Potato Leaf Disease Classification using Deep

Learning: A Convolutional Neural Network

Approach
Utkarsh Yashwant Tambe1, ⃰ Dr. A. Shobanadevi2 Dr. A. Shanthini3
Department of Data Science and Department of Data Science and Department of Data Science and
Business Systems Business Systems Business Systems
SRM Institute of Science and SRM Institute of Science and SRM Institute of Science and
Technology Technology Technology
Kattankulathur (Chennai), India Kattankulathur (Chennai), India Kattankulathur (Chennai), India
[email protected] [email protected] [email protected]

l Hsiu-Chun Hsu4
Department of Information
Management
National Chung Cheng University
Chiayi County, Taiwan R.O.C.
[email protected]

Abstract - In this study, a Convolutional Neural Network

(CNN) is used to classify potato leaf illnesses using Deep
Learning. The suggested approach entails preprocessing the leaf
image data, training a CNN model on that data, and assessing
the model's success on a test set. The experimental findings show
that the CNN model, with an overall accuracy of 99.1%, is highly
accurate in identifying two kinds of potato leaf diseases,
including Early Blight, Late Blight, and Healthy. The suggested
method may offer a trustworthy and effective remedy for
identifying potato diseases, which is essential for maintaining
food security and minimizing financial losses in agriculture. The
model can accurately recognize the various disease types even
when there are severe infections present. This work highlights
the potential of deep learning methods for categorizing potato
diseases, which can help with effective and automated disease Figure 1: Potato (Solatium Tuberosum). Adapted from [3]
management in potato farming. In this study, we suggest a deep learning method for
categorizing potato leaf illnesses. In particular, we investigate
Keywords - image processing, computer vision, plant disease
the application of convolutional neural networks (CNNs) for
diagnosis, agricultural technology, crop protection, AI in
agriculture the automated detection and classification of various potato
leaf diseases. The primary goal is to develop a precise and
I. INTRODUCTION useful method for the early detection and classification of
Potato is a vital food crop globally and a significant potato leaf diseases so that farmers can act quickly and
source of income for farmers. However, potato crops are control the disease. The suggested method has the potential
susceptible to various diseases that can cause significant yield to greatly increase the effectiveness and accuracy of
and quality losses. Leaf diseases are among the most common managing potato leaf disease and increasing crop yields while
and destructive of these diseases [1]. Early detection and minimizing financial losses.
accurate diagnosis of these diseases are essential for effective As, due to its capacity to extract intricate patterns and
disease management and control. Traditional methods of features from sizable databases, deep learning has become
disease diagnosis rely on visual inspections, which are increasingly popular in recent years. Deep neural networks
subjective and time-consuming. In recent years, there has known as convolutional neural networks (CNNs) are
been a growing interest in the use of deep learning techniques frequently employed for picture classification and
for automatic disease detection and classification. recognition tasks. They can learn hierarchical representations
The potato, Solatium tuberosum, as seen in fig. 1 is the of images, which makes them especially suitable for this job.
fourth-ranked food crop used to feed an increasing global It is built with the intention of detecting or identifying the
population because of its adaptability in cultivars and high various diseases that can affect potato leaves, because
complex carbohydrate content. For the dining and the Convolutional Neural Network (CNN) can identify them with
processing market, potatoes can be kept for a long time, but ease while the human eye cannot. Unbelievably, some pre-
storage disease issues are common [2]. Wherever potatoes are trained Neural Network Architectures have error rates of
produced, diseases both in the field and during storage can be about 3%, which is even lower than the best 5% error rate of
a limiting factor in the ability to produce them sustainably human vision. 5.1% was determined to be the human top-5
and profitably. error on large-scale images, which is higher than pre-trained
networks [4].
