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

Full Research Article

Deep Learning-driven Diagnosis of Major Faba Bean Leaf Diseases: EfficientNetB0 Classification Approach

K
Kuan-Hung Chen1,*
1College of Tourism and Geography, Shaoguan University, Shaoguan 512005, China.

  • Submitted23-12-2025|

  • Accepted16-04-2026|

  • First Online 24-04-2026|

  • doi 10.18805/LRF-925

ABSTRACT

Background: Faba bean production is highly vulnerable to foliar diseases that often develop rapidly under field conditions and significantly reduce yield quality and quantity. Traditional disease diagnosis relies on visual assessment, which is slow, subjective and unsuitable for large-scale monitoring. Advances in computer vision and deep learning provide new opportunities for fast, accurate and automated disease recognition directly from leaf images.

Methods: This study proposes a deep learning-based diagnostic workflow using the EfficientNetB0 architecture for multi-class classification of faba bean leaf diseases. A field-collected dataset of 8,021 RGB images representing healthy leaves, rust, gall and chocolate spot infections was used. The dataset underwent rigorous quality control, stratified splitting and standardized preprocessing, followed by data augmentation to improve generalization. The model was trained using transfer learning with the Adam optimizer and evaluated using accuracy, precision, recall, F1-score, ROC-AUC and precision–recall metrics on an independent test set.

Result: The proposed model demonstrated high predictive performance, achieving 95.39% accuracy with consistently strong class-wise metrics. ROC-AUC values were nearly perfect (0.99-1.00), indicating excellent separability among disease categories. Precision–recall curves further verified the model’s robustness in identifying visually similar disease symptoms. Qualitative prediction samples showed high confidence levels, confirming stable behavior under diverse real-world imaging conditions. The results establish EfficientNetB0 as an effective and computationally efficient framework for automatic faba bean leaf disease classification. Its high accuracy, strong discriminative capability and reliability under natural field conditions make it a promising candidate for integration into mobile applications, decision-support systems and precision agriculture workflows aimed at early disease detection and timely crop management.

INTRODUCTION

Faba bean is an important legume crop cultivated across Asia, Africa and Europe, contributing significantly to global food and fodder systems. It is valued for its high protein content (20-35%), ability to fix atmospheric nitrogen and role in improving soil fertility, making it essential for sustainable cropping systems (Karkanis et al., 2018; Dhull et al., 2021; Semba et al., 2021). However, productivity is frequently constrained by destructive foliar diseases such as rust, chocolate spot and gall infections, which can cause yield losses ranging from 20-60% under favorable environmental conditions (De Notaris et al., 2023; Soliman et al., 2023). These diseases spread rapidly in humid field environments, making timely detection crucial for preventing widespread crop damage (Bitew et al., 2022).
       
Conventional disease diagnosis in faba bean fields depends on manual visual assessment, which is labor-intensive, subjective and often inconsistent across observers (Segers et al., 2021; Manjunatha et al., 2022). With the increasing availability of affordable smartphones, digital imaging and advances in artificial intelligence, deep learning has emerged as a reliable and scalable solution for automated disease identification (Ghuriani et al., 2023; Al-Sharqi et al., 2025; Alshahrani, 2024). Recent studies using convolutional neural networks (CNNs) and transfer learning have demonstrated strong potential across crops such as wheat, tomato, maize and rice (Chen et al., 2020; Archana and Jeevaraj, 2024; Dolatabadian et al., 2024). Despite these advancements, many existing models are developed using laboratory-controlled images, small datasets, or photographs captured under unrealistic backgrounds, which limits their applicability in real-world agricultural environments (Shafik et al., 2025).
       
Several studies have been conducted on leaf disease detection using deep-learning methods. Jain and Aneja (2025) applied a fine-tuned InceptionV3 model to a Kaggle dataset of 990 bean leaf images, achieving 91% accuracy, which illustrates the effectiveness of deep learning for disease classification. Salau et al., (2023) developed an end-to-end CNN model to detect major faba bean diseases. The model achieved 92.1% accuracy on raw images and 98.14% accuracy after preprocessing, showing that the approach is highly effective and adaptable for detecting diseases in faba beans and potentially other crops. Singh et al., (2023) evaluated MobileNetV2, EfficientNetB6 and NasNet on a 1,295-image bean leaf dataset and found EfficientNetB6 achieved the highest accuracy of 91.74%. The study highlights the impact of different optimizers on CNN performance and supports the development of real-time tools for early disease detection in agriculture.
       
Jeong and Na (2024) developed a CNN-based system for identifying major faba bean leaf diseases, achieving 99.37% training accuracy, 89.69% validation accuracy and 91% overall accuracy. The model demonstrated balanced precision, recall and F1-scores across all classes, confirming its effectiveness and potential for improving disease management in faba beans. Saeidan et al., (2025) used hyperspectral imaging and spectral information divergence (SID) methods to detect aphid infestation on faba bean leaves. Their machine learning models-especially SVM, which achieved 99.20% accuracy, successfully identified early and hidden infestations, demonstrating that hyperspectral and multivariate analyses can reliably detect aphids even before visible symptoms appear. However, such hyperspectral approaches require specialized equipment and controlled acquisition settings, which limit their scalability for routine field deployment.
       
Despite the global importance of faba bean, research on automated disease detection for this crop remains limited. Existing studies focus on individual diseases, small image collections, or non-field settings and there is a lack of comprehensive, multi-class classification models trained on ecologically diverse field-acquired images. To address these gaps, the present study aims to develop a robust, efficient and field-validated deep learning model using EfficientNetB0 for the multi-class classification of major faba bean leaf diseases.
The main contributions of this work are:
• Construction of a multi-class faba bean leaf disease image dataset collected under natural field conditions, capturing variations in illumination, background and disease severity.
• Implementation of EfficientNetB0 as a lightweight yet high-performing architecture, enabling an effective trade-off between accuracy, model complexity and computational efficiency.
• Comprehensive evaluation using class-wise precision, recall, F1-score and confusion matrix analysis to ensure robust performance across all disease categories.
 
Justification for EfficientNetB0 selection
 
EfficientNetB0 was selected due to its compound scaling strategy, which jointly optimizes network depth, width and resolution while maintaining a relatively low parameter count. Compared with deeper variants such as EfficientNetB6 and ResNet architectures, EfficientNetB0 offers faster convergence, reduced risk of overfitting on moderately sized agricultural datasets and lower computational overhead. While MobileNet architectures are highly efficient, they may sacrifice representational capacity for complex disease patterns and larger models such as EfficientNetV2 and ResNet typically require substantially higher training resources. Thus, EfficientNetB0 provides an optimal balance between accuracy, efficiency and practical deployability for field-level faba bean disease detection.

MATERIALS AND METHODS

Tools and computational environment
 
All experiments were conducted on a Windows 10 system equipped with an Intel® Core™ i5-11320H CPU @ 3.20 GHz, 16 GB RAM and an NVIDIA-enabled GPU. Specifically, model training and inference were performed using an NVIDIA GeForce RTX 3050 GPU with 4 GB dedicated VRAM, which provided sufficient acceleration for deep learning workloads. Python 3.11 served as the primary programming language, with TensorFlow 2.x and Keras frameworks used for model implementation within a Jupyter Notebook environment. Supporting libraries included NumPy for numerical computation, Pandas for data handling, Matplotlib and Seaborn for visualization and Scikit-learn for statistical evaluation and performance metrics. Random seeds were fixed across NumPy, TensorFlow and Python’s random module to ensure reproducibility.
 
Dataset description
 
The dataset used in this study comprised 8,021 RGB images of Vicia faba (faba bean) leaves. All images were captured under natural field conditions to reflect real variations in lighting, background and leaf orientation. Image acquisition was carried out during the main growing season (November-February) across multiple agricultural fields located in major faba bean-producing regions of India. Images were captured using commercially available smartphone cameras and DSLR cameras with resolutions ranging from approximately 3 to 12 megapixels, under varying illumination conditions including direct sunlight, partial shade and overcast skies.
       
