2019 6th IEEE International Conference on Engineering Technologies and Applied Sciences (ICETAS)

Detecting Sugarcane Diseases through Adaptive

Deep Learning Models of Convolutional Neural
Network
Sammy V. Militante Bobby D. Gerardo
University of Antique West Visayas State University
Sibalom, Antique, Philippines Lapaz, Iloilo City, Philippines
[email protected] [email protected]

Abstract— Sugarcane disease causes a threat in the Convolutional Neural Network. The CNN method includes a
sugarcane industry that leads to massive demolition of infected deep learning method that uses artificial neural networks [1].
crops, decreases cultivation and financial loss of farmers. Early With the introduction of deep learning in the field of
detection of the diseases using machine learning technology can agriculture [2], and in plant disease identification [3], has
avoid such loss and disaster. Subsequently, deep learning is an become popular now a days due to the performance and
interesting technology to solve this problem through machine accuracy. The Convolutional Neural Networks (CNNs) is the
learning. CNN has proven to be a popular technique in deep learning methods and has potentials in handling
recognizing and detecting plant diseases through deep difficult process and doing pattern recognition using images
learning. This paper aims to integrate various CNN
[4].
architectures of deep learning models to achieve the highest
accuracy rate in detecting and recognizing sugarcane diseases. This paper trains different models of CNN to predict the
The models were trained using 14,725 images of a healthy type of sugarcane disease based on the features of leaves and
sugarcane leaves and infested sugarcane diseases and achieves to compare the best model in detecting and recognizing
a maximum 95.40% accuracy rate during training. Three sugarcane diseases. The dataset is composed of 7 classes of
architectures of CNN; StridedNet, LeNet, and VGGNet were sugarcane images in which 6 belongs to the type of diseases
used in the conduct of this study. VGGNet model tops the other and one to the healthy class. Initial analysis of sugarcane to
two models achieving an accuracy rate of 95.40%, LeNet model
prevent damage to the crop is the appropriate method to
achieves the rate of 93.65%, while StridedNet model achieves
apply and to sustain healthy cultivation.
90.10% accuracy rate during training. The three models
trained serve its purpose by classifying images of a healthy The paper is arrange as follows. Section 1 gives the
sugarcane leaves and diseased infested sugarcane leaves based introductory part and the importance of system design for
on the leaf features. This paper also bids a successful result for early detection of sugarcane leaf disease. Section 2 consist of
the researchers who are in the development of recognition the related works done in this area. Section 3 portrays the
system for sugarcane diseases. CNN design. The next section is the proposed methodology
of various models of CNN. Section 5 discusses the
Keywords— Sugarcane disease detection, Sugarcane disease
experimental settings. Section 6 discusses the results and
recognition, Convolutional Neural Network, Deep Learning
analysis and the last section concludes the paper and future
works.
I. INTRODUCTION
Diseases from a sugarcane plant has a wide scientific II. RELATED WORKS
area of study and may focus on the biological characteristics
of the disease. Disease detection and identification of plants To identify the type of sugarcane disease, the basic
turn out to be inspiring and requires special attention. Plant requirement is to identify the crop species. A. Kadir in his
diseases endangers food safety and disastrous to small-scale study used color features of leaves and consolidated these
features through grey level co-occurrence matrix or GLCM
farmers whose income depends on healthy cultivation. Initial
finding of these diseases facilitates high-quality of to distinguish leaf image texture and produces specific values
production of sugarcane as for the benefit of both the farmers through statistical computations on pairs of pixels and spatial
and consumers. To restrain from greater damage preventive occurrence in relation to the image [5].
measures through early findings of sugarcane diseases are Naik and Sivappagari implements a neural network in
implemented. Traditional practice for disease detection, their study in detecting plant diseases [6]. Additional
disease recognition and dealing with these diseases in processing is done to recognize the type of diseases when the
sugarcane is usually done manually. The recent development system detects an infected leaf. Genetic algorithm and SVM
of automated systems to do these tasks for the users will is used to optimize the loss and classify the type of disease.
prevent inconveniences and reduce the time consumed in A study by Olmschenk et al. introduces the Generative
doing it manually. Advancement of the internet in the Adversarial Networks (GAN) using semi-supervise learning
computer vision field is possible to provide an impactful to classify images [7]. The discriminator is transformed into
solution to this problem. multi-class classifier and the generator is used for the
There are several methods have been used to counteract training of discriminator. Paper [8] used transfer learning for
problems with plant diseases and one of these methods is

