Faba bean (Vicia faba L.) is one of the oldest cultivated legume crops. It is an important grain legume cultivated for human consumption, animal feed and soil fertility enhancement (Etemadi et al., 2019). It represents a valuable source of plant-based protein, carbohydrates, vitamins and minerals and plays a central role in sustainable cropping systems due to its nitrogen-fixing ability (Dhull et al., 2021). In many regions, particularly in Asia and Africa, faba bean contributes significantly to food security and smallholder farmer income (Acheampong et al., 2025). Beyond its nutritional importance, the crop improves soil structure and reduces dependency on synthetic nitrogen fertilizers, thereby supporting environmentally sustainable agricultural practices (Chetto et al., 2025). Na et al., (2024) demonstrated that integrating remote sensing with AI enables effective monitoring of legume crop health and growth using satellite and sensor data. Pei et al., (2026) highlighted that traditional methods were insufficient for controlling peanut diseases and pests due to genetic and environmental challenges. Potharaju et al., (2026) proposed an explainable Vision Transformer (ViT-B/16) model with transfer learning for bean leaf disease classification, achieving 97.52% accuracy. Their approach used GradCAM++ to provide visual explanations, improving model transparency and trust for real-world agricultural applications.

Despite its agronomic and economic value, faba bean production is frequently constrained by foliar diseases that reduce both yield and grain quality. Among the most destructive diseases are rust (Uromyces viciae-fabae), chocolate spot (Botrytis fabae) and faba bean gall (Zewide and Ademe, 2025). Rust disease is characterized by reddish-brown pustules on leaf surfaces and spreads rapidly under warm and humid conditions. Chocolate spot appears as dark necrotic lesions that expand and may coalesce, leading to premature leaf senescence. Gall disease causes abnormal swelling and distortion of leaf tissues, which interferes with normal physiological processes (Segers et al., 2020; Villegas-Fernández et al.,  2023). Under severe infection, these diseases substantially reduce photosynthetic activity and biomass accumulation, resulting in significant yield losses that may exceed 30% in epidemic years (Phiri et al., 2024).

Conventional disease diagnosis in faba bean fields relies primarily on visual inspection performed by farmers or agricultural experts. Although this method is widely practiced, it is subjective and depends heavily on individual experience (Zeleke et al., 2023). Early-stage symptoms are often subtle and may resemble nutrient deficiencies or abiotic stress effects. Furthermore, large-scale farms make manual scouting time-consuming and labor-intensive. Misclassification can lead to inappropriate fungicide application, increased production costs, environmental contamination and the development of pathogen resistance (Shafik et al., 2025). Consequently, there is a growing need for accurate, rapid and automated disease detection systems that can assist decision-making in integrated pest management programs.

Recent advances in computer vision and artificial intelligence have opened new possibilities for automated crop health monitoring (Shehu et al., 2025). Convolutional Neural Networks (CNNs) have demonstrated remarkable performance in image classification tasks by automatically learning hierarchical feature representations from raw pixel data. In agricultural applications, CNN-based approaches have been successfully applied to detect diseases in crops such as tomato, potato, maize and grape (Dhanya et al., 2022; Ibrahim et al., 2025). These models eliminate the need for handcrafted feature extraction and can capture complex visual patterns in colour, texture and lesion morphology.

MobileNet is a lightweight CNN architecture designed to achieve high efficiency with reduced computational complexity (Zhao et al., 2022). It employs depthwise separable convolutions, which decompose standard convolution into depthwise and pointwise operations, thereby significantly reducing the number of parameters and floating-point operations (Paek et al., 2026; Souza et al., 2026; Sriram and Kumari, 2026). Although MobileNet has been widely used with pretrained weights in agricultural studies, limited research has explored its performance when trained from scratch on field-acquired faba bean datasets (Mohammad et al., 2026; Özçelik et al., 2026). Training without pretrained ImageNet weights allows the network to learn feature representations exclusively from domain-specific leaf images, potentially enhancing sensitivity to disease-related visual patterns (Shi, 2024).