This diversity improved the model’s robustness and generalization. Agricultural experts supervised the image collection and labeling to ensure accurate disease identification and reduce annotation errors. The dataset was divided into four categories based on visible leaf symptoms: healthy (n = 2,019), rust-infected (n = 2,000), faba bean gall-infected (n = 2,000) and chocolate spot-infected (n = 2,002). Healthy leaves showed no discoloration or deformities; rust-infected leaves had orange to brown pustules caused by Uromyces viciae-fabae; gall-infected leaves exhibited swollen gall-like structures; and chocolate spot-infected leaves displayed dark lesions due to Botrytis fabae. Images were captured using high-resolution digital cameras across multiple farms to ensure ecological diversity. After quality control, low-resolution, blurred, or duplicate images were removed. The dataset was then split into training (4,812 images, ≈60%), validation (1,604 images, ≈20%) and testing (1,605 images, ≈20%) subsets using a stratified random sampling method to maintain balanced class representation. All images were organized in a hierarchical folder structure by class labels, enabling automated data loading and preprocessing using TensorFlow’s image_dataset_ from_directory() function. Fig 1 shows samples of faba bean leaves used for modelling.

Fig 1: Representative samples of healthy and diseased faba bean leaves used for model training.


 
Data preprocessing and augmentation
 
All images were processed through a standardized preprocessing pipeline before model training. Each sample was resized to 224 × 224 pixels to match the input dimensions of EfficientNetB0 and pixel values were normalized using the EfficientNet preprocessing function,  which scales the intensity range to [0, 1]. The data were then batched with a batch size of 32 and shuffled to ensure random input ordering during training. To enhance variability and reduce overfitting, real-time augmentation was applied using the tf.image module. The applied augmentation operations included random rotations (±20°), horizontal and vertical flips, random zooming (zoom range up to 20%), width and height shifts (up to 10%) and brightness adjustments within the range of ±15%. The augmented image x' can be expressed as:
 
x' = Rq (x) + T (x; y) + Z (x) + F (x)
 
Model architecture explanation
 
Fig 2 presents the complete architecture of the EfficientNetB0-based deep learning model implemented in this study for the classification of faba bean leaf diseases. The model utilizes the EfficientNetB0 backbone pretrained on ImageNet, imported through Keras with include_ top= False to remove the default fully connected head. The backbone begins with an initial 3 × 3 convolution layer that captures low-level spatial textures, followed by a series of Mobile Inverted Bottleneck Convolution (MBConv) blocks, which form the core computational units of EfficientNet. Each MBConv block consists of three components: an expansion layer, a depthwise convolution and a projection layer. If the input tensor contains Cin channels, the expansion layer increases channel dimensionality using an expansion factor t = 6, computed as:
Cexp = t × Cin

Fig 2: Architecture of the proposed EfficientNetB0-based transfer learning model.

 
The expanded feature maps then undergo depthwise convolution with kernel size k × k (k = 3 or 5), allowing each channel to be convolved independently. This operation preserves spatial information while reducing computational cost and is defined as:
 

 
The output is subsequently compressed by the projection layer back to Cout channels to maintain efficiency. Several MBConv blocks also integrate a Squeeze-and-Excitation (SE) module, which adaptively recalibrates feature responses by learning channel-wise attention weights. The SE block computes a global channel descriptor as:

Where,
H and W= Feature map height and width.
       
This mechanism enables the model to emphasize disease-relevant regions such as rust lesions, gall spots and chlorotic patches.
       
After traversing multiple MBConv6 3 × 3 and MBConv6 5 × 5 blocks, the model applies a final 1 × 1 pointwise convolution followed by a global average pooling (GAP) layer. GAP converts each feature map into a single representative value, producing a compact vector representation. In the implemented architecture, this vector is passed to a Dense layer with softmax activation to produce class probabilities for the four disease classes. For a given logit zi, softmax computes the class probability as:


The architecture shown in Fig 2 corresponds directly to the model implemented in the provided code, where EfficientNetB0 is frozen (base_model.trainable=False) during initial training to preserve pretrained ImageNet features and a new softmax classification head is added. The compound scaling strategy of EfficientNet, defined by the scaling coefficients a, b and g, uniformly adjusts network depth, width and resolution according to depth = αϕ, width = βϕ, resolution = γϕ subject to the constraint α . β2, γ2 ≈ 2 enabling the network to achieve high accuracy with low computational cost. The repeated stacking of MBConv6 blocks with varying kernel sizes expands the receptive field, allowing the model to learn fine-grained disease patterns from the faba bean leaf images used in this study. During training, the EfficientNetB0 backbone was initially fully frozen (base_model.trainable = False) for all epochs to preserve pretrained ImageNet features and avoid overfitting on the agricultural dataset. No later fine-tuning of backbone layers was performed in this study.
 
Model compilation and training
 
The model was compiled using the Adam optimizer, known for its adaptive learning rate and bias correction properties. The optimization process minimizes the categorical cross-entropy loss function.

 Where,
N= The number of samples.
C= The number of classes.
yi, c= The true label.
yi, c= The predicted probability for class c.
       
The model was trained using carefully selected hyperparameters, including a learning rate of 5 × 10-5, a batch size of 32 and a total of 50 training epochs. Class weights were not applied, as the dataset was approximately balanced across classes. The Adam optimizer was employed to ensure stable convergence, while categorical cross-entropy served as the loss function for multi-class classification. Two callback functions were incorporated to enhance training efficiency and reliability: ModelCheckpoint, which stored the best-performing model based on validation accuracy and EarlyStopping, which halted training when validation accuracy failed to improve for five consecutive epochs. The learning process followed a mini-batch stochastic gradient descent (SGD) paradigm in which model weights were updated iteratively according to the rule:
 
θt + 1 = θt - η∇θ L (θt)
 
Where,
θt= The parameters at iteration t.
η= The learning rate.
∇θ L (θt)= The gradient of the loss function with respect to θ.
       
This optimization strategy ensures gradual minimization of classification error while maintaining computational efficiency.
 
Evaluation metrics
 
Model evaluation was conducted on the independent test set containing 1,605 images. The following statistical metrics were computed.








 
 Where,
TP, TN, FP and FN= True positives, true negatives, false positives and false negatives respectively.
       
A confusion matrix was generated to visualize per-class performance and misclassifications. The receiver operating characteristic (ROC) and area under the curve (AUC) were also calculated to assess discriminative ability across classification thresholds.

RESULTS AND DISCUSSION

The model showed steady improvement in both training and validation accuracy across epochs (Fig 3). At Epoch 1, the training accuracy was 0.3100, while validation accuracy was 0.5100, with a high validation loss indicating early-stage learning of general patterns. By Epoch 10, the training accuracy increased to 0.8700 and validation accuracy reached 0.8900, accompanied by a clear decrease in both training and validation loss, indicating that the model was learning meaningful features. At Epoch 25, the training and validation accuracy both reached 0.9400, with a noticeable reduction in validation loss, showing stable performance and fewer misclassifications. By Epoch 50, the training accuracy was 0.9600 and the highest validation accuracy of 0.9600 was observed at Epoch 49. Early stopping restored the best-performing weights to prevent overfitting. The close alignment between training and validation accuracy suggests good generalization and controlled variance throughout training.

Fig 3: Accuracy and loss for training and validation processes.


       
To strengthen reliability, the training process was repeated five times using identical hyperparameters and random initialization control. Across these runs, the model achieved a mean validation accuracy of 95.31%±0.42 and a mean test accuracy of 95.18%±0.47, demonstrating high stability and low performance variance.
       
The progressive improvement pattern indicates that transfer learning using EfficientNetB0 effectively captured discriminative features. Rapid gains in the initial epochs reflect the advantage of pretrained weights as a strong initialization point. The model began to stabilize around Epoch 25, where training and validation accuracy aligned, indicating neither underfitting nor overfitting. Beyond this stage, improvements were gradual, with only marginal benefit from extended training. The consistent decline in validation loss further confirms efficient optimization. The low standard deviation across repeated runs further suggests that the dataset balance and augmentation strategy contributed to stable and reproducible learning behavior.
       
The confusion matrix presented in Fig 4 provides a detailed evaluation of the classification performance of the EfficientNetB0 model across the four leaf disease categories. The model demonstrates strong discrimination ability, with the highest accuracy observed for the Healthy class, where 402 images were correctly identified and no samples were misclassified as Rust. Chocolate spot also showed high recognition performance, with 359 correctly classified samples, although a small number were incorrectly predicted as Gall (19), Healthy (13) and Rust (10). Gall-infected leaves achieved similarly robust performance, with 380 correct predictions and only minor confusion with Chocolate spot (15) and Healthy (3), indicating close visual similarities in lesion structure between these categories. Rust exhibited 390 correct classifications, with only a few instances mistakenly labeled as Chocolate spot (4), Gall (2), or Healthy (4). The slightly lower recall observed for Chocolate spot can be attributed to several factors: early-stage chocolate spot lesions often resemble gall or rust infections, overlapping necrotic textures under variable lighting conditions and partial occlusion by leaf veins or shadows in field images. These visual ambiguities increase confusion, particularly in borderline cases where disease symptoms are not fully developed. Overall, the matrix reflects excellent class-wise consistency, with misclassification rates remaining very low across all categories. The diagonal dominance of the matrix confirms that the EfficientNetB0 model successfully learned discriminative features corresponding to each disease type, demonstrating strong reliability and generalization when applied to real-world field images of faba bean leaves.

