Plant Disease Detection

Using Machine Learning

Mrs. Nidhi Sharma
Paras Dwivedi Ram Sharma Devesh Rai
Department of CSE
Department of CSE Department of CSE Department of CSE
G L Bajaj Institute of
G L Bajaj Institute of G L Bajaj Institute of G L Bajaj Institute of
Technology and Technology and
Technology and Technology and
Management Management
Management Management
Greater Noida, India
Greater Noida, India Greater Noida, India Greater Noida, India
[email protected]
[email protected] [email protected] [email protected]

Abstract-- Plant diseases significantly threaten global crop comprehensive dataset covering 38 distinct plant disease
yields, impacting both food security and farmer’s incomes. classes, including healthy leaves. To extend its practical
Accurate and early detection of plant diseases is crucial for value, the model is deployed through a Flask-powered web
effective intervention and management. This paper explores
application that enables users to upload plant leaf images for
a deep learning-based approach for plant disease
real-time diagnosis. The system supplements its predictions
with symptom descriptions, recommended treatments, and
classification using a Convolutional Neural Network (CNN)
direct marketplace links by integrating structured data from
model. The model was trained using a large dataset
curated Excel files. This end-to-end framework offers an
consisting of 61,486 plant leaf images, covering 38 distinct
efficient, accessible, and scalable solution for early plant
classes of plant diseases, including healthy leaves. We tested
disease detection and management.
the model on a separate dataset of 54,306 images, achieving a
classification accuracy of 88.62% and a macro-averaged F1-
II. LITERATURE SURVEY
score of 0.84. Performance evaluation through a confusion
matrix and classification report shows that the model The early methods of plant disease detection heavily relied
performs strongly across most disease categories. This model
on manual inspection and standardized physical assessments.
aims to assist farmers and agricultural experts in early
Stubbs et al. provided a fundamental foundation through the
disease detection, thereby improving crop health, boosting
yield, and minimizing agricultural losses. Cereal Disease Methodology Manual, detailing visual-based
and field inspection methods to identify cereal crop diseases.
Keywords-Machine Learning (ML), Plant Disease However, such traditional techniques were time-consuming,
Detection, Plant Leaf Analysis, Image Classification, highly subjective, and limited in scope, particularly for large-
Convolutional Neural Network (CNN). scale applications. [1]

I. INTRODUCTION Recognizing the broader impact of plant diseases, Strange

and Scott (2005) emphasized that diseases pose a severe
Plant diseases pose a serious threat to global agriculture, threat to global food security. They advocated for more
causing significant yield losses and impacting food
efficient detection and management strategies, stressing that
security. These diseases stem from various biological
agents—such as fungi, bacteria, and viruses—as well as increasing global agricultural demands would require a shift
environmental stressors like drought and nutrient from manual techniques to more automated and scalable
deficiencies. Their symptoms often overlap, including approaches. [2]
spots, blights, wilting, and discoloration, making early and
accurate diagnosis challenging. Różewicz et al. (2021) further discussed the persistent
problem of fungal diseases in cereals and the need for
Traditional diagnostic methods rely on expert analysis and
integrated management solutions, emphasizing how
lab testing, which are time-consuming, costly, and often
inaccessible to farmers in rural regions. Moreover, visual environmental factors, fertilization practices, and crop
assessment is prone to error, especially in the early stages rotations influence disease prevalence and severity. [3]
when multiple diseases appear similar.
With the limitations of manual detection apparent, molecular
Recent advances in artificial intelligence (AI), particularly diagnostic techniques became pivotal. Schaad et al. (2003)
in deep learning, have enabled automated solutions for reviewed the advancements in molecular-based diagnostics
plant disease detection. Convolutional Neural Networks such as PCR and real-time PCR technologies. These methods
(CNNs) are especially effective for image classification
greatly improved the accuracy, sensitivity, and speed of
due to their ability to learn complex visual patterns directly
from raw data. pathogen detection and helped enforce biosecurity and
phytosanitary measures worldwide. [4]
This study introduces a CNN-based approach trained on a
Despite their effectiveness, molecular techniques required Similarly, Van Klompenburg et al. (2020) reviewed the use
specialized laboratories, equipment, and expertise, making of ML in crop yield prediction, stressing that robust datasets
them less accessible for small-scale farmers or field-level capturing environmental and biological variability are crucial
operations, thus creating a need for more portable, not just for yield prediction but also for reliable disease
automated solutions. detection models. [11]