In this context, the present study develops a MobileNet-based convolutional neural network trained from scratch for the classification of faba bean leaf diseases. The dataset comprises 8,021 RGB images collected under natural field conditions and verified by agricultural experts. Four categories are considered: healthy leaves, rust-infected leaves, faba bean gall-infected leaves and chocolate spot-infected leaves. The model is evaluated using comprehensive performance metrics and confusion matrix analysis.
 
Research gap and objective
 
Despite these advancements, limited studies focus on faba bean disease detection under real-field conditions using lightweight models. Most existing methods rely on pretrained networks or controlled datasets, which may not be suitable for practical use. There is also limited work on training models from scratch using crop-specific data. This highlights a research gap in developing efficient and field-applicable solutions for faba bean disease detection.

The primary objectives and contributions of this study are summarized as follows:
· To investigate whether a lightweight MobileNet architecture trained from scratch can achieve reliable disease classification performance under realistic field conditions.
· To contribute to the limited body of deep learning research focused specifically on underrepresented crops such as faba bean.
· To develop a scalable and computationally efficient disezse detection framework disease detection framework suitable for precision agriculture applications.

Experimental setup
 
All experiments were performed on a Windows 10 workstation equipped with an Intel® Core™ i5-11320H processor operating at 3.20 GHz and 16 GB RAM. The implementation was carried out using Python 3.11 within a Jupyter Notebook environment to ensure flexibility in experimentation and result visualization. The deep learning framework TensorFlow 2.x with the Keras API was employed for model development, training and evaluation. NumPy and Pandas were used for numerical operations and data handling, while Matplotlib and Seaborn supported graphical analysis of learning curves and performance metrics. Scikit-learn was utilized for computation of classification metrics and the confusion matrix. To ensure reproducibility, random seeds were fixed for NumPy, TensorFlow and Python’s random module (np. random. seed(42), tf.random.set_seed(42), random. seed(42). A fixed random state (42) was also used during dataset splitting to ensure consistent results across runs. The overall methodological workflow followed in this study is illustrated in Fig 1, which summarizes the sequential stages from data acquisition to model evaluation.


 
Dataset description
 
The dataset used in this study consists of 8,021 RGB images of faba bean leaves. Data collection was conducted across multiple farms located in Riyadh, Saudi Arabia. Images were captured using portable handheld devices, including smartphones integrated with the Open Data Kit (ODK) platform, at varying angles and illumination levels to reflect realistic agricultural environments. Environmental conditions during image acquisition included natural daylight, partial shading and varying background complexity typical of real field settings. Disease labeling was performed and verified by agricultural experts to ensure annotation reliability. Four classes were considered: healthy leaves (2,019 images), rust-infected leaves (2,000 images), faba bean gall-infected leaves (2,000 images) and chocolate spot-infected leaves (2,002 images). The dataset was divided into training, validation and testing subsets using an 80:10:10 ratio. Accordingly, 4,812 images were allocated for training, 1,604 for validation and 1,605 for testing. This stratified split preserves class proportions across subsets and enables unbiased performance evaluation.
 
Image preprocessing
 
Prior to model training, all input images were resized to 224 x 224 pixels to ensure compatibility with the standard input dimension of the MobileNet architecture. Consequently, each image was represented as a three-dimensional tensor of size 224×224×3, where the three channels correspond to the red, green and blue (RGB) components. Standardizing the spatial resolution ensures uniformity across samples and enables efficient batch processing during training. The original pixel intensity values ranged from 0 to 255. To improve numerical stability and facilitate efficient gradient-based optimization, pixel normalization was performed according to equation 1.
 

 
Where,
x = Original pixel intensity.
x_norm = Normalized value within the range [0, 1].

