A New Approach for Leaf Disease Detection using

Multilayered Convolutional Neural Network

2023 3rd International conference on Artificial Intelligence and Signal Processing (AISP) | 979-8-3503-2074-9/23/$31.00 ©2023 IEEE | DOI: 10.1109/AISP57993.2023.10134843

Vivek Shukla Sweta Rani Ramesh Kumar Mohapatra

Department of Computer Science Department of Computer Science Department of Computer Science
National Institute of Technology National Institute of Technology National Institute of Technology
Rourkela, Odisha, India Rourkela, Odisha, India Rourkela, Odisha, India
[email protected] [email protected] [email protected]

Abstract—Leaf diseases reduce agricultural yield by 35% improves the agricultural industry. The optimal amount and
annually in India, where agriculture is the main sector. Manual timing of agricultural inputs are used after harvest. Precision
detection of the type of disease present in leaves takes a long agriculture makes use of tools such as Global Positioning
time since laboratories lack the necessary tools and expertise
to recognise early leaf diseases. Diseases include early blight, System (GPS), remote sensing, and Geographic Information
late blight, black root, bacterial spot, mould leaf, healthy leaf, Systems (GIS). This method has enabled the secure collection,
etc. Automated leaf disease detection systems are beneficial for analysis, and deployment of field data. The field’s spatial
spotting disease symptoms on plant leaves as soon as they appear, and temporal diversity indicates patterns, linkages, and the
which helps to ease the time-consuming process of monitoring importance of this strategy. Fungi, viruses, bacteria, and certain
large agricultural farms. Recent advances in deep learning (DL)
and computer vision models have highlighted the importance of nematodes are mainly responsible for the majority of leaf
developing autonomous leaf disease detection algorithms based on diseases. Leaf disease symptoms include a major change in
visual symptoms on leaves. The fundamental goal of the proposed the colour or the leaf’s form, scabs, wilting, rust, blotches
model is to tackle the problem of leaf disease diagnosis using the and mouldy coatings caused by the pathogen. For example,
most basic strategy while utilising the fewest computer resources tomato powdery mildew first appears as light yellow patches
to generate results comparable to state-of-the-art techniques.
We used the multilayered architecture of convolutional neural on the leaves. The spots quickly fill with white spores, giving
networks (CNN) to identify and overcome leaf diseases. When the leaves the appearance of being sprinkled using flour.
compared to existing techniques, our proposed method can The once-pale leaf segments turn black and shrink, becoming
successfully train an image classification model to achieve 98.5% brittle and dry as the fungus infection worsens. Separating
accuracy. defective regions, extracting relevant features, and categorising
Index Terms—Plant Leaf Disease Detection, Images Processing,
CNN, Agriculture, Deep Learning the data using machine learning algorithms are all part of the
pre-processing stages for computer vision-based modules for
classifying leaf illnesses.
I. I NTRODUCTION
Training numerous spectral image inputs and defining their
Agriculture is the primary source of income for India’s objectives are the main issues faced by CNN. With CNN clas-
massive rural population. Agriculture is the principal source sifier applications, classifying the variations from the provided
of income for 58% of the population. Raising the threshold input data has become even more difficult.
in agriculture will very certainly raise it for rural residents.
Numerous reports and studies, however, show that our Indian A. Motivation
farmers continue to use old farming techniques and are un- The use of CNN in agriculture in recent years, its growing
aware of contemporary equipment and strategies for battling popularity and success in resolving multiple agricultural chal-
leaf diseases. As a result, both the quality and quantity of agri- lenges, and the fact that there are currently numerous research
cultural products suffer. The demand for a reliable food supply projects employing CNN to discuss various agricultural topics
develops in tandem with the world population. It is critical are driving forces behind the research. Due to its remarkable
to producing food in a clean environment. This is possible in result, CNN is perhaps the most well-known and widely used
large quantities on a plantation with modern technologies. Pre- methodology in the agricultural field, and the fact that there
harvest and post-harvest planning, application of agricultural are currently numerous research projects employing CNN to
input resources such as fertilisers, early detection of pests discuss various agricultural topics is the driving force behind
and illnesses, and weed identification are just a couple of the research. In terms of image analysis, the current study
areas for image processing and computer vision in agriculture. focuses on a specific branch of Deep Learning models and
Plants in the dataset include tomato, apple, blueberry, grape, approaches because there aren’t many studies of this type in
peach, cherry, corn, orange, strawberry, pepper, squash, potato, agriculture, particularly with CNN use. As a result, it would be
raspberry, and soybean. The data set contains 18,345 training beneficial to offer and analyse essential information to assist
images and 3656 test images. Precision agriculture technology researchers in conducting a more thorough examination.

