Physical Education, Health and Social Sciences 116

[Link] E-ISSN: 2958-5996

P-ISSN: 2958-5988

Automated Classification of Healthy and Infected Grape Leaves Using a Deep

Convolutional Neural Network

Muhammad Zia1, Muhammad Uzair Khan1, Dawar Awan1, Muhammad Lais1*, Saadia
Tabassum2, Rohail Ali Khan1, Muhammad Awais Khan1
1
Department of Electrical Engineering Technology, Shuhada-e-APS University of Technology,
Nowshera
2
Department of Electronics Engineering Technology, Shuhada-e-APS University of Technology,
Nowshera
Corresponding Author: *Muhammad Lais, Email: lais@[Link]
DOI: [Link]
Abstract
This study proposes a deep learning-based method for the automated identification of grape leaf
diseases, focusing on the classification of infected and healthy grape leaves. A Deep Convolutional
Neural Network (CNN) model was designed and trained on a curated dataset comprising 1,180
diseased and 1,000 healthy grape leaf images. Data augmentation techniques, including rotation,
flipping, scaling, noise injection, gamma correction and principal component analysis (PCA), were
employed to improve model generalization and reduce overfitting. The model was trained using
optimized hyperparameters such as epoch, batch size, and dropout. Experimental results achieved
a classification accuracy of 97.25%, with precision, recall, and F1-score of 95.16%, 100%, and
97.52% for infected leaves, and 100%, 94%, and 96.91% for Healthy leaves, respectively. The
proposed model demonstrates reliable performance and can be integrated into precision agriculture
systems for early disease detection and crop management.

Keywords: Grape Leaf Disease Detection, Convolutional Neural Network, Deep Learning, Plant
Disease Identification, Automated Diagnosis, Transfer Learning, Machine Learning in Agriculture

Introduction
Grape stands among the world’s most significant fruit crops, cultivated across approximately 7.2
million hectares globally and valued at over USD 70 billion annually [1], Global grape production
reaches 74.7 million tonnes. After accounting for a significant loss of 4.4 million tonnes, 70.3
million tonnes are available for use. This total is divided into 34.0 million tonnes of pressed grapes
and 36.3 million tonnes of unpressed grapes. The pressed grapes are used to produce 237 million
hectoliters of wine and 2.7 million tonnes of musts and juices (which is equal to 22.0 million
hectoliters). The unpressed grapes are primarily allocated for fresh consumption as table grapes
(31.7 million tonnes) and for drying into 1.2 million tonnes of dried grapes, a process that requires
4.6 million tonnes of fresh grapes as shown in figure 1.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 117

Figure 1: Global Grapes Balance Sheet

In 2025 China accounting for about 49% of the total global fresh grape production, followed by
India and Turky as major producers as shown in figure 2.

Figure 2: Evaluation of world fresh grape production

Of the thousands of grapevine varieties known over 10,000 globally about 3,000 are actively
cultivated for wine, table fruit, juice, and dried formats. Despite its economic importance, grape
production remains highly susceptible to fungal, bacterial, and viral diseases, including powdery
mildew (Uncinula necator), downy mildew (Plasmopara viticola), Botrytis bunch rot (Botrytis
cinerea), anthracnose, and leaf spot diseases [2, 3]. These pathogens affect leaves, stems, and fruit,
leading to chlorosis, necrosis, defoliation, and decay; severe outbreaks can result in crop losses
exceeding 70% [4]. Traditional detection and management manual field scouting and lab testing
are time-consuming, labor-intensive, and often subjective, with visible symptoms appearing only
after disease establishment. This underscores the need for early, accurate, and automated detection
systems.
Advancements in deep learning, particularly CNNs, have paved the way for automated plant
disease detection. By employing data augmentation techniques, CNN-based systems can robustly
classify grape leaves as healthy or infected, outperforming traditional methods that rely on
handcrafted image features.
Moreover, modern sensor technologies including hyperspectral imaging, multispectral cameras,
RGB imaging, chlorophyll fluorescence, and thermal imaging can detect subtle physiological
changes in grape leaves before visible symptoms emerge [5, 6]. The overall risk of disease
occurrence in grapevines can be expressed mathematically as a function of host susceptibility,
tissue availability, inoculum pressure, and environmental conditions, as shown in Equation (1):
R(t) = S(t). A(t). I(t). W(t) (1)
Here, S(t) represents host susceptibility at time t, A(t) is the area of susceptible tissue, I(t) denotes
inoculum pressure, and W(t) captures the effect of weather favorability. Together, these factors
provide a framework for quantifying disease risk and emphasize why early detection systems are
essential for timely and efficient vineyard management.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 118