This scaling reduces large gradient fluctuations and supports faster convergence during backpropagation by maintaining consistent feature magnitudes across layers. In addition to image preprocessing, categorical labels were encoded using one-hot representation to match the multi-class classification framework. For a four-class problem, the target vector yis defined as
 

 

 
Where only one element of the vector equals 1, indicating the true class and the remaining elements are set to 0. This encoding scheme enables the use of categorical cross-entropy loss during model training and ensures proper gradient computation for multi-class probability outputs.
 
Data augmentation
 
To improve model generalization and reduce overfitting, data augmentation was applied to the training set. Field-acquired images exhibit natural variability in orientation, scale and perspective. Augmentation simulates such variations and enhances robustness. The following transformations were applied randomly during training: horizontal flipping, random rotation within ±10% and random zoom up to 10%. If I(x,y) denotes the original image, an augmented image I' (x,y) can be expressed as
 

 

 
Where,
T = Represents a geometric transformation operator including rotation, scaling, or reflection.

These transformations preserve class identity while increasing the effective diversity of training samples. Augmentation was not applied to validation or testing datasets to ensure objective performance assessment. The integration of preprocessing and augmentation steps into the training pipeline ensures that the proposed model learns invariant features relevant to disease classification under real field conditions.
 
MobileNet architecture
 
In this study, the MobileNet architecture was implemented without pretrained ImageNet weights (weights=None) and all network parameters were initialized randomly. This design choice was made because the target dataset consists of domain-specific agricultural images (faba bean leaves), which differ significantly from the natural images in the ImageNet dataset. Using pretrained weights could introduce irrelevant feature biases, whereas training from scratch allows the model to learn task-specific and disease-relevant features directly from the dataset. Additionally, the relatively balanced dataset and sufficient sample size support effective training without reliance on transfer learning. This approach ensures that the feature representations are learned exclusively from the faba bean leaf dataset rather than being influenced by features extracted from unrelated natural image domains. The architecture follows the standard MobileNet design, which is built upon depthwise separable convolutions to reduce computational complexity while maintaining classification performance. In conventional convolution, the computational cost is determined by equation 4.
 

 

 
Where,
D_k = Kernel size.
M = Number of input channels.
N = Number of output channels.
Df = Spatial dimension of the feature map.

In contrast, depthwise separable convolution decomposes this operation into a depthwise convolution followed by a pointwise convolution, resulting in the reduced computational cost shown in equation 5.
 

 

 
This factorization substantially lowers the number of parameters and floating-point operations compared to standard convolution. As a result, MobileNet is computationally efficient and suitable for deployment in resource-constrained agricultural environments. Following the convolutional backbone, a Global Average Pooling (GAP) layer was introduced to reduce the spatial dimensions of the feature maps and minimize overfitting by replacing fully connected layers with spatial averaging. The final classification stage consists of a Dense layer with four neurons corresponding to the four disease categories. The Softmax activation function was applied to generate normalized class probabilities, as defined in equation 6.
 

 

 
Where,
z_i = Represents the logit for class i.

The denominator ensures that the predicted probabilities sum to one across all four classes. Fig 2 illustrates the architecture of the proposed MobileNet model trained from scratch.


 
Training strategy
 
The proposed model was compiled using the Adam optimizer, which combines adaptive learning rate estimation with momentum-based updates. The initial learning rate was set to
 

 

 
To ensure stable convergence during training from scratch. The categorical cross-entropy loss function was employed to measure the discrepancy between true labels and predicted probabilities. The loss function is defined as:
 

 

 
Where,
yi = Denotes the true class label.
i  = Represents the predicted probability for class i.

Training was conducted with a batch size of 32 for a maximum of 50 epochs. To prevent overfitting and improve generalization, early stopping was applied with a patience value of 5 epochs based on validation performance. The best-performing model weights were automatically restored after training. This strategy ensured that the final model corresponded to the optimal validation accuracy achieved during the learning process.