Authorized licensed use limited to: Graphic Era (Deemed to be University). Downloaded on October 25,2023 at 06:58:29 UTC from IEEE Xplore. Restrictions apply.
B. Research Contribution clustering for NN and segmentation for classification using a
In this study, deep convolutional networks are used to create texture feature set. Images of potato leaves have been classified
a way to figure out what’s wrong with a plant-based on how as healthy or sick using the support vector machine classifier
its leaves look. The suggested model is capable of detecting [7]. This research aims to provide a disease management tool
38 distinct forms of plant diseases in different plant species that can automatically identify areas of potato leaves that have
by distinguishing plant leaves from their surroundings. been impacted by early blight or late blight. The deep learning
technique has been used for leaf disease identification by the
• To use the dataset available on Kaggle named “New Plant
authors [8].
Disease Dataset” for leaf disease detection on different The CNN and RNN methods detect several disease types
plants. in plants. A total of thirteen different leaf diseases may be
• To Propose an efficient model using CNN in order to
identified and categorized automatically using this approach.
detect the leaf disease. There are 10 distinct illnesses that may be detected using
• To achieve good accuracy using a multilayered CNN
CNN on roughly 500 photos of rice plant leaves. In this case,
architecture. the author used cross-validation with a 10-fold sample size.
Rice leaf diseases are described by the authors in [6]. Leaf
II. R ELATED W ORK
disease detection segmentation has been studied in relation to
Different sorts of literature have recommended various preprocessing methods. Normal and sick leaves are separated
approaches and techniques for identifying leaf diseases in using a histogram plot. Features of both form and colour
plants. The illnesses in tea leaves have been identified using are derived. The PCA approach [9] is used to extract shape
a Support Vector Machine classifier. An important contributor features, while colour features are retrieved using colour-based
to the economic development of countries such as India and grid moments. Every one of these strategies relies on some
Bangladesh is the manufacturing of tea. The authors [1] have combination of texture, shape, and colour traits in order to
dedicated their efforts to creating a model that can identify extract relevant information. It’s also difficult to say how many
the most common tea leaf diseases. As a result, the country’s neurons the NN classifier uses, and training it requires a very
output and rate of growth benefit. In order to categorise the limited sample.
various leaf diseases, machine learning methods have been In terms of classification, and accuracy, Support Vector Ma-
used. The identification of leaf diseases [2] has been addressed chines are superior to both Random Forest and Logistic Re-
by several authors who have employed KNN, Decision Tree, gression. A mechanism for automatically detecting early blight
SVM, and neural networks. Identifying diseases in leaves This and late blight damage in plant leaves has been implemented
article focuses on three major leaf diseases: early black spot, in this study. To better identify leaf diseases, deep learning
late black spot, and black rot. methods might be used. Different types of leaf diseases may
Understanding which parts of a leaf are diseased and which be detected using CNN and RNN approaches. This approach
are healthy is the primary goal. There have been attempts to may be used to automatically detect and categorize thirteen
employ image processing methods for this purpose. Imple- distinct leaf diseases. Ten distinct illnesses were identified in
menting classification and image analysis algorithms for leaf over 500 photos of rice plant leaves using CNN. In this case,
disease detection is the primary focus of this effort. In order to the cross validation method has been used 10 times. Very few
distinguish between healthy and diseased leaf photos, the SVM test photos are used in the report despite a large number of
classification algorithm has been used. Neural Networks(NN) investigations. Approximately 15,000 leaves, both healthy and
have been suggested by the authors [3] for the diagnosis and harmful, are considered. The presence of plant diseases in the
categorization of grape leaf diseases. The input for this system leaf has been identified using three machine learning methods:
is a picture of a grape leaf. Green pixels are masked using random forest classifier, logistic regression, and support vector
thresholding. Noise cancellation is achieved by anisotropic machine [10]. These models have been compared in depth with
diffusion. The grape leaf disease is then segmented using K- all algorithms.
means clustering. The afflicted organ is identified by the use
of neural networks. In [4], the authors presented a method III. BACKGROUND D ETAILS
for identifying leaf spots caused by illness using colour One of the most widely used AI techniques today among
transforms. This research provides an analysis of how various researchers looking at automation and the Internet of Things
colour spaces affect the identification of disease spots. is deep learning (DL). This is because of significant im-
Comparisons are made between the HSI, CIELAB, and provements in processing speed and data accessibility, which
YCbCr colour spaces, and the CIELAB colour model is have enabled researchers in numerous domains to come to
ultimately implemented through the A component. When meaningful conclusions. CNN algorithms have developed to
smoothing a picture, the median filter is often used. Finally, the point where they can understand images in a manner
the threshold is determined by applying the Otsu technique similar to that of human brains.
to the colour part of the image. The authors [5] established a To begin with, DL systems need a tonne of high-quality data
rapid and thorough method for diagnosing and classifying leaf in order to properly learn and predict the future. Therefore, the
diseases. In this approach, the author [6] employs K-means initial stage in applying DL to image processing is to assess the