The paper is organized as follows: Section 2 reviews the related literature, Section 3 details the
materials and methodology adopted for plant disease identification, Section 4 reports and discusses
the experimental results, and Section 5 concludes the study with suggestions for future research.

Literature review
Grapevine (Vitis vinifera) is one of the most widely cultivated fruit crops, serving as the basis for
fresh fruit, raisins, juice, and wine. However, grapevines are highly vulnerable to a variety of
fungal, bacterial, and viral diseases that threaten both yield and quality. Among the most
destructive are powdery mildew (Erysiphe necator), downy mildew (Plasmopara viticola), and
Botrytis bunch rot (Botrytis cinerea) [7]. These pathogens attack leaves, stems, and fruits, leading
to chlorosis, necrosis, premature leaf drop, and reduced fruit quality. Uncontrolled outbreaks may
result in yield losses exceeding 70%.
As shown in figure 3, symptoms of powdery mildew and downy mildew are visible on grape leaves
and berries, manifesting as white powdery spots and yellow lesions, respectively. Such visual
symptoms highlight the destructive nature of these diseases and the challenges of relying on
manual identification.

Figure 3: Grapes and Leaves affected by Powdery mildew.

Beyond fungal infections, grapevines are also affected by trunk diseases, viral infections, and
secondary bacterial pathogens [8]. These challenges highlight the necessity of early detection and
monitoring systems, since latent infections can predispose berries to opportunistic organisms and
degrade grape quality even before visible symptoms appear [9].

Traditional Approaches to Disease Detection

Historically, vineyard management relied on manual field scouting and visual inspection to detect
disease symptoms [10]. While this method remains common, it is subjective, labor-intensive, and
prone to delays, as symptoms often become visible only after pathogens have spread significantly.
Figure 4 illustrates a vineyard worker visually inspecting grape leaves for symptoms of mildew.
While this method remains widely practiced, it is labor-intensive and prone to subjectivity, which
underscores the limitations of traditional disease detection strategies.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 119

Figure 4: Traditional Disease Detection Method

Laboratory-based methods such as ELISA (enzyme-linked immunosorbent assay) and PCR
(polymerase chain reaction) have provided accurate and reliable pathogen identification,
particularly for viruses and bacteria [11]. However, these techniques are expensive, time-
consuming, and impractical for large-scale vineyard monitoring. As vineyards continue to expand
globally, such limitations underscore the need for automated, scalable, and non-destructive
alternatives.
2.1 Imaging Technologies for Plant Disease Detection
Recent years have seen significant advances in imaging-based technologies for crop health
monitoring. These include:
 RGB Imaging: Widely available and cost-effective, RGB cameras capture visible
symptoms such as leaf spots, discoloration, or mildew patches. However, they perform
poorly in varying light conditions and cannot detect latent infections [12].
 Multispectral and Hyperspectral Imaging: Hyperspectral imaging captures reflectance
across hundreds of wavelengths, enabling early detection of physiological stress prior to
symptom development. For example, Calderón et al. demonstrated its effectiveness in
identifying downy mildew under vineyard conditions [13].
 Thermal Imaging: Pathogen-induced stomatal closure alters transpiration and canopy
temperature, which can be detected thermographically [14].
 Chlorophyll Fluorescence Imaging: Detects early photosynthetic disruptions caused by
pathogens, providing insight into plant stress before visible symptoms appear [15].
As depicted in figure 5, RGB cameras capture visible symptoms such as leaf spots, whereas
multispectral and hyperspectral sensors detect stress before symptoms appear. Similarly, thermal
imaging identifies stomatal closure, and chlorophyll fluorescence highlights early photosynthetic
disruption. Together, these technologies offer complementary insights into plant health.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 120