To evaluate the robustness and generalization performance of the proposed model, 5-fold cross-validation was conducted on the training dataset. The dataset was first loaded and converted into NumPy arrays, after which it was partitioned into five subsets using the KFold method with shuffling enabled and a fixed random seed (42) to ensure reproducibility.
 
Evaluation metrics
 
The performance of the proposed MobileNet model was comprehensively evaluated using confusion matrix analysis, standard classification metrics, Receiver Operating Characteristic (ROC) curves, Precision-Recall (PR) curves and prediction confidence scores. These complementary measures provide a detailed assessment of model behavior across all four disease classes.

Fig 3 presents the training and validation accuracy curves of the MobileNet model over 50 epochs. The model exhibited rapid convergence during the initial training phase, with accuracy increasing from approximately 40% to above 90% within the first 10 epochs. This sharp improvement indicates effective early-stage feature extraction and stable gradient updates.



As training progressed, both training and validation accuracy curves gradually stabilized and continued to improve at a slower rate, eventually reaching 97.36% training accuracy and 96.32% validation accuracy at Epoch 50. The minimal gap of approximately 1% between training and validation accuracy demonstrates strong generalization performance and suggests the absence of significant overfitting. The close alignment of the two curves throughout the training process further confirms the robustness of the optimization strategy and the suitability of the selected hyperparameters. The plateau observed after approximately 30 epochs indicates convergence of the learning process, where additional epochs contributed marginal but consistent improvements. Notably, despite being trained from scratch without pretrained weights, the MobileNet architecture achieved high validation performance, demonstrating its capability to learn discriminative features effectively using depthwise separable convolutions.

Table 1 presents the confusion matrix of the MobileNet model for four-class classification. Out of 1605 total samples, 1527 were correctly classified, resulting in an overall accuracy of 95.14%, which confirms strong predictive performance. The diagonal elements show high true positive rates for all classes, indicating effective discrimination among disease categories. The Healthy class achieved the highest correct predictions (403 out of 404), with only one misclassification, demonstrating excellent recognition capability for healthy leaves. The Gall class recorded 377 correct predictions out of 400, with minor confusion mainly toward Chocolate Spot and Rust. Chocolate Spot showed 360 correct predictions out of 401 samples, with some misclassification primarily into Rust (24 samples). Similarly, Rust achieved 387 correct predictions out of 400, with limited confusion toward Chocolate Spot and other classes. The concentration of values along the diagonal and relatively small off-diagonal errors indicate that the model learned discriminative features effectively, with only minor overlap between visually similar disease symptoms. Overall, the confusion matrix confirms strong class-wise performance and reliable generalization across all categories.



The classification metrics demonstrate strong and balanced performance of the MobileNet model across all four classes (Fig 4). An overall accuracy of 95.14% was achieved on 1605 test samples, confirming high predictive reliability. The healthy class recorded the highest recall (99.75%) and F1-score (98.05%), indicating near-perfect identification with minimal false negatives. The gall class also showed robust performance, with a precision of 97.42% and an F1-score of 95.81%, reflecting low false positive and false negative rates. Although chocolate spot exhibited slightly lower recall (89.78%), its precision remained high (94.24%), suggesting that most predicted cases were correct despite some missed detections. Similarly, rust achieved balanced performance with a recall of 96.75% and an F1-score of 94.62%, indicating effective recognition with minor misclassification. The macro and weighted averages (approximately 95.1% across precision, recall and F1-score) confirm consistent performance across classes without significant bias toward any specific category. Overall, the results indicate stable, well-generalized multi-class classification performance.



Fig 5 shows ROC curves and PR curves for fava bean disease classes. The model achieved strong class separability, as evidenced by high AUC values for all classes: chocolate spot (0.9980), gall (0.9994), healthy (1.0000) and rust (0.9991). Precision-recall analysis further confirmed this performance, with average precision (AP) values of 0.9941, 0.9982, 0.9999 and 0.9977, respectively. The 5-fold cross-validation results demonstrate consistent model performance across different data splits. Specifically, the model achieved accuracies of 0.9460, 0.9460, 0.9168, 0.9366 and 0.9449 for folds 1 through 5, respectively. The overall mean cross-validation accuracy was 0.9381 with a standard deviation of 0.0112, indicating stable and reliable generalization capability.