978-1-7281-4082-7 /19/$31.00 ©2019 IEEE

Authorized licensed use limited to: University of Melbourne. Downloaded on June 23,2020 at 04:39:01 UTC from IEEE Xplore. Restrictions apply.
content-based image recommendation and incorporates Convolutional Neural Network is a supervised deep
neural networks using Inception-V3 model. learning technique that plays an innovative push on various
computer vision and image-based applications. CNN is
Hinton and other scientists [9] introduces a deep neural commonly applied to face recognition, object detection,
architectures in obtaining complex attributes. Their method
image classification, etc. The components of CNN model are
was able to extract highly abstract features that can preserve presented in Fig. 1. The usages of these components vary
the maximum information of raw data thus complicated depending on the network architecture.
correlation can be condensed to fewer elements. Alex
Krizhevsky [10] introduces ImageNet as an improvement in Three CNN architectures were analyzed in this study to
deep learning methods using Deep Convolutional Neural identify diseases from sugarcane plant based on the leaf
Networks in their work and improve the performance of images were the following: (i) StridedNet [22], (ii) LeNet
ConvNets. The ConvNet has surpassed the capability of [23], and (iii), VGG [24]. The three models also apply Keras
humans in certain parts. The research of Alexey Dosovitskiy and OpenCV for machine learning computational
which turn out to be a state-of-the-art result for unsupervised framework.
feature learning implements a ConvNets with data
augmentation [11]. A. Convolution Layer
Other studies suggest a model trained with a server-side This layer convolves the filter of an input image to
component integrated in an intelligent device that recognizes produce an output by dividing it to small blocks, performs
diseases of a captured leaf image over a camera of several an element-wise multiplication and summation of all the
plants [12] [13]. Moreover, datasets could be enlarged by element-wise products [24]. A sequence of mathematical
adding aerial photographs on wider lands [14]. operation of convolution is explained in equation (1).
𝐹𝑙𝑘 = (𝐼𝑥,𝑦 ∗ 𝐾𝑙𝑘 ) (1)
III. CONVOLUTIONAL NEURAL NETWORK The image input is denoted by , x,y shows spatial
Deep Learning engages in developing artificial human
intelligence and automated systems. It uses the processor or locality where represents the Ith convolutional kernel of
the embedded video processor of the computer to control the kth layer. The division of an image into small blocks
large number of neurons in a neural network which is helps in extracting locally correlated pixel values. This
characterized as a single node [15], [16], [17]. locally aggregated information is also known as feature
motifs. A different set of features is extracted by sliding
With the popularity of deep learning, it is used in several
applications in the field of agriculture like detection of convolutional kernel on the image with the same set of
diseases [18], naming and plant categorization [19], fruit weights [1].
grading of images [20]. Images captured from a camera or
attached on a robot [21]. Mohanty et al classifies diseases of
leaves using deep learning [19] while Amara et al introduces B. Activation Layer
LeNet, AlexNet, GoogleNet models in their study [21]. The This layer helps in learning a complex image pattern and
paper of Pawara et al. compares VGGNet and AlexNet serves as a decision function. It accelerates the learning
models in classifying plant diseases [16]. In their study, process by selecting the suitable activation function. The
VGG achieved the best result while classifying 17,548 image activation function for the convolved feature map is defined
datasets with an accuracy of 99.53%. in equation (2).
𝑇𝑙𝑘 = 𝑓𝐴 (𝐹𝑙𝑘 ) (2)
C. Pooling Layer
Pooling layers are used for the downsampling to lower
neuron size and reduce overfitting. A downsampling layer is
applied after the activation layer to reduce the spatial
dimension with no change in-depth. A filter of size 2x2 is
applied to the pooling layer to produce an output based on
the type of pooling although filters with different sizes may
also use. The pooling layer reduces the size of the feature
map which reduces the number of parameters and weighs,
decreases the training time and cuts the rate of computation
Fig. 1. CNN Architecture and in general, it controls overfitting [20].
Artificial neural networks are mathematical simulations Overfitting is a state where the model achieves a
that imitate the human brain function through their neurons maximum of 100% on the training dataset but an average of
and synapses. The crucial feature is the ability of ANN to be 50% on the test data. This can be overcome by activating a
trained throughout the supervised learning process. CNN [4] random set of dropped out setting the value to 0. Dropout is
is driven generally on applications with repeated patterns in a function that increases simplification by learning
diverse areas like image recognition or image detection. numerous representations of patterns. In this study, a max
ANNs key characteristic is that, with the procedure used in pooling filter with stride 2 has been applied. ReLU and max
the layering, it reduces the required number of neurons in pooling are applied to get a reduced dimension feature map
contrast to traditional feedforward neural networks. [21]. Max pooling is done by selecting the largest value in