Figure 5: A visual comparison of RGB, multispectral, and thermal imaging used for
analyzing plants.
The integration of these imaging systems with UAVs (drones) has allowed large-scale vineyard
monitoring, combining high spatial resolution with rapid coverage [15]. While effective in
controlled experiments, adoption at commercial scale remains limited due to sensor cost, data
processing complexity, and computational demands [16].
2.2 Machine Learning and Deep Learning Approaches
The introduction of artificial intelligence (AI), particularly deep learning (DL), has revolutionized
plant disease detection. Traditional machine learning algorithms such as SVMs and Random
Forests required manual feature engineering based on color, texture, or shape. In contrast, CNNs
automatically extract hierarchical features, providing superior accuracy and robustness [17].
CNNs have achieved groundbreaking results in grape disease classification. Mohanty et al. [18]
showed that CNNs could classify multiple plant diseases with over 99% accuracy, while
Ferentinos [19] reported high accuracy across diverse datasets including grapevine leaves. Recent
innovations include:
 Lightweight CNNs optimized for mobile and edge devices, enabling real-time field
applications without requiring cloud-based computation [9].
 Attention-enhanced models (e.g., YOLOv5-CA) that improve detection of grape downy
mildew under variable lighting conditions [20].
 Ensemble models combining CNN feature extractors with classifiers like Random Forests,
which enhance robustness in heterogeneous vineyard environments [10].
 Vision Transformers (ViTs) integrated with CNNs to capture both spatial and global
contextual features, further improving accuracy in complex disease scenarios [8].
Transfer learning and fine-tuning of pretrained networks (such as ResNet, Inception, and
MobileNet) have also shown promising results in grapevine disease datasets, particularly where
data availability is limited [21].
Figure 6 presents a schematic of a CNN applied for grape leaf disease classification. CNNs
automatically extract hierarchical features such as edges, color patterns, and lesion shapes,
eliminating the need for handcrafted features. This architecture has enabled accuracies above
95% for grape disease classification in recent studies.

Figure 6: Plant Disease Detection using CNN architecture

Challenges in Automated Detection

Despite significant advancements, several challenges persist in grapevine disease detection
systems:
 Generalization – CNN models often lose accuracy when applied across different cultivars,
climates, or lighting conditions [19].
 Early Detection – Most imaging systems still rely on visible symptoms, while detecting
asymptomatic infections remains difficult [13].
 Multi-Disease Classification – Grapevines frequently face co-infections (e.g., powdery
mildew and Botrytis), complicating disease differentiation [20].
 Field Deployment – High costs of hyperspectral sensors and the computational intensity
of CNNs limit real-time use in commercial vineyards [16].
Explain-ability – CNNs are often treated as “black boxes.” Recent work on explainable AI
(XAI), such as Grad-CAM visualization, aims to make model predictions interpretable for
viticulturists [8].

Methodology
Data Acquisition and Pre-processing
The dataset consisted of 2,180 grape leaf images, including 1,180 infected leaves and 1,000 healthy
leaves as shown in figure 7. Images were collected from publicly available datasets and field

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 121

sources to ensure variability in environmental conditions, orientations, and backgrounds. All

images were resized to 150 × 150 pixels for uniformity and computational efficiency. The dataset
was split into 70% training (1,526 images), 20% validation (436 images), and 10% testing (218
images) using stratified sampling to maintain class balance. To enhance the generalization
capability of the model and reduce overfitting, data augmentation was applied to the training set.
The augmentation techniques included in the following table 1:
Table 1: Parameters of the proposed Deep CNN model
Parameters Value
Rotation +30
Width and height shift 20%
Shear and zoom 20
Horizontal flip 180
Brightness adjustment 0.8-1.2
Noise injection and gamma 20%
correction
PCA based augmentation 20%

Figure 7: Healthy and Infected Leaves

Image transformations are applied to expand the dataset and minimize overfitting by introducing
variations of the original images. Techniques such as flipping, gamma correction, noise injection,
PCA color augmentation, rotation, and scaling are employed to generate augmented images for
training.