 II. LITERATURE SURVEY for spotting pests and diseases in blueberry plants. Since there
Singh et al. (2015) [5] - performed a review of the was no openly available database for this kind of fruit, the
literature on various image processing techniques for authors created their own database of images. The research
detecting leaf disease. The writers wanted to speed up used a variety of filters, including addWeighted and
identification and detection of plant diseases while lowering gaussianBlur to enhance the details in the images and
the subjectivity that comes with naked-eye observation. They medianBlur and gaussianBlur to remove noise. Utilizing both
presented an algorithm that uses picture segmentation to normalised and unnormalized versions of algorithms like
automatically identify and categorise plant leaf diseases. The HOG and LBP, traits were extracted. Using Deep Learning,
impact of HSI, CIELAB, and YCbCr colour spaces on disease the findings revealed an accuracy index of 84%. The report
spot detection was examined by the authors. To identify the does not, however, contain a literature review.
disease spot, the Otsu technique was applied to the colour Iqbal and Talukder (2020) [11] - investigated the use of
component after an image was smoothed using a median filter. image processing and machine learning methods in
The suggested method was not put to the test on any datasets. identifying and classifying diseases of potato leaf. In order to
He et al. (2015) [6] - deep residual networks were find diseased regions, the authors processed 450 images of
discussed in relation to image recognition challenges. As healthy and diseased potato leaves from the freely accessible
shallow representations for image retrieval and classification, Plant Village database using image segmentation methods. To
VLAD and Fisher Vector are cited in the paper's literature differentiate between the sick and healthy leaves, they used
review as related concepts. The writers also covered the seven classifier algorithms, with the Random Forest classifier
advantages of encoding residual vectors over original vectors having the best accuracy (97%) rate. The study also included
for vector quantization. The Multigrid method was also a review of prior studies that used machine learning and image
mentioned as a method for solving partial differential processing to identify plant diseases.
equations (PDEs) by reformulating systems as subproblems at III. PROPOSED METHODOLOGY
multiple scales, where each subproblem is accountable for the
residual solution between a coarser and a finer scale. This
method is used in low-level vision and computer graphics.
Islam et al. (2017) [7] - suggested a method to identify
diseases from images of potato plant leaves by combining
image processing and machine learning. The authors
classified diseases using the 'Plant Village' database of openly
accessible plant images. Using the suggested method, the
research classified 300 images of diseases with a 95%
accuracy. The paper also highlights how crucial contemporary
phenotyping and plant disease detection are to assuring food
security and sustainable agriculture.
Velmurugan and Renukadevi (2017) [8] - describe
research on the creation of software for the automatic
detection and classification of plant leaf diseases. Only a few
methods were suggested for particular databases, such as
satellite images, leaf sets, maps, faces, fingers, etc., the authors
discovered after conducting a literature review. The research
suggests a novel algorithm that, with 94% accuracy, can
identify and categorise the diseases under study. The
processing process consists of four major steps: colour
transformation, masking of green pixels, segmentation, and
computation of texture statistics from the SGDM matrices.
The suggested method was tried on a database of about 500
plant leaves, and the outcomes support the robustness of the
suggested algorithm.
Dasgupta et al. (2019) [9] - address the issue of food
scarcity in developing countries, particularly in India, due to
potato production being affected by diseases such as Early
Blight and Late Blight. To tackle this problem, the authors
propose the use of deep learning and transfer learning
techniques to develop a model capable of accurately detecting
these diseases in potato plants. The literature review section
of the paper delves into the fungus Alternaria Solani that is
responsible for the early Blight disease. The disease has a
polycyclic life cycle and requires free water to germinate.
Early Blight primarily causes the premature defoliation of Figure 2: Methodology Flowchart
potato plants.