Authorized licensed use limited to: University of Melbourne. Downloaded on June 23,2020 at 04:39:01 UTC from IEEE Xplore. Restrictions apply.
the sub-windows and this value is transferred to in a new B. Pre-processing of Images
matrix. Preprocessing includes a reduction in image size to a
specified size and cropping the image to its region of
D. Fully Connected Layer interest. It enhances and process the needed color scale of
the image. Compared to a segmented and grayscale type of
This layer identifies very high-level features that highly
image the RGB images gives a greater presentation. The
correlate to an object or class. Input to a fully connected
proposed work uses a 64x64 resolution of colored images
layer is a set of features for classifying images without
for processing.
considering the spatial structure of the images Fully
connected layers are used to calculate the class probabilities
or scores and its result can be the input of the classifier. A C. Feature Extraction
well-known and used classifier is a softmax classifier. The convolutional layers extract features from the
Recognition and classification are implemented in this layer. resized images. CNN learns weights and biases of different
feature maps to specific feature extractors compared to
IV. PROPOSED METHODOLOGY generally crafted featured extractors like SIFT, Gabor filter,
etc. The Rectified non-linear activation function (ReLU)
The workflow diagram showing the experimental design applied after convolution which presents the non-linearity to
of CNN to predict whether the sugarcane plant is infected or CNN. Different types of pooling are max pooling [4] and
not by monitoring the leaf images presented in Fig. 2. average pooling which decreases the dimension of the
Different stages in processing include the following: extracted features. The convolution and pooling together act
Start as a filter to produce features. The extracted features
generally require convolution and pooling.
Build dataset
forexperimentation D. Classification
Create the model to be use Three blocks namely, convolutional, activation and max-
pooling layer followed by fully connected layers and
Train and test the network with softmax activation are used in this architecture. For feature
parameters initialized extraction it uses convolutional and pooling layers. The
classification of classifying the sugarcane leaves if it is
Save the model created infected with the disease or not is done in this layer.

