Current IssueSpecial IssueEditorial BoardOnline First ArticlesArchive
Current IssueSpecial IssueEditorial BoardOnline First ArticlesArchive

Full Research Article

A Modified EfficientNetB0-based Deep Learning Model for Accurate Detection and Classification of Groundnut Leaf Diseases

J
Jie-Shin Lin1
Y
Yu-Nan Tai2
S
Suh Chen Hsiao3
L
Luke H.C. Hsiao1,*
1Department of Public Policy and Management, I-Shou University, Taiwan.
2Program in Intelligent Tourism and Hospitality Management, I-Shou University, Taiwan.
3Director and Professor of Social Work Practicum Education, University of Southern California Suzanne Dworak-Peck School of Social Work, USA.

  • Submitted18-03-2026|

  • Accepted20-08-2026|

  • First Online 29-08-2026|

  • doi 10.18805/LRF-948

ABSTRACT

Background: Groundnut farming is affected by several leaf diseases that reduce crop yield. Farmers often rely on visual inspection, which can be inaccurate and time-consuming. Early and precise identification of leaf diseases is essential for effective crop management. This study presents a deep learning-based solution to automate disease detection.

Methods: A modified EfficientNetB0 architecture is proposed for classifying five types of groundnut leaf conditions: healthy, leaf spot (early and late), alternaria leaf spot, rust and rosette. The dataset is sourced from the Mendeley database that was collected from Ramchandrapur village in West Bengal, India, under natural lighting. A total of 1,720 images were captured using a DSLR camera. After verification and cleaning, the dataset was split into 1,204 training and 516 testing images. All images were resized, normalized and label-encoded. Data augmentation techniques such as rotation, flipping and zoom were used to improve generalization. Regularization was applied to reduce overfitting. The model was trained for 100 epochs using the RMSprop optimizer and early stopping.

Result: The model achieved a test accuracy of 99.22%. Evaluation metrics confirm high performance across all classes. The model outperformed existing methods such as ResNet50 (82.3%), CNN with progressive resizing (96.12%) and LeafNet (97.23%). It also maintained low training and validation loss throughout training. These results highlight the model’s robustness, accuracy and potential for real-time field applications. The approach is lightweight and suitable for mobile-based disease detection tools.

INTRODUCTION

Groundnut (Arachis hypogaea) is an important legume cultivated across tropical and subtropical regions. It serves as a rich source of edible oil, comprising about 45-50% of the seed and protein (»25%) (ICRISAT, 2023; FAO, 2023). Globally, groundnut is grown on approximately 32.7 million hectares, producing nearly 53.9 million tonnes annually (FAO, 2023). China and India are the two largest producers of groundnuts. In 2023, China produced around 18.4 million tonnes, followed by India with approximately 7.38 million tonnes (ReportLinker, 2023). China, India and Nigeria together produced nearly 60% of the world’s groundnuts during 1961-2022 (Nayak, 2021). The estimated production for 2023-24 in India reached 8.66 million tonnes (NEDFi, 2024).

Groundnut is vital to Indian agriculture. It is cultivated in both Kharif (May-June) and Rabi (October) seasons and contributes significantly to vegetable oil production (FAO, 2023; APEDA, 2023). Gujarat is the largest producing state (»42%), followed by Rajasthan (17%), Tamil Nadu (11%), Andhra Pradesh (9%) and Karnataka (6%) (APEDA, 2023). Despite its economic and nutritional importance, India’s average yield (~1.3 t/ha) remains low compared to the global average (ICRISAT, 2023; ReportLinker, 2023). Constraints such as erratic rainfall, limited irrigation and the prevalence of foliar diseases affect productivity.

Among biotic stresses, foliar diseases significantly reduce groundnut yield and quality. The most common diseases include early and late leaf spot, Alternaria leaf spot, Rust and Rosette (Pooniya et al., 2020; Shasidhar et al., 2019). Early leaf spot occurs one month after sowing and appears as circular lesions that later darken. Late leaf spot appears after 6-7 weeks, showing necrotic lesions with faint halos. Alternaria leaf spot develops after 5-6 weeks and causes water-soaked chlorotic spots and leaf curling. Rust typically emerges six weeks post-sowing and presents as reddish-brown pustules. Rosette disease occurs within 2-3 weeks of sowing and causes yellowing, curling and stunted growth, giving leaves a rosette-like structure (Sasmal et al., 2024).

These diseases disrupt photosynthesis and reduce both seed quality and marketable yield. Currently, field inspection is the standard approach for diagnosis, but it is time-consuming, labor-intensive and prone to subjectivity. Misdiagnosis can lead to improper treatment and crop loss. In recent years, artificial intelligence (AI) and deep learning (DL) have gained attention as scalable tools for automated plant disease detection (Cho, 2024; AlZubi, 2023). CNNs are capable of classifying leaf diseases with high precision. Transfer learning, using architectures such as EfficientNet, ResNet and DenseNet, enables high performance even on moderately sized datasets (Anbumozhi and Shanthini, 2024; Sivaganesan, 2023).

Recent advances in DL methods for agriculture show promising results. Models like ResNet, DenseNet and EfficientNet have been used for leaf classification tasks across various crops (Mahto and Mathew, 2025; Desfita et al., 2025). Transfer learning, which uses pre-trained models on large-scale datasets such as ImageNet, has made it possible to achieve high accuracy even with moderately sized agricultural datasets (Kishore and Senthilvel, 2025; Hai and Duong, 2024; Semara et al., 2024; Maltare et al., 2023). M.P and Reddy, 2023 proposed an ensemble method for groundnut leaf disease classification using a tri-CNN architecture composed of pre-trained DenseNet169, Inception and Xception models. These base models were trained on the ImageNet dataset (Mohammad et al., 2026; Özçelik et al.,  2026; Paek et al., 2026; Souza et al., 2026; Sriram and Kumari, 2026). The method was tested on a real-world groundnut leaf dataset and achieved a high classification accuracy of 98.46%, demonstrating the effectiveness of ensemble deep learning in plant disease detection. Vardhan et al.  (2024) evaluated three deep learning models, CNN, MobileNetV2 and InceptionResNetV2, for groundnut disease detection. Janani and Jebakumar (2022) developed a system called Nutrient Range Analysis based on Greenness (NRAG) to detect and classify nitrogen levels in groundnut leaves using RGB images. They applied a median filter to reduce noise and extracted greenness features, which were directly linked to nitrogen content. These features were fed into a CNN-based Holding Vector Network (HVN) model. The model classified leaves into three nitrogen categories: low, normal and excess. It achieved 95% training accuracy and 92% validation accuracy, offering a scalable and cost-effective solution for real-time use by farmers.

In this study, we propose a modified EfficientNetB0 model for the multi-class classification of groundnut leaf diseases. This work utilizes a dataset from the Mendeley Data Repository, comprising field-collected images that provide a practical basis for model training. The EfficientNetB0 architecture is enhanced with additional dense layers, batch normalization and regularization to improve performance. The model is evaluated using key metrics and is designed with scalability in mind, making it suitable for mobile-based plant health applications. This work contributes to the development of intelligent decision-support tools. These tools can assist groundnut farmers in detecting diseases at an early stage. Early detection supports timely intervention and improved crop management. The approach also promotes sustainable agriculture, especially in regions with limited resources.