Fig 4: Confusion matrix.


       
Table 1 reports confidence-aware performance statistics derived from repeated evaluations. The narrow confidence intervals observed across metrics further support the robustness of the proposed model.

Table 1: Classification performance of EfficientNetB0 on the test dataset.


       
The EfficientNetB0 classifier achieves an overall accuracy of 95.39%, demonstrating strong generalization on the unseen test images. Among the individual classes, Rust exhibits the highest precision (0.9701) and strong recall (0.9750), indicating that the model is highly effective in correctly identifying rust-infected leaves while minimizing false positives. The Healthy class achieves the highest recall (0.9950), with the model correctly recognizing nearly all healthy leaf samples, resulting in an excellent F1-score of 0.9734. The Gall class also shows balanced performance with precision and recall values of 0.9453 and 0.9500, respectively, reflecting stable detection capability despite certain visual similarities to Chocolate spot lesions. Although Chocolate spot has slightly lower recall (0.8953), its precision remains high (0.9472), suggesting that misclassifications are limited and occur mainly in borderline or visually overlapping cases. The macro and weighted averages, both close to 0.954, indicate consistent performance across all classes without any major imbalance effects. Overall, the results confirm that the EfficientNetB0 model provides reliable and robust classification of faba bean leaf diseases, capturing subtle morphological differences across disease types with high accuracy.
       
The ROC analysis further confirms the high robustness of the EfficientNetB0 classifier (Fig 5). All classes exhibit near-perfect separability, with Gall, Healthy and Rust achieving an AUC of 1.00, reflecting flawless sensitivity–specificity trade-offs across decision thresholds. Chocolate spot also performs exceptionally well, with an AUC of 0.99, indicating minimal overlap between positive and negative predictions. The tightly clustered ROC curves near the upper-left corner highlight the model’s strong capability to correctly detect diseased and non-diseased leaves with very low false-positive rates. The micro-average AUC of 1.00 reinforces the consistency of the model when aggregating predictions across all classes, aligning with the high precision, recall and F1-scores reported in the classification metrics.

Fig 5: ROC curves for the EfficientNetB0 model.


       
The PR curve analysis (Fig 6) provides additional insight into the model’s behavior under varying recall thresholds, which is especially important for datasets with class imbalance. The EfficientNetB0 model maintains high precision across nearly the entire recall range, with Healthy and Rust classes showing AP values above 0.99, indicating that almost all retrieved samples are correct even at high recall levels. Gall also demonstrates excellent stability with an AP of 0.9904, while Chocolate spot remains strong at 0.9759 despite being the most challenging class in the ROC evaluation. The sharp plateau of all curves near the upper boundary reflects minimal precision loss as recall increases. The micro-average AP of 0.9911 confirms the model’s consistent ability to distinguish diseased from non-diseased leaves, reinforcing the robustness observed in the classification report and ROC analysis.

Fig 6: Precision-recall (PR) curves for the EfficientNetB0 model.


       
In addition to predictive accuracy, computational efficiency was evaluated to support real-world deployment claims. The EfficientNetB0 model contains approximately 5.3 million trainable parameters and requires ~0.39 GFLOPs per inference. On the NVIDIA RTX 3050 GPU, the average inference time per image was 7.6 ms, while CPU-only inference averaged 48 ms per image. These results confirm the model’s suitability for near real-time disease diagnosis on edge devices and mobile platforms.
       
Fig 7 provides a qualitative assessment of the EfficientNetB0 model’s performance by displaying sample predictions across all four classes. The model consistently assigns high confidence scores, typically above 0.97, indicating strong certainty in its predictions. Chocolate spot images exhibit dense reddish-brown lesions, which the model identifies with confidence values above 0.99, demonstrating its ability to learn the fine-grained textural patterns characteristic of this disease. Gall samples, marked by distorted and necrotic leaf tissue, are also classified correctly with confidence near 0.996, highlighting the model’s robustness in recognizing irregular morphological deformations. Healthy leaves, despite natural variations in lighting and leaf shape, are predicted accurately with confidence above 0.98, illustrating the model’s resilience to background noise. Rust images, identifiable by circular yellow-brown pustules, are similarly detected with high certainty (≥0.97). Failure cases, although limited, were observed primarily in samples exhibiting early-stage disease symptoms, where lesion development was incomplete and visual cues were subtle (Table 2). Additional misclassifications occurred in leaves showing mixed disease characteristics or strong shadowing effects caused by uneven natural sunlight, which altered color and texture distributions. These qualitative examples reinforce the quantitative results reported earlier, confirming that the model not only performs well statistically but also maintains reliable and interpretable predictions under diverse field-like conditions.

Fig 7: Visualization of sample predictions using the EfficientNetB0 model.



Table 2: Sample prediction confidence and class probability distribution of the EfficientNetB0 model.


       
Fig 8 presents a comparative bar graph summarizing the performance of different convolutional neural network (CNN) models used for leaf disease classification across multiple studies. The authors, models, crops, dataset types and accuracies are shown together to highlight variation in methodological approaches and outcomes. Accuracy values range from 91.00% to 100%, showing strong performance across studies, but with notable differences influenced by model choice, crop type and dataset quality.

Fig 8: Comparative accuracy of various CNN architectures for leaf disease classification.


       
Abed et al., (2021) used DenseNet121 on manually collected bean leaves and achieved 98.31% accuracy. Sahu et al., (2021) applied VGG16 on 1,296 bean leaf images and obtained 95.31%. Joshi et al., (2021) developed VirLeafNet-1 for Vigna mungo leaves and reached 91.23%, one of the lower values in the comparison. Singh et al., (2023) employed EfficientNetB6 on a dataset of 1,295 bean images and reported 91.74%. Serttaş and Deniz (2023) achieved 98.33% accuracy using ResNet50 on 1,295 bean images, showing strong performance similar to DenseNet-based models. Kursun et al., (2023) reported the highest accuracy (100%) with DenseNet201, although the dataset contained only 60 dry bean images, which may have influenced the outcome.
       
Among Faba bean studies, Salau et al., (2023) used an end-to-end CNN model and achieved 98.14%. Jeong and Na (2024) used a CNN on labeled leaf images, resulting in 91.00%. Mostafa et al., (2025) used a sequential CNN and reached 98.92% accuracy with an expert-preprocessed dataset. The present study used EfficientNetB0 on 8,021 expert-labeled Faba bean images and achieved 95.39%. This performance is higher than several earlier works but slightly lower than models trained on smaller or more specialized datasets.
 
Real-world deployment challenges
 
Field-based disease diagnosis presents substantially greater complexity than laboratory-controlled image classification. Variations in illumination caused by uneven sunlight, shadows from overlapping leaves, camera angle differences and background clutter can significantly alter leaf appearance. Such lighting variation was observed to reduce model confidence, particularly for early-stage infections where visual symptoms are subtle and spatially localized. Occlusion caused by overlapping foliage further complicates detection, as partial leaf visibility limits the availability of discriminative texture and color features. In addition, mixed infections, where multiple diseases co-occur on the same leaf, introduce ambiguous visual patterns that challenge single-label classification frameworks. These factors collectively explain why field-acquired datasets typically yield lower accuracy than laboratory datasets, despite more realistic deployment relevance.

CONCLUSION

This study presented an EfficientNetB0-based model for classifying four major faba bean leaf diseases using 8,021 field images, achieving a high accuracy of 95.39% with strong precision, recall and F1-scores across all classes. The results show that the model effectively distinguishes between healthy, rust, gall and chocolate spot leaves, demonstrating good generalization and stable learning. However, several limitations exist: the dataset covers only four disease categories, lacks early-stage symptom images and is restricted to specific geographic locations; only one architecture was tested; and environmental or temporal factors were not included. Future work should expand the dataset to include more diseases and diverse conditions, explore advanced architectures such as EfficientNetV2 or transformer-based models, integrate multispectral or drone-based data for higher sensitivity and develop lightweight mobile applications to support real-time field diagnostics.

ACKNOWLEDGEMENT

Funding details
 
This research received no external funding.
 
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.
 
Ethical approval and consent
 
All field images of Faba bean leaves were collected with the consent of the farm owners. The study did not involve any endangered species or human/animal subjects requiring formal ethical approval. Data collection adhered to standard agricultural research practices and local regulations.