Machine Learning (ML) introduced a paradigm shift by Hasan et al. (2024) further emphasized dataset importance in
enabling the automation of plant disease detection through their work on weed detection benchmarks. They curated
data-driven models. detailed, annotated datasets at the object level for precision
agriculture, noting that without diverse and realistic datasets,
Nturambirwe and Opara (2020) showcased ML even the best models fail under variable field conditions. [12]
applications for non-destructive defect detection in
horticultural products, utilizing sensor data combined with Future research should focus on developing large-scale,
algorithms like Support Vector Machines (SVM) and diverse, field-captured datasets, exploring transfer learning,
Neural Networks. Their review highlighted the potential of improving model interpretability, and designing hybrid
ML to achieve real-time, scalable quality assessment, multi-modal diagnostic systems.
reducing dependency on human expertise. [5]
III. METHODOLOGY
Similarly, Waldamichael et al. (2022) focused specifically
on cereal crops, reviewing how ML techniques like Dataset: For this research, a comprehensive image dataset
Random Forests, SVMs, and Decision Trees have been of plant leaves was used for training and testing the model.
employed for early disease detection. However, they also The dataset comprises over 61,486 images spanning 38
identified significant gaps, such as the lack of cereal- distinct plant disease classes, including healthy leaf
specific datasets and generalization issues across different categories.
environmental conditions. [6] Images in the dataset represent various plants such as apple,
grape, tomato, corn, and potato, and capture a diverse range
Deep Learning (DL), particularly Convolutional Neural of diseases including fungal, bacterial, and viral infections.
Networks (CNNs), further advanced plant disease Additionally, a separate independent test set of 54,306
recognition by eliminating the need for manual feature images was used for final model evaluation to ensure
extraction. Saleem et al. (2019) reviewed the application of unbiased performance assessment.
CNNs in detecting various plant diseases, noting superior
performance over traditional ML methods due to CNNs' To extend the functionality of the trained model, a web
ability to automatically learn complex patterns directly application was developed using Flask to provide an
from raw images. [7] accessible user interface. Users can upload images of
infected plant leaves directly to the application, which are
Expanding on this, Abade et al. (2020) systematically processed in real-time using the trained CNN model. Upon
reviewed the use of CNNs for plant disease recognition. predicting the disease, the application accesses structured
They highlighted the dominance of architectures like data from integrated Excel files to enhance decision support.
AlexNet, VGG, Inception, and ResNet across research Specifically, it retrieves the list of associated symptoms
studies. However, they pointed out common challenges, from disease_info.xlsx, while supplements_info.xlsx
such as model overfitting, dependency on synthetic provides recommended treatments along with marketplace
datasets like PlantVillage, and lack of robustness under links or product names for acquiring the suggested items.
real-world field conditions. [8] This approach allows the system not only to diagnose the
disease but also to inform the user about symptoms, suggest
Nagaraju and Chawla (2020) provided additional insights appropriate supplements, and direct them to verified sources,
into DL applications, emphasizing the importance of high- thereby delivering a complete and practical plant health
quality, annotated image datasets and proposing future support tool.
integration of hyperspectral data analysis for improved
performance. [9] Data Preprocessing : The Plant Village dataset, comprising
61,486 images of healthy and diseased leaves, was standardized
Recognizing the critical role of data availability, Hughes before training. Images were resized to 255×255 pixels and
and Salathé (2015) created the PlantVillage dataset — an center-cropped to 224×224 to ensure uniform input dimensions.
open-access repository of over 50,000 images of healthy Each image was converted into a tensor and normalized to the
and diseased plant leaves. This dataset democratized
[0,1] range using PyTorch’s transformation tools. The dataset
access to training data for researchers worldwide,
facilitating rapid advancements in mobile-based disease was randomly shuffled and split into training (59.5%), validation
diagnostic tools. [10] (25.5%), and testing (15%) subsets to support balanced training
and robust evaluation. insights beyond overall accuracy (Table 4.1).
 Macro-Averaged F1-Score: To evaluate balanced
Modal Architecture: The proposed convolutional neural performance across all classes regardless of class
network (CNN) was carefully designed to balance depth, frequency.
complexity, and computational efficiency, with the goal of
Additionally, a progress bar was implemented during
achieving high classification accuracy across a diverse set of
evaluation to monitor the batch-wise testing process
plant disease categories. The architecture consists of four efficiently.
convolutional blocks, each comprising two convolutional
layers followed by rectified linear unit (ReLU) activations and We exported the confusion matrix results to an Excel file
batch normalization layers. The use of batch normalization (confusion_matrix.xlsx), and classification details were
helps stabilize and accelerate training by normalizing stored in a text file (classification_report.txt) for detailed
analysis.
activations at each layer, while ReLU introduces non-linearity
that enables the network to learn complex features. Tools and Libraries: The model was developed and
evaluated using the following software environment:
After each pair of convolutional layers within a block, a max
pooling operation with a 2×2 kernel is applied to reduce the  Programming Language: Python 3.8
spatial dimensions by a factor of two. This progressive
reduction in size not only lowers computational overhead but  Web Framework: Flask
also facilitates hierarchical feature learning—starting from
low-level features such as edges and textures in the earlier  Frontend: HTML5, CSS3
layers, to more abstract representations in the deeper layers.
 Excel Export: Pandas (for saving confusion matrix
as an Excel sheet)
The convolutional backbone transforms the input image into a
dense, high-level feature map of shape 256×14×14. This is  Deep Learning Framework: PyTorch
then flattened into a one-dimensional vector of size 50,176 and
passed through two fully connected (dense) layers. The first  Image Processing: TorchVision
dense layer consists of 1,024 neurons with ReLU activation,
and dropout regularization is applied with a dropout rate of 0.4  Evaluation Metrics: scikit-learn (classification
to prevent overfitting. The final layer maps the output to K report, confusion matrix, F1-score calculations)
neurons, corresponding to the number of plant disease classes,
producing raw class scores (logits) for classification.