MATERIALS AND METHODS

This study uses a publicly available dataset of groundnut leaf images obtained from the Mendeley Data Repository (Sasmal et al., 2024). The images were captured under real-world agricultural conditions to reflect natural variation in appearance. Data collection was conducted in Ramchandrapur village, located in the Purba Medinipur district of West Bengal, India. A DSLR camera was used to capture high-resolution images, each with a size of 4624 × 3472 pixels. All images were taken in natural lighting conditions without artificial enhancements. To introduce variability, photographs were captured at different times of day and from multiple plants. This helped incorporate changes in illumination, leaf angles and background complexity, making the dataset more robust and realistic. In total, 1,720 images were collected. The dataset covers five distinct classes: healthy, leaf spot (early and late), alternaria leaf spot, rust and rosette. Each image was organized into one of five class-specific folders. The folder names correspond exactly to the disease categories. All images underwent manual verification. Each file was reviewed by trained observers to confirm the accuracy of its classification. Images that were misclassified, blurry, or duplicated were removed to maintain dataset quality. The resulting dataset serves as a reliable foundation for training deep learning models. A few images from the dataset are given in Fig 1.

Fig 1: Images from the dataset for training and testing of the model.


 
Image preprocessing, label encoding and augmentation
 
Each image in the dataset was resized to 224 x 224 pixels. This resizing ensured uniform input dimensions for the model. Pixel values were then normalized to a range between 0 and 1. This was done by dividing each pixel intensity by 255. Label encoding was applied using the LabelEncoder tool from Scikit-learn. This converted the class names from strings to numerical values. The dataset was then split into training and testing sets using a 70:30 ratio. As a result, 1,204 images were used for training and 516 for testing. To improve model generalization, data augmentation techniques were applied. These included random rotation up to 20 degrees. Width and height were shifted by up to 10%. The zoom range was set between 0.8 and 1.2. A shear transformation of 0.2 was used. Horizontal flipping was also applied. The fill mode was set to ‘nearest’ to handle missing pixels after transformation. These techniques introduced variation in image orientation and scale. This helped the model learn more robust features and reduced the risk of overfitting.
 
Model architecture
 
In this paper, the proposed model is based on the EfficientNetB0 architecture (Fig 2). This model is well known for its balance between performance and computational efficiency. EfficientNetB0 was used as the feature extractor in the present system. It was initialized with ImageNet pre-trained weights. The top layer of the base model was removed to allow customization for the classification task. An input layer was defined to accept images of size 224 x 224 x 3. This matched the size used during preprocessing. The input was passed through the EfficientNetB0 base. The output from the base model was a 1,280-dimensional feature vector. A custom classifier head was added after the base model. This head consisted of three dense (fully connected) layers. The first dense layer had 128 units and used the ReLU activation function. It was followed by Batch Normalization and a Dropout layer with a dropout rate of 0.3. The second dense layer had 64 units. Like the first, it also used ReLU, batch normalization and dropout. The third dense layer had 32 units and followed the same structure.

Fig 2: Working procedure of the model with architecture of the proposed groundnut leaf disease classification model using EfficientNetB0.



L1 regularization was applied to each dense layer. The regularization parameter was set to 0.0001. This helped prevent overfitting by penalizing large weights. The final output layer was a dense layer with 5 units, one for each class. This layer used a Softmax activation function. It converted the output vector into class probabilities. Mathematically, for an input image x, the final probability output is given by
 
 
Where,
zc  = Logit for class c.
K=5 = Total number of classes.

The total number of parameters in the model was 4,224,929. Among these, 4,182,465 parameters were trainable. The rest were non-trainable, as they belonged to the frozen layers of the pre-trained base model. This architecture was built using the Keras Functional API. The use of batch normalization and dropout after each dense layer made the model more stable and robust during training. This architecture was chosen for its ability to extract detailed features and to generalize well with limited data.
 
Model training and evaluation
 
The model was compiled using the RMSprop optimizer. The learning rate was set to 0.0001 and the loss function used was Sparse Categorical Crossentropy, as labels were encoded as integers. The training process used an early stopping callback to monitor the validation loss. If no improvement was observed for 10 consecutive epochs, training was halted and the best weights were restored. The model was trained for a maximum of 100 epochs using a batch size of 32. The model was evaluated using four performance metrics. These are defined as follows.
 




 
  
Where,
TP = True positives.
FP = False positives.
FN = False negatives.

After training, the model was saved in HDF5 format for future inference or fine-tuning. Table 1 summarizes hyperparameters used for training the efficientB0 model. 

Table 1: Hyperparameters used for model training.

RESULTS AND DISCUSSION

The proposed model was trained for 100 epochs using the EfficientNetB0 backbone and custom dense layers (Fig 3). The training and validation performance improved steadily over time. In the initial phase (Epoch 1), the training accuracy was low at 19.86% and the loss was high at 2.9853. The validation accuracy at this point was 20.35%, with a loss of 2.3289. By Epoch 25, the model had significantly improved. It achieved a training accuracy of 98.28% and a validation accuracy of 98.64%. Corresponding training and validation losses were 0.6777 and 0.6316, respectively. This shows that the model was able to learn useful features and generalize well by the mid-point of training.

Fig 3: Training and validation accuracy and loss curves over 100 epochs.



At Epoch 50, the model continued to perform well. The training accuracy reached 99.75% and the validation accuracy improved to 98.84%. The training loss reduced to 0.4466 and the validation loss dropped to 0.4522. This indicated that the model was not overfitting and maintained stability across epochs. By Epoch 75, the training accuracy was 99.66% and validation accuracy remained at 98.84%. Training and validation losses further reduced to 0.3011 and 0.3186, respectively. This confirmed consistent performance over time. At the final epoch (Epoch 100), the model reached 99.95% accuracy on the training set. The validation accuracy peaked at 99.22%, showing excellent generalization. Final losses were 0.1818 for training and 0.1963 for validation. The gap between training and validation losses remained small, which indicates minimal overfitting. Overall, the model demonstrated strong learning capacity and robustness. The EfficientNetB0 backbone with the added dense layers contributed significantly to high accuracy and low validation loss. The early stopping callback helped prevent overfitting and ensured optimal weight retention.

The confusion matrix in Fig 4 highlights the performance of the proposed EfficientNetB0-based model on the groundnut leaf disease classification task. Each row represents the actual class and each column indicates the predicted class. The model correctly classified 130 samples of Alternaria Leaf Spot without any misclassifications. All 179 healthy leaf images were accurately identified. For the Leaf Spot (Early and Late) category, 139 images were correctly classified, while 3 were mistakenly predicted as Alternaria Leaf Spot. In the case of Rosette disease, 20 images were correctly classified and only one was misclassified as Healthy. Rust disease was perfectly predicted for all 44 samples. These results demonstrate strong class-wise performance, with very few errors occurring mainly between visually similar diseases. Overall, the model exhibits high reliability in distinguishing between multiple groundnut leaf diseases.

​

Fig 4: Confusion matrix showing true and false classifications for five groundnut leaf categories.



The classification report further validates the effectiveness of the proposed model across all five groundnut leaf classes (Table 2). For Alternaria Leaf Spot, the model achieved a precision of 0.9774 and a perfect recall of 1.0, resulting in an F1-score of 0.9886 across 130 test images. Healthy leaves were classified with extremely high precision (0.9944) and perfect recall, yielding an F1-score of 0.9972 over 179 samples. The Leaf Spot (Early and Late) class recorded a precision of 1.0 and a recall of 0.9789, producing an F1-score of 0.9893 for 142 samples. In the Rosette category, the model achieved perfect precision but a slightly lower recall of 0.9524, which led to an F1-score of 0.9756 for 21 images. Rust classification was flawless with 1.0 scores across all metrics for 44 samples. The model attained an overall accuracy of 99.22% on the test dataset. The macro average precision, recall and F1-score were 0.9944, 0.9863 and 0.9901, respectively, indicating balanced performance across all classes. The weighted average, which accounts for the number of samples per class, remained consistent at 0.9922 for both recall and F1-score. These results confirm that the model is robust, consistent and well-suited for multi-class classification of groundnut leaf diseases.

