Proceedings of the 5th International Conference on Inventive Research in Computing Applications (ICIRCA 2023)

IEEE Xplore Part Number: CFP23N67-ART; ISBN: 979-8-3503-2142-5

A Comparative Study on Algorithms for Plant

Disease Detection using Transfer Learning
Heenakausar Pendhari Rasika Virkar Betsy Edakkalathur
Department of Computer Engineering Department of Computer Engineering Department of Computer Engineering
2023 5th International Conference on Inventive Research in Computing Applications (ICIRCA) | 979-8-3503-2142-5/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICIRCA57980.2023.10220597

Fr.CRCE, Mumbai Fr.CRCE, Mumbai Fr.CRCE, Mumbai

[email protected] [email protected] [email protected]

Akshata Jadhav
Department of Computer Engineering
Fr.CRCE, Mumbai
[email protected]

Abstract—There are many species of plants in this world. Some intervention and mitigation strategies. Thus, timely detec-
of which are known and some are unknown. Each variety of tion enables farmers, gardeners, and conservationists to take
plants has different kinds of diseases. Recognising these diseases appropriate measures, such as applying targeted treatments,
can help the plants from not getting affected further. This
process requires the help of experts. But visually observing and adjusting cultivation practices, or removing infected plants to
diagnosing each diseases requires lots of time and resources. prevent the spread of diseases.
In this study, the authors make use of MobileNetV2, VGG16,
VGG19, NASNet and Inception Resnet to process data of II. LITERATURE REVIEW
different plant image samples to get fast analysis of the various Ferentinos et al., [1] included 29 classes of plant dis-
diseases. This knowledge of plant diseases collected can then help
gardeners, horticulturists, environmentalists, farmers, eases containing 87,848 images. AlexNet, AlexNetOWTBn,
researchers, etc for different needs. GoogleNet, Overfeat and VGG were the models used in this
Index Terms—Transfer Learning, Deep Learning, Convolu- study. The main drawback was the images used were not in
tional Network Layers, Plant Disease Detection, VGG16, Mo- real cultivational setup but instead it was in laboratory setup.
bilenetV2, NASNet Mobile Also the performance on original images were better than
preprocessed images.
I. INTRODUCTION
Mohanty et al., [2] used AlexNet and GoogleNet deep
The decline in plant species and overall flora is a concerning learning modules for diagnosing plant diseases through smart-
issue with wide-ranging impacts on ecosystems and human phone assisted program. The study uses transfer learning as
well-being. Recognizing the importance of preserving plants, well as training from scratch on 3 different image types.
people have been actively engaged in conservation efforts to : Coloured, Grayscaled and Segmented. The models were
save the flora of the planet. Protecting plants from diseases is tested on different distribution of training-testing datasets. The
a crucial aspect of these conservation endeavours. Models were able to identify images upto 99.35%.
To effectively save plants from diseases, it is essential to In [3], [4] and [7], all uses SVM classifer showing accuracy
detect and identify the diseases affecting them. One common upto 95.71%, 80% and 40.33% respectively.
approach is to focus on analysing the plant’s leaves since it Utkarsha et al., [5] presents 99.32% classification accuracy
often shows visible symptoms of diseases. Researchers and over 12673 images distributed among 4 classes. The classi-
conservationists have created datasets specifically designed for fication accuracy for color imaged is better than grayscale
disease detection in plants. These datasets consist of images and segmented images. The drawbacks is that the graph is
and associated labels indicating the presence or absence of overfitting.
diseases. Serawork et al., [6] shows accuracy of 89.75% on test
Deep neural networks, computer vision techniques, and images and 93.22% on test images. The module used is of
object recognition methods have been employed to develop DenseNet with 15 classes of plant diseases.
disease detection models. By leveraging these technologies, The study [8], has implemented a Flask based web
researchers have trained models to automatically analyse im- application to detect plant disease achieving accuracy upto 93%
ages of plant leaves and identify diseases with a high degree & F1 score of 0.93.
of accuracy. The models learn from the labelled dataset, rec- Pandian et al.[9], explains about the use of Deep Convolu-
ognizing patterns indicative of various diseases. Once trained, tional Neural Network (DCNN) for image classification which
these models can be used to process new images of plant leaves is a mix of various layers fully connected, pooled for the
and classify them as healthy or diseased. purpose of a frequent monitoring system on agricultural fields.
The implementation of these advanced techniques aids in The overall accuracy of this algorithm was 92% for tomato
early identification of diseases, which is crucial for prompt

