0.3 to reduce overfitting. This was followed by a second dense layer of 64 units, again with batch normalization and dropout. A third dense layer with 32 units was added, also with batch normalization and dropout layers. Each dense layer used the ReLU activation function and included L2 regularization to further enhance generalization and stabilize training. The final output layer was a softmax layer with five units, corresponding to the five target classes. This architecture allowed the model to learn abstract and hierarchical feature representations specific to each disease class.
 
Model training and evaluation
 
The model was compiled using the RMSprop optimizer with a learning rate of 0.0001. The loss function used was sparse_categorical_crossentropy, which is appropriate when labels are provided as integers. The model was trained for a maximum of 75 epochs using a batch size of 32. The key hyperparameters used in this study include learning rate (0.0001), batch size (32), number of epochs (75), dropout rate (0.3) and L2 regularization to prevent overfitting and improve generalization. Early stopping was implemented to prevent overfitting, monitoring the validation loss with a patience value of 10 epochs. If no improvement was observed for 10 consecutive epochs, training was halted and the best model weights were restored.
model.compile [optimizer=RM sprop (learning_rate= 0.0001), loss=’sparse_categorical_crossentropy’, metrics= (‘accuracy’)]
       
After training, the model was evaluated using the test set. Metrics such as accuracy, precision, recall and F1-score were computed using Scikit-learn’sclassification_ report to evaluate class-wise performance. The confusion matrix was plotted using Seaborn’s heatmap to visualize true versus predicted classifications.The final trained model was saved in HDF5 format, making it suitable for future deployment or further research extensions in mobile or web-based agricultural advisory systems.
 
Hardware and software systems
 
All experiments were executed on a Windows 10-based workstation equipped with an Intel® Core™ i5-11320H CPU operating at 3.20 GHz and 16 GB of RAM. No dedicated GPU was used; all training processes were performed on the CPU, which may increase training time but does not affect model accuracy. The experimental pipeline was implemented using Python 3.11, chosen for its compatibility with modern deep learning frameworks and numerical libraries. Model development, training and evaluation were carried out using TensorFlow 2.x with the Keras API. Supporting libraries included NumPy, OpenCV, Matplotlib, Seaborn and Scikit-learn. All experiments were conducted in a Jupyter Notebook environment to facilitate iterative experimentation, real-time monitoring of training progress and visualization of results. To ensure reproducibility, random seeds were fixed across all libraries: NumPy [np.random.seed(42)], TensorFlow [tf.random.set_ seed(42)] and Python’s random module (random.seed (42)], ensuring consistent results across multiple runs.

The section outlines the performance of the InceptionV3 model after training and testing. The dataset initially comprised 1,720 high-resolution images; however, after addressing class imbalance and applying preprocessing techniques, 1,204 images were used for model development. Of these, 688 images were used for training and 516 for testing. The training set included 315 images of alternaria leaf spot, 420 healthy leaves, 315 leaf spot (Early and late), 70 rosette and 84 rust. The test set consisted of 135 alternaria leaf spot, 180 healthy, 135 leaf spot (Early and late), 30 rosette and 36 rust images. Although the model was configured to train for 75 epochs, it stopped early at epoch 47 due to an early stopping mechanism that monitored validation loss, indicating optimal convergence.
       
The accuracy and loss measurements during training the model are plotted in Fig 3. In the first epoch, the model recorded a low training accuracy of 24.17% and validation accuracy of 25.39%, with high loss values (2.0032 and 1.8009, respectively), indicating the learning process had just begun. By epoch 25, the training accuracy increased significantly to 98.57% and the validation accuracy rose to 96.71%, with a considerable reduction in loss values (training loss: 0.1079, validation loss: 0.1473). At epoch 47, the model reached near-optimal performance with training accuracy of 99.84% and validation accuracy of 98.45%, accompanied by low loss values (training loss: 0.0580, validation loss: 0.1342), indicating excellent convergence and generalization.


       
The confusion matrix provides a clear view of the model’s classification accuracy across the five peanut leaf disease categories (Fig 4). The model correctly classified all 135 samples of alternaria leaf spot, all 30 samples of rosette and all 36 samples of rust, demonstrating perfect prediction performance for these classes. In the healthy category, 179 out of 180 samples were correctly classified, with 1 sample misclassified as alternaria leaf spot. The leaf spot (Early and late) class showed slightly more misclassification, with 127 out of 135 samples correctly identified, while 6 were misclassified as alternaria leaf spot and 2 as healthy. These minor misclassifications suggest some visual similarity between Leaf Spot symptoms and those of alternaria leaf spot or healthy leaves. Despite this, the overall distribution indicates strong class separability and high reliability of the model, particularly for underrepresented classes like Rosette and Rust, which achieved 100% accuracy.