Model Implementation: The model was implemented in

PyTorch using the nn.Sequential API for modular design.
Convolutional layers increased in depth, paired with batch
normalization and non-linear activations. Max pooling
provided translation invariance and reduced computational
cost. The output from the final convolutional block was
flattened into a vector of size 50,176 and processed
through fully connected layers. Dropout was used during
training for regularization. Training employed the Adam
optimizer and cross-entropy loss, with performance
monitored on the validation set to ensure generalization.

Model Evaluation : We evaluated the model using a

separate test dataset consisting of 54,306 images.
Evaluation Metrics included:
 Accuracy: The percentage of correctly
classified images over the total number of test
images.
 Confusion Matrix: To visualize per-class Figure 3.1 - Block Diagram of proposed model
performance, understand class-wise confusion,
and identify classes with higher
misclassification rates (Table 4.2).
 Classification Report: We provide precision,
recall, and F1-scores for each class to offer
 IV. RESULT AND DISCUSSION

Index Crop Status/Disease Precision Recall F1-Score Support

0 Apple Apple scab 0.78 0.74 0.76 630
1 Apple Black rot 0.89 0.83 0.86 621
2 Apple Cedar apple rust 0.85 0.92 0.88 275
3 Apple Healthy 0.94 0.78 0.85 1645
4 Background Without leaves 0.08 1 0.15 1
5 Blueberry Healthy 0.8 0.99 0.88 1502
6 Cherry Powdery mildew 0.88 0.97 0.92 1052
7 Cherry Healthy 0.91 0.98 0.94 854
8 Corn Cercospora leaf spot Gray leaf spot 0.81 0.79 0.8 513
9 Corn Common rust 1 0.97 0.98 1192
10 Corn Northern Leaf Blight 0.91 0.89 0.9 985
11 Corn Healthy 1 0.97 0.98 1162
12 Grape Black rot 0.93 0.63 0.75 1180
13 Grape Esca (Black Measles) 0.80 0.97 0.88 1383
14 Grape Leaf blight (Isariopsis Leaf Spot) 0.99 0.85 0.92 1076
15 Grape Healthy 0.98 0.79 0.87 423
16 Orange Huanglongbing (Citrus greening) 1 0.97 0.98 5507
17 Peach Bacterial spot 0.94 0.88 0.91 2297
18 Peach Healthy 0.85 0.96 0.9 360
19 Pepper bell Bacterial spot 0.92 0.71 0.8 997
20 Pepper bell Healthy 0.93 0.92 0.92 1478
21 Potato Early blight 1 0.38 0.55 1000
22 Potato Late blight 0.88 0.77 0.82 1000
23 Potato Healthy 0.96 0.79 0.87 152
24 Raspberry Healthy 0.97 0.71 0.82 371
25 Soybean Healthy 0.97 0.97 0.97 5090
26 Squash Powdery mildew 0.95 0.96 0.96 1835
27 Strawberry Leaf scorch 0.99 0.94 0.96 1109
28 Strawberry Healthy 0.91 0.75 0.82 456
29 Tomato Bacterial spot 0.92 0.85 0.88 2127
30 Tomato Early blight 0.82 0.58 0.68 1000
31 Tomato Late blight 0.71 0.86 0.78 1909
32 Tomato Leaf Mold 0.49 0.96 0.65 952
33 Tomato Septoria leaf spot 0.66 0.9 0.76 1771
34 Tomato Spider mites Two-spotted spider mite 0.69 0.95 0.8 1676
35 Tomato Target Spot 0.88 0.65 0.75 1404
36 Tomato Yellow Leaf Curl Virus 0.96 0.94 0.95 5357
37 Tomato Mosaic virus 0.94 0.97 0.95 373
38 Tomato Healthy 0.95 0.97 0.96 1591
Overall Accuracy 0.8862 54306
Macro F1 Score 0.8408
Table 4.1 : Classification Report
 Table 4.2 : Confusion Matrix