979-8-3503-2142-5/23/$31.00 ©2023 IEEE 528

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:54:18 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 5th International Conference on Inventive Research in Computing Applications (ICIRCA 2023)
IEEE Xplore Part Number: CFP23N67-ART; ISBN: 979-8-3503-2142-5

yellow leaf curl disease detection. VGGNet and Resnet were our models. Diseases of following plants were used : Apple,
used to improve the classification model of this algorithm. Tomato, Potato, Pepper Bell, Strawberry, Orange, Grape, Corn,
Basic Image Manipulation (BIM) and Neural Style Transfer Soybean. The images are resized to 224 x 224 pixels. The
(NST) are used for image augmentation. images are having 3 channels (RGB) and had been already
Rehan et al.[10], computer vision makes it easier to check augmented. The dataset’s example photos are as shown in Fig.
on crop and food security through the hues of AI. This 1.
algorithm uses the SIFT method for leaf shape and type
detection.
Nishant et al.[11], 95.6% accuracy was achieved through
training 50 epochs of the model. The CNN model is used
to detect both sick and healthy leaves. Collecting dataset,
preprocessing and augmentation of dataset, building CNN
and thereafter creating an android lightweight app were the
procedures involved in this algorithm.
P.Loganathan et al.[12], shows that the model has only
achieved 34% significant accuracy for 4x4 pixel resolution.
For rose plant leaves the precision was 95% and for cucumber
plant leaves the precision was 86%. From this paper it was
discovered that ResNet performed better than the ResNet152
V2 model.
Ramkumar et al.[13], in this paper explains specifically Fig. 1. Sample Images in Dataset.
https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset
about Cercospora disease affecting spinach leaves. The re-
search paper gives an emphasis on image segmentation con-
cept for the resolution of images. The paper also explains B. Architecture
the timeline of the spreading of the disease. The use of These CNN architectures that were used in this work are
Artificial Neural Network (ANN) is used in the algorithm MobileNet V2, NASNet Mobile, VGG 16, VGG 19 and
implementation. The methodology talks about usage of the Interception Resnet. They are well-known for their light-
ResNet50 module. weight models.
Adedamola et al.[14], showed this algorithm achieved an MobileNet V2 is a type of convolutional neural network that
accuracy rate of 99.31% using the NasNet mobile CNN model. has a depth of 53 layers, which includes 32 filter layers and 19
There has been a wide usage of lightweight Convolutional residual bottleneck layers. A pre-trained version of the network
Neural Network. This paper shows classification for plant leaf is available, which has been trained on a dataset of over one
disease recognition systems using traditional machine learning million images from the ImageNet database. This pre-trained
and deep learning as well. Steps such as max pooling, convo- network is capable of categorizing images into 1000 object
lutions, average pooling, separable convolution, and identity categories, such as animals and plants.
mapping are involved. NASNet Mobile is a deep neural network architecture that
Rajaneni et al.[15], proposed that the paper is based on was developed using Neural Architecture Search (NAS) by
using KNN along with convolutional neural networking for Google AI researchers. The architecture of NASNet Mobile
image classification. The study explains about diseases with is optimized for mobile devices with low memory and pro-
their database images and descriptions. The output is provided cessing power, while still maintaining high accuracy in image
through the mean square error method. Epoch of 2000 is given classification tasks. It is 26 layers deep.
to obtain more accuracy in leaf disease detection. VGG16 is a CNN architecture that achieved success in
Emanuel Cortes[16], the algorithm made use of a public winning the 2014 ImageNet Large Scale Visual Recognition
database consisting of 86,147 images of healthy and diseased Challenge. VGG16 employs convolution layers with a 3x3
plant leaves. The algorithm makes use of deep convolutional filter and stride 1, with consistent padding, instead of a large
networks and semi-supervised learning methods. There are 57 number of hyper-parameters. VGG16 also utilizes a maxpool
disease classes considered in this research paper. There is a use layer with a 2x2 filter and stride 2, which reduces the size of
of HSV features and SVM too. The algorithm scored above feature maps while retaining the most important information.
80% under 5 epochs with a learning rate of le-5. VGG-19 is a deep convolutional neural network with 19
layers. It is possible to use a pre-trained version of the network,
III. METHODOLOGY
which was trained on over a million images from the ImageNet
A. Dataset database, to classify images into 1000 different categories. The
The study utilized the ”New Plant Diseases Dataset” from network is designed to accept images with a size of 224-by-
the Kaggle website. The original dataset were classified into 38 224 as input.
classes depending on different types of plants whether they are The Inception-ResNet architecture is a type of deep convo-
healthy or diseased. However, we have used only 31 classes for lutional neural network that combines Inception and residual