Experimental Setup
The experiments were conducted on a system with the following configuration:
 Processor: Intel Core i7
 RAM: 16 GB
 GPU: NVIDIA CUDA-enabled (if available)
 Framework: TensorFlow (Keras API)
The hyperparameters used were:
 Batch size: 64
 Learning rate: 0.001
 Dropout rate: 0.2
 Epochs: 30
 Loss function: Categorical cross-entropy
 Metrics: Accuracy

CNN Model Architecture

In this research a sequential CNN designed for the binary classification of grape leaf images into
"healthy" or "infected" categories. The model processes input images of 150x150 pixels in RGB
format (3 channels) and utilizes a series of convolutional and pooling layers to automatically learn
and extract relevant features for accurate classification.
The network begins with an input layer that accepts RGB images, followed by three sequential
convolutional layers that progressively increase the feature depth while reducing spatial
dimensions as shown in figure 8. The first convolutional layer (Conv1) employs 32 filters of size
3×3, generating 32 feature maps (148 × 148 × 32), and is followed by a max-pooling layer (Pool1)
that reduces the dimensions to 74 × 74 × 32. The second convolutional layer (Conv2) uses 64
filters (3×3) producing 72 × 72 × 64 feature maps, followed by Pool2 (36 × 36 × 64). The third

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 122

convolutional layer (Conv3) applies 128 filters (3×3), producing 34 × 34 × 128 feature maps, and
Pool3 further reduces them to 17 × 17 × 128. These hierarchical convolution-pooling operations
allow the network to progressively learn low- to high-level features such as edges, textures, and
disease-related patterns. The output from Pool3 is flattened into a 1D feature vector of 36,992
elements, which is connected to a fully connected dense layer of 128 neurons with a dropout of
0.2 to prevent overfitting as shown in figure 9.
Parameter = (Filter H × Filter W × Input Channels + 1(Bias)) (2)
× Numberof Filters
Finally, a softmax output layer with two neurons provides class probabilities for the healthy and
infected classes. The entire network comprises 4,828,610 trainable parameters (18.42 MB), all
optimized using the Adam optimizer and categorical cross-entropy loss during training. Formula
used for finding the parameter is given in equation 2.

Figure 8: shows the original and visual representation of the different layers’ outputs of the
proposed model.

a) Input image

b) Conv layer 1

c) Pool 1

d) Conv layer 2

e) Pool 2

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 123

f) Conv layer 3

g) Pool 3

h) 1’D features vector

Figure 9: Convolution layers
Down-sampling is performed by the pooling layer along the spatial dimensions, which serves to
reduce the model's parameter count. The proposed architecture employs max pooling, a method
demonstrated to achieve superior performance over average pooling. Furthermore, a dropout layer
is integrated for regularization, mitigating overfitting by randomly removing network units.
Empirical evaluation of dropout probabilities (ranging from 0.2 to 0.8) indicated that a rate of 0.2
produced the optimal validation accuracy, shown in figure 10.

Figure 10: Accuracy at different dropout values

The impact of batch size on model performance was systematically evaluated as shown in figure
11. A batch size of 16 yielded the highest performance, with a training accuracy of 99.67% and a
validation accuracy of 99.54%, demonstrating exceptional learning efficacy and generalizability.
Employing a batch size of 32 also resulted in strong performance, achieving a training accuracy of
99.48% and a validation accuracy of 99.37%, though marginally lower than the 16-unit batch. A
further increase to a batch size of 64 led to a noticeable decline in both training (99.15%) and
validation (99.08%) accuracy. Notably, the largest tested batch size of 128 exhibited a substantial
performance drop, with a marked disparity between training accuracy (97.71%) and validation
accuracy (94.43%). This significant gap indicates a pronounced tendency towards overfitting and
reduced model generalization at larger batch sizes.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 124

