Crop Disease Detection Using Deep Learning

Omkar Kulkarni
Dept. of Information Technology,
Pune Institute of Computer Technology, Savitribai Phule Pune University,
Pune, India.
Email: omkar.omkar.135@gmail.com

Abstract—In recent times, drastic climate changes and lack of II. RELATED WORK
immunity in crops has caused substantial increase in growth of
crop diseases. This causes large scale demolition of crops, The identification of the crop species is the basic
decreases cultivation and eventually leads to financial loss of requirement before we identify the class of disease. Abdul
farmers. Due to rapid growth in variety of diseases and adequate Kadir in his research work has used color features like mean,
knowledge of farmer, identification and treatment of the disease standard deviation, skewness and kurtosis are made on the
has become a major challenge. The leaves have texture and visual pixel values of the leaves. He has consolidated features
similarities which attributes for identification of disease type. coming through grey level co-occurrence matrix (GLCM)
Hence, computer vision employed with deep learning provides
functions which identifies the texture of an image. It generates
the way to solve this problem. This paper proposes a deep
learning-based model which is trained using public dataset a GLCM that produce statistical measures by calculating how
containing images of healthy and diseased crop leaves. The model frequently pairs of pixel with specific values and in a specified
serves its objective by classifying images of leaves into diseased spatial relation occur in an image [1].
category based on the pattern of defect In paper [2], neural network is used to detect the disease. If
leaf is infected, further processing is done to identify the
Keywords—Crop disease, Deep learning, Transfer Learning, disease. Genetic algorithm is used along with SVM to
MobileNet, InceptionV3, Image classification. optimize loss and identify the type of disease. This paper
states method for optimization of loss function using genetic
I. INTRODUCTION algorithm, which is similar to theory of natural selection
where only strong parameters survive.
Agriculture is the main source of food, raw material and In paper [3], semi-supervise learning is used in the form of
fuel which contributes to the economic development of a Generative Adversarial Networks to classify images. In such
nation. Nearly 66% of the population depends on agriculture type of learning discriminator is transformed into a multi-class
directly or indirectly. As there is a rapid growth in global classifier and generator is used only for training of
population, agriculture is struggling to fulfill its necessity. The discriminator. In paper [4], transfer learning is used for
food security remains threatened by various circumstances content based image recommendation and Inception-V3 model
including climate change, the decline in pollinators, crop is used for building of neural
diseases, lack of irrigation, etc. Crop disease alleviates the
production and also the quality of food. Crop diseases not only
affect the food security at the global level, but it also has III. METHODOLOGY
adverse consequences for small scale farmers whose income In this paper the implementation is done in different phases
depends on healthy cultivation. There is an advantage that the in the following manner: collecting the dataset, pre-processing
crop diseases can be controlled by identifying the diseases as the dataset, training the Convolutional Neural Network (CNN)
soon as it develops on crops. Due to advancement of internet, model to identify the type of crop, training CNN model to
field of computer vision it has been possible to provide detect the disease, validation of model through obtained
impactful solution to this problem. results.
The objective of this paper is to provide an application
which predicts the type of crop disease based on textural A. Dataset
similarity of leaves. Publicly available dataset containing Plantvilla which is an open source dataset is used which
healthy and diseased crop leaves is used to train the model. contains 54,306 images of crop leaves classified in 38
The early diagnosis of crop disease can be used to prevent different classes. The dataset covers 13 type crop species and
further damage that can be done to the crops which is helpful 26 types of diseases. Each class has pair of fields containing
for sustaining the cultivation. the name of crop and the name of disease. All these images are
The remaining part of the paper is organized as follows. segmented and resized to 224 x 224 size and are converted
Section II includes the related work done in this particular into grayscale images before further processing.
area. Section III deals with the proposed model and the
methodology that has been used to solve the problem. Section B. Preprocessing of image
IV deals with experiments performed to on the given problem
and results obtained for the same. Finally, the results are The images present in the dataset have varying background
concluded which includes comparison between different and non-uniform lighting which affects accuracy of the
models and future scope. application. The pre-processing of image is essential for
removing noise and segmentation of the image which helps in
improving the accuracy of CNN model. Hence, to handle the