E. Compile and Train the Model

Calculate the performance using
accuracy, loss and size metrics A ratio of 80:20 is created for training and testing set of
datasets [20] [24]. 80% used for training set and 20% used
for testing set.
Fine tuning of
Achieve the model varying the
highest iterations and V. EXPERIMENTAL SETTINGS
accuracy and No
dropout parameters To increase the dataset, data augmentation has been used
less error?
at random by rotating the images on 25 degrees, horizontal
Yes flipping, vertical and horizontal shifting of images. The
Prediction is done on the highest optimization was carried out using Adam optimizer with
accuracy obtained categorical cross-entropy as the loss function. It trained 40
epochs with 32 as the batch size for the three models. The
End initial learning rate has been set to 0.01 and it is reduced by a
Fig. 2. Workflow diagram for prediction
factor of 0.3 on a plateau where the loss stops decreasing.
Early stopping has also been used in order to monitor the
A. Image Dataset Acquisition validation loss and stop the training process once it increases.
Dataset is captured manually with different digital All the experiments were performed on Dell Inspiron 14-
camera models, enhances and segments it if required and 3476 with the 8th generation Intel Core i5 processor and 16
saves it in a folder distinguishing the different diseased GB memory.
images and healthy images for sugarcane plant leaves.
Stored images can be either colored, grayscales or VI. RESULTS AND ANALYSIS
segmented images. The acquired dataset consists of 14,725 The highest validation accuracy of 95.40% obtained over
images belonging to seven different classes. The identified 40 epochs of training using VGGNet model. LeNet model
dataset includes images of important types of sugarcane leaf achieves the rate of 93.65%, and 90.10% was achieved by
diseases that can possibly affect the plant's growth. Each the StridedNet model. The plot in Fig. 3 displays the LeNet
captured image uses RGB color and stored in JPG format or model visualization of accuracy and loss. Fig. 4 presents the
PNG format. visualization of accuracy and loss on StridedNet model and
Fig. 5 displays the VGGNet model accuracy and loss.

Authorized licensed use limited to: University of Melbourne. Downloaded on June 23,2020 at 04:39:01 UTC from IEEE Xplore. Restrictions apply.
 The results show that the three models performs well on
the classification of sugarcane leaf diseases. Fig. 6 shows an
accuracy rate of 53.59% sensing and recognizing as a
sugarcane plant infected with 45.47% rust disease using the
StridedNet model. Fig. 7 shows the VGG model output with
an evaluation result of 100% distinguishing a sugarcane plant
and it is infected with a rust disease with an accuracy rate of
100%. While Fig. 8 shows a LeNet model output with an
evaluation result of an image recognized as 91.64%
sugarcane plant infected with 3.68% rust disease.

Fig. 5. Shows the VGGNet model visualization of accuracy and loss against
epochs.

Fig. 3. LeNet model plot of accuracy and loss against epochs.

Fig. 6. Shows the StridedNet model evaluation result of an image

recognized as 53.59% sugarcane plant and rust infected disease of 45.47%.

Fig. 4. Plots of accuracy and loss against epochs on StridedNet model.

Fig. 7. Shows the evaluation result of an image recognized as 100%

sugarcane plant infected with 100% rust disease using the VGGNet model.