Table 2: Classification report for each groundnut leaf disease class using the proposed EfficientNetB0-based model.



The ROC curve in the image shows excellent model performance for all classes (Fig 5). Each class achieves an AUC close to or equal to 1.0. Specifically, Healthy, Rosette and Rust have an AUC of 1.0000, indicating perfect class separation. Alternaria Leaf Spot and Leaf Spot (Early and Late) also perform well, with AUC values of 0.9999. The curves rise sharply toward the top-left corner, confirming high true positive rates and low false positive rates across all categories.

​

Fig 5: Multi-class ROC curve illustrating AUC values for each groundnut leaf class.



The precision-recall (PR) curve shows that the model performs well across all classes (Fig 6). Average precision (AP) for Healthy, Rosette and Rust is 1.0000, indicating perfect balance between precision and recall. Alternaria Leaf Spot has an AP of 0.9998 and Leaf Spot (Early and Late) achieves 0.9997. The curves remain close to the top-right corner, reflecting high confidence in predictions and minimal false positives or false negatives.

​

Fig 6: Precision-recall curve showing average precision (AP) values for each class.



The F1-score vs. threshold plot demonstrates stable and high F1-scores for all five classes across most threshold values (Fig 7). Scores remain above 0.9 between thresholds of 0.1 and 0.95, showing the model maintains a good balance between precision and recall. A sharp drop is observed near thresholds 0 and 1, which is expected. All classes, Alternaria Leaf Spot, Healthy, Leaf Spot (Early and Late), Rosette and Rust, exhibit similar trends, confirming model consistency.

​

Fig 7: F1-score versus threshold curve for each groundnut leaf disease class using the proposed EfficientNetB0-based model.



Fig 8 presents the classification outcomes of groundnut leaf samples using the proposed EfficientNetB0-based model. In the first row, the model accurately predicted a healthy leaf with 99.99% confidence, showing its ability to recognize disease-free foliage. It also correctly identified a Rosette-infected leaf with 99.98% confidence, characterized by leaf curling and discoloration. An Alternaria Leaf Spot case, showing distinct dark lesions, was also correctly predicted with high certainty (99.85%).

​

Fig 8: Sample prediction results for groundnut leaf disease classes showing actual and predicted labels with corresponding confidence scores.



In the second row, three additional Alternaria Leaf Spot cases were correctly classified with confidences ranging from 99.95% to 99.98%, demonstrating the model’s consistent ability to detect different symptom severities and lighting conditions. A second healthy sample was also accurately predicted with 99.97% confidence, reinforcing the model’s robustness in classifying non-infected leaves. In the final row, a Leaf Spot (early and late) case showing circular lesions was recognized with 99.99% confidence, indicating the model’s sensitivity to early-stage symptoms. Another Alternaria-infected leaf, despite dense clustering of lesions, was accurately classified with 99.95% confidence. Lastly, a leaf affected by Rust was identified correctly, supported by a high prediction confidence of 99.94%. Overall, these results highlight the model’s reliability and precision across diverse leaf diseases.

Table 3 presents a comparison of the presented work with the existing literature. The modified EfficientNetB0 model proposed in this study achieved an impressive accuracy of 99.22% using a real-world dataset collected under natural lighting conditions (Sasmal et al., 2024), surpassing the performance of several existing approaches. Abhilasha et al., (2023) used ResNet50 on images from the Durgapur Agriculture Research Centre and reported only 82.30% accuracy. Bama and Priyadharsini (2022) achieved higher accuracies, with SVM reaching 99.84%; however, their dataset included PlantVillage images, which may lack real-world variability. Rakholia et al., (2022) employed a CNN with progressive resizing and achieved 96.12% accuracy using a dataset from Gujarat. Similarly, Maheswaran et al., (2022) trained a CNN on multi-disease samples and reached 96.50% accuracy. Paramanandham et al., (2024) proposed the LeafNet architecture, which attained 97.23% on a large dataset of 10,361 images. Despite the effectiveness of these models, the customized EfficientNetB0 used in this study, with added dense layers, regularization and batch normalization, proved to be both accurate and lightweight. This makes it especially suitable for real-time field applications and mobile-based disease detection tools.

Table 3: Comparison of existing groundnut leaf disease detection models with the proposed model.



Furthermore, the proposed EfficientNetB0-based model demonstrates strong potential for real-time deployment in mobile-based disease detection applications. The model can be integrated with a user-friendly mobile interface, enabling farmers to capture leaf images and receive instant disease predictions. In addition, the classification output can be linked to an expert system module, which associates each detected disease with appropriate management recommendations, including pesticide application, cultural practices and preventive measures. Such integration would enhance the practical utility of the model by providing decision support to farmers, thereby facilitating timely intervention and improving crop productivity. This combined framework of deep learning and expert systems can serve as an efficient tool for precision agriculture and sustainable crop management.

CONCLUSION

This study proposed a DL model based on a modified EfficientNetB0 architecture for classifying groundnut leaf diseases. It achieved a high test accuracy of 99.22% across five disease categories. The use of regularization, dropout and batch normalization improved generalization and stability. Evaluation metrics such as accuracy and ROC curves confirmed the model’s strong and consistent performance. Compared to prior studies, the proposed model outperformed traditional CNN architectures, LeafNet and various transfer learning approaches. However, the study has limitations. The dataset was region-specific, which may affect model generalization in other agro-climatic zones. Disease severity levels were not differentiated and the model was not tested in real-time mobile environments, though it remains lightweight. Future work will address these limitations by collecting data from multiple regions and seasons. Disease severity scoring will be integrated to support early intervention. Additionally, real-time testing using mobile applications will be explored. Lightweight model optimization techniques will be applied to ensure the model runs efficiently on low-resource edge devices, making it suitable for field deployment.

ACKNOWLEDGEMENT

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.

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.
 
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.

CONFLICT OF INTEREST

The authors declare that they have no conflict of interest.

REFERENCES


  1. Abhilasha, A., Vyas, V. and Chaplot, N. (2023). Disease identification of groundnut leaves by deep learning technology. Neuro Quantology. 21(5): 891-892. https://doi.org/10.48047/ nq.2023.21.5.NQ222082.

  2. Agricultural and Processed Food Products Export Development Authority (APEDA). (2023). Indian Groundnut Market and Export Potential. https://apeda.gov.in.

  3. AlZubi, A.A. (2023). Artificial intelligence and its application in the prediction and diagnosis of animal diseases: A review. Indian Journal of Animal Research. 57(10): 1265-1271. doi: 10.18805/IJAR.BF-1684. 

  4. Anbumozhi, A. and Shanthini, A. (2024). GLDICCNN Model: Groundnut leaf diseases identification and classification for multiclass classification using deep learning. Journal of Advances in Information Technology. 15(7): 812-821. https://doi. org/10.12720/jait.15.7.812-821.

  5. Bama, K.A. and Priyadharsini, S.S. (2022). Groundnut Leaves and Their Disease, Deficiency and Toxicity Classification using a Machine Learning Approach. In Elsevier eBooks (pp. 253-275). https://doi.org/10.1016/b978-0-12-824410-4.00003-9.

  6. Cho, O.H. (2024). An evaluation of various machine learning approaches for detecting leaf diseases in agriculture. Legume Research. 47(4): 619-627. doi: 10.18805/LRF-787.