978-1-5386-5257-2/18/$31.00 ©2018 IEEE

Authorized licensed use limited to: Canara Engineering College - BANTWAL. Downloaded on February 28,2025 at 08:47:30 UTC from IEEE Xplore. Restrictions apply.
varying background problem, segmentation is performed that
extracts only relevant part of the image. So, after performing
the segmentation all the leaf images with black background
are obtained. Further, to handle the non-uniform lighting
conditions segmented images are converted to grey scale
images and are passed on for further processing. Images
obtained after preprocessing are shown in fig 1.

Fig 2. Architecture of InceptionV3 [4].

a.) Original b) Grey Scale c) Segmented

Fig 1. Preprocessing of leaf

C. Crop disease detection and classification

Classification of the disease is performed in two steps
where the first step is to detect the type of crop and seconds
step is to detect the type of disease. To perform these task
deep convolutional neural networks is used. Transfer learning
is used to build the deep learning model and is trained using
the ImageNet dataset. Transfer learning is a machine learning
technique in which a model is trained on one task is re-
purposed on another related task. It is a technique in which
pre-trained neural networks are used to build the neural
network for similar kind of task to implant rapid and
sustainable progress for solving the problem. These pre-
trained networks are developed by training on large datasets
which contains enormous number of diversified images.
Several research organizations build such kind of models
which take weeks to train on latest high-end hardware. These
are released under a permissive license for reuse that are
directly incorporated to build new model for solving similar
type of problems. These pre-trained models can be fine-tuned
by using new dataset, if its nature is similar to the dataset on
which network is trained. In such type of cases only last layer
of network is trained, then tuned network can be directly used
Fig 3. Architecture of MobileNet [5].
to solve problem. If the size of dataset is large enough then the
pre-trained model can be retrained using new data and, in such
Transfer learning is used to build deep learning model
case, the neural network is initialized with weights of the pre-
using MobileNet and InceptionV3 pre-trained models. These
trained model.
models are fine-tuned by using image dataset of 5 different
Among the various pre-trained models such as Xception,
types of crops with 5277 images. These models are initialized
VGG16, VGG19, MobileNet, ResNet50, InceptionV3, etc.
with the weights of the pre-trained model. Dataset is
InceptionV3 and MobileNet are used for implementation
preprocessed and divided into 80%-20% training and testing
through transfer learning. The InceptionV3 architecture as
data. Label encoding is used for conversion of output to
shown in Fig 2 has 23,851,784 parameters and depth of 159
categorical type. Image data generator is used to introduce
layers and MobileNet architecture as shown in Fig 3 has
variation in the input images. A dense layer is appended with
4,253,864 parameters and depth of 88 layers. the softmax activation function to extract the result from
model. Adam optimizer is used with the categorical cross
entropy as loss function. These pre-trained models are then re-
trained for 10 epochs with batch size of 8 by using new
training dataset to create required model. Dropout of 1e-3 is
added to overcome the problem of overfitting. Generated
model is tested using testing dataset to find out validation
accuracy. Performance of both the models is measured on the
basis of training accuracy, training loss and validation
accuracy, validation loss per epoch.

2018 Fourth International Conference on Computing Communication Control and Automation (ICCUBEA)
Authorized licensed use limited to: Canara Engineering College - BANTWAL. Downloaded on February 28,2025 at 08:47:30 UTC from IEEE Xplore. Restrictions apply.
 A dedicated deep convolutional neural network is used to
detect disease related to individual crop which incorporates
similar methodology of plant detection. There are multiple
models generated for disease detection each corresponding to
a particular type of crop. The type of plant is detected using
previously mentioned model which determines to which
disease detection model the image should be passed on for
detection of the type of disease. The models build for
identification of disease are evaluated using testing dataset on
the basis of training accuracy, training loss and validation
accuracy, validation loss per epoch.