Data collection, pre-processing of the data, data
Cecilia et al. (2019) [10] - share a research on the use of augmentation, and disease classification make up the main
image processing methods and machine learning algorithms
four stages of the methodology described in this paper. Fig. 2 technique for modifying data without changing its initial
mentions the graphical representation of procedure for meaning. 5100 datasets are still insufficient for this research
classifying different diseases of Potato leaf, step by step. So, to achieve optimal performance, necessitating the use of data
the four mentioned steps are: augmentation. The augmentation factors in this study are
generated by the automatic application of simple geometric
A. Data Acquisition/Data Collection and Description transformations like translations, rotations, scale changes,
Different image resolutions and sizes have been obtained shearing, vertical and horizontal flips.
from dataset uploaded on ‘Kaggle’, known as Plant Village by
Arjun Tejaswi which contains plant diseases of around 15 But for this dataset, only the rescaling and resizing of
types [12]. The dataset contains about 20,600 images of which images (1/255), horizontal flip, vertical flip and random-
2152 images belong to plant ‘Potato’ and categorised as rotation of 0.2 degrees is used as the leaves do not have high
Healthy, Early-Blight, and Late-Blight as in fig. 3. The dataset number of features to be considered.
is arranged in Joint Photographic Experts Group data format D. Disease Classification using CNN
(.jpg file).
ML (Machine learning) in Artificial Intelligence (AI)
includes DL (Deep Learning), also referred to as deep neural
learning or deep neural network, have more levels than
machine learning. Its state-of-the-art abilities in a variety of
fields, including object detection, speech recognition, object
categorization, and picture classification, have been enhanced
by deep learning techniques [16]. Convolutional Neural
Networks are one of the most popular deep learning choice. In
some studies, CNN have been used to detect plant diseases
based on the condition of the leaves [17]. CNNs, or
Convolutional Neural Networks, usually have a few multi-
layer convolutional layers that are arranged according to
functions. One or more fully connected layers that are typical
of a neural network commonly follow the subsampling layer.
The input for the following feature layer is a feature collection
that is contained on the preceding layer in a small area. CNN
Figure 3: Images of Potato Leaves with different Diseases has a lot of expertise with computer vision issues. Compared
to ANN (Artificial Neural Network), sharing parameters is
B. Data Preprocessing one of its features, which lowers the number of parameters
The potato-disease dataset is divided into three classes. needed for the model. Additionally, CNN's features are of
These classes include Early-blight class (1000 images), very high standard content. Softmax is the final output layer
Healthy class (152 images) and Late-blight class (1000 as we perform multiclass categorization. The sample model of
images). Hence, the dataset was divided into 46.4%, 7.06% CNN can be visualized in fig. 4 as given below.
and 46.4% respectively.
The splitting of the dataset for training (70%), testing
(15%) and validation (15%) is carried out. Before splitting the
dataset, images of dataset were divided into batches for the
ease of splitting the dataset. Also, specific functions like:
‘prefetch’ - loads next batch of images into CPU when GPU
is busy training the images, ‘cache’ - saves time while reading
the images; were applied for better use of resources of the
system being used.
The CNN model is practiced on the train set. The Figure 4: Convolution Neural Network (CNN)
validation set is used to fine-tune the hyperparameters (i.e.,
In our research, we created a feature map by applying
the design) of a classifier, while the test set provides an
convolution to the input picture through a convolution filter.
objective assessment of the final model fit on the training As a convolution filter, a 3x3 filter is employed. For this
dataset [13]. Noise in the picture can first be reduced by procedure, padding = "valid" is used. When padding is "valid,"
removing any unnecessary parts. If there is too much it must reflect zero padding. 5 convolution layers (C1, C2, C3,
disturbance in the image, it won't be used [14]. Pictures C4, C5) with a kernel/filter size of 3x3 and a subsampling
collected from various sources and of varying sizes must be layer (S1) with a max-pooling frame of 2x2 are included in the
resized to 256x256 pixels for the dataset's standardized input CNN model.
photos. The batch size is of 32.
