See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/340565584

Identification and recognition of rice diseases and pests using convolutional

neural networks

Article in Biosystems Engineering · June 2020

DOI: 10.1016/j.biosystemseng.2020.03.020

CITATIONS READS

437 1,688

7 authors, including:

Rafeed Rahman Preetom Saha Arko

Bangladesh University of Engineering and Technology Bangladesh University of Engineering and Technology
25 PUBLICATIONS 548 CITATIONS 2 PUBLICATIONS 437 CITATIONS

SEE PROFILE SEE PROFILE

Mohammed Eunus Ali Mohammad Ashik Iqbal Khan

Bangladesh University of Engineering and Technology Bangladesh Rice Research Institute
144 PUBLICATIONS 1,921 CITATIONS 103 PUBLICATIONS 1,252 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mohammad Ashik Iqbal Khan on 11 May 2022.

The user has requested enhancement of the downloaded file.

 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/issn/15375110

Research Paper

Identification and recognition of rice diseases and

pests using convolutional neural networks

Chowdhury R. Rahman c,*, Preetom S. Arko a, Mohammed E. Ali a,

Mohammad A. Iqbal Khan b, Sajid H. Apon a, Farzana Nowrin b,
Abu Wasif a
a
Bangladesh University of Engineering and Technology, Dhaka, Bangladesh
b
Bangladesh Rice Research Institute, Gazipur, Bangladesh
c
United International University, Dhaka, Bangladesh

article info abstract

Article history: Accurate and timely detection of diseases and pests in rice plants can help farmers in
Received 14 May 2019 applying timely treatment on the plants and thereby can reduce the economic losses
Received in revised form substantially. Recent developments in deep learning-based convolutional neural networks
18 March 2020 (CNN) have greatly improved image classification accuracy. Being motivated by the success
Accepted 21 March 2020 of CNNs in image classification, deep learning-based approaches have been developed in
this paper for detecting diseases and pests from rice plant images. The contribution of this
paper is two fold: (i) State-of-the-art large scale architectures such as VGG16 and Incep-
Keywords: tionV3 have been adopted and fine tuned for detecting and recognising rice diseases and
Rice disease pests. Experimental results show the effectiveness of these models with real datasets. (ii)
Pest Since large scale architectures are not suitable for mobile devices, a two-stage small CNN
Convolutional neural network architecture has been proposed, and compared with the state-of-the-art memory efficient
Dataset CNN architectures such as MobileNet, NasNet Mobile and SqueezeNet. Experimental re-
Memory efficient sults show that the proposed architecture can achieve the desired accuracy of 93.3% with a
Two stage training significantly reduced model size (e.g., 99% smaller than VGG16).
© 2020 IAgrE. Published by Elsevier Ltd. All rights reserved.

because of 10 major diseases of rice in Bangladesh

1. Introduction (Mahmud, Hossain, & Ahmad, 2016). Timely detection of rice
plant diseases and pests is one of the major challenges in
Rice occupies about 70 percent of the gross crop area and agriculture. Hence, there is a need for automatic rice disease
accounts for 93 percent of total cereal production in detection using readily available mobile devices in rural
Bangladesh (Coelli, Rahman, & Thirtle, 2002). Rice also en- areas.
sures food security of over half the world population (Calpe, Deep learning techniques have shown great promise in
2002). Researchers have observed 10e15% average yield loss image classification. In recent years, these techniques have

* Corresponding author.
E-mail address: [email protected] (C.R. Rahman).
https://doi.org/10.1016/j.biosystemseng.2020.03.020
1537-5110/© 2020 IAgrE. Published by Elsevier Ltd. All rights reserved.
 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0 113