  7. Desfita, S., Sari, W., Wahyuni, D., Putri, F., Pramesyanti Pramono, A., Pato, U., Pratiwi, D., Grzelczyk, J. and Budryn, G. (2025). Synergistic effects of multi-strain probiotic and prebiotic combinations on immune recovery in aging populations. International Journal of Probiotics and Prebiotics. 20: 10-18. https://doi.org/10.37290/ijpp2641-7197.

  8. Food and Agriculture Organization. (2023). FAOSTAT: Crops and Livestock Products-Peanuts. Rome, Italy: FAO. n.p. Available at: https://www.fao.org/faostat/en/#data/QCL.

  9. Hai, N.T. and Duong, N.T. (2024). An improved environmental management model for assuring energy and economic prosperity. Acta Innovations. 52: 9-18. https://doi.org/ 10.62441/ActaInnovations.52.2.

  10. International Crops Research Institute for the Semi-Arid Tropics (ICRISAT). (2023). Groundnut: A Key Crop for Food and Income Security. https://www.icrisat.org.

  11. Janani, M. and Jebakumar, R. (2022). Detection and classification of groundnut leaf nutrient level extraction in RGB images. Advances in Engineering Software. 175: 103320. https:/ /doi.org/10.1016/j.advengsoft.2022.103320.

  12. Kishore, N. and Senthilvel, P.G. (2025). An efficient system to improve the accuracy of growth rate in Indian agriculture crops using Densenet and ResNet-50 algorithms. Melville, NY, USA: AIP Publishing. https://doi.org/10.1063/5.0264896.

  13. Maheswaran, S., Indhumathi, N., Dhanalakshmi, S., Nandita, S., Shafiq, I.M. and Rithka, P. (2022). Identification and classification of groundnut leaf disease using convolutional neural network. In IFIP Advances in Information and Communication Technology. pp. 251-270. https://doi. org/10.1007/978-3-031-16364-7_19.

  14. Mahto, I.C. and Mathew, J. (2025). Compact deep learning models for leaf disease classification and recognition in precision agriculture. Neural Computing and Applications. 37(22): 18379-18399. https://doi.org/10.1007/s00521-025-114 00-5.

  15. Maltare, N.N., Sharma, D. and Patel, S. (2023). An exploration and prediction of rainfall and groundwater level for the District of Banaskantha, Gujrat, India. International Journal of Environmental Sciences. 9(1): 1-17.

  16. MÿP.A. and Reddy, P. (2023). Ensemble of CNN models for classification of groundnut plant leaf disease detection. Smart Agricultural Technology. 6: 100362. https://doi.org/10.1016/j.atech.2023.100362.

  17. Mohammad, A.A.S., Mohammad, S.I., Vasudevan, A. and Heindric, A. (2026). Intelligent decision support system for crop management using data analytics. Journal of Experimental Biology and Agricultural Sciences. 14(3): 9-14. https://jebas.org/ojs/index.php/jebas/article/view/3796.

  18. Nayak, A. (2021). Growth and instability Analysis of groundnut production in India and Karnataka. Economic Affairs. 66(1). https://doi.org/10.46852/0424-2513.1.2021.8.

  19. North Eastern Development Finance Corporation Ltd. (NEDFi). (2024). Groundnut production overview-India and Northeast. Guwahati, India: NEDFi. n.p. https://databank.nedfi.com.

  20. Open Knowledge Repository-FAO. (2022). Crop Monitoring and Market Analysis for Legumes in Asia. https://open knowledge. fao.org.

  21. Özçelik, H., Saðlam, B. and Öcal, A.S. (2026). New natural methods for combating mold fungi in agricultural products: The Isparta example. Bulletin of Pure and Applied Sciences- Botany. 45B(2): 1-17. https://bpasjournals.com/botany/ index. php/journal/ article/view/174/150.

  22. Paek, J.J., Kim, J., Bae, S., Paek, K. and Lee, Y. (2026). Chive- fortified fermentation enhances gamma-aminobutyric acid production by Levilactobacillus brevis PL9014. International Journal of Probiotics and Prebiotics. 21(1): 1-6. https://doi.org/10.37290/ijpp.v21i1.12.

  23. Paramanandham, N., Sundhar, S. and Priya, P. (2024). Enhancing disease detection with weight initialization and residual connections using LeafNet for groundnut leaf diseases. IEEE Access. 12: 91511-91526. https://doi.org/10.1109/ access.2024.3422311.

  24. Pooniya, S.K., Yadav, S., Rathore, M.S., Tiwari, S., Sikarwar, R.S. and Tripathi, M.K. (2020). Field evaluation of early and late leaf spot diseases in advanced breeding lines of groundnut (Arachis hypogaea L.). International Journal of Current Microbiology and Applied Sciences. 9(7): 3910-3919. https://doi.org/10.20546/ijcmas. 2020.907.458.

  25. Rakholia, R.M., Tailor, J.H., Saini, J.R., Kaur, J. and Pahuja, H. (2022). Groundnuts leaf disease recognition using neural network with progressive resizing. International Journal of Advanced Computer Science and Applications. 13(6): 83-88. https://doi.org/10.14569/ijacsa.2022.0130611.

  26. ReportLinker (2023). Global Peanut Market Report: Trends, forecasts and Insights. New York, USA: ReportLinker. n.p. https://www.reportlinker.com

  27. Sasmal, B., Das, A., Dhal, K.G., Saheb, S.B., Khurma, R.A. and Castillo, P.A. (2024). A novel groundnut leaf dataset for detection and classification of groundnut leaf diseases. Data in Brief. 55: 110763. https://doi.org/10.1016/j.dib. 2024.110763.

  28. Semara, I.M.T., Sunarta, I.N., Antara, M., Arida, I.N.S. and Wirawan, P.E. (2024). Tourism sites and environmental reservation. International Journal of Environmental Sciences. 10(1): 44-55.

  29. Shasidhar, Y., Variath, M.T., Vishwakarma, M.K., Manohar, S.S., Gangurde, S.S., Sriswathi, M., Sudini, H.K. et al. (2019). Improvement of three popular Indian groundnut varieties for foliar disease resistance and high oleic acid using SSR markers and SNP array in marker-assisted backcrossing. The Crop Journal. 8(1): 1-15. https://doi.org/10.1016/ j.cj.2019.07.001.

  30. Sivaganesan, D. (2023). Deep learning approaches for disease detection in groundnut crops using CNN models. Journal of Soft Computing Paradigm. 5(4): 404-416. https://doi. org/10.36548/jscp.2023.4.006.

  31. Souza, K.F., Gonçalves, L.C.O. and Espindola, F.S. (2026). A broad overview of the multifaceted royal jelly and its applicability in health and performance. Current Topics in Nutraceutical Research. 24(1): 4-14. https://doi.org/10.37290/ctnr. v24i1.8.

  32. Sriram, P. and Kumari, Y.S. (2026). Game theory in urban birds: A Nash equilibrium analysis using crows and pigeons. Bio-Science Research Bulletin. 42(1): 1-6. https://bpas journals.com/life-sciences/index.php/journal/article/view/150/140.