data.
The proposed CNN-based model was evaluated on a large
independent test set comprising 54,306 images across 38 Table 4.2 provides the numeric confusion matrix generated
plant disease classes. The model achieved an overall after evaluating the model on the test set. In this matrix,
classification accuracy of 88.62%, demonstrating strong each row represents the actual disease class (by index), and
generalization capabilities over diverse crop and disease each column represents the class predicted by the model.
types. Values along the diagonal indicate correctly classified
instances, while values off the diagonal highlight instances
Table 4.1 presents a detailed classification report that
where the model confused one disease class for another.
summarizes the model’s performance across all 38 plant
This matrix is instrumental in diagnosing systematic
disease classes. Each row corresponds to a specific disease
prediction errors. For example, it helps identify specific
class, identified by its index, along with the related crop
classes that are frequently misclassified into one another —
and status/disease name. The table includes four key
such as the confusion observed between index 21 and
evaluation metrics for each class:
indices 31 and 33, which was supported both numerically
 Precision: The proportion of correct positive
and visually. Such insights are essential for future
predictions out of all predicted positives.
improvements in model accuracy, including refining class
 Recall: The proportion of actual positives correctly
boundaries or enhancing feature extraction methods.
identified by the model.
 F1-Score: The harmonic mean of precision and High-Performing Classes : Several classes show excellent
recall, providing a balanced view. classification performance with precision, recall, and F1-
 Support: The number of actual test samples scores above 0.95.
present in each class. Example 1: Corn — Healthy (Index 11, Support: 1162):
 Precision: 1.00
This tabular format enables a clear comparison of model
 Recall: 0.97
performance across different disease categories. High-
 F1-Score: 0.98
scoring classes reflect strong learning and clear feature
distinction, while classes with lower precision or recall
Example 2: Soybean — Healthy (Index 25, Support:
suggest areas where misclassifications occurred, likely due
5090):
to visual similarity or under-representation in the training
 Precision: 0.97
  Recall: 0.97  Support: 1000
 F1-Score: 0.97  Correct Predictions: 577
 Misclassified as:
Example 3: Cherry — Powdery mildew (Index 6, Support: o Tomato — Late blight (Index 31): 138 times
1052): o Tomato — Leaf Mold (Index 32): 89 times
 Precision: 0.88 o Tomato — Septoria leaf spot (Index 33): 76
 Recall: 0.97 times
 F1-Score: 0.92

Example 4: Grape — Esca (Index 13, Support: 1383):

 Precision: 0.80
 Recall: 0.97
 F1-Score: 0.88

These classes are both well-represented and visually

distinct, which may contribute to high learning accuracy Figure 4.2 : Index 30, 31, 32 and 33 (in order)
and model confidence.
Tomato Early Blight (1st) is often misclassified as Late
Misclassification and Weak Classes : Some classes, Blight (2nd), Leaf Mold (3rd), or Septoria Leaf Spot (4th)
however, show lower performance — not necessarily due to because all four diseases produce dark spots or patches on
dataset size, but due to visual similarity with other diseases. leaves that may look alike to a machine learning model.
Early Blight typically shows circular lesions with concentric
Example 1: Potato — Early blight (Index 21) rings, but in poor lighting or with aging leaves, these rings
 Support: 1000 can blend in. Late Blight creates irregular brown patches,
 Correct Predictions: 382 Leaf Mold forms pale yellow spots that darken over time,
and Septoria shows many small, dark specks. When
 Misclassified as:
symptoms overlap in shape, color, and distribution, the
o Tomato — Septoria leaf spot (Index 33): 362
CNN may struggle to tell them apart. More precise training
times
data, better contrast, and highlighting lesion edges could
o Tomato — Late Blight (Index 31): 129 times
help the model make clearer distinctions.