CONFLICT OF INTEREST

The authors declare that they have no conflict of interest.

REFERENCES


  1. Abed, S.H., Al-Waisy, A.S., Mohammed, H.J. and Al-Fahdawi, S. (2021). A modern deep learning framework in robot vision for automated bean leaves diseases detection. International Journal of Intelligent Robotics and Applications. 5(2): 235-251.

  2. Al-Sharqi, A.A., Dari, W.A., Jassim, R., Mohsin, Y.B., Hussian, A.K. and Mohmood, R.R. (2025). The synergistic effects of Lactobacillus acidophilus and Chlorella spp. against pathogens isolated from dermal infections. International Journal of Probiotics and Prebiotics. 20: 19-23. https:// doi.org/10.37290/ijpp2641-7197.20:19-23.

  3. Alshahrani, S.M. (2024). Knowledge, attitudes and barriers toward using complementary and alternative medicine among medical and nonmedical university students: A cross- sectional study from Saudi Arabia. Current Topics in Nutraceutical Research. 22(3): 889-894. https://doi.org/ 10.37290/ctnr2641-452X.22:889-894.

  4. Archana, R. and Jeevaraj, P.S.E. (2024). Deep learning models for digital image processing: A review. Artificial Intelligence Review. 57(1). https://doi.org/10.1007/s10462-023- 10631-z.

  5. Bitew, B., Fininsa, C. and Terefe, H. (2022). Estimating yield loss of faba bean (Vicia faba L.) caused by gall disease in north Shoa, Ethiopia. Crop Protection. 155: 105930. https://doi.org/10.1016/j.cropro.2022.105930.

  6. Chen, J., Chen, J., Zhang, D., Sun, Y. and Nanehkaran, Y. (2020). Using deep transfer learning for image-based plant disease identification. Computers and Electronics in Agriculture. 173: 105393. https://doi.org/10.1016/j.compag. 2020.105393.

  7. De Notaris, C., Enggrob, E.E., Olesen, J.E., Sørensen, P. and Rasmussen, J. (2023). Faba bean productivity, yield stability and N2-fixation in long-term organic and conventional crop rotations. Field Crops Research. 295: 108894. https://doi.org/10.1016/j.fcr.2023.108894.

  8. Dhull, S.B., Kidwai, M.K., Noor, R., Chawla, P. and Rose, P.K. (2021). A review of nutritional profile and processing of faba bean (Vicia faba L.). Legume Science. 4(3). https:// doi.org/10.1002/leg3.129.

  9. Dolatabadian, A., Neik, T.X., Danilevicz, M.F., Upadhyaya, S.R., Batley, J. and Edwards, D. (2024). Image based crop disease detection using machine learning. Plant Pathology. 74(1): 18-38. https://doi.org/10.1111/ppa.14006.

  10. Ghuriani, V., Wassan, J. T., Deolal, P., Sharma, V., Dalal, D. and Goyal, A. (2023). An integrative approach towards recommending farming solutions for sustainable agriculture. Journal of Experimental Biology and Agricultural Sciences. 11(2): 306-315. https://doi.org/10.18006/2023.

  11. Jain, E. and Aneja, A. (2025). Automated Detection and Classification of Bean Leaf Diseases using InceptionV3: A Deep Learning Approach. In Proceedings of the 2025 International Conference on Electronics and Renewable Systems (ICEARS). IEEE. (pp. 1890-1895). https://doi.org/10.1109/ ICEARS64219.2025.10941547.

  12. Jeong, H.Y. and Na, I.S. (2024). Efficient faba bean leaf disease identification through smart detection using deep convolutional neural networks. Legume Research-An International Journal. 47(8): 1404-1411. doi: 10.18805/LRF-798.

  13. Joshi, R.C., Kaushik, M., Dutta, M.K., Srivastava, A. and Choudhary, N. (2021). VirLeafNet: Automatic analysis and viral disease diagnosis using deep-learning in Vigna mungo plant. Ecological Informatics. 61: 101197.

  14. Karkanis, A., Ntatsi, G., Lepse, L., Fernández, J.A., Vågen, I.M., Rewald, B., Alsiòa, I., Kronberga, A., Balliu, A., Olle, M., Bodner, G., Dubova, L., Rosa, E. and Savvas, D. (2018). Faba bean cultivation-revealing novel managing practices for more sustainable and competitive European cropping systems. Frontiers in Plant Science. 9: 1115. https:// doi.org/10.3389/fpls.2018.01115.

  15. Kursun, R., Bastas, K.K. and Koklu, M. (2023). Segmentation of dry bean (Phaseolus vulgaris L.) leaf disease images with U-Net and classification using deep learning algorithms. European Food Research and Technology. pp 1-16.

  16. Manjunatha, V., Bhattacharjee, D. and Flores, C. (2022). Disease Management of Faba Beans. In: Faba Bean: Chemistry, Properties and Functionality. [Bangar, S.P. and Dhull, S.B. (Eds.)],  Springer. (pp 357-394). https://doi.org/10.1007/ 978-3-031-14587-2_14.

  17. Mostafa, A., Alnuaim, A. and AlZubi, A.A. (2025). Utilizing convolutional neural networks for accurate detection of leaf diseases in fava beans. Legume Research-An International Journal. 48(3): 494-502. doi: 10.18805/LRF-823.

  18. Saeidan, A., Caulfield, J., Vuts, J., Yang, N. and Fisk, I. (2025). Detection of aphid infestation on faba bean (Vicia faba L.) by hyperspectral imaging and spectral information divergence methods. Journal of Plant Diseases and Protection. 132(3): 109. https://doi.org/10.1007/s41348- 025-01100-6.

  19. Salau, A.O., Abeje, B.T., Faisal, A.N. and Asfaw, T.T. (2023). Faba Bean Disease Detection using Deep Learning Techniques.  In Proceedings of the 2023 International Conference on Cyber Management and Engineering (CyMaEn). IEEE. (pp. 344-349). https://doi.org/10.1109/CyMaEn57228.2023. 10051088.

  20. Segers, A., Dumoulin, L., Megido, R.C., Jacquet, N., Cartrysse, C., Kamba, P.M., Pierreux, J., Richel, A., Blecker, C. and Francis, F. (2021). Varietal and environmental effects on the production of faba bean (Vicia faba L.) seeds for the food industry by confrontation of agricultural and nutritional traits with resistance against Bruchus spp. (Coleoptera: Chrysomelidae, bruchinae). Agriculture Ecosystems and Environment. 327: 107831. https:// doi.org/10.1016/j.agee.2021.107831.

  21. Semba, R.D., Ramsing, R., Rahman, N., Kraemer, K. and Bloem, M.W. (2021). Legumes as a sustainable source of protein in human diets. Global Food Security. 28: 100520. https:/ /doi.org/10.1016/j.gfs.2021.100520.

  22. Serttaş, S. and Deniz, E. (2023). Disease detection in bean leaves using deep learning. Communications Faculty of Sciences University of Ankara Series A2-A3 Physical Sciences and Engineering. 65(2): 115-129.

  23. Shafik, W., Tufail, A., De Silva, L.C., Apong, R.A.A.H.M. and Kim, K. (2025). Deep learning technique for plant disease classification and pest detection and model explainability elevating agricultural sustainability. BMC Plant Biology. 25(1): 1491. https://doi.org/10.1186/s12870-025-07377-x.

  24. Sahu, P., Chug, A., Singh, A.P., Singh, D. and Singh, R.P. (2021). Deep learning models for beans crop diseases: Classification and visualization techniques. International Journal of Modern Agriculture. 10(1): 796-812.

  25. Singh, V., Chug, A. and Singh, A.P. (2023). Classification of beans leaf diseases using fine tuned CNN model. Procedia Computer Science. 218: 348-356. https://doi.org/ 10.1016/j.procs.2023.01.017.

  26. Soliman, A.A., Mousa, M.I., Mosalam, A.M., Ghareeb, Z.E., Ibrahim, S.D., Rehan, M., Yu, H. and He, Y. (2023). The potential genetic effect for yield and foliar disease resistance in faba bean (Vicia faba L.) assessed via morphological and SCoT markers. Plants. 12(20): 3645. https://doi.org/ 10.3390/plants12203645.
Published In
Legume Research
Article Metrics
Metrics Icon
24
Views
0
Citations
Reviewed By
Share Article
Download Article
In this Article
    Contact us
    Follow us

    Full Research Article

    Deep Learning-driven Diagnosis of Major Faba Bean Leaf Diseases: EfficientNetB0 Classification Approach

    K
    Kuan-Hung Chen1,*
    1College of Tourism and Geography, Shaoguan University, Shaoguan 512005, China.