  33. Vardhan, B.B.S.H., Paluvadi, H., Srihari, V.K.S. and R.CB. (2024). Disease Detection in Groundnut Crop using Deep Learning Models. In Proceedings of the 2024 4th International Conference on Artificial Intelligence and Signal Processing (AISP). pp. 1-5. IEEE. https://doi.org/10.1109/AISP61711. 2024.10870722.
Published In
Legume Research
Article Metrics
Metrics Icon
64
Views
0
Citations
Reviewed By
Share Article
Download Article
In this Article
    Contact us
    Follow us

    Full Research Article

    A Modified EfficientNetB0-based Deep Learning Model for Accurate Detection and Classification of Groundnut Leaf Diseases

    J
    Jie-Shin Lin1
    Y
    Yu-Nan Tai2
    S
    Suh Chen Hsiao3
    L
    Luke H.C. Hsiao1,*
    1Department of Public Policy and Management, I-Shou University, Taiwan.
    2Program in Intelligent Tourism and Hospitality Management, I-Shou University, Taiwan.
    3Director and Professor of Social Work Practicum Education, University of Southern California Suzanne Dworak-Peck School of Social Work, USA.

    • Submitted18-03-2026|

    • Accepted20-08-2026|

    • First Online 29-08-2026|

    • doi 10.18805/LRF-948

    ABSTRACT

    Background: Groundnut farming is affected by several leaf diseases that reduce crop yield. Farmers often rely on visual inspection, which can be inaccurate and time-consuming. Early and precise identification of leaf diseases is essential for effective crop management. This study presents a deep learning-based solution to automate disease detection.

    Methods: A modified EfficientNetB0 architecture is proposed for classifying five types of groundnut leaf conditions: healthy, leaf spot (early and late), alternaria leaf spot, rust and rosette. The dataset is sourced from the Mendeley database that was collected from Ramchandrapur village in West Bengal, India, under natural lighting. A total of 1,720 images were captured using a DSLR camera. After verification and cleaning, the dataset was split into 1,204 training and 516 testing images. All images were resized, normalized and label-encoded. Data augmentation techniques such as rotation, flipping and zoom were used to improve generalization. Regularization was applied to reduce overfitting. The model was trained for 100 epochs using the RMSprop optimizer and early stopping.

    Result: The model achieved a test accuracy of 99.22%. Evaluation metrics confirm high performance across all classes. The model outperformed existing methods such as ResNet50 (82.3%), CNN with progressive resizing (96.12%) and LeafNet (97.23%). It also maintained low training and validation loss throughout training. These results highlight the model’s robustness, accuracy and potential for real-time field applications. The approach is lightweight and suitable for mobile-based disease detection tools.

    INTRODUCTION

    Groundnut (Arachis hypogaea) is an important legume cultivated across tropical and subtropical regions. It serves as a rich source of edible oil, comprising about 45-50% of the seed and protein (»25%) (ICRISAT, 2023; FAO, 2023). Globally, groundnut is grown on approximately 32.7 million hectares, producing nearly 53.9 million tonnes annually (FAO, 2023). China and India are the two largest producers of groundnuts. In 2023, China produced around 18.4 million tonnes, followed by India with approximately 7.38 million tonnes (ReportLinker, 2023). China, India and Nigeria together produced nearly 60% of the world’s groundnuts during 1961-2022 (Nayak, 2021). The estimated production for 2023-24 in India reached 8.66 million tonnes (NEDFi, 2024).

    Groundnut is vital to Indian agriculture. It is cultivated in both Kharif (May-June) and Rabi (October) seasons and contributes significantly to vegetable oil production (FAO, 2023; APEDA, 2023). Gujarat is the largest producing state (»42%), followed by Rajasthan (17%), Tamil Nadu (11%), Andhra Pradesh (9%) and Karnataka (6%) (APEDA, 2023). Despite its economic and nutritional importance, India’s average yield (~1.3 t/ha) remains low compared to the global average (ICRISAT, 2023; ReportLinker, 2023). Constraints such as erratic rainfall, limited irrigation and the prevalence of foliar diseases affect productivity.

    Among biotic stresses, foliar diseases significantly reduce groundnut yield and quality. The most common diseases include early and late leaf spot, Alternaria leaf spot, Rust and Rosette (Pooniya et al., 2020; Shasidhar et al., 2019). Early leaf spot occurs one month after sowing and appears as circular lesions that later darken. Late leaf spot appears after 6-7 weeks, showing necrotic lesions with faint halos. Alternaria leaf spot develops after 5-6 weeks and causes water-soaked chlorotic spots and leaf curling. Rust typically emerges six weeks post-sowing and presents as reddish-brown pustules. Rosette disease occurs within 2-3 weeks of sowing and causes yellowing, curling and stunted growth, giving leaves a rosette-like structure (Sasmal et al., 2024).

    These diseases disrupt photosynthesis and reduce both seed quality and marketable yield. Currently, field inspection is the standard approach for diagnosis, but it is time-consuming, labor-intensive and prone to subjectivity. Misdiagnosis can lead to improper treatment and crop loss. In recent years, artificial intelligence (AI) and deep learning (DL) have gained attention as scalable tools for automated plant disease detection (Cho, 2024; AlZubi, 2023). CNNs are capable of classifying leaf diseases with high precision. Transfer learning, using architectures such as EfficientNet, ResNet and DenseNet, enables high performance even on moderately sized datasets (Anbumozhi and Shanthini, 2024; Sivaganesan, 2023).

    Recent advances in DL methods for agriculture show promising results. Models like ResNet, DenseNet and EfficientNet have been used for leaf classification tasks across various crops (Mahto and Mathew, 2025; Desfita et al., 2025). Transfer learning, which uses pre-trained models on large-scale datasets such as ImageNet, has made it possible to achieve high accuracy even with moderately sized agricultural datasets (Kishore and Senthilvel, 2025; Hai and Duong, 2024; Semara et al., 2024; Maltare et al., 2023). M.P and Reddy, 2023 proposed an ensemble method for groundnut leaf disease classification using a tri-CNN architecture composed of pre-trained DenseNet169, Inception and Xception models. These base models were trained on the ImageNet dataset (Mohammad et al., 2026; Özçelik et al.,  2026; Paek et al., 2026; Souza et al., 2026; Sriram and Kumari, 2026). The method was tested on a real-world groundnut leaf dataset and achieved a high classification accuracy of 98.46%, demonstrating the effectiveness of ensemble deep learning in plant disease detection. Vardhan et al.  (2024) evaluated three deep learning models, CNN, MobileNetV2 and InceptionResNetV2, for groundnut disease detection. Janani and Jebakumar (2022) developed a system called Nutrient Range Analysis based on Greenness (NRAG) to detect and classify nitrogen levels in groundnut leaves using RGB images. They applied a median filter to reduce noise and extracted greenness features, which were directly linked to nitrogen content. These features were fed into a CNN-based Holding Vector Network (HVN) model. The model classified leaves into three nitrogen categories: low, normal and excess. It achieved 95% training accuracy and 92% validation accuracy, offering a scalable and cost-effective solution for real-time use by farmers.

    In this study, we propose a modified EfficientNetB0 model for the multi-class classification of groundnut leaf diseases. This work utilizes a dataset from the Mendeley Data Repository, comprising field-collected images that provide a practical basis for model training. The EfficientNetB0 architecture is enhanced with additional dense layers, batch normalization and regularization to improve performance. The model is evaluated using key metrics and is designed with scalability in mind, making it suitable for mobile-based plant health applications. This work contributes to the development of intelligent decision-support tools. These tools can assist groundnut farmers in detecting diseases at an early stage. Early detection supports timely intervention and improved crop management. The approach also promotes sustainable agriculture, especially in regions with limited resources.