The input picture with varying sizes is convolved with 32
C. Data Augmentation filters with no padding in the first convolution (C1), producing
In contrast to the shallow networks used in machine 32 features that are 254x254 in size. During the training
learning, deep learning (Deep Network) needs a lot of data. process, the algorithms will learn these parameters. Rectified
Machine learning and deep learning frequently encounter Linear Unit (reLU) are utilised as the activation function in
issues with a lack of training data and an imbalance in the each convolution layer. It makes the pictures' nonlinearity
quantity of data for each class [15]. Data augmentation is the greater. These characteristics are produced by the C1 layer.
approach taken to solve this issue. Data augmentation is a
 64 convolution filters are applied to the features map from RMSProp (Root Mean Square Propagation) and AdaGrad
the preceding layer in the second convolution layer (C2) to (Adaptive Gradient Algorithm). The five-layer Convolutional
create 64 convolved features maps, each measuring 3x3. The Neural Network architecture model was used for the dataset
second subsampling layer (S2), which employs max-pooling learning exercise. The study's period specification called for
and a window size of 2x2, receives these features. As a result, 50 epochs, a 32-batch size, and a learning rate of 0.01 to
18496 shared feature maps, each measuring 125x125, are enhance the model's performance. The suggested method's
produced. algorithm searches for values in the image dataset as part of
the learning process in order to recognise new images. Each
The third convolution layer (C3) uses 64 filters to create epoch phase's value must correspond to the value of the
64 feature maps, each measuring 3x3. The features are passed picture being trained. To calculate the loss and accuracy
to S3, the third subsampling layer, which also employs max values, the outcomes of the epochs are recorded. The accuracy
pooling with a window size of 2x2. The result is the creation value indicates how accurately the system classifies objects,
of 36928 shared feature maps, each measuring 60x60. and the value of the loss gained must be close to or equal to
There are 64 filters used in the fourth convolution layer zero. Loss is an indication that the model has a poor value.
(C4), producing 64 feature maps, each of which is 3x3. With In both plots shown in fig. 5, the findings of the Accuracy and
a window size of 2x2, the fourth subsampling layer (S4),
Loss values are displayed. obtaining values that are near to one
which employs maximum pooling, receives these generated and other to zero. While training, the early and later epochs
features. Consequently, an image of 36928 pooled features show more fluctuation values. After the 50th epoch, the
with 14x14 pixels is produced. training dataset's accuracy, which includes the original
64 filters are used to create 64 feature maps, each images, achieves 98.189%. The CNN needed 7000 seconds,
measuring 3x3 pixels, in the fifth and final convolution layer or approximately 1.9 hours, to complete 50 epochs at an
is (C5). These features are passed to the fifth subsampling average speed of 104 seconds per epoch.
layer (S5), which also employs maximum pooling and a
window size of 2x2, using these features. The result is the
creation of 36928 shared feature maps, each measuring 3x3.
So, no dropout layer is employed. Dropout is a regularisation
method that lowers overfitting. Only the model's training step
uses dropout. Dropout disables some randomly selected
neurons' capacity to activate downstream neurons on the
forward pass and does not apply any weight updates to the
neurons during the backward run. This ignores the neurons
during the training phase in line with the dropout rate.
The model adds a flatten layer as it flattens the data into a
1-dimensional array for entry into the following layer [18]. Figure 5: Accuracy and Loss recorded while Training the
Following that, a dense layer with 64 nodes receives the Dataset
produced features. A dense layer with 64 nodes feeds these
generated characteristics. There are 1,47,520 parameters total B. Testing Process
here. Finally, for the classification, we used a second dense
layer with softmax activation and 3 nodes. Thus, there will be
195 factors in total.
The algorithm has a total of 277,891 trainable factors and the
same are shown in table 1, below.
TABLE I
DEPICTION OF LAYERS AND ITS PARAMETERS INVOLVED DURING
TRAINING THE MODEL
Layers Parameters
Conv2D (C1) 896
Conv2D (C2) 18496
Conv2D (C3) 36928
Conv2D (C4) 36928
Conv2D (C5) 36928
Dense (1) 147520
Dense (2) 195
Total 277891

IV. RESULTS AND DISCUSSIONS

A. Training Process
The CNN model is finally created, and then the Adam
optimizer is used to build the model. It is a highly effective
optimisation method used for deep neural network training. It
incorporates the benefits of two optimisation techniques, Figure 6: Testing the Trained Model on Test Dataset for
Accurate Classification of Disease
 Following the training procedure using the generated enabling the development of more effective and sustainable
dataset, we retrieve the remaining part of the dataset i.e., agricultural practices.
testing dataset to be tested with the model. The testing
accuracy is recorded as 99.18% and testing loss as 2.48%. REFERENCES
Refer, to the fig. 6 for results. [1] J. F. Kreuze, J. A. C. Souza-Dias, A. Jeevalatha, A. R. Figueira, J. P.