979-8-3503-2142-5/23/$31.00 ©2023 IEEE 529

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:54:18 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 5th International Conference on Inventive Research in Computing Applications (ICIRCA 2023)
IEEE Xplore Part Number: CFP23N67-ART; ISBN: 979-8-3503-2142-5

networks. Every inception layer follows filter expansion. The

input size is 299 x 299.

C. Approach
The Datasource originally contained a large number of
images that were categorized into 38 different groups. The
dataset was then modified by removing irrelevant images and
was eventually transformed into a dataset that included only 31
categories of plant diseases. The dataset was initially split into
two parts - the training set and the validation set - with a ratio
of 80:20. Afterwards, additional images were incorporated into
the dataset specifically for the purpose of testing. The images
were subsequently adjusted in size and scale to ensure they
were compatible with the appropriate models.
To implement transfer learning, pre-existing architecture Fig. 2. Performance of all models.
models such as NASNetMobile, MobileNetV2, Inception-
Resnet, VGG16, and VGG19 were obtained from Keras ap-
plications. These models had already been trained on large
datasets, and their learned weights and features were utilized
to enhance the accuracy of the new model being developed.
After obtaining the pre-trained architecture models, they
were compiled and configured with a set of performance
metrics. These metrics included accuracy, loss, precision,
recall, and f1 score, which were used to evaluate the
performance of the models during the training process.
Then confusion matrix was generated to assess the
performance of the models. This matrix provided information
about the true positive, true negative, false positive, and false
negative values for each classification. Finally, the models
were used to make predictions on new images to test their
effectiveness in classifying plant diseases.
IV. RESULTS
The performance metrics for different transfer learning Fig. 3. Inception-Resnet Accuracy-Loss Graph.
models, including VGG16, VGG19, MobileNetV2, Inception-
RestNetV2, and NasNet Mobile, were evaluated on the plant
disease dataset. The results were tabulated, and it was observed
that VGG19 had the highest overall accuracy among all the
models. The table showed in fig 2 also displayed the accuracy,
loss, recall, precision, and F1 score values for each model,
which could be used to determine the optimal transfer learning
model for this specific dataset. These findings can be used to
inform the selection of an appropriate model for future work
with similar datasets.
Figures 3, 4, 5, 6, 7 shows how the accuracy and loss curves
looks like for different models. These curves are commonly
used to evaluate the performance of machine learning models.
The accuracy curve depicts how well the model is performing
over time. It shows the percentage of correctly classified
samples in the training set. And the loss, on the other hand,
represents the model’s error during the training process and is
calculated here with categorical cross-entropy. The validation
accuracy (represented by the red line) tracks the training
accuracy (represented by the orange line). While the validation Fig. 4. NASNet Accuracy-Loss Graph.
loss is represented by green and training loss is represented
by blue color.