    MATERIALS AND METHODS

    This study uses a publicly available dataset of groundnut leaf images obtained from the Mendeley Data Repository (Sasmal et al., 2024). The images were captured under real-world agricultural conditions to reflect natural variation in appearance. Data collection was conducted in Ramchandrapur village, located in the Purba Medinipur district of West Bengal, India. A DSLR camera was used to capture high-resolution images, each with a size of 4624 × 3472 pixels. All images were taken in natural lighting conditions without artificial enhancements. To introduce variability, photographs were captured at different times of day and from multiple plants. This helped incorporate changes in illumination, leaf angles and background complexity, making the dataset more robust and realistic. In total, 1,720 images were collected. The dataset covers five distinct classes: healthy, leaf spot (early and late), alternaria leaf spot, rust and rosette. Each image was organized into one of five class-specific folders. The folder names correspond exactly to the disease categories. All images underwent manual verification. Each file was reviewed by trained observers to confirm the accuracy of its classification. Images that were misclassified, blurry, or duplicated were removed to maintain dataset quality. The resulting dataset serves as a reliable foundation for training deep learning models. A few images from the dataset are given in Fig 1.

    Fig 1: Images from the dataset for training and testing of the model.


     
    Image preprocessing, label encoding and augmentation
     
    Each image in the dataset was resized to 224 x 224 pixels. This resizing ensured uniform input dimensions for the model. Pixel values were then normalized to a range between 0 and 1. This was done by dividing each pixel intensity by 255. Label encoding was applied using the LabelEncoder tool from Scikit-learn. This converted the class names from strings to numerical values. The dataset was then split into training and testing sets using a 70:30 ratio. As a result, 1,204 images were used for training and 516 for testing. To improve model generalization, data augmentation techniques were applied. These included random rotation up to 20 degrees. Width and height were shifted by up to 10%. The zoom range was set between 0.8 and 1.2. A shear transformation of 0.2 was used. Horizontal flipping was also applied. The fill mode was set to ‘nearest’ to handle missing pixels after transformation. These techniques introduced variation in image orientation and scale. This helped the model learn more robust features and reduced the risk of overfitting.
     
    Model architecture
     
    In this paper, the proposed model is based on the EfficientNetB0 architecture (Fig 2). This model is well known for its balance between performance and computational efficiency. EfficientNetB0 was used as the feature extractor in the present system. It was initialized with ImageNet pre-trained weights. The top layer of the base model was removed to allow customization for the classification task. An input layer was defined to accept images of size 224 x 224 x 3. This matched the size used during preprocessing. The input was passed through the EfficientNetB0 base. The output from the base model was a 1,280-dimensional feature vector. A custom classifier head was added after the base model. This head consisted of three dense (fully connected) layers. The first dense layer had 128 units and used the ReLU activation function. It was followed by Batch Normalization and a Dropout layer with a dropout rate of 0.3. The second dense layer had 64 units. Like the first, it also used ReLU, batch normalization and dropout. The third dense layer had 32 units and followed the same structure.

    Fig 2: Working procedure of the model with architecture of the proposed groundnut leaf disease classification model using EfficientNetB0.



    L1 regularization was applied to each dense layer. The regularization parameter was set to 0.0001. This helped prevent overfitting by penalizing large weights. The final output layer was a dense layer with 5 units, one for each class. This layer used a Softmax activation function. It converted the output vector into class probabilities. Mathematically, for an input image x, the final probability output is given by
     
     
    Where,
    zc  = Logit for class c.
    K=5 = Total number of classes.

    The total number of parameters in the model was 4,224,929. Among these, 4,182,465 parameters were trainable. The rest were non-trainable, as they belonged to the frozen layers of the pre-trained base model. This architecture was built using the Keras Functional API. The use of batch normalization and dropout after each dense layer made the model more stable and robust during training. This architecture was chosen for its ability to extract detailed features and to generalize well with limited data.
     
    Model training and evaluation
     
    The model was compiled using the RMSprop optimizer. The learning rate was set to 0.0001 and the loss function used was Sparse Categorical Crossentropy, as labels were encoded as integers. The training process used an early stopping callback to monitor the validation loss. If no improvement was observed for 10 consecutive epochs, training was halted and the best weights were restored. The model was trained for a maximum of 100 epochs using a batch size of 32. The model was evaluated using four performance metrics. These are defined as follows.
     




     
      
    Where,
    TP = True positives.
    FP = False positives.
    FN = False negatives.

    After training, the model was saved in HDF5 format for future inference or fine-tuning. Table 1 summarizes hyperparameters used for training the efficientB0 model. 

    Table 1: Hyperparameters used for model training.

    RESULTS AND DISCUSSION

    The proposed model was trained for 100 epochs using the EfficientNetB0 backbone and custom dense layers (Fig 3). The training and validation performance improved steadily over time. In the initial phase (Epoch 1), the training accuracy was low at 19.86% and the loss was high at 2.9853. The validation accuracy at this point was 20.35%, with a loss of 2.3289. By Epoch 25, the model had significantly improved. It achieved a training accuracy of 98.28% and a validation accuracy of 98.64%. Corresponding training and validation losses were 0.6777 and 0.6316, respectively. This shows that the model was able to learn useful features and generalize well by the mid-point of training.

    Fig 3: Training and validation accuracy and loss curves over 100 epochs.



    At Epoch 50, the model continued to perform well. The training accuracy reached 99.75% and the validation accuracy improved to 98.84%. The training loss reduced to 0.4466 and the validation loss dropped to 0.4522. This indicated that the model was not overfitting and maintained stability across epochs. By Epoch 75, the training accuracy was 99.66% and validation accuracy remained at 98.84%. Training and validation losses further reduced to 0.3011 and 0.3186, respectively. This confirmed consistent performance over time. At the final epoch (Epoch 100), the model reached 99.95% accuracy on the training set. The validation accuracy peaked at 99.22%, showing excellent generalization. Final losses were 0.1818 for training and 0.1963 for validation. The gap between training and validation losses remained small, which indicates minimal overfitting. Overall, the model demonstrated strong learning capacity and robustness. The EfficientNetB0 backbone with the added dense layers contributed significantly to high accuracy and low validation loss. The early stopping callback helped prevent overfitting and ensured optimal weight retention.

    The confusion matrix in Fig 4 highlights the performance of the proposed EfficientNetB0-based model on the groundnut leaf disease classification task. Each row represents the actual class and each column indicates the predicted class. The model correctly classified 130 samples of Alternaria Leaf Spot without any misclassifications. All 179 healthy leaf images were accurately identified. For the Leaf Spot (Early and Late) category, 139 images were correctly classified, while 3 were mistakenly predicted as Alternaria Leaf Spot. In the case of Rosette disease, 20 images were correctly classified and only one was misclassified as Healthy. Rust disease was perfectly predicted for all 44 samples. These results demonstrate strong class-wise performance, with very few errors occurring mainly between visually similar diseases. Overall, the model exhibits high reliability in distinguishing between multiple groundnut leaf diseases.

    ​

    Fig 4: Confusion matrix showing true and false classifications for five groundnut leaf categories.



    The classification report further validates the effectiveness of the proposed model across all five groundnut leaf classes (Table 2). For Alternaria Leaf Spot, the model achieved a precision of 0.9774 and a perfect recall of 1.0, resulting in an F1-score of 0.9886 across 130 test images. Healthy leaves were classified with extremely high precision (0.9944) and perfect recall, yielding an F1-score of 0.9972 over 179 samples. The Leaf Spot (Early and Late) class recorded a precision of 1.0 and a recall of 0.9789, producing an F1-score of 0.9893 for 142 samples. In the Rosette category, the model achieved perfect precision but a slightly lower recall of 0.9524, which led to an F1-score of 0.9756 for 21 images. Rust classification was flawless with 1.0 scores across all metrics for 44 samples. The model attained an overall accuracy of 99.22% on the test dataset. The macro average precision, recall and F1-score were 0.9944, 0.9863 and 0.9901, respectively, indicating balanced performance across all classes. The weighted average, which accounts for the number of samples per class, remained consistent at 0.9922 for both recall and F1-score. These results confirm that the model is robust, consistent and well-suited for multi-class classification of groundnut leaf diseases.