been used to analyse diseases of tea (Karmokar, Ullah, Sid- used. These studies played an important role for automatic
diquee, & Alam, 2015), apple (Wang, Sun, & Wang, 2017), and accurate recognition and classification of plant diseases.
tomato (Fuentes, Yoon, Kim, & Park, 2017), grapevine, But their focus was not on modifying the training method for
peach, and pear (Sladojevic, Arsenovic, Anderla, Culibrk, & the models that they had constructed and used. Moreover,
Ste-fanovic, 2016). Babu and Rao (2007) proposed a feed they did not consider the impact of the large number of pa-
forward back propagation neural network from scratch in rameters of these high performing CNN models in real life
order to detect the species of plant from leaf images. Neural mobile application deployment.
network ensemble (NNE) was used by Karmokar, Ullah, In this research, two state-of-the-art CNN architectures,
Siddiquee, and Alam (2015) to recognise five different dis- VGG16 and InceptionV3, have been tested in various set-
eases of tea plant from leaves. Bhagawati, Bhagawati, tings. Fine tuning, transfer learning and training from
Singh, Nongthombam, Sarmah, and Bhagawati (2015) scratch have been implemented to assess their perfor-
trained a neural network with weather parameters such mance. In both the architectures, fine tuning the model
as temperature, relative humidity, rainfall and wind speed while training has shown the best performance. Though
to forecast rice blast disease. Mohanty, Hughes, and Salathe these deep learning-based architectures perform well in
(2016) used deep CNN to detect disease from leaves using practice, a major limitation of these architectures is that
54,306 images of 14 crop species representing 26 diseases, they have a large number of parameters, a problem similar
while Sladojevic, Arsenovic, Anderla, Culibrk, and to previously conducted researches. For example, there are
Stefanovic (2016) used CaffeNet model to recognize 13 about 138 million parameters in VGG16 (Simonyan &
different types of plant diseases. Wang, Sun, and Wang Zisserman, 2014). In remote areas of developing countries,
(2017) worked on detecting four severity stages of apple farmers do not have internet connectivity or have slow
black rot disease using PlantVillage dataset. They used CNN internet speed. So, a mobile application capable of running
architectures with different depths and implemented two CNN-based model offline is needed for rice disease and pest
different training methods on each of them. A real time detection. So, a memory efficient CNN model with reason-
tomato plant disease detector was built using deep learning ably good classification accuracy is required. Since the
by Fuentes, Yoon, Kim, and Park (2017). Brahimi, Boukhalfa, reduction of the number of parameters in a CNN model
and Moussaoui (2017) used fine-tuned AlexNet and Goo- reduces its learning capability, one needs to make a trade-
gleNet to detect nine diseases of tomatoes. Cruz, Luvisi, De off between memory requirement and classification accu-
Bellis, and Ampatzidis (2017) injected some texture and racy to build such a model.
shape features to the fully connected layers placed after the To address the above issue, in this paper, a new training
convolutional layers so that the model can detect Olive method called two stage training has been proposed. A CNN
Quick Decline Syndrome effectively from the limited data- architecture, named Simple CNN, has been proposed which
set. Instead of resizing images to a smaller size and training achieves high accuracy leveraging two stage training in spite
a model end-to-end, DeChant, Wiesner-Hanks, Chen, of its small number of parameters. Experimental study shows
Stewart, Yosinski, and Gore (2017) used a three stage ar- that the proposed Simple CNN model outperforms state-of-
chitecture (consisting of multiple CNNs) and trained the the-art memory efficient CNN architectures such as Mobile-
stage-one model on full scaled images by dividing a single Net, NasNet Mobile and SqueezeNet in recognising rice plant
image into many smaller images. Barbedo (2018) used diseases and pests.
transfer learning on GoogleNet to detect 56 diseases All training and validation have been conducted on a rice
infecting 12 plant species. Using a dataset of 87,848 images dataset collected in real life scenario as part of this research. A
of leaves captured both in laboratory and in the field, rice disease may show different symptoms based on various
Ferentinos (2018) worked with 58 classes containing 25 weather and soil conditions. Similarly, pest attack can show
different plants. Liu, Zhang, He, and Li (2018) built a CNN different symptoms at different stages of an attack. Moreover,
combining the ideas of AlexNet and GoogLeNet to detect the diseases and pests can occur at any part of the plant which
four diseases of apple. Images of individual lesions and include leaf, stem and grain. Images can