Figure 11: Batch size comparison

The experimental results indicate that the model maintained stable performance at 10 and 15
epochs, achieving training and validation accuracies of approximately 96% and 97%, respectively.
The highest performance was observed at 20 epochs, where the validation accuracy reached
98.62%, demonstrating strong generalization capability. However, extending the training to 25 and
30 epochs led to a decline in validation accuracy to 92.43% and 93.34%, despite consistently high
training accuracy. This trend suggests the onset of overfitting, indicating that 20 epochs provided
the most optimal balance between training accuracy and validation performance. A comparison of
the validation performance for different batch sizes is presented in figure 12.

Figure 12: Accuracy at Different Epochs

Results and discussion

The dataset consisted of 1000 healthy and 1180 infected grape leaf images, divided into 70%
training, 20% validation, and 10% testing. To improve robustness, six augmentation methods
(flipping, gamma correction, noise injection, PCA color augmentation, rotation, and scaling) were
applied to the training data.
Figure 13 shows transfer learning models with the proposed CNN. AlexNet (87.34%) performed
lowest, while ResNet and VGG16 (92.56%) and Inception v3 (94.32%) performed better. The
proposed CNN achieved 98.25% validation accuracy, surpassing all transfer learning methods due
to its task-specific design, augmented dataset, and balanced class distribution.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 125

Figure 13: Validation accuracy of the popular transfer learning approaches

As shown in figure 14, conventional classifiers such as SVM (50.69%), Decision Tree (72.23%),
Logistic Regression (80.99%), and KNN (87.86%) achieved moderate accuracies. In contrast, the
proposed CNN model reached 97.25% testing accuracy, demonstrating the advantage of automatic
feature extraction and augmented data in capturing complex disease patterns.

Figure 14: Average testing accuracy of the different models

Precision is defined as the ratio of correctly predicted positive instances (True Positives, TP) to
the total instances predicted as positive (TP + FP). Its value ranges between 0 and 1 and is
expressed mathematically as given in equation 3:
TP (3)
Precision =
TP + FP

Precision measures the proportion of correctly identified positive cases among all predicted
positives. As illustrated in figure 15, the proposed Deep CNN model demonstrates significantly
higher Precision compared to the other machine learning techniques.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 126

Figure 15: Precision values of the different models

Recall is defined as the ratio of correctly predicted positive instances (TP) to all actual positive
instances (TP + FN), as given in equation 4:

TP (4)
Recall =
TP + FN

Recall indicates the proportion of actual positive samples that are correctly identified by the model.
As shown in figure 16, the proposed Deep CNN achieves higher recall values compared to the
other models on the testing dataset.

Figure 16: Recall values of the different models

F1-score is a widely adopted metric for evaluating machine learning models, defined as the
harmonic mean of Precision and Recall. Its value ranges between 0 and 1 and is expressed as in
equation 5:
2TP (5)
F1 =
2TP + FP + FN

The F1-score represents the balance between Precision and Recall, indicating the proportion of
instances correctly classified by the model. As illustrated in figure 17, the proposed Deep CNN
model achieves a substantially higher F1-score compared to the other algorithms.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 127

Figure 17: F1 Score of the different Models

Overall, these results confirm that the proposed approach outperforms conventional machine
learning methods in plant disease identification using leaf images.

Conclusion
This research introduced a deep learning-based framework for the automatic detection of grape
leaf diseases through a customized CNN. The model was trained on a balanced dataset of healthy
and diseased leaf images, where augmentation techniques were applied to enhance generalization
and reduce the risk of overfitting. Experimental analysis showed that the proposed CNN achieved
a testing accuracy of 97.25%, outperforming conventional machine learning algorithms as well as
transfer learning models including AlexNet, ResNet, VGG16, and Inception v3. In addition, the
model recorded higher Precision, Recall, and F1-scores, demonstrating its robustness and
reliability in disease classification.
Overall, the results suggest that the proposed CNN can be a valuable tool in precision agriculture,
enabling early disease detection and supporting efficient crop management. Future research will
aim to expand the framework toward multi-class disease recognition, incorporate real-time
detection capabilities, and validate the system on larger and more diverse field datasets.