979-8-3503-2142-5/23/$31.00 ©2023 IEEE 530

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:54:18 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Inventive Research in Computing Applications (ICIRCA 2023)
IEEE Xplore Part Number: CFP23N67-ART; ISBN: 979-8-3503-2142-5

The confusion matrix for the new plant disease dataset

was generated after training the transfer learning models and
making predictions on the test set. The confusion matrix
revealed that the models had high accuracy and performed well
in classifying most of the plant diseases. However, some mis-
classifications were also observed, particularly among classes
that had similar visual characteristics. Certain plant diseases
do exhibit similar symptoms, making it difficult for machine
learning models to accurately differentiate between them.
For example, the “Apple Scab” and “Cedar Apple Rust”
classes of diseases both affect apple trees and can cause
similar visual patterns on the leaves, such as brown spots or
lesions. This can lead to confusion between the two classes
of diseases by the machine learning models. Additionally,
it is not uncommon for a single plant to be affected by
multiple diseases simultaneously, which further complicates
Fig. 5. VGG16 Accuracy-Loss Graph. the problem of disease detection. Confusion Matrix also offers
a thorough breakdown of the model’s predictions, including
the number of true positive, false positive, true negative, and
false negative values for every class. By carefully examining
the confusion matrix, researchers can gain insights into the
model’s performance for each class in the dataset. It helps
identify specific classes where the model may struggle or
exhibit higher rates of false positives or false negatives. These
insights can guide further improvements in the model or help
develop strategies to address specific challenges in disease
classification.

Fig. 6. VGG19 Accuracy-Loss Graph.

Fig. 8. VGG16 Confusion Matrix.

Fig. 7. MobileNet V2 Accuracy-Loss Graph.

979-8-3503-2142-5/23/$31.00 ©2023 IEEE 531

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:54:18 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Inventive Research in Computing Applications (ICIRCA 2023)
IEEE Xplore Part Number: CFP23N67-ART; ISBN: 979-8-3503-2142-5

Fig. 9. NASNet Mobile Confusion Matrix Fig. 11. Inception-Resnet Confusion Matrix

Fig. 10. VGG19 Confusion Matrix Fig. 12. MobileNet V2 Confusion Matrix

979-8-3503-2142-5/23/$31.00 ©2023 IEEE 532

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:54:18 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 5th International Conference on Inventive Research in Computing Applications (ICIRCA 2023)
IEEE Xplore Part Number: CFP23N67-ART; ISBN: 979-8-3503-2142-5

V. CONCLUSION AND FUTURE WORK [13] M. O. Ramkumar, S. Sarah Catharin, V. Ramachandran, and A. Sak-
thikumar, “Cercospora Identification in Spinach Leaves Through Resnet-
The study highlights the crucial role of machine learning, 50 Based Image Processing,” Journal of Physics: Conference Series, vol.
especially deep learning, in detecting plant diseases. Failure to 1717, issue 1, 2021.
[14] Adedamola O. Adedoja, Pius A. Owolawi, Temitope Mapayi, and
accurately identify these diseases can significantly impact crop Chunling Tu, “Intelligent Mobile Plant Disease Diagnostic System
yield and contribute to long-term problems like climate change Using NASNet-Mobile Deep Learning,” IAENG International Journal
and potential food shortages. The positive aspect is that, it of Computer Science, vol. 49, no.1, pp. 216–231, 2022.
[15] Rajaneni Deepika Sai Chowdry, Banupriya .N, Yogeshwari V., Varsha,
can be incredibly helpful for individuals such as gardeners, and Swathy C., “Plant Disease Detection Using Image Processing And
farmers, horticulturists, and scientists, providing them with the Machine Learning Algorithm,” Journal of Xidian University, vol. 14,
means to accurately identify and address plant diseases. 2022.
[16] Emanuel Cortes, “Plant Disease Classification Using Convolutional
Networks and Generative Adverserial Networks,” Stanford University,
2018.
For future work, diversity of the dataset can be increased,
which may result in more accurate predictions and improved
performance of the models. Also, the hyperparameters of the
models can be fine-tuned to optimize their performance. And
the trained models can be used to develop a web or mobile
application that can help farmers and gardeners identify plant
diseases quickly and accurately.