       
The classification report highlights the strong predictive performance of the proposed InceptionV3-based model across all peanut leaf disease categories (Table 1). The Alternaria Leaf Spot class achieved a precision of 0.9507, recall of 1.0000 and an F1-score of 0.9747, indicating that all 135 actual cases were correctly identified, although a few samples from other classes were misclassified as this class. The healthy class recorded a precision of 0.9890, recall of 0.9944 and F1-score of 0.9917 across 180 test samples, showing minimal misclassification. The leaf spot (Early and late) class had the lowest recall (0.9407), with a precision of 1.0000 and F1-score of 0.9695, reflecting some confusion with similar classes. Both Rosette and Rust achieved perfect scores across all metrics (precision, recall, F1-score = 1.000), despite having fewer test samples (30 and 36, respectively), underscoring the model’s robustness even for underrepresented categories. The model achieved an overall accuracy of 98.26%, with a macro average F1-score of 0.9872 and weighted average F1-score of 0.9825, confirming its high effectiveness in multi-class disease classification.


       
Fig 5 illustrates a set of representative prediction outputs from the trained model, showing the true class, the predicted class and the associated confidence score (%) for each image. All examples shown were correctly classified by the model, highlighting its robustness and high confidence across multiple peanut leaf disease categories.The predictions for leaf spot (Early and late) (top-left and bottom-left images) were made with high confidence, scoring 99.40% and 99.54%, respectively. These images clearly show visible symptoms such as dark lesions and necrotic areas, which the model correctly identified. The images classified as alternaria leaf spot (middle-right and bottom-right groups) achieved confidence scores ranging from 99.51% to 99.69%, reflecting the model’s reliability in detecting fine-grained variations in leaf texture and spot patterning.


       
Meanwhile, healthy leaves were also correctly classified with confidence values between 97.99% and 99.39%. These leaves appear free from visible disease symptoms and the model’s high confidence suggests that it effectively learned the distinction between healthy and infected samples. The consistent alignment between the true class, predicted class and high confidence scores across different leaf conditions demonstrates the model’s strong generalization ability and suitability for real-world application in peanut disease diagnosis.
       
The multi-class ROC curve illustrates the model’s excellent ability to distinguish between the five peanut leaf disease categories (Fig 6). The Area Under the Curve (AUC) values were exceptionally high for all classes: Healthy, rosette and rust each achieved a perfect score of 1.0000, while Alternaria leaf spot and leaf spot (Early and late) recorded AUCs of 0.9992 and 0.9987, respectively. These results indicate that the model has a very high true positive rate with minimal false positives across all classes. The near-perfect AUC values confirm the model’s strong discriminative capability and robust generalization in multi-class classification tasks, further supporting its suitability for practical deployment in peanut disease detection systems.


       
The precision-recall (PR) curve provides a deeper insight into the model’s classification performance, especially for imbalanced datasets. As shown in Fig 7, the model achieved nearly perfect average precision (AP) scores across all five peanut leaf disease classes. Rosette and rust achieved an AP of 1.0000, indicating flawless precision and recall balance. The healthy class also performed exceptionally well with an AP of 0.9999, followed by alternaria leaf spot and leaf spot (Early and late) with AP scores of 0.9978 and 0.9965, respectively. These consistently high AP values confirm that the model maintains excellent prediction quality across all classes, effectively minimizing both false positives and false negatives. The PR curves further validate the model’s reliability and robustness in real-world disease classification tasks, particularly in scenarios where class imbalance may affect performance.


       
To further validate the performance and reliability of the proposed model, several advanced evaluation metrics were computed. Cohen’s kappa score was 0.9762, indicating a near-perfect agreement between the predicted and true class labels, well beyond chance level. The log loss, which penalizes incorrect classifications with high confidence, was low at 0.0656, reflecting the model’s well-calibrated probability estimates and high certainty in its predictions. Additionally, the matthews correlation coefficient (MCC), a balanced measure that takes into account true and false positives and negatives, was 0.9765, further confirming the model’s strong overall predictive performance across all classes. These complementary metrics reinforce the findings from the confusion matrix, ROC and PR curves and demonstrate the robustness, reliability and clinical utility of the model in multi-class peanut leaf disease classification.
       