T. Valkonen, and R. A. C. Jones, “Viral Diseases in Potato,” The Potato
Also, the model trained and tested is capable of signifying Crop, pp. 389–430, 2020, doi: https://doi.org/10.1007/978-3-030-
the confidence up to what extent it is able to classify or predict 28683-5_11
the disease. [2] Tijjani Ahmadu, A. E. Abdullahi, and K. Ahmad, “The Role of Crop
Protection in Sustainable Potato (Solanum tuberosum L.) Production to
When assessing the model's accuracy with a leaf image Alleviate Global Starvation Problem: An Overview,” IntechOpen
object, it is crucial to keep in mind that image noise will result eBooks, Dec. 2021, doi: https://doi.org/10.5772/intechopen.100058
in subpar classification results. By cutting out everything, but [3] L. Blankers, “Brown Potato Lot,” Unsplash, Feb. 18, 2018. Accessed:
the desired leaf object and ensuring that there is only one leaf Nov. 25, 2022. [Online]. Available:
https://unsplash.com/photos/brown-potato-lot-B0s3Xndk6tw
in each frame, noise can be removed from a picture.
[4] J. Boulent, S. Foucher, J. Théau, and P.-L. St-Charles, “Convolutional
V. CONCLUSION AND FUTURE WORK Neural Networks for the Automatic Identification of Plant Diseases,”
Frontiers in Plant Science, vol. 10, Jul. 2019, doi:
In this research, we suggested a deep learning strategy for https://doi.org/10.3389/fpls.2019.00941
categorising potato leaf diseases using a CNN (Convolutional [5] V. Singh, Varsha, and A. K. Misra, “Detection of unhealthy region of
Neural Network). Pre-processing the leaf images, training a plant leaves using image processing and genetic algorithm,” 2015
CNN model on the pre-processed data, and assessing the International Conference on Advances in Computer Engineering and
Applications, Mar. 2015, doi:
model's performance on a test set are all steps in the suggested https://doi.org/10.1109/icacea.2015.7164858
approach. With an overall accuracy of 99.18%, the [6] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for
experimental findings showed that the CNN model was highly Image Recognition,” 2016 IEEE Conference on Computer Vision and
accurate in classifying two different potato leaf diseases, Pattern Recognition (CVPR), Jun. 2016, doi:
including Early Blight and Late Blight. https://doi.org/10.1109/cvpr.2016.90
[7] M. Islam, Anh Dinh, K. Wahid, and P. Bhowmik, “Detection of potato
The suggested method may offer a trustworthy and diseases using image segmentation and multiclass support vector
effective means of diagnosing potato disease, which is machine,” 2017 IEEE Canadian Conference on Electrical and
essential for ensuring food security and minimising financial Computer Engineering (CCECE), Apr. 2017, doi:
losses in agriculture. This approach can be further extended to https://doi.org/10.1109/CCECE.2017.7946594
other plant species, enabling early detection of diseases, and [8] P. Velmurugan and M. Renukadevi. “Detection of Unhealthy Region
of Plant Leaves and Classification of Plant Leaf Diseases using Texture
facilitating prompt interventions to prevent further spread. Based Clustering Features,” Artificial Intelligent Systems and Machine
Learning, vol. 9, no. 1, pp. 8-10, Feb. 2017. Available at:
Overall, the findings of this research show the efficacy of http://www.ciitresearch.org/dl/index.php/aiml/article/view/AIML012
deep learning-based methods for identifying plant diseases 017003
and their potential to revolutionise the agricultural technology [9] S. R. Dasgupta, S. Rakshit, D. Mondal, and D. K. Kole, “Detection of
industry [19]. The proposed approach can assist farmers and Diseases in Potato Leaves Using Transfer Learning,” Computational
experts in the early detection of diseases and timely Intelligence in Pattern Recognition, vol. 999, pp. 675–684, Aug. 2019,