IV. RESULTS Fig 6. Accuracy Vs Epoch for crop detection for InceptionV3

In this section results and observations of the

experimentation performed on both the models are mentioned.
After the experimentation on trained model it is found that
model trained using segmented images perform better than
model trained using color and gray-scale images. In case of
experimentation for detection of crop type, as per Fig 4-7 both
MobileNet and InceptionV3 models perform well with
99.62% and 99.74% accuracy respectively. Significant growth
in accuracy is observed in initial stages which get converged
later on. Exponential drop in loss function signifies faster
learning in the initial stage. It is observed that InceptionV3
model performs better than MobileNet in the task of crop
detection
Fig 7. Loss Vs Epoch for crop detection for InceptionV3

In similar way for crop disease detection as per Fig 8-11

both MobileNet and InceptionV3 models shows steady growth
with 99.04% and 99.45% accuracy respectively. Slight
decrement can be observer in the accuracy from 6th epoch to
7th epoch in case of InceptionV3 model. Value of loss
function at the end of 10th epoch supports the better
performance of InceptionV3 in the task of disease detection.

Fig 4. Accuracy Vs Epoch for crop detection for MobileNet

Fig 8. Accuracy Vs Epoch for disease detection for MobileNet

Fig 5. Loss Vs Epoch for crop detection for MobileNet

2018
Authorized licensed use limited Fourth
to: Canara International
Engineering CollegeConference on Computing
- BANTWAL. Downloaded Communication
on February Control
28,2025 at 08:47:30 andIEEE
UTC from Automation (ICCUBEA)
Xplore. Restrictions apply.
 REFERENCES
[1] Kadir, A., “A Model of Plant Identification System Using GLCM,
Lacunarity and Shen Features,” Research Journal of Pharmaceutical,
Biological, and Chemical Sciences Vol.5(2) 2014.
[2] Naik, M.R., Sivappagari, C., “Plant Leaf and Disease Detection by Using
HSV Features and SVM,” IJESC, Volume 6 Issue No.12, 2016
[3] Greg Olmschenk, Hao Tang, Zhigang Zhu, “Crowd Counting with
Minimal Data Using Generative Adversarial Networks for Multiple
Target Regression, Applications of Computer Vision (WACV), 2018
IEEE Winter Conference on Computer Vision (WACV), 2018.

[4] Surbhi Jain, Joydip Dhar, ”Image based search engine using deep learn-
ing”, 2017 Tenth International Conference on Contemporary Computing
(IC3), August 2017

Fig 9. Loss Vs Epoch for disease detection for MobileNet [5] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko,
Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam, ”Mo-
bileNets: Efficient Convolutional Neural Networks for Mobile Vision
Applications”, arXiv preprint arXiv:1704.04861v1, 2017.

Fig 10. Accuracy Vs Epoch for disease detection for InceptionV3

Fig 11. Loss Vs Epoch for disease detection for InceptionV3

V. CONCLUSION
In this paper, respectively, the applications of Deep
Convolutional Neural Networks have been formulated with
the goal of classifying both crop species and identity of
disease on images. The proposed methodology was tested on
five classes of crops and three types of crop diseases for each
class. The experimental results show that the InceptionV3
model performs better than the MobileNet model in terms of
accuracy and validation loss.
An extension of this work will include the classification of
images that are not captured in a controlled environment and
images that have multiple orientation. Also, the number of
classes of crops and its diseases can be further increased. This
methodology can be integrated with smart phone applications
that would provide user friendly GUI and simplicity for its
usage.

2018
Authorized licensed use limited Fourth
to: Canara International
Engineering CollegeConference on Computing
- BANTWAL. Downloaded Communication
on February Control
28,2025 at 08:47:30 andIEEE
UTC from Automation (ICCUBEA)
Xplore. Restrictions apply.