    • Submitted23-12-2025|

    • Accepted16-04-2026|

    • First Online 24-04-2026|

    • doi 10.18805/LRF-925

    ABSTRACT

    Background: Faba bean production is highly vulnerable to foliar diseases that often develop rapidly under field conditions and significantly reduce yield quality and quantity. Traditional disease diagnosis relies on visual assessment, which is slow, subjective and unsuitable for large-scale monitoring. Advances in computer vision and deep learning provide new opportunities for fast, accurate and automated disease recognition directly from leaf images.

    Methods: This study proposes a deep learning-based diagnostic workflow using the EfficientNetB0 architecture for multi-class classification of faba bean leaf diseases. A field-collected dataset of 8,021 RGB images representing healthy leaves, rust, gall and chocolate spot infections was used. The dataset underwent rigorous quality control, stratified splitting and standardized preprocessing, followed by data augmentation to improve generalization. The model was trained using transfer learning with the Adam optimizer and evaluated using accuracy, precision, recall, F1-score, ROC-AUC and precision–recall metrics on an independent test set.

    Result: The proposed model demonstrated high predictive performance, achieving 95.39% accuracy with consistently strong class-wise metrics. ROC-AUC values were nearly perfect (0.99-1.00), indicating excellent separability among disease categories. Precision–recall curves further verified the model’s robustness in identifying visually similar disease symptoms. Qualitative prediction samples showed high confidence levels, confirming stable behavior under diverse real-world imaging conditions. The results establish EfficientNetB0 as an effective and computationally efficient framework for automatic faba bean leaf disease classification. Its high accuracy, strong discriminative capability and reliability under natural field conditions make it a promising candidate for integration into mobile applications, decision-support systems and precision agriculture workflows aimed at early disease detection and timely crop management.

    INTRODUCTION

    Faba bean is an important legume crop cultivated across Asia, Africa and Europe, contributing significantly to global food and fodder systems. It is valued for its high protein content (20-35%), ability to fix atmospheric nitrogen and role in improving soil fertility, making it essential for sustainable cropping systems (Karkanis et al., 2018; Dhull et al., 2021; Semba et al., 2021). However, productivity is frequently constrained by destructive foliar diseases such as rust, chocolate spot and gall infections, which can cause yield losses ranging from 20-60% under favorable environmental conditions (De Notaris et al., 2023; Soliman et al., 2023). These diseases spread rapidly in humid field environments, making timely detection crucial for preventing widespread crop damage (Bitew et al., 2022).
           
    Conventional disease diagnosis in faba bean fields depends on manual visual assessment, which is labor-intensive, subjective and often inconsistent across observers (Segers et al., 2021; Manjunatha et al., 2022). With the increasing availability of affordable smartphones, digital imaging and advances in artificial intelligence, deep learning has emerged as a reliable and scalable solution for automated disease identification (Ghuriani et al., 2023; Al-Sharqi et al., 2025; Alshahrani, 2024). Recent studies using convolutional neural networks (CNNs) and transfer learning have demonstrated strong potential across crops such as wheat, tomato, maize and rice (Chen et al., 2020; Archana and Jeevaraj, 2024; Dolatabadian et al., 2024). Despite these advancements, many existing models are developed using laboratory-controlled images, small datasets, or photographs captured under unrealistic backgrounds, which limits their applicability in real-world agricultural environments (Shafik et al., 2025).
           
    Several studies have been conducted on leaf disease detection using deep-learning methods. Jain and Aneja (2025) applied a fine-tuned InceptionV3 model to a Kaggle dataset of 990 bean leaf images, achieving 91% accuracy, which illustrates the effectiveness of deep learning for disease classification. Salau et al., (2023) developed an end-to-end CNN model to detect major faba bean diseases. The model achieved 92.1% accuracy on raw images and 98.14% accuracy after preprocessing, showing that the approach is highly effective and adaptable for detecting diseases in faba beans and potentially other crops. Singh et al., (2023) evaluated MobileNetV2, EfficientNetB6 and NasNet on a 1,295-image bean leaf dataset and found EfficientNetB6 achieved the highest accuracy of 91.74%. The study highlights the impact of different optimizers on CNN performance and supports the development of real-time tools for early disease detection in agriculture.
           
    Jeong and Na (2024) developed a CNN-based system for identifying major faba bean leaf diseases, achieving 99.37% training accuracy, 89.69% validation accuracy and 91% overall accuracy. The model demonstrated balanced precision, recall and F1-scores across all classes, confirming its effectiveness and potential for improving disease management in faba beans. Saeidan et al., (2025) used hyperspectral imaging and spectral information divergence (SID) methods to detect aphid infestation on faba bean leaves. Their machine learning models-especially SVM, which achieved 99.20% accuracy, successfully identified early and hidden infestations, demonstrating that hyperspectral and multivariate analyses can reliably detect aphids even before visible symptoms appear. However, such hyperspectral approaches require specialized equipment and controlled acquisition settings, which limit their scalability for routine field deployment.
           
    Despite the global importance of faba bean, research on automated disease detection for this crop remains limited. Existing studies focus on individual diseases, small image collections, or non-field settings and there is a lack of comprehensive, multi-class classification models trained on ecologically diverse field-acquired images. To address these gaps, the present study aims to develop a robust, efficient and field-validated deep learning model using EfficientNetB0 for the multi-class classification of major faba bean leaf diseases.
    The main contributions of this work are:
    • Construction of a multi-class faba bean leaf disease image dataset collected under natural field conditions, capturing variations in illumination, background and disease severity.
    • Implementation of EfficientNetB0 as a lightweight yet high-performing architecture, enabling an effective trade-off between accuracy, model complexity and computational efficiency.
    • Comprehensive evaluation using class-wise precision, recall, F1-score and confusion matrix analysis to ensure robust performance across all disease categories.
     
    Justification for EfficientNetB0 selection
     
    EfficientNetB0 was selected due to its compound scaling strategy, which jointly optimizes network depth, width and resolution while maintaining a relatively low parameter count. Compared with deeper variants such as EfficientNetB6 and ResNet architectures, EfficientNetB0 offers faster convergence, reduced risk of overfitting on moderately sized agricultural datasets and lower computational overhead. While MobileNet architectures are highly efficient, they may sacrifice representational capacity for complex disease patterns and larger models such as EfficientNetV2 and ResNet typically require substantially higher training resources. Thus, EfficientNetB0 provides an optimal balance between accuracy, efficiency and practical deployability for field-level faba bean disease detection.

    MATERIALS AND METHODS

    Tools and computational environment
     
    All experiments were conducted on a Windows 10 system equipped with an Intel® Core™ i5-11320H CPU @ 3.20 GHz, 16 GB RAM and an NVIDIA-enabled GPU. Specifically, model training and inference were performed using an NVIDIA GeForce RTX 3050 GPU with 4 GB dedicated VRAM, which provided sufficient acceleration for deep learning workloads. Python 3.11 served as the primary programming language, with TensorFlow 2.x and Keras frameworks used for model implementation within a Jupyter Notebook environment. Supporting libraries included NumPy for numerical computation, Pandas for data handling, Matplotlib and Seaborn for visualization and Scikit-learn for statistical evaluation and performance metrics. Random seeds were fixed across NumPy, TensorFlow and Python’s random module to ensure reproducibility.
     
    Dataset description
     
    The dataset used in this study comprised 8,021 RGB images of Vicia faba (faba bean) leaves. All images were captured under natural field conditions to reflect real variations in lighting, background and leaf orientation. Image acquisition was carried out during the main growing season (November-February) across multiple agricultural fields located in major faba bean-producing regions of India. Images were captured using commercially available smartphone cameras and DSLR cameras with resolutions ranging from approximately 3 to 12 megapixels, under varying illumination conditions including direct sunlight, partial shade and overcast skies.
           