    Table 2: Classification report for each groundnut leaf disease class using the proposed EfficientNetB0-based model.



    The ROC curve in the image shows excellent model performance for all classes (Fig 5). Each class achieves an AUC close to or equal to 1.0. Specifically, Healthy, Rosette and Rust have an AUC of 1.0000, indicating perfect class separation. Alternaria Leaf Spot and Leaf Spot (Early and Late) also perform well, with AUC values of 0.9999. The curves rise sharply toward the top-left corner, confirming high true positive rates and low false positive rates across all categories.

    ​

    Fig 5: Multi-class ROC curve illustrating AUC values for each groundnut leaf class.



    The precision-recall (PR) curve shows that the model performs well across all classes (Fig 6). Average precision (AP) for Healthy, Rosette and Rust is 1.0000, indicating perfect balance between precision and recall. Alternaria Leaf Spot has an AP of 0.9998 and Leaf Spot (Early and Late) achieves 0.9997. The curves remain close to the top-right corner, reflecting high confidence in predictions and minimal false positives or false negatives.

    ​

    Fig 6: Precision-recall curve showing average precision (AP) values for each class.



    The F1-score vs. threshold plot demonstrates stable and high F1-scores for all five classes across most threshold values (Fig 7). Scores remain above 0.9 between thresholds of 0.1 and 0.95, showing the model maintains a good balance between precision and recall. A sharp drop is observed near thresholds 0 and 1, which is expected. All classes, Alternaria Leaf Spot, Healthy, Leaf Spot (Early and Late), Rosette and Rust, exhibit similar trends, confirming model consistency.

    ​

    Fig 7: F1-score versus threshold curve for each groundnut leaf disease class using the proposed EfficientNetB0-based model.



    Fig 8 presents the classification outcomes of groundnut leaf samples using the proposed EfficientNetB0-based model. In the first row, the model accurately predicted a healthy leaf with 99.99% confidence, showing its ability to recognize disease-free foliage. It also correctly identified a Rosette-infected leaf with 99.98% confidence, characterized by leaf curling and discoloration. An Alternaria Leaf Spot case, showing distinct dark lesions, was also correctly predicted with high certainty (99.85%).

    ​

    Fig 8: Sample prediction results for groundnut leaf disease classes showing actual and predicted labels with corresponding confidence scores.



    In the second row, three additional Alternaria Leaf Spot cases were correctly classified with confidences ranging from 99.95% to 99.98%, demonstrating the model’s consistent ability to detect different symptom severities and lighting conditions. A second healthy sample was also accurately predicted with 99.97% confidence, reinforcing the model’s robustness in classifying non-infected leaves. In the final row, a Leaf Spot (early and late) case showing circular lesions was recognized with 99.99% confidence, indicating the model’s sensitivity to early-stage symptoms. Another Alternaria-infected leaf, despite dense clustering of lesions, was accurately classified with 99.95% confidence. Lastly, a leaf affected by Rust was identified correctly, supported by a high prediction confidence of 99.94%. Overall, these results highlight the model’s reliability and precision across diverse leaf diseases.

    Table 3 presents a comparison of the presented work with the existing literature. The modified EfficientNetB0 model proposed in this study achieved an impressive accuracy of 99.22% using a real-world dataset collected under natural lighting conditions (Sasmal et al., 2024), surpassing the performance of several existing approaches. Abhilasha et al., (2023) used ResNet50 on images from the Durgapur Agriculture Research Centre and reported only 82.30% accuracy. Bama and Priyadharsini (2022) achieved higher accuracies, with SVM reaching 99.84%; however, their dataset included PlantVillage images, which may lack real-world variability. Rakholia et al., (2022) employed a CNN with progressive resizing and achieved 96.12% accuracy using a dataset from Gujarat. Similarly, Maheswaran et al., (2022) trained a CNN on multi-disease samples and reached 96.50% accuracy. Paramanandham et al., (2024) proposed the LeafNet architecture, which attained 97.23% on a large dataset of 10,361 images. Despite the effectiveness of these models, the customized EfficientNetB0 used in this study, with added dense layers, regularization and batch normalization, proved to be both accurate and lightweight. This makes it especially suitable for real-time field applications and mobile-based disease detection tools.

    Table 3: Comparison of existing groundnut leaf disease detection models with the proposed model.



    Furthermore, the proposed EfficientNetB0-based model demonstrates strong potential for real-time deployment in mobile-based disease detection applications. The model can be integrated with a user-friendly mobile interface, enabling farmers to capture leaf images and receive instant disease predictions. In addition, the classification output can be linked to an expert system module, which associates each detected disease with appropriate management recommendations, including pesticide application, cultural practices and preventive measures. Such integration would enhance the practical utility of the model by providing decision support to farmers, thereby facilitating timely intervention and improving crop productivity. This combined framework of deep learning and expert systems can serve as an efficient tool for precision agriculture and sustainable crop management.

    CONCLUSION

    This study proposed a DL model based on a modified EfficientNetB0 architecture for classifying groundnut leaf diseases. It achieved a high test accuracy of 99.22% across five disease categories. The use of regularization, dropout and batch normalization improved generalization and stability. Evaluation metrics such as accuracy and ROC curves confirmed the model’s strong and consistent performance. Compared to prior studies, the proposed model outperformed traditional CNN architectures, LeafNet and various transfer learning approaches. However, the study has limitations. The dataset was region-specific, which may affect model generalization in other agro-climatic zones. Disease severity levels were not differentiated and the model was not tested in real-time mobile environments, though it remains lightweight. Future work will address these limitations by collecting data from multiple regions and seasons. Disease severity scoring will be integrated to support early intervention. Additionally, real-time testing using mobile applications will be explored. Lightweight model optimization techniques will be applied to ensure the model runs efficiently on low-resource edge devices, making it suitable for field deployment.

    ACKNOWLEDGEMENT

    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.

    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.
     
    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.

    CONFLICT OF INTEREST

    The authors declare that they have no conflict of interest.

    REFERENCES


    1. Abhilasha, A., Vyas, V. and Chaplot, N. (2023). Disease identification of groundnut leaves by deep learning technology. Neuro Quantology. 21(5): 891-892. https://doi.org/10.48047/ nq.2023.21.5.NQ222082.

    2. Agricultural and Processed Food Products Export Development Authority (APEDA). (2023). Indian Groundnut Market and Export Potential. https://apeda.gov.in.

    3. AlZubi, A.A. (2023). Artificial intelligence and its application in the prediction and diagnosis of animal diseases: A review. Indian Journal of Animal Research. 57(10): 1265-1271. doi: 10.18805/IJAR.BF-1684. 

    4. Anbumozhi, A. and Shanthini, A. (2024). GLDICCNN Model: Groundnut leaf diseases identification and classification for multiclass classification using deep learning. Journal of Advances in Information Technology. 15(7): 812-821. https://doi. org/10.12720/jait.15.7.812-821.

    5. Bama, K.A. and Priyadharsini, S.S. (2022). Groundnut Leaves and Their Disease, Deficiency and Toxicity Classification using a Machine Learning Approach. In Elsevier eBooks (pp. 253-275). https://doi.org/10.1016/b978-0-12-824410-4.00003-9.

    6. Cho, O.H. (2024). An evaluation of various machine learning approaches for detecting leaf diseases in agriculture. Legume Research. 47(4): 619-627. doi: 10.18805/LRF-787.