ACKNOWLEDGMENT
The authors gratefully acknowledge the guidance of our
mentor Prof. Heenakausar Pendhari, C.R.C.E., for her valu-
able technical guidance and suggestions towards improving
this project. We also extend our appreciation to the faculty
of C.R.C.E, for their encouragement and for providing the
necessary infrastructure for us to successfully carry out this
project.

REFERENCES
[1] Ferentinos, K. P., “Deep learning models for plant disease detection
and diagnosis,” Computers and Electronics in Agriculture, vol. 145, pp.
311–318, 2018.
[2] Mohanty, S. P., Hughes, D. P., and Salathé, M., “Using Deep Learning
for Image-Based Plant Disease Detection,” Frontiers in Plant Science,
vol. 7, 2016.
[3] Singh, V., and Misra, A., “Detection of plant leaf diseases using image
segmentation and soft computing techniques,” Information Processing
in Agriculture, vol. 4(1), pp. 41–49, 2016.
[4] Utkarsha N. Fulari, Rajveer K. Shastri, and Anuj N. Fulari, “Leaf
Disease Detection Using Machine Learning,” Journal of Seybold Report,
2020.
[5] Serawork A. Wallelign, Mihai Polceanu, and Cédric Buche, “Soybean
Plant Disease Identification Using Convolutional Neural Network,”
FLAIRS-31, pp. 146–151 〈hal-01807760〉, May 2018.
[6] “Vegetable Leaf Disease Detection Using Deep Learning,” International
Journal of Research in Engineering and Science (IJRES), vol. 11, pp.
548–555, 2023.
[7] Shima R. Maniyath, Vinod P.V, Niveditha M, Pooja R., Prasad N.,
Shashank N., and Ramachandra H., “Plant Leaf Disease Detection using
Machine Learning,” International Conference on Design Innovations for
3Cs Compute Communicate Control (ICDI3C), 2018.
[8] Pranesh K., Atharva K., Tejas K., Soham K., Akshay J., and Medha
W., “Plant Disease Detection Using Image Processing and Machine
Learning,” 2021.
[9] J. A. Pandian, V. D. Kumar, O. Geman, M. Hnatiuc, M. Arif, and
K. Kanchanadevi, “Plant Disease Detection Using Deep Convolutional
Neural Network,” Applied Sciences, vol. 12, no. 14, pp. 6982, Jul. 2022.
[10] Rehan Ullah K., Khalil K., Waleed A., and Ali Mustafa Qamar, “Image-
Based Detection of Plant Diseases: From Classical Machine Learning
to Deep Learning Journey,” 2023.
[11] Nishant Shelar, Suraj S. , Shubham S. , Shreyash D. , and Kausar F.,
“Plant Disease Detection Using Cnn,” ITM Web Conf., vol. 44, 2022.
[12] P.Loganathan, Dr.R.Karthikeyan, “Residual Neural Network (ResNet)
Based Plant Leaf Disease Detection and Classification,” Turkish Online
Journal of Qualitative Inquiry (TOJQI), vol. 12, issue 6, June 2021.

979-8-3503-2142-5/23/$31.00 ©2023 IEEE 533

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:54:18 UTC from IEEE Xplore. Restrictions apply.