In addition to accuracy-based metrics, the computational efficiency of the model was evaluated using inference time. The model required approximately 22 seconds to process 516 test images, resulting in an average inference time of about 0.043 seconds (~43 ms) per image on a CPU-based system.
       
Several studies have explored peanut or groundnut leaf disease classification using both classical machine learning and deep learning techniques (Table 2). Xu et al., (2023) reported high accuracy (99.69%) using an improved Xception-based model with attention mechanisms, while Patayon and Crisostomo (2022) achieved 98% accuracy using DenseNet-169. Vaishnnave et al., (2020) also demonstrated strong performance (99.88%) using a DCNN on the PlantVillage dataset. In contrast, studies using real-world datasets, such as Aishwarya and Reddy (2023), reported slightly lower accuracy (98.46%), highlighting the challenges of field conditions.


       
Traditional machine learning approaches, as reported by Bama and Priyadharsini (2022), achieved high accuracy (~100%); however, these methods rely heavily on manual feature extraction and may lack generalization across diverse datasets. Similarly, Chetan et al., (2024) showed comparatively lower performance (92.26%) using ResNet50V2, indicating variability across architectures and datasets. More recent work by Kaur et al., (2024) using fine-tuned InceptionV3 achieved 97.30% accuracy, which is slightly lower than the performance obtained in the present study. In comparison, the proposed model achieved 98.26% accuracy, along with strong AUC (≥0.9987), F1-score (≥0.96) and Cohen’s Kappa (0.9762). While slightly lower than some complex ensemble or attention-based models, the proposed approach provides a good balance between accuracy and computational efficiency. Overall, the results indicate that the proposed model can serve as a reliable and efficient reference for real-world agricultural applications.

This study has certain limitations. The dataset size, although sufficient for model training, is relatively limited and sourced from a specific geographic region. That may affect the generalizability of the model to diverse field conditions. Additionally, the study focuses on a single deep learning architecture (InceptionV3), without comparative evaluation of multiple models. Future work will aim to incorporate larger and more diverse datasets from multiple open-access sources to improve generalization. Furthermore, comparative analysis with advanced architectures and optimization techniques will be conducted to enhance model performance and scalability for real-world agricultural applications.

The fine-tuned InceptionV3 model achieved high accuracy (98.26%) in classifying five groundnut leaf conditions, including rare classes like rosette and rust. It demonstrated strong precision, recall and F1-scores (≥0.96), along with high AUC (≥0.9987) and a cohen’s kappa of 0.9762, confirming its robustness and reliability. Compared to prior studies, the model outperformed several conventional CNNs and closely matched the performance of more complex ensemble or attention-based models. Its simplicity, interpretability and computational efficiency make it suitable for real-world deployment in agricultural settings. However, the dataset used was region-specific, limiting its global applicability. Classes like Rosette and Rust had fewer samples, which may affect generalization. The model was also tested only on static images, without accounting for environmental variables like lighting changes or background noise. In future work, the dataset can be expanded to include samples from various regions and seasons. Testing on real-time video or mobile data will support field deployment. Attention modules or ensemble methods could further improve robustness. Integrating the model with GIS tools and drone imagery may also enable large-scale, automated disease monitoring. Overall, the model presents a reliable solution for early disease diagnosis in groundnut crops and supports precision agriculture efforts.

This paper was supported by the Projects of Talents Recruitment of Guangdong University of Petrochemical Technology, China. We thank Guangdong University of Petrochemical Technology for its support in Technical Education and Law Training in this paper.
 
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 accept any liability for any direct or indirect losses resulting from the use of this content.
 
Authors’ contributions
 
All authors contributed toward data analysis, drafting and revising the paper and agreed to be responsible for all the aspects of this work.
 
Data availability
 
The data analysed/generated in the present study will be made available from corresponding authors upon reasonable request.
 
Availability of data and materials
 
Not applicable.
 
Use of artificial intelligence
 
Not applicable.
 
Declarations
 
Authors declare that all works are original and this manuscript has not been published in any other journal.

The authors declare that they have no conflict of interest.