    7. Desfita, S., Sari, W., Wahyuni, D., Putri, F., Pramesyanti Pramono, A., Pato, U., Pratiwi, D., Grzelczyk, J. and Budryn, G. (2025). Synergistic effects of multi-strain probiotic and prebiotic combinations on immune recovery in aging populations. International Journal of Probiotics and Prebiotics. 20: 10-18. https://doi.org/10.37290/ijpp2641-7197.

    8. Food and Agriculture Organization. (2023). FAOSTAT: Crops and Livestock Products-Peanuts. Rome, Italy: FAO. n.p. Available at: https://www.fao.org/faostat/en/#data/QCL.

    9. Hai, N.T. and Duong, N.T. (2024). An improved environmental management model for assuring energy and economic prosperity. Acta Innovations. 52: 9-18. https://doi.org/ 10.62441/ActaInnovations.52.2.

    10. International Crops Research Institute for the Semi-Arid Tropics (ICRISAT). (2023). Groundnut: A Key Crop for Food and Income Security. https://www.icrisat.org.

    11. Janani, M. and Jebakumar, R. (2022). Detection and classification of groundnut leaf nutrient level extraction in RGB images. Advances in Engineering Software. 175: 103320. https:/ /doi.org/10.1016/j.advengsoft.2022.103320.

    12. Kishore, N. and Senthilvel, P.G. (2025). An efficient system to improve the accuracy of growth rate in Indian agriculture crops using Densenet and ResNet-50 algorithms. Melville, NY, USA: AIP Publishing. https://doi.org/10.1063/5.0264896.

    13. Maheswaran, S., Indhumathi, N., Dhanalakshmi, S., Nandita, S., Shafiq, I.M. and Rithka, P. (2022). Identification and classification of groundnut leaf disease using convolutional neural network. In IFIP Advances in Information and Communication Technology. pp. 251-270. https://doi. org/10.1007/978-3-031-16364-7_19.

    14. Mahto, I.C. and Mathew, J. (2025). Compact deep learning models for leaf disease classification and recognition in precision agriculture. Neural Computing and Applications. 37(22): 18379-18399. https://doi.org/10.1007/s00521-025-114 00-5.

    15. Maltare, N.N., Sharma, D. and Patel, S. (2023). An exploration and prediction of rainfall and groundwater level for the District of Banaskantha, Gujrat, India. International Journal of Environmental Sciences. 9(1): 1-17.

    16. MÿP.A. and Reddy, P. (2023). Ensemble of CNN models for classification of groundnut plant leaf disease detection. Smart Agricultural Technology. 6: 100362. https://doi.org/10.1016/j.atech.2023.100362.

    17. Mohammad, A.A.S., Mohammad, S.I., Vasudevan, A. and Heindric, A. (2026). Intelligent decision support system for crop management using data analytics. Journal of Experimental Biology and Agricultural Sciences. 14(3): 9-14. https://jebas.org/ojs/index.php/jebas/article/view/3796.

    18. Nayak, A. (2021). Growth and instability Analysis of groundnut production in India and Karnataka. Economic Affairs. 66(1). https://doi.org/10.46852/0424-2513.1.2021.8.

    19. North Eastern Development Finance Corporation Ltd. (NEDFi). (2024). Groundnut production overview-India and Northeast. Guwahati, India: NEDFi. n.p. https://databank.nedfi.com.

    20. Open Knowledge Repository-FAO. (2022). Crop Monitoring and Market Analysis for Legumes in Asia. https://open knowledge. fao.org.

    21. Özçelik, H., Saðlam, B. and Öcal, A.S. (2026). New natural methods for combating mold fungi in agricultural products: The Isparta example. Bulletin of Pure and Applied Sciences- Botany. 45B(2): 1-17. https://bpasjournals.com/botany/ index. php/journal/ article/view/174/150.

    22. Paek, J.J., Kim, J., Bae, S., Paek, K. and Lee, Y. (2026). Chive- fortified fermentation enhances gamma-aminobutyric acid production by Levilactobacillus brevis PL9014. International Journal of Probiotics and Prebiotics. 21(1): 1-6. https://doi.org/10.37290/ijpp.v21i1.12.

    23. Paramanandham, N., Sundhar, S. and Priya, P. (2024). Enhancing disease detection with weight initialization and residual connections using LeafNet for groundnut leaf diseases. IEEE Access. 12: 91511-91526. https://doi.org/10.1109/ access.2024.3422311.

    24. Pooniya, S.K., Yadav, S., Rathore, M.S., Tiwari, S., Sikarwar, R.S. and Tripathi, M.K. (2020). Field evaluation of early and late leaf spot diseases in advanced breeding lines of groundnut (Arachis hypogaea L.). International Journal of Current Microbiology and Applied Sciences. 9(7): 3910-3919. https://doi.org/10.20546/ijcmas. 2020.907.458.

    25. Rakholia, R.M., Tailor, J.H., Saini, J.R., Kaur, J. and Pahuja, H. (2022). Groundnuts leaf disease recognition using neural network with progressive resizing. International Journal of Advanced Computer Science and Applications. 13(6): 83-88. https://doi.org/10.14569/ijacsa.2022.0130611.

    26. ReportLinker (2023). Global Peanut Market Report: Trends, forecasts and Insights. New York, USA: ReportLinker. n.p. https://www.reportlinker.com

    27. Sasmal, B., Das, A., Dhal, K.G., Saheb, S.B., Khurma, R.A. and Castillo, P.A. (2024). A novel groundnut leaf dataset for detection and classification of groundnut leaf diseases. Data in Brief. 55: 110763. https://doi.org/10.1016/j.dib. 2024.110763.

    28. Semara, I.M.T., Sunarta, I.N., Antara, M., Arida, I.N.S. and Wirawan, P.E. (2024). Tourism sites and environmental reservation. International Journal of Environmental Sciences. 10(1): 44-55.

    29. Shasidhar, Y., Variath, M.T., Vishwakarma, M.K., Manohar, S.S., Gangurde, S.S., Sriswathi, M., Sudini, H.K. et al. (2019). Improvement of three popular Indian groundnut varieties for foliar disease resistance and high oleic acid using SSR markers and SNP array in marker-assisted backcrossing. The Crop Journal. 8(1): 1-15. https://doi.org/10.1016/ j.cj.2019.07.001.

    30. Sivaganesan, D. (2023). Deep learning approaches for disease detection in groundnut crops using CNN models. Journal of Soft Computing Paradigm. 5(4): 404-416. https://doi. org/10.36548/jscp.2023.4.006.

    31. Souza, K.F., Gonçalves, L.C.O. and Espindola, F.S. (2026). A broad overview of the multifaceted royal jelly and its applicability in health and performance. Current Topics in Nutraceutical Research. 24(1): 4-14. https://doi.org/10.37290/ctnr. v24i1.8.

    32. Sriram, P. and Kumari, Y.S. (2026). Game theory in urban birds: A Nash equilibrium analysis using crows and pigeons. Bio-Science Research Bulletin. 42(1): 1-6. https://bpas journals.com/life-sciences/index.php/journal/article/view/150/140.

    33. Vardhan, B.B.S.H., Paluvadi, H., Srihari, V.K.S. and R.CB. (2024). Disease Detection in Groundnut Crop using Deep Learning Models. In Proceedings of the 2024 4th International Conference on Artificial Intelligence and Signal Processing (AISP). pp. 1-5. IEEE. https://doi.org/10.1109/AISP61711. 2024.10870722.
    In this Article
      Contact us
      Follow us
      Published In
      Legume Research

      Editorial Board

      View all (0)