Authorized licensed use limited to: University of Melbourne. Downloaded on June 23,2020 at 04:39:01 UTC from IEEE Xplore. Restrictions apply.
 [5] 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.
[6] Naik, M.R., Sivappagari, C., “Plant Leaf and Disease Detection by
Using HSV Features and SVM,” IJESC, Volume 6 Issue No.12, 2016.
[7] 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.
[8] Surbhi Jain, Joydip Dhar, ”Image-based search engine using deep
learning”, 2017 Tenth International Conference on Contemporary
Computing (IC3), August 2017
[9] G. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
deep belief nets, Neural Computation, vol.18, no.7, pp. 1527-1554,
2006.
[10] Krizhevsky A, Sutskever I, Hinton G E. “Imagenet classification with
Fig. 8. Shows the evaluation result of an image recognized as 91.64%
sugarcane plant infected with 3.68% rust disease using the LeNet model. deep convolutional neural networks,” Advances in Neural
Information Processing Systems, 2012.
[11] Dosovitskiy A, Springenberg J T, Riedmiller M, et al.
VII. CONCLUSION “Discriminative Unsupervised Feature Learning with Convolutional
Neural Networks,” Eprint Arxiv1406, 2014.
Three deep learning models applied in this study to
[12] Y. Tian, P. Zheng, and R. Shi, "The Detection System for Greenhouse
detect and identify sugarcane leaf diseases. The architecture Tomato Disease Degree Based on Android Platform," in 2016 3rd
used in the study compares LeNet, StridedNet, and VGGNet International Conference on Information Science and Control
models. It also verifies which model has the utmost capacity Engineering (ICISCE), Beijing, 2016.
in categorizing and identifying the sugarcane leaf diseases. [13] A. K. Hase, P. S. Aher and S. K. Hase, "Detection, categorization,
and suggestion to cure infected plants of tomato and grapes by using
The VGGNet model achieves the highest accuracy rate on OpenCV framework for Android environment," in 2017 2nd
identifying sugarcane leaf diseases while stridedNet International Conference for Convergence in Technology (I2CT),
achieves the lowest rate. For the future work, an increase Mumbai, 2017.
number of datasets to validate and increase the learning [14] S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D.
rates and optimizers may be used for the models above- Stefanovic, "Deep Neural Networks Based Recognition of Plant
Diseases by Leaf Image Classification," Computational Intelligence
mentioned. Also includes the experimentation of a newer and Neuroscience, 2016.
and different architectures for a better performance of the [15] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T.
model on the training set. Thus, this study can be used as a Weyand, M. Andreetto, and H. Adam, “MobileNets: Efficient
decision tool to assist and provide farmers the familiarity in Convolutional Neural Networks for Mobile Vision Applications,”
identifying and recognizing the sugarcane diseases that can 2017.
be found in the study. [16] K. P. Ferentinos, “Deep learning models for plant disease detection
and diagnosis,” Comput. Electron. Agric., vol. 145, no. September
2017, pp. 311–318, 2018.
[17] A. Kamilaris and F. X. Prenafeta-Boldú, “Deep learning in
ACKNOWLEDGMENT agriculture: A survey,” Comput. Electron. Agric., vol. 147, no. July
Aside from God Almighty, the researchers would like to 2017, pp. 70–90, 2018.
express their gratitude to their families, relatives, and loved [18] M. Brahimi, K. Boukhalfa, and A. Moussaoui, “Deep Learning for
Tomato Diseases: Classification and Symptoms Visualization,”
ones whose involvement made a significant impact on the Application of Artificial Intelligence, vol. 31, no. 4, pp. 299–315,
successful completion of this project. 2017
[19] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using Deep Learning
for Image-Based Plant Disease Detection,” Frontier Plant Science,
vol. 7, no. September, pp. 1–10, 2016
[20] S. Militante, “Fruit Grading of Garcinia Binucao (Batuan) using
REFERENCES Image Processing”, International Journal of Recent Technology and
Engineering (IJRTE), vol. 8 issue 2, pp. 1829- 1832, 2019
[21] J. Amara, B. Bouaziz, and A. Algergawy, “A Deep Learning-based
[1] LeCun, Y., Bengio, Y., Hinton, G., 2015,” Deep learning”, Nature Approach for Banana Leaf Diseases Classification”, BTW, pp. 79–88,
521, 436–444. http://dx.doi.org/10.1038/nature14539. 2017.
[2] Carranza-Rojas, J., Goeau, H., Bonnet, P., Mata-Montero, E., Joly, [22] Simonyan, K., Zisserman, A. 2014. “Very deep convolutional
A.,” Going deeper in the automated identification of Herbarium networks for large-scale image recognition”. arXiv:1409.1556.
specimens”, BMC Evol. Biol, 2017.
[23] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., et
http://dx.doi.org/10.1186/s12862-017-1014-z.
al. 2015. Going deeper with convolutions. Proc. of the IEEE
[3] Lee, S.H., Chan, C.S., Wilkin, P., Remagnino, P., “Deep-plant: Plant Conference on Computer Vision and Pattern Recognition.
identification with convolutional neural networks”, 2015 IEEE Intl
[24] He K., Zhang X., Ren S., and Sun, J. “Delving Deep into Rectifiers:
Conf. on Image Processing, pp. 452–456, 2015.
Surpassing Human-Level Performance on ImageNet Classification”,
[4] LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., “Gradient-based arXiv:1502.01852v1.
learning applied to document recognition”, Proceedings IEEE 86 (11),
[25] Bouvrie, J., ”Introduction Notes on Convolutional Neural
2278–2324, 1998.
Networks”.,(2006).doi:http://dx.doi.org/10.1016/j.protcy.2014.09.007

Authorized licensed use limited to: University of Melbourne. Downloaded on June 23,2020 at 04:39:01 UTC from IEEE Xplore. Restrictions apply.