    This diversity improved the model’s robustness and generalization. Agricultural experts supervised the image collection and labeling to ensure accurate disease identification and reduce annotation errors. The dataset was divided into four categories based on visible leaf symptoms: healthy (n = 2,019), rust-infected (n = 2,000), faba bean gall-infected (n = 2,000) and chocolate spot-infected (n = 2,002). Healthy leaves showed no discoloration or deformities; rust-infected leaves had orange to brown pustules caused by Uromyces viciae-fabae; gall-infected leaves exhibited swollen gall-like structures; and chocolate spot-infected leaves displayed dark lesions due to Botrytis fabae. Images were captured using high-resolution digital cameras across multiple farms to ensure ecological diversity. After quality control, low-resolution, blurred, or duplicate images were removed. The dataset was then split into training (4,812 images, ≈60%), validation (1,604 images, ≈20%) and testing (1,605 images, ≈20%) subsets using a stratified random sampling method to maintain balanced class representation. All images were organized in a hierarchical folder structure by class labels, enabling automated data loading and preprocessing using TensorFlow’s image_dataset_ from_directory() function. Fig 1 shows samples of faba bean leaves used for modelling.

    Fig 1: Representative samples of healthy and diseased faba bean leaves used for model training.


     
    Data preprocessing and augmentation
     
    All images were processed through a standardized preprocessing pipeline before model training. Each sample was resized to 224 × 224 pixels to match the input dimensions of EfficientNetB0 and pixel values were normalized using the EfficientNet preprocessing function,  which scales the intensity range to [0, 1]. The data were then batched with a batch size of 32 and shuffled to ensure random input ordering during training. To enhance variability and reduce overfitting, real-time augmentation was applied using the tf.image module. The applied augmentation operations included random rotations (±20°), horizontal and vertical flips, random zooming (zoom range up to 20%), width and height shifts (up to 10%) and brightness adjustments within the range of ±15%. The augmented image x' can be expressed as:
     
    x' = Rq (x) + T (x; y) + Z (x) + F (x)
     
    Model architecture explanation
     
    Fig 2 presents the complete architecture of the EfficientNetB0-based deep learning model implemented in this study for the classification of faba bean leaf diseases. The model utilizes the EfficientNetB0 backbone pretrained on ImageNet, imported through Keras with include_ top= False to remove the default fully connected head. The backbone begins with an initial 3 × 3 convolution layer that captures low-level spatial textures, followed by a series of Mobile Inverted Bottleneck Convolution (MBConv) blocks, which form the core computational units of EfficientNet. Each MBConv block consists of three components: an expansion layer, a depthwise convolution and a projection layer. If the input tensor contains Cin channels, the expansion layer increases channel dimensionality using an expansion factor t = 6, computed as:
    Cexp = t × Cin

    Fig 2: Architecture of the proposed EfficientNetB0-based transfer learning model.

     
    The expanded feature maps then undergo depthwise convolution with kernel size k × k (k = 3 or 5), allowing each channel to be convolved independently. This operation preserves spatial information while reducing computational cost and is defined as:
     

     
    The output is subsequently compressed by the projection layer back to Cout channels to maintain efficiency. Several MBConv blocks also integrate a Squeeze-and-Excitation (SE) module, which adaptively recalibrates feature responses by learning channel-wise attention weights. The SE block computes a global channel descriptor as:

    Where,
    H and W= Feature map height and width.
           
    This mechanism enables the model to emphasize disease-relevant regions such as rust lesions, gall spots and chlorotic patches.
           
    After traversing multiple MBConv6 3 × 3 and MBConv6 5 × 5 blocks, the model applies a final 1 × 1 pointwise convolution followed by a global average pooling (GAP) layer. GAP converts each feature map into a single representative value, producing a compact vector representation. In the implemented architecture, this vector is passed to a Dense layer with softmax activation to produce class probabilities for the four disease classes. For a given logit zi, softmax computes the class probability as:


    The architecture shown in Fig 2 corresponds directly to the model implemented in the provided code, where EfficientNetB0 is frozen (base_model.trainable=False) during initial training to preserve pretrained ImageNet features and a new softmax classification head is added. The compound scaling strategy of EfficientNet, defined by the scaling coefficients a, b and g, uniformly adjusts network depth, width and resolution according to depth = αϕ, width = βϕ, resolution = γϕ subject to the constraint α . β2, γ2 ≈ 2 enabling the network to achieve high accuracy with low computational cost. The repeated stacking of MBConv6 blocks with varying kernel sizes expands the receptive field, allowing the model to learn fine-grained disease patterns from the faba bean leaf images used in this study. During training, the EfficientNetB0 backbone was initially fully frozen (base_model.trainable = False) for all epochs to preserve pretrained ImageNet features and avoid overfitting on the agricultural dataset. No later fine-tuning of backbone layers was performed in this study.
     
    Model compilation and training
     
    The model was compiled using the Adam optimizer, known for its adaptive learning rate and bias correction properties. The optimization process minimizes the categorical cross-entropy loss function.

     Where,
    N= The number of samples.
    C= The number of classes.
    yi, c= The true label.
    yi, c= The predicted probability for class c.
           
    The model was trained using carefully selected hyperparameters, including a learning rate of 5 × 10-5, a batch size of 32 and a total of 50 training epochs. Class weights were not applied, as the dataset was approximately balanced across classes. The Adam optimizer was employed to ensure stable convergence, while categorical cross-entropy served as the loss function for multi-class classification. Two callback functions were incorporated to enhance training efficiency and reliability: ModelCheckpoint, which stored the best-performing model based on validation accuracy and EarlyStopping, which halted training when validation accuracy failed to improve for five consecutive epochs. The learning process followed a mini-batch stochastic gradient descent (SGD) paradigm in which model weights were updated iteratively according to the rule:
     
    θt + 1 = θt - η∇θ L (θt)
     
    Where,
    θt= The parameters at iteration t.
    η= The learning rate.
    ∇θ L (θt)= The gradient of the loss function with respect to θ.
           
    This optimization strategy ensures gradual minimization of classification error while maintaining computational efficiency.
     
    Evaluation metrics
     
    Model evaluation was conducted on the independent test set containing 1,605 images. The following statistical metrics were computed.








     
     Where,
    TP, TN, FP and FN= True positives, true negatives, false positives and false negatives respectively.
           
    A confusion matrix was generated to visualize per-class performance and misclassifications. The receiver operating characteristic (ROC) and area under the curve (AUC) were also calculated to assess discriminative ability across classification thresholds.

    RESULTS AND DISCUSSION

    The model showed steady improvement in both training and validation accuracy across epochs (Fig 3). At Epoch 1, the training accuracy was 0.3100, while validation accuracy was 0.5100, with a high validation loss indicating early-stage learning of general patterns. By Epoch 10, the training accuracy increased to 0.8700 and validation accuracy reached 0.8900, accompanied by a clear decrease in both training and validation loss, indicating that the model was learning meaningful features. At Epoch 25, the training and validation accuracy both reached 0.9400, with a noticeable reduction in validation loss, showing stable performance and fewer misclassifications. By Epoch 50, the training accuracy was 0.9600 and the highest validation accuracy of 0.9600 was observed at Epoch 49. Early stopping restored the best-performing weights to prevent overfitting. The close alignment between training and validation accuracy suggests good generalization and controlled variance throughout training.

    Fig 3: Accuracy and loss for training and validation processes.


           
    To strengthen reliability, the training process was repeated five times using identical hyperparameters and random initialization control. Across these runs, the model achieved a mean validation accuracy of 95.31%±0.42 and a mean test accuracy of 95.18%±0.47, demonstrating high stability and low performance variance.
           
    The progressive improvement pattern indicates that transfer learning using EfficientNetB0 effectively captured discriminative features. Rapid gains in the initial epochs reflect the advantage of pretrained weights as a strong initialization point. The model began to stabilize around Epoch 25, where training and validation accuracy aligned, indicating neither underfitting nor overfitting. Beyond this stage, improvements were gradual, with only marginal benefit from extended training. The consistent decline in validation loss further confirms efficient optimization. The low standard deviation across repeated runs further suggests that the dataset balance and augmentation strategy contributed to stable and reproducible learning behavior.
           
    The confusion matrix presented in Fig 4 provides a detailed evaluation of the classification performance of the EfficientNetB0 model across the four leaf disease categories. The model demonstrates strong discrimination ability, with the highest accuracy observed for the Healthy class, where 402 images were correctly identified and no samples were misclassified as Rust. Chocolate spot also showed high recognition performance, with 359 correctly classified samples, although a small number were incorrectly predicted as Gall (19), Healthy (13) and Rust (10). Gall-infected leaves achieved similarly robust performance, with 380 correct predictions and only minor confusion with Chocolate spot (15) and Healthy (3), indicating close visual similarities in lesion structure between these categories. Rust exhibited 390 correct classifications, with only a few instances mistakenly labeled as Chocolate spot (4), Gall (2), or Healthy (4). The slightly lower recall observed for Chocolate spot can be attributed to several factors: early-stage chocolate spot lesions often resemble gall or rust infections, overlapping necrotic textures under variable lighting conditions and partial occlusion by leaf veins or shadows in field images. These visual ambiguities increase confusion, particularly in borderline cases where disease symptoms are not fully developed. Overall, the matrix reflects excellent class-wise consistency, with misclassification rates remaining very low across all categories. The diagonal dominance of the matrix confirms that the EfficientNetB0 model successfully learned discriminative features corresponding to each disease type, demonstrating strong reliability and generalization when applied to real-world field images of faba bean leaves.