Example 3: Grape — Black rot (Index 12)

 Support: 1180
 Correct Predictions: 743
 Misclassified as:
o Grape — Esca (Index 13): 322 times

Figure 4.1 : Index 21, Index 33 and Index 31 (in order)

Potato Early Blight is often mistaken by the model for

Tomato Septoria Leaf Spot or Tomato Late Blight because
they all create dark spots that look quite similar at a glance.
Early Blight usually forms brown spots with ring-like
patterns, but under poor lighting or on young leaves, those
rings can fade or blend in. That makes it easy to confuse
with Septoria’s tiny dark specks or the uneven patches of
Figure 4.3 : Index 12 and 13 (in order)
Late Blight. Since potato and tomato are closely related
plants, their leaves and disease symptoms often resemble
The misclassification of Grape Black Rot as Grape Esca
each other, which adds to the confusion. To help the model
likely occurs due to similar brown lesion colors,
make better predictions, it would be useful to train it on a
overlapping textures, and comparable leaf shapes, which
wider variety of images and use tools that help focus on the
confuse the CNN. Low-resolution images and label noise
details of each spot, like its edges or texture.
can further worsen this. To reduce such errors,
segmentation of lesion areas, targeted data augmentation,
Example 2: Tomato — Early blight (Index 30)
attention mechanisms, and use of lesion masks or class-
weighted loss can help the model better distinguish between
the two diseases.

Example 4: Tomato — Target Spot (Index 35)

 Support: 1404 images
 Correct Predictions: 912
 Misclassified as:
o Two-spotted spider mite (Index 34): 349 times

Figure 4.4 : Index 35 and 34 (in order)

Figure 4.5 : Home Page
The Tomato Target Spot (left) is often misclassified as
Two-Spotted Spider Mite damage (right) because they can As shown in Figure 4.5, the home page provides a visually
look surprisingly similar—both show yellowing and patchy appealing layout that allows users to browse by crop type. A
areas on the leaf. While Target Spot usually forms round, variety of fruits and vegetables such as tomato, potato, corn,
ring-like lesions, and spider mites leave tiny dots or faded apple, grape, and others are displayed with images to guide
patches, these differences can blur under poor lighting or in the user. This helps simplify disease identification and
low-quality images. The model likely sees the shared shows the model’s coverage.
yellowish, uneven patterns and assumes they’re the same.
To help the model tell them apart more accurately, clearer
images, better contrast, and focusing only on the actual
lesion areas would make a big difference.

Macro-Level Insights : From the classification report:

 Macro F1-score: 0.84 (balanced performance)
 Weighted Avg F1-score: 0.89
 High recall for most classes shows the model is
good at capturing true positives.

Web Implementation : To translate the trained CNN

model into a practical solution, a web application was
developed using Flask for backend processing and
HTML/CSS for a responsive frontend. The application
enables end-users—such as farmers, horticulturists, or Figure 4.6 : AI Engine
agronomists—to interact with the AI system in a simple and
intuitive way without technical knowledge. Figure 4.2 demonstrates the core functionality of the AI
engine. This page includes:
 A brief educational note about the importance of
plant disease detection
 A file upload field where users can select an image
of an affected leaf
 A submit button that triggers real-time prediction
using the backend model
 A right-side panel listing preventive measures for
disease management
 recommendations, extends its impact beyond prediction to
actionable guidance for users.

Looking forward, future improvements may involve

enhancing the dataset with more field-sourced images,
applying targeted data augmentation, and fine-tuning model
performance through hyperparameter optimization and
transfer learning. These efforts could further boost
accuracy, reduce confusion among similar classes, and
make the system even more robust and field-ready.