Fig 6 shows representative test sample predictions with corresponding true (T) and predicted (P) labels and confidence scores. The majority of samples are correctly classified with high confidence values exceeding 98%, demonstrating strong model certainty and reliable feature learning. A few misclassifications are present, such as a gall leaf predicted as rust with 75.46% confidence and a rust sample predicted with comparatively lower confidence (69.77%), indicating overlapping visual characteristics between certain disease categories. Overall, the high confidence levels for correct predictions further validate the model’s discriminative capability and classification reliability.



For baseline comparison, additional experiments were conducted using MobileNet (trained from scratch), ResNet50 and EfficientNetB0 under identical training conditions. The proposed MobileNet model achieved a test accuracy of 95.14%, demonstrating effective learning of domain-specific features without relying on pretrained weights. In comparison, ResNet50 and EfficientNetB0 achieved test accuracies of 93.70% and 90.90%, respectively, showing comparable performance but with higher computational complexity, particularly in the case of ResNet50 (23.59M parameters) and EfficientNetB0 (4.05M parameters).

These results indicate that MobileNet provides an efficient and effective solution for this task while maintaining competitive accuracy. The model size comparison further highlights its advantage (3.23M parameters) over ResNet50 (23.59M) and EfficientNetB0 (4.05M).

Several recent studies have demonstrated the effectiveness of lightweight deep learning architectures, particularly MobileNet variants, for plant leaf disease classification across different crops (Table 2). Zaki et al., (2020) fine-tuned MobileNetV2 for tomato leaf disease detection using the PlantVillage dataset and achieved over 90% accuracy. Elfatimi et al., (2022) applied MobileNet and MobileNetV2 for bean leaf disease classification and reported more than 97% training accuracy and over 92% test accuracy. Similarly, Jeong and Na (2024) developed a deep CNN model for Faba bean diseases and obtained 99.37% training accuracy with an overall accuracy of 91%, while Salau et al., (2023) showed that preprocessing significantly improved CNN performance from 92.1% to 98.14% training accuracy. Sireesha et al., (2025) further demonstrated the superiority of MobileNetV2 with transfer learning for potato leaf disease detection, achieving 98.33% accuracy. Appati et al., (2026) enhanced MobileNet with custom layers and ensemble techniques for tomato early blight detection, attaining 94.5% accuracy on field data and up to 100% under controlled conditions.



In comparison with these studies, the present work implemented MobileNet trained from scratch for four-class Faba bean leaf disease classification and achieved a test accuracy of 95.14% on 1,605 samples. Unlike several prior approaches that relied heavily on transfer learning or ensemble methods, this study demonstrated that a carefully optimized lightweight architecture can achieve competitive and balanced performance without pretrained weights.
 
Limitations and practical implications
 
Despite the strong performance, some limitations were observed. Minor misclassifications occurred between visually similar classes, particularly chocolate spot and rust, likely due to overlapping lesion patterns, color similarities and variations in illumination conditions. These challenges are common in field-acquired datasets where environmental variability affects feature consistency. Additionally, while the model achieved high accuracy, its performance may vary when applied to unseen datasets with different backgrounds or acquisition conditions.

From a practical perspective, the proposed MobileNet model is computationally efficient, with a relatively small size (3.23M parameters) and low inference time (~11.34 ms per image), making it suitable for deployment on mobile and edge devices. This enables real-time disease detection in field conditions, supporting precision agriculture and early intervention. However, further optimization and validation on embedded hardware platforms are recommended for large-scale deployment.

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.
 
Code availability
 
The implementation code and detailed experimental setup will be made available upon request from the corresponding author.
 
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.

Authors declare that they have no conflict of interest.