References
1. Lu, G., et al., Grapevine double cropping: a magic technology. Frontiers in Plant Science,
2023. 14: p. 1173985.
2. Compendium of grape diseases, disorders, and pests. 2015.
3. Dufour, R., Grapes: organic production. United States: ATTRA-National Sustainable
Agriculture Information Service, 2006.
4. Ficke, A., D.M. Gadoury, and R.C. Seem, Ontogenic resistance and plant disease
management: A case study of grape powdery mildew. Phytopathology, 2002. 92(6): p. 671-
675.
5. Savary, S. and L. Willocquet, Modeling the impact of crop diseases on global food security.
Annual review of phytopathology, 2020. 58(1): p. 313-341.
6. Schaad, N.W., J.B. Jones, and W. Chun, Laboratory guide for the identification of plant
pathogenic bacteria. 2001.
7. Xie, X., et al., A deep-learning-based real-time detector for grape leaf diseases using
improved convolutional neural networks. Frontiers in plant science, 2020. 11: p. 751.
8. Kunduracioglu, I. and I. Pacal, Advancements in deep learning for accurate classification
of grape leaves and diagnosis of grape diseases. Journal of Plant Diseases and Protection,
2024. 131(3): p. 1061-1080.
9. Karim, M.J., et al., Enhancing agriculture through real-time grape leaf disease
classification via an edge device with a lightweight CNN architecture and Grad-CAM.
Scientific Reports, 2024. 14(1): p. 16022.
10. Ishengoma, F.S. and N.N. Lyimo, Ensemble model for grape leaf disease detection using
CNN feature extractors and random forest classifier. Heliyon, 2024. 10(12).
11. Praveen, S.P., et al., A novel classification approach for grape leaf disease detection based
on different attention deep learning techniques. International Journal of Advanced
Computer Science and Applications (IJACSA), 2023. 14(6): p. 2023.
12. Kamilaris, A. and F.X. Prenafeta-Boldú, Deep learning in agriculture: A survey.
Computers and electronics in agriculture, 2018. 147: p. 70-90.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025
 128

13. Sutanudjaja, E., et al., Using ERS spaceborne microwave soil moisture observations to
predict groundwater head in space and time. Remote sensing of environment, 2013. 138:
p. 172-188.
14. Wang, S., et al., Advances in deep learning applications for plant disease and pest
detection: A review. Remote Sensing, 2025. 17(4): p. 698.
15. Perich, G., et al., Crop nitrogen retrieval methods for simulated sentinel-2 data using in-
field spectrometer data. Remote sensing, 2021. 13(12): p. 2404.
16. Ye, W., et al., Detection of pesticide residue level in grape using hyperspectral imaging
with machine learning. Foods, 2022. 11(11): p. 1609.
17. Knoll, F.J., et al., Improving efficiency of organic farming by using a deep learning
classification approach. Computers and electronics in agriculture, 2018. 153: p. 347-356.
18. Mohanty, S.P., D.P. Hughes, and M. Salathé, Using deep learning for image-based plant
disease detection. Frontiers in plant science, 2016. 7: p. 215232.
19. Ferentinos, K.P., Deep learning models for plant disease detection and diagnosis.
Computers and electronics in agriculture, 2018. 145: p. 311-318.
20. Zámková, M., et al., Consumers’ behavior in the field of organic agriculture and food
products during the COVID-19 pandemic in the Czech Republic: Focus on a comparison
of hyper-, super-and farmers’ markets and direct purchases from producers. Agriculture,
2023. 13(4): p. 811.
21. Too, E.C., et al., A comparative study of fine-tuning deep learning models for plant disease
identification. Computers and Electronics in Agriculture, 2019. 161: p. 272-279.

__________________________________________________________________________________________________________________
Volume 3, No. 3 July – September, 2025