REFERENCES

[1] R. W. Stubbs, J. M. Prescott, E. E. Saari, and H. J. Dubin, “Cereal

disease methodology manual,” Cimmyt.org, 2025.
https://repository.cimmyt.org/entities/publication/f1339c96-a580-
43e4-a7ff-f6698f5af65d (accessed May 02, 2025)
[2] R. N. Strange and P. R. Scott, “Plant disease: a threat to global
food security,” Annual Review of Phytopathology, vol. 43, pp. 83–
116, 2005, doi:
https://doi.org/10.1146/annurev.phyto.43.113004.133839.
[3] M. Różewicz, M. Wyzińska, and J. Grabiński, “The Most
Important Fungal Diseases of Cereals—Problems and Possible
Solutions,” Agronomy, vol. 11, no. 4, p. 714, Apr. 2021, doi:
https://doi.org/10.3390/agronomy11040714.
[4] N. W. Schaad et al., “ADVANCES IN MOLECULAR-BASED
DIAGNOSTICS IN MEETING CROP BIOSECURITY AND
Figure 4.7 : Prediction Results and Recommendations PHYTOSANITARY ISSUES,” Annual Review of Phytopathology,
vol. 41, no. 1, pp. 305–324, Nov. 2003, doi:
https://doi.org/10.1146/annurev.phyto.41.052002.095435.
Once the image is submitted, the system performs [5] J. F. I. Nturambirwe and U. L. Opara, “Machine learning
prediction and redirects to a results page (Figure 4.3), which applications to non-destructive defect detection in horticultural
includes: products,” Biosystems Engineering, vol. 189, pp. 60–83, Jan. 2020,
doi: https://doi.org/10.1016/j.biosystemseng.2019.11.011.
 The predicted disease name and image
[6] F. G. Waldamichael, T. G. Debelee, F. Schwenker, Y. M. Ayano,
 A brief scientific description of the disease and S. R. Kebede, “Machine Learning in Cereal Crops Disease
 A list of symptoms Detection: A Review,” Algorithms, vol. 15, no. 3, p. 75, Feb. 2022,
doi: https://doi.org/10.3390/a15030075.
 Recommended prevention steps
[7] Saleem, Potgieter, and Mahmood Arif, “Plant Disease Detection
 Suggested supplements or fungicides, along with a and Classification by Deep Learning,” Plants, vol. 8, no. 11, p.
"Buy Product" button linked to an external 468, Oct. 2019, doi: https://doi.org/10.3390/plants8110468.
marketplace. [8] A. S. Abade, P. A. Ferreira, and Vidal, “Plant Diseases recognition
on images using Convolutional Neural Networks: A Systematic
V. CONCLUSION Review,” arXiv.org, 2020. https://arxiv.org/abs/2009.04365
[9] M. Nagaraju and P. Chawla, “Systematic review of deep learning
techniques in plant disease detection,” International Journal of
This study illustrates the effectiveness of a Convolutional System Assurance Engineering and Management, vol. 11, no. 3, pp.
Neural Network (CNN) in accurately detecting and 547–560, May 2020, doi: https://doi.org/10.1007/s13198-020-00972-
1.
classifying plant diseases from image data. Trained on a [10] D. P. Hughes and M. Salathe, “An open access repository of
large and diverse dataset, the model achieved a images on plant health to enable the development of mobile
commendable accuracy of 88.62% when tested on an disease diagnostics,” arXiv:1511.08060 [cs], Apr. 2016, Available:
independent set of 54,306 images, successfully identifying https://arxiv.org/abs/1511.08060
38 distinct plant disease classes, including healthy [11] T. van Klompenburg, A. Kassahun, and C. Catal, “Crop yield
leaves. prediction using machine learning: A systematic literature
review,” Computers and Electronics in Agriculture, vol. 177, no. 1, p.
105709, Oct. 2020, doi:
While the majority of predictions were made with high https://doi.org/10.1016/j.compag.2020.105709.
precision and recall, the analysis revealed some [12] ASM Mahmudul Hasan, D. Diepeveen, H. Laga, Michael G.K.
misclassifications—primarily among diseases with Jones, and Ferdous Sohel, “Object-level benchmark for deep
overlapping visual symptoms. This highlights a common learning-based detection and classification of weed species,” Crop
challenge in image-based diagnosis, especially when Protection, vol. 177, pp. 106561–106561, Mar. 2024, doi:
https://doi.org/10.1016/j.cropro.2023.106561.
diseases exhibit similar patterns or affect the same crop
type.

Nonetheless, the model proved highly capable of

automating plant disease detection, offering significant
potential for real-world agricultural use, particularly in
regions where expert access is limited. Its integration into a
web-based interface, along with supplementary