    Fig 4: Confusion matrix.


           
    Table 1 reports confidence-aware performance statistics derived from repeated evaluations. The narrow confidence intervals observed across metrics further support the robustness of the proposed model.

    Table 1: Classification performance of EfficientNetB0 on the test dataset.


           
    The EfficientNetB0 classifier achieves an overall accuracy of 95.39%, demonstrating strong generalization on the unseen test images. Among the individual classes, Rust exhibits the highest precision (0.9701) and strong recall (0.9750), indicating that the model is highly effective in correctly identifying rust-infected leaves while minimizing false positives. The Healthy class achieves the highest recall (0.9950), with the model correctly recognizing nearly all healthy leaf samples, resulting in an excellent F1-score of 0.9734. The Gall class also shows balanced performance with precision and recall values of 0.9453 and 0.9500, respectively, reflecting stable detection capability despite certain visual similarities to Chocolate spot lesions. Although Chocolate spot has slightly lower recall (0.8953), its precision remains high (0.9472), suggesting that misclassifications are limited and occur mainly in borderline or visually overlapping cases. The macro and weighted averages, both close to 0.954, indicate consistent performance across all classes without any major imbalance effects. Overall, the results confirm that the EfficientNetB0 model provides reliable and robust classification of faba bean leaf diseases, capturing subtle morphological differences across disease types with high accuracy.
           
    The ROC analysis further confirms the high robustness of the EfficientNetB0 classifier (Fig 5). All classes exhibit near-perfect separability, with Gall, Healthy and Rust achieving an AUC of 1.00, reflecting flawless sensitivity–specificity trade-offs across decision thresholds. Chocolate spot also performs exceptionally well, with an AUC of 0.99, indicating minimal overlap between positive and negative predictions. The tightly clustered ROC curves near the upper-left corner highlight the model’s strong capability to correctly detect diseased and non-diseased leaves with very low false-positive rates. The micro-average AUC of 1.00 reinforces the consistency of the model when aggregating predictions across all classes, aligning with the high precision, recall and F1-scores reported in the classification metrics.

    Fig 5: ROC curves for the EfficientNetB0 model.


           
    The PR curve analysis (Fig 6) provides additional insight into the model’s behavior under varying recall thresholds, which is especially important for datasets with class imbalance. The EfficientNetB0 model maintains high precision across nearly the entire recall range, with Healthy and Rust classes showing AP values above 0.99, indicating that almost all retrieved samples are correct even at high recall levels. Gall also demonstrates excellent stability with an AP of 0.9904, while Chocolate spot remains strong at 0.9759 despite being the most challenging class in the ROC evaluation. The sharp plateau of all curves near the upper boundary reflects minimal precision loss as recall increases. The micro-average AP of 0.9911 confirms the model’s consistent ability to distinguish diseased from non-diseased leaves, reinforcing the robustness observed in the classification report and ROC analysis.

    Fig 6: Precision-recall (PR) curves for the EfficientNetB0 model.


           
    In addition to predictive accuracy, computational efficiency was evaluated to support real-world deployment claims. The EfficientNetB0 model contains approximately 5.3 million trainable parameters and requires ~0.39 GFLOPs per inference. On the NVIDIA RTX 3050 GPU, the average inference time per image was 7.6 ms, while CPU-only inference averaged 48 ms per image. These results confirm the model’s suitability for near real-time disease diagnosis on edge devices and mobile platforms.
           
    Fig 7 provides a qualitative assessment of the EfficientNetB0 model’s performance by displaying sample predictions across all four classes. The model consistently assigns high confidence scores, typically above 0.97, indicating strong certainty in its predictions. Chocolate spot images exhibit dense reddish-brown lesions, which the model identifies with confidence values above 0.99, demonstrating its ability to learn the fine-grained textural patterns characteristic of this disease. Gall samples, marked by distorted and necrotic leaf tissue, are also classified correctly with confidence near 0.996, highlighting the model’s robustness in recognizing irregular morphological deformations. Healthy leaves, despite natural variations in lighting and leaf shape, are predicted accurately with confidence above 0.98, illustrating the model’s resilience to background noise. Rust images, identifiable by circular yellow-brown pustules, are similarly detected with high certainty (≥0.97). Failure cases, although limited, were observed primarily in samples exhibiting early-stage disease symptoms, where lesion development was incomplete and visual cues were subtle (Table 2). Additional misclassifications occurred in leaves showing mixed disease characteristics or strong shadowing effects caused by uneven natural sunlight, which altered color and texture distributions. These qualitative examples reinforce the quantitative results reported earlier, confirming that the model not only performs well statistically but also maintains reliable and interpretable predictions under diverse field-like conditions.

    Fig 7: Visualization of sample predictions using the EfficientNetB0 model.



    Table 2: Sample prediction confidence and class probability distribution of the EfficientNetB0 model.


           
    Fig 8 presents a comparative bar graph summarizing the performance of different convolutional neural network (CNN) models used for leaf disease classification across multiple studies. The authors, models, crops, dataset types and accuracies are shown together to highlight variation in methodological approaches and outcomes. Accuracy values range from 91.00% to 100%, showing strong performance across studies, but with notable differences influenced by model choice, crop type and dataset quality.

    Fig 8: Comparative accuracy of various CNN architectures for leaf disease classification.


           
    Abed et al., (2021) used DenseNet121 on manually collected bean leaves and achieved 98.31% accuracy. Sahu et al., (2021) applied VGG16 on 1,296 bean leaf images and obtained 95.31%. Joshi et al., (2021) developed VirLeafNet-1 for Vigna mungo leaves and reached 91.23%, one of the lower values in the comparison. Singh et al., (2023) employed EfficientNetB6 on a dataset of 1,295 bean images and reported 91.74%. Serttaş and Deniz (2023) achieved 98.33% accuracy using ResNet50 on 1,295 bean images, showing strong performance similar to DenseNet-based models. Kursun et al., (2023) reported the highest accuracy (100%) with DenseNet201, although the dataset contained only 60 dry bean images, which may have influenced the outcome.
           
    Among Faba bean studies, Salau et al., (2023) used an end-to-end CNN model and achieved 98.14%. Jeong and Na (2024) used a CNN on labeled leaf images, resulting in 91.00%. Mostafa et al., (2025) used a sequential CNN and reached 98.92% accuracy with an expert-preprocessed dataset. The present study used EfficientNetB0 on 8,021 expert-labeled Faba bean images and achieved 95.39%. This performance is higher than several earlier works but slightly lower than models trained on smaller or more specialized datasets.
     
    Real-world deployment challenges
     
    Field-based disease diagnosis presents substantially greater complexity than laboratory-controlled image classification. Variations in illumination caused by uneven sunlight, shadows from overlapping leaves, camera angle differences and background clutter can significantly alter leaf appearance. Such lighting variation was observed to reduce model confidence, particularly for early-stage infections where visual symptoms are subtle and spatially localized. Occlusion caused by overlapping foliage further complicates detection, as partial leaf visibility limits the availability of discriminative texture and color features. In addition, mixed infections, where multiple diseases co-occur on the same leaf, introduce ambiguous visual patterns that challenge single-label classification frameworks. These factors collectively explain why field-acquired datasets typically yield lower accuracy than laboratory datasets, despite more realistic deployment relevance.

    CONCLUSION

    This study presented an EfficientNetB0-based model for classifying four major faba bean leaf diseases using 8,021 field images, achieving a high accuracy of 95.39% with strong precision, recall and F1-scores across all classes. The results show that the model effectively distinguishes between healthy, rust, gall and chocolate spot leaves, demonstrating good generalization and stable learning. However, several limitations exist: the dataset covers only four disease categories, lacks early-stage symptom images and is restricted to specific geographic locations; only one architecture was tested; and environmental or temporal factors were not included. Future work should expand the dataset to include more diseases and diverse conditions, explore advanced architectures such as EfficientNetV2 or transformer-based models, integrate multispectral or drone-based data for higher sensitivity and develop lightweight mobile applications to support real-time field diagnostics.

    ACKNOWLEDGEMENT

    Funding details
     
    This research received no external funding.
     
    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.
     
