Plant Disease Detection using CNN

Sivaranjani B Saisuma N Saipriyanka M

Department of CSE Department of CSE Department of CSE
CEG, Anna University CEG, Anna University CEG, Anna University
Chennai, India Chennai, India Chennai, India

Abstract—India loses 35% of the annual crop yield due to plant of deep neural network (DNN) inspired by the human visual
diseases. Early detection of plant diseases remains difficult till system, used for processing images. CNNs have been used for
date. Convolutional neural networks (CNN) have demonstrated plant image analysis since the early days of their evolution.
great performance in object recognition and image classification
problems. The model is designed to perform the plant disease CNNs consist of convolutional layers, which are sets of image
classification. Deep learning with convolutional neural networks filters convoluted to images or feature maps, along with
(CNNs) has achieved great success in the classification of various other (e.g., pooling) layers. In image classification, feature
plant diseases.Revealing the CNN to extract the learned feature maps are extracted through convolution and other processing
as an interpretable form not only ensures its reliability but layers repetitively and the network eventually outputs a label
also enables the validation of the model authenticity and the
training dataset by human intervention.A variety of neuron-wise indicating an estimated class. Given a training dataset, CNN,
and layer-wise visualization methods were applied using a CNN, unlike traditional machine learning techniques that use hand-
trained with a publicly available plant disease image dataset. crafed features , optimizes the weights and filter parameters
We showed that neural networks can capture the colors and in the hidden layers to generate features suitable to solve the
textures of lesions specific to respective diseases upon diagnosis, classification problem.
which resembles human decision-making.The implemented model
achieves 84.54% classification accuracy which shows clearly that
CNN can extract important features and classify plant diseases II. L ITERATURE R EVIEW
from images taken in the natural environment.
Index Terms—Deep Learning, Object Detection, Image Clas- Prior to the advent of deep learning, image processing
sification and machine learning techniques have been used to classify
different plant diseases.
I. I NTRODUCTION First digital images are acquired using digital camera. Then
Plant disease has long been one of the major threats to image processing techniques, such as image enhancement,
food security because it dramatically reduces the crop yield segmentation, color space conversion and filtering, are applied
and compromises its quality. Therefore, early detection and to make the images suitable for the next steps. Then important
identification of plant diseases plays the utmost important features are extracted from the image and used as an input for
role to take timely measures. Accurate and precise diagnosis the classifier.
of diseases has been a significant challenge. Traditionally, The overall classification accuracy is therefore dependent on
identification of plant diseases has relied on human annotation the type of image processing and feature extraction techniques
by visual inspection. Nowadays, digital image acquisition have used. However, latest studies have shown that state of the
allowed the introduction of an array of image-based diagnosis art performance can be achieved with networks trained using
methods at a practical level. However, as the acquired image generic data.
encloses condensed information that is extremely difficult for CNNs are multi-layer supervised networks which can learn
the computer to process, it requires a preprocessing step to features automatically from datasets. For the last few years,
extract a certain feature (e.g., color and shape) that is manually CNNs have achieved state-of-the-art performance in almost
predefined by experts. In such situations, deep learning is all important classification tasks. It can perform both feature
typically used because it allows the computer to autonomously extraction and classification under the same architecture.
learn the most suitable feature without human intervention. A CNN is a special kind of neural networks that has been
The diseases may exhibit symptoms on different parts of the widely applied to a variety of pattern recognition problems,
plant, i.e. leaves, stem, fruits/seeds etc. This research focuses such as computer vision, speech recognition, etc. CNNs com-
on detection and classification of plant diseases based on the bine three architectural ideas to ensure some degree of shift,
symptoms of the diseases that show signs on the leaves of the scale, and distortion invariance: local receptive fields, shared
plant. Datasets of plant diseases differ from others by both the weights and spatial or temporal subsampling. Various CNN
variation and size of the features required for classifcation. architectures were proposed to be used for object recognition
Among various network architectures used in deep learning, eg. LeNet, AlexNet, GoogLeNet etc.
convolutional neural networks (CNN) are widely used in The LeNet architecture is the first CNN introduced by
image recognition. A convolutional neural network is a type LeCun et al. to recognize hand written digits. It consists of
two convolutional layers and two subsampling layers followed A dataset containing 12,673 leaf images is downloaded. The
by a fully connected MLP. samples per class of the dataset is summarized in Table 1.
Few researchers proposed the use of CNN for leaf recog-
No Type of Disease Number
nition and plant disease classification. Atabay (Atabay 2016b) 1 Healthy Leaf 6234
designed a convolutional neural network architecture to iden- 2 Diseased Leaf 6439
tify plants based on leaf images. The proposed architecture Table 1 : Dataset used for the classification.
consists five layers. After each convolutional layer a Rectified
Linear Unit (ReLU) or Exponential Linear Unit (ELU) activa-
tion function is used and for each pooling layer, MaxPooling
approach is applied. The proposed system is applied on Flavia
(Wu et al. 2007) and Swedish (S¨oderkvist 2001) leaf datasets
containing 32 plant species with 1907 samples and 15 species
with 1125 samples respectively. The images in the dataset
are pictures of a single leaf taken at uniform background.
All the input images are 160x160 pixel grayscale images.
The model achieved a classification accuracy of 97.24% and
99.11% accuracy for each dataset. The results showed that
the proposed architecture for CNN-based leaf classification Fig. 1. healthy leaf image.
is closely competing with the latest extensive approaches on
devising leaf features and classifiers.
Angie K. Reyes et al, used a deep learning approach
in which the complete system was learned without hand-
engineered components. The designed system has 5 Conv
layers followed by 2 fully connected layers. The CNN is
trained using 1.8 million images from ILSVRC 2012 dataset
and used a finetuning strategy to transfer learned recognition
capabilities from general domains to the specific challenge
of Plant Identification task. The dataset is combination of
images of a plant or part of a plant taken both under a
Fig. 2. Diseased leaf image.
controlled environment as well as in the natural environment.
They obtained an average precision of 0.486.
Sharada P. Mohanty et al, used the existing deep CNN
architectures, i.e AlexNet and GoogLeNet to classify plant
diseases. Using a public dataset of 54,306 images of diseased
and healthy plant leaves collected under controlled conditions,
the CNN was trained to identify 14 crop species and 26
diseases (or absence thereof). The models achieved 99.35%
accuracy. When tested on a set of images taken at a different
environment than the images used for the training, however,
the model’s accuracy dropped to 31.4%. Overall the result
demonstrates the feasibility of deep CNN for plant disease
classification.