Authorized licensed use limited to: Graphic Era (Deemed to be University). Downloaded on October 25,2023 at 06:58:29 UTC from IEEE Xplore. Restrictions apply.
images, followed by annotation and generalisation. To develop
the model, used Convolutional Neural Network.
The methods for categorising images are listed in Table I.

TABLE I
D IFFERENT I MAGE P ROCESSING T ECHNIQUES

Sr.No. Techniques
1 Convolutional Neural Network (CNN)
2 Random Forest Classifier (RF)
3 K-Nearest Neighbors (KNN)
4 Probabilistic Neural Network (PNN)
5 Naive Bayes (NB)
6 Artificial Neural Network (ANN)
7 Support Vector Machine (SVM)

Fig. 1. Workflow of Model

A. Deep Learning
A subset of machine learning entitled “deep learning” is
fully supported by artificial neural networks. Deep learning A histogram must be used to plot the regions independently.
is a type of human brain mimic because neural networks are HOG consists of the following actions: The first step is to
built to resemble the human brain. Not everything has to be read the image from the folder, and then the OpenCV library
explicitly programmed into deep learning. The premise of deep converts RGB images to grayscale. HOG is then taken from
learning is not new. It has been in existence for some time. each image after it has been resized to a format that may be
We aren’t equipped with as much processing power or data as used. The NumPy array, which contains the images and titles,
we previously did, therefore it’s more frequent now. is then used to construct the array. Fig. 1 depicts the suggested
model’s workflow.
B. Neural Network (NN)
A. Steps
Biological neural networks serve as the foundation for arti-
1) The images are loaded, their size is downsized to 128
ficial neural networks. These systems do not use task-oriented
× 128 (256 × 256 images take longer to process), and
rules; instead, they learn to perform tasks by being exposed to
then they are turned into tensors.
a variety of examples and facts. It is believed that the system,
2) 30% of the overall dataset is used to create a validation
which is not pre-programmed with a pre-coded comprehension
dataset, and 70% is used for the training dataset.
of such datasets, produces distinguishing characteristics from
3) Data loading with batches.
the data that is provided to it. Threshold logic computational
4) Combination of Multilayered CNN architecture is used
models underpin neural networks. The foundations of neural
to train the model.
networks are the study of the brain or the application of neural
5) Model development and evaluation using test data.
networks to artificial intelligence.
B. Data Preparation
C. Convolutional Neural Network (CNN)
Data preparation is used to clean, standardize, and enrich
CNN, or Convolutional neural network, is a subset of deep the raw image to make it ready for training the model.
learning neural networks that are used to analyse ordered When data is acquired in the form of images, it should
arrays of data, like photos. CNN is extremely good at recognis- be ensured that there are enough features for the learning
ing design elements in the input image, such as arcs, gradients, model to be correctly taught and trained. Generally, more data
rings, or even faces and people. always yields better outcomes. The leaf image of the plant
IV. P ROPOSED W ORK is clipped in order to isolate the affected regions, and the
image is then smoothed. Each image then goes through image
The proposed model uses deep convolutional networks to preprocessing techniques. In order to improve image quality,
develop a method for recognizing leaf diseases based on preprocessing is applied to a collection of images that are in
the classification of leaf images. Out of 14 different plants, various dimensions. Both unwanted distortion and background
the proposed model can identify 38 different types of leaf noise are suppressed by it. This image is first downsized to
illnesses. Following the collection of the dataset, basic data 128 × 128, after which thresholding is carried out to extract
analysis and feature extraction methods are applied. It helps us all green colour components.
focus on the image’s most valuable elements while dismissing
the rest. For feature extraction, a feature descriptor called the C. Understanding The Data
Histogram of Oriented Gradients (HOG) is used. The first The dataset referred to as ”New Plant Diseases Dataset”
step is to recognise the edges and orientation of pixels in could be accessed at https://www.kaggle.com/vipoooool/new-
an image. To do this, gradient and orientation are computed. plant-diseases-dataset. This dataset was obtained from Kaggle.