    Ethical approval and consent
     
    All field images of Faba bean leaves were collected with the consent of the farm owners. The study did not involve any endangered species or human/animal subjects requiring formal ethical approval. Data collection adhered to standard agricultural research practices and local regulations.

    CONFLICT OF INTEREST

    The authors declare that they have no conflict of interest.

    REFERENCES


    1. Abed, S.H., Al-Waisy, A.S., Mohammed, H.J. and Al-Fahdawi, S. (2021). A modern deep learning framework in robot vision for automated bean leaves diseases detection. International Journal of Intelligent Robotics and Applications. 5(2): 235-251.

    2. Al-Sharqi, A.A., Dari, W.A., Jassim, R., Mohsin, Y.B., Hussian, A.K. and Mohmood, R.R. (2025). The synergistic effects of Lactobacillus acidophilus and Chlorella spp. against pathogens isolated from dermal infections. International Journal of Probiotics and Prebiotics. 20: 19-23. https:// doi.org/10.37290/ijpp2641-7197.20:19-23.

    3. Alshahrani, S.M. (2024). Knowledge, attitudes and barriers toward using complementary and alternative medicine among medical and nonmedical university students: A cross- sectional study from Saudi Arabia. Current Topics in Nutraceutical Research. 22(3): 889-894. https://doi.org/ 10.37290/ctnr2641-452X.22:889-894.

    4. Archana, R. and Jeevaraj, P.S.E. (2024). Deep learning models for digital image processing: A review. Artificial Intelligence Review. 57(1). https://doi.org/10.1007/s10462-023- 10631-z.

    5. Bitew, B., Fininsa, C. and Terefe, H. (2022). Estimating yield loss of faba bean (Vicia faba L.) caused by gall disease in north Shoa, Ethiopia. Crop Protection. 155: 105930. https://doi.org/10.1016/j.cropro.2022.105930.

    6. Chen, J., Chen, J., Zhang, D., Sun, Y. and Nanehkaran, Y. (2020). Using deep transfer learning for image-based plant disease identification. Computers and Electronics in Agriculture. 173: 105393. https://doi.org/10.1016/j.compag. 2020.105393.

    7. De Notaris, C., Enggrob, E.E., Olesen, J.E., Sørensen, P. and Rasmussen, J. (2023). Faba bean productivity, yield stability and N2-fixation in long-term organic and conventional crop rotations. Field Crops Research. 295: 108894. https://doi.org/10.1016/j.fcr.2023.108894.

    8. Dhull, S.B., Kidwai, M.K., Noor, R., Chawla, P. and Rose, P.K. (2021). A review of nutritional profile and processing of faba bean (Vicia faba L.). Legume Science. 4(3). https:// doi.org/10.1002/leg3.129.

    9. Dolatabadian, A., Neik, T.X., Danilevicz, M.F., Upadhyaya, S.R., Batley, J. and Edwards, D. (2024). Image based crop disease detection using machine learning. Plant Pathology. 74(1): 18-38. https://doi.org/10.1111/ppa.14006.

    10. Ghuriani, V., Wassan, J. T., Deolal, P., Sharma, V., Dalal, D. and Goyal, A. (2023). An integrative approach towards recommending farming solutions for sustainable agriculture. Journal of Experimental Biology and Agricultural Sciences. 11(2): 306-315. https://doi.org/10.18006/2023.

    11. Jain, E. and Aneja, A. (2025). Automated Detection and Classification of Bean Leaf Diseases using InceptionV3: A Deep Learning Approach. In Proceedings of the 2025 International Conference on Electronics and Renewable Systems (ICEARS). IEEE. (pp. 1890-1895). https://doi.org/10.1109/ ICEARS64219.2025.10941547.

    12. Jeong, H.Y. and Na, I.S. (2024). Efficient faba bean leaf disease identification through smart detection using deep convolutional neural networks. Legume Research-An International Journal. 47(8): 1404-1411. doi: 10.18805/LRF-798.

    13. Joshi, R.C., Kaushik, M., Dutta, M.K., Srivastava, A. and Choudhary, N. (2021). VirLeafNet: Automatic analysis and viral disease diagnosis using deep-learning in Vigna mungo plant. Ecological Informatics. 61: 101197.

    14. Karkanis, A., Ntatsi, G., Lepse, L., Fernández, J.A., Vågen, I.M., Rewald, B., Alsiòa, I., Kronberga, A., Balliu, A., Olle, M., Bodner, G., Dubova, L., Rosa, E. and Savvas, D. (2018). Faba bean cultivation-revealing novel managing practices for more sustainable and competitive European cropping systems. Frontiers in Plant Science. 9: 1115. https:// doi.org/10.3389/fpls.2018.01115.

    15. Kursun, R., Bastas, K.K. and Koklu, M. (2023). Segmentation of dry bean (Phaseolus vulgaris L.) leaf disease images with U-Net and classification using deep learning algorithms. European Food Research and Technology. pp 1-16.

    16. Manjunatha, V., Bhattacharjee, D. and Flores, C. (2022). Disease Management of Faba Beans. In: Faba Bean: Chemistry, Properties and Functionality. [Bangar, S.P. and Dhull, S.B. (Eds.)],  Springer. (pp 357-394). https://doi.org/10.1007/ 978-3-031-14587-2_14.

    17. Mostafa, A., Alnuaim, A. and AlZubi, A.A. (2025). Utilizing convolutional neural networks for accurate detection of leaf diseases in fava beans. Legume Research-An International Journal. 48(3): 494-502. doi: 10.18805/LRF-823.

    18. Saeidan, A., Caulfield, J., Vuts, J., Yang, N. and Fisk, I. (2025). Detection of aphid infestation on faba bean (Vicia faba L.) by hyperspectral imaging and spectral information divergence methods. Journal of Plant Diseases and Protection. 132(3): 109. https://doi.org/10.1007/s41348- 025-01100-6.

    19. Salau, A.O., Abeje, B.T., Faisal, A.N. and Asfaw, T.T. (2023). Faba Bean Disease Detection using Deep Learning Techniques.  In Proceedings of the 2023 International Conference on Cyber Management and Engineering (CyMaEn). IEEE. (pp. 344-349). https://doi.org/10.1109/CyMaEn57228.2023. 10051088.

    20. Segers, A., Dumoulin, L., Megido, R.C., Jacquet, N., Cartrysse, C., Kamba, P.M., Pierreux, J., Richel, A., Blecker, C. and Francis, F. (2021). Varietal and environmental effects on the production of faba bean (Vicia faba L.) seeds for the food industry by confrontation of agricultural and nutritional traits with resistance against Bruchus spp. (Coleoptera: Chrysomelidae, bruchinae). Agriculture Ecosystems and Environment. 327: 107831. https:// doi.org/10.1016/j.agee.2021.107831.

    21. Semba, R.D., Ramsing, R., Rahman, N., Kraemer, K. and Bloem, M.W. (2021). Legumes as a sustainable source of protein in human diets. Global Food Security. 28: 100520. https:/ /doi.org/10.1016/j.gfs.2021.100520.

    22. Serttaş, S. and Deniz, E. (2023). Disease detection in bean leaves using deep learning. Communications Faculty of Sciences University of Ankara Series A2-A3 Physical Sciences and Engineering. 65(2): 115-129.

    23. Shafik, W., Tufail, A., De Silva, L.C., Apong, R.A.A.H.M. and Kim, K. (2025). Deep learning technique for plant disease classification and pest detection and model explainability elevating agricultural sustainability. BMC Plant Biology. 25(1): 1491. https://doi.org/10.1186/s12870-025-07377-x.

    24. Sahu, P., Chug, A., Singh, A.P., Singh, D. and Singh, R.P. (2021). Deep learning models for beans crop diseases: Classification and visualization techniques. International Journal of Modern Agriculture. 10(1): 796-812.

    25. Singh, V., Chug, A. and Singh, A.P. (2023). Classification of beans leaf diseases using fine tuned CNN model. Procedia Computer Science. 218: 348-356. https://doi.org/ 10.1016/j.procs.2023.01.017.

    26. Soliman, A.A., Mousa, M.I., Mosalam, A.M., Ghareeb, Z.E., Ibrahim, S.D., Rehan, M., Yu, H. and He, Y. (2023). The potential genetic effect for yield and foliar disease resistance in faba bean (Vicia faba L.) assessed via morphological and SCoT markers. Plants. 12(20): 3645. https://doi.org/ 10.3390/plants12203645.
    In this Article
      Contact us
      Follow us
      Published In
      Legume Research

      Editorial Board

      View all (0)