III. M ATERIALS AND M ETHODS

To classify plant diseases a large collection of the plant’s
leaf images is required. The images are downloaded from
the PlantVillage database. In this section the methodology
Fig. 3. Gray Scale Image.
followed is discussed in detail.
To prepare the dataset for the training, the images originally
A. Dataset
at different resolution are re-sized to 128x128 pixels. Since the
Proper and large dataset is required for all classification re- images were taken in the uncontrolled environment the differ-
search during the training and the testing phase. The dataset for ent lighting condition and background in the training images
the experiment is downloaded from the PlantVillage database may bias the neural network. To test this, the experiment was
which contains different plant leaf images and their labels. It also performed using the grayscale and the segmented version
contains a collection of images taken at different environment. of the database.
B. The proposed CNN model [4] S. P. Mohanty, D. P. Hughes, and M. Salath´e, “Using deep learning for
image-based plant disease detection,” Frontiers in Plant Science, vol. 7,
CNN architectures vary with the type of the problem at p. 1419, 2016.
hand. The proposed model consists of three convolutional [5] J. Amara, B. Bouaziz, and A. Algergawy, “A deep learning-based
approach for banana leaf diseases classifcation,” in Proceedings of the
layers each followed by a maxpooling layer. The final layer Datenbanksysteme fur Business, Technologie und Web ¨ (BTW ’17) -
is fully connected MLP. ReLu activation function is applied Workshopband, 2017.
to the output of every convolutional layer and fully connected [6] K. P. Ferentinos, “Deep learning models for plant disease detection
and diagnosis,” Computers and Electronics in Agriculture, vol. 145, pp.
layer. 311–318, 2018.
The first convolutional layer filters the input image with 32 [7] S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D. Stefanovic,
kernels of size 3x3. After maxpooling is applied, the output “Deep neural networks based recognition of plant diseases by leaf image
classifcation,” Computational Intelligence and Neuroscience, vol. 2016,
is given as an input for the second convolutional layer with Article ID 3289801, 11 pages, 2016.
64 kernels of size 4x4. The last convolutional layer has 128 [8] G. Wang, Y. Sun, and J. Wang, “Automatic image-based plant disease
kernels of size 1x1 followed by a fully connected layer of 512 severity estimation using deep learning,” Computational Intelligence and
Neuroscience, vol. 2017, Article ID 2917536, 8 pages, 2017.
neurons. The output of this layer is given to softmax function [9] A. Ramcharan, K. Baranowski, P. McCloskey, B. Ahmed, J. Legg, and D.
which produces a probability distribution of the four output P. Hughes, “Deep learning for image-based cassava disease detection,”
classes. Frontiers in Plant Science, vol. 8, p. 1852, 2017.
[10] M. Brahimi, K. Boukhalfa, and A. Moussaoui, “Deep learning for tomato
The model is trained using adaptive moment estimation diseases: classifcation and symptoms visualization,” Applied Artifcial
(Adam) with batch size of 100 for 50 epochs. Intelligence, vol. 31, no. 4, pp. 299–315, 2017.

C. Experimental Results
The dataset is divided 70% for the training, 10% for
validation and 20% for testing. Different models with different
architectures and learning rate are tested. The parameters of
the network like the kernel size, filter size, learning parameter
were selected by trial and error. ReLu activation function is
used since researches have shown that ReLU result in faster
training.
The classification accuracy from the color images is better
than the gray scale and the segmented images. This shows
the color feature is important to extract important features
for classification. The model that provides good classification
accuracy contains three convolutional layers each followed by
max pooling layer. ReLu activation function is used for each
layer.
Since the dataset is too small when compared to the
total number of trainable parameters of the model, the first
experiment we did is to increase the training data by rotating,
flipping, re-scaling of the images.The validation accuracy got
is 99.32%.

D. Conclusions
In this study convolutional neural network is used to detect
and classify plant diseases. The Network is trained using the
images taken in the natural environment and achieved 99.32%
classification ability. This shows the ability of CNN to extract
important features in the natural environment which is required
for plant disease classification.

R EFERENCES
[1] R. Balodi, S. Bisht, A. Ghatak, and K. H. Rao, “Plant disease diagnosis:
Technological advancements and challenges,” Indian Phytopathology,
vol. 70, no. 3, pp. 275–281, 2017.
[2] F. Martinelli, R. Scalenghe, S. Davino et al., “Advanced methods of plant
disease detection. A review,” Agronomy for Sustainable Development,
vol. 35, no. 1, pp. 1–25, 2015.
[3] A. Johannes, A. Picon, A. Alvarez-Gila et al., “Automatic plant disease
diagnosis using mobile capture devices, applied on a wheat use case,”
Computers and Electronics in Agriculture, vol. 138, pp. 200–209, 2017.