Authorized licensed use limited to: Graphic Era (Deemed to be University). Downloaded on October 25,2023 at 06:58:29 UTC from IEEE Xplore. Restrictions apply.
When data is gathered in the form of images, it’s important that might not otherwise be noticeable. Fig. 4 illustrates how
to make sure there are enough features so that the trained the image was permuted so that the pixel image became
model can be properly taught and trained. 87,867 images are reversed.
gathered to build the dataset requisite to apply the model.
Only 22,001 of the 87,867 images are used to test and train
the model. There are images of various healthy and diseased

Fig. 4. Single image visualization

E. Building the CNN Model Architecture

The proposed model has been implemented using multilay-
Fig. 2. Sample Images of New Plant Diseases Dataset ered CNN architecture. Initially, a base image classification
class is defined that has features for each batch of data
crop leaves, as shown in Fig. 2. A train and valid directory training and data validation. This will save us from repeatedly
are created by dividing the dataset. The folders present in the developing these routines and allow us to build multiple CNN
train directory are as follows: models. For training, the model batch of images is used shown
Tomato Target Spot, Corn (maize) healthy, in Fig. 5. A CNN architecture is created by inheriting the basic
Tomato Bacterial spot, Strawberry healthy, Blue- class.
berry healthy, Tomato Tomato mosaic virus, Straw- CNN architecture provided for this model is given below:
berry Leaf scorch, Tomato Leaf Mold, Potato healthy, • Building a multilayer CNN model from scratch using
Grape Esca (Black Measles), etc. Conv2D
Number of Unique Plants present in the dataset: 14 • Relu and Maxpool 2D and Linear Layers.
Unique Plants: [‘Strawberry’, ‘Blueberry’, ‘Potato’, ‘Pepper’, Data loading into a GPU device and model creation: For fast
‘Apple’, ‘Tomato’, ‘Peach’, ‘Soybean’, ‘Grape’, ‘Squash’, processing of the images, the data must be loaded from the
‘Corn’, ‘Cherry’, ‘Raspberry’, ‘Orange’] CPU to the GPU. Computing on a CPU will be exceedingly
To process images more quickly, a batch of the dataset is time consuming due to the size of the data.
created. As long as it fits in memory, the batch size of 64
images is chosen according to the need of the model. For
quicker processing of the images, a batch of the dataset has
been created as shown in Fig. 3. Once the directory structure

Fig. 3. Batch of Images Fig. 5. CNN Architecture

has been verified, the dataset is established and images of

size 128 × 128 are loaded as PyTorch tensors. Applying V. R ESULTS AND D ISCUSSION
transform to the images in the directory allows us to load To determine how well the suggested method worked, our
them as Pytorch tensors. The test dataset will be created in model ran a number of tests utilising databases of photos
the valid folder, while the training and validation datasets of both healthy and diseased tomato leaves. The leaves with
will be created in the train folder. various diseases are quite similar to one another, which makes
disease identification and classification difficult for this study.
D. Visualizing a single image Because of this similarity, certain leaves may be folded into
By putting data in a visual context and trying to understand the incorrect groups. The accuracy of the model on test data
it, data visualisation studies patterns, trends, and connections is 98.5%.

Authorized licensed use limited to: Graphic Era (Deemed to be University). Downloaded on October 25,2023 at 06:58:29 UTC from IEEE Xplore. Restrictions apply.
A. Result analysis of proposed model VI. C ONCLUSION AND F UTURE D IRECTION
In this proposed model, a multilayered Convolutional Neu-
ral Network architecture is used for the classification and
detection of different leaf diseases provided in the given
dataset. The collection is comprised of 22,001 pictures of
various types of leaves. Each image in the dataset has had
a distinct input matrix developed for its R, G, and B channels.
These matrices have also been constructed. This model is
capable of effectively training an image classification model
with multilayered CNN architectures, as demonstrated by the
model’s test accuracy of 98.5%.
The majority of the studies used CNN techniques, and also
underlined that pre-training models, rather than creating new
models from starting with leaf image datasets, can quickly
enhance performance accuracy, especially if there is enough
Fig. 6. Accuracy of model after applying CNN architecture data for each class to train the models. However, a critical
future impact would be the development of extremely efficient
detection algorithms using vast datasets containing various
B. Prediction on some single image of dataset plant leaf diseases. Requiring large generalised datasets would
assist to balance out the class imbalance. In future, image
localization techniques can be used in order to pinpoint the
precise location of the leaf’s damaged portion. To put this
model into action, a Flask application programming interface
(API) might be constructed.
R EFERENCES
[1] Newlin Shebiah Russel and Arivazhagan Selvaraj. Leaf species and
disease classification using multiscale parallel deep CNN architecture.
Neural Computing and Applications, 34:19217–19237, 2022.
[2] Heba Al-Hiary, Sulieman Bani-Ahmad, M Reyalat, Malik Braik, and
Zainab Alrahamneh. Fast and accurate detection and classification of
plant diseases. International Journal of Computer Applications, 17:31–
38, 2011.
[3] Srdjan Sladojevic, Marko Arsenovic, Andras Anderla, Dubravko Culi-
brk, and Darko Stefanovic. Deep neural networks based recognition of
plant diseases by leaf image classification. Computational intelligence
and neuroscience, 2016.
[4] R Chavan, A Deoghare, R Dugar, and P Karad. Iot based solution
for grape disease prediction using convolutional neural network and
farm monitoring. International Journal of Scientific Research and
Engineering Development, 2019.
[5] T Suman and T Dhruvakumar. Classification of paddy leaf diseases
using shape and color features. IJEEE, 7:239–250, 2015.
[6] Jayme Garcia Arnal Barbedo. Digital image processing techniques for
detecting, quantifying and classifying plant diseases. SpringerPlus, 2:1–
12, 2013.
[7] Changjian Zhou, Sihan Zhou, Jinge Xing, and Jia Song. Tomato leaf
disease identification by restructured deep residual dense network. IEEE
Access, 9:28822–28831, 2021.
[8] Archana Chaudhary, Savita Kolhe, and Raj Kamal. An improved random
forest classifier for multi-class classification. Information Processing in
Agriculture, 3:215–222, 2016.
[9] Anisha Kumari, Satya Prakash Sahoo, Ranjan Kumar Behera, and
Bibhudatta Sahoo. Supervised machine learning for link prediction
using path-based similarity features. In 2020 IEEE 17th India Council
International Conference (INDICON), pages 1–7. IEEE, 2020.
[10] Ananda S Paymode, Shyamsundar P Magar, and Vandana B Malode.
Tomato leaf disease detection and classification using convolution neural
Fig. 7. Prediction on test dataset network. In 2021 International Conference on Emerging Smart Com-
puting and Informatics (ESCI), pages 564–570. IEEE, 2021.

The above Fig. 7 shows the prediction of our model. In

the figure, the label means the actual name of the leaf and
the predicted means the output of our model. Some of the
prediction and label name is not matching.

Authorized licensed use limited to: Graphic Era (Deemed to be University). Downloaded on October 25,2023 at 06:58:29 UTC from IEEE Xplore. Restrictions apply.