implementation of control measures, thereby improving crop doi: https://doi.org/10.1007/978-981-13-9042-5_58
yields, reducing losses, and contributing to global food [10] C. Sullca, C. Molina, C. Rodríguez, and T. Fernández, “Diseases
security. Detection in Blueberry Leaves using Computer Vision and Machine
Learning Techniques,” International Journal of Machine Learning and
Furthermore, the proposed deep learning-based approach Computing, vol. 9, no. 5, pp. 656–661, Oct. 2019, doi:
https://doi.org/10.18178/ijmlc.2019.9.5.854
offers several advantages over traditional methods of plant
disease diagnosis. Unlike manual inspection, which is time- [11] Md. A. Iqbal and K. H. Talukder, “Detection of Potato Disease Using
Image Segmentation and Machine Learning,” 2020 International
consuming and prone to human error, deep learning-based Conference on Wireless Communications Signal Processing and
approaches can provide accurate and consistent results with Networking (WiSPNET), Aug. 2020, doi:
minimal human intervention. Additionally, this approach can https://doi.org/10.1109/wispnet48689.2020.9198563
process large amounts of data in a short amount of time, [12] A. Tejaswi, “Plant Village Dataset,” Kaggle, version 1, Mar. 2020.
making it ideal for real-time monitoring and detection of plant Accessed: Sept. 18, 2022. Available:
diseases. https://www.kaggle.com/datasets/arjuntejaswi/plant-village
[13] Wikipedia, “Training, validation, and test data sets,” Wikipedia, Jun.
While the proposed approach achieved high accuracy in 07, 2019. Accessed: Nov. 26, 2022. [Online]. Available:
identifying potato leaf diseases, further research is necessary https://en.wikipedia.org/wiki/Training,_validation,_and_test_data_set
s
to investigate its performance on other plant species and under
different environmental conditions. Additionally, the [14] R. A. Sholihati, I. A. Sulistijono, A. Risnumawan, and E. Kusumawati,
“Potato Leaf Disease Classification Using Deep Learning Approach,”
proposed method can be further improved by incorporating 2020 International Electronics Symposium (IES), Sep. 2020, doi:
additional data sources and developing more efficient models https://doi.org/10.1109/ies50839.2020.9231784
on these datasets. [15] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using Deep Learning
for Image-Based Plant Disease Detection,” Frontiers in Plant Science,
In conclusion, this study demonstrates the potential of vol. 7, Sep. 2016, doi: https://doi.org/10.3389/fpls.2016.01419
deep learning-based approaches for plant disease diagnosis [16] Y. LeCun, Y. Bengio, and G. Hinton, “Deep Learning,” Nature Review
and management. Early detection of potato leaf diseases is Articles, vol. 521, no. 7553, pp. 436–444, May 2015, doi:
made possible by the suggested technique, which can have a https://doi.org/10.1038/nature14539
significant impact on crop protection and global food security. [17] E. Fujita, Y. Kawasaki, H. Uga, S. Kagiwada, and H. Iyatomi, “Basic
With further research and development, this approach can be Investigation on a Robust and Practical Plant Diagnostic System,” 2016
extended to other crops and environmental conditions, IEEE International Conference on Machine Learning and Applications
(ICMLA), Dec. 2016, doi: https://doi.org/10.1109/ICMLA.2016.0178
[18] MathWorks, “flattenLayer,” The MathWorks, 2019. Accessed: Nov. [19] Agrivi, “Improvements in AgTech are Revolutionizing the Farming
27, 2022. [Online]. Available: Industry,” AGRIVI, Sep. 29, 2015. Accessed: Nov. 28, 2022. [Online].
https://in.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.flattenl Available: https://www.agrivi.com/blog/technologies-changing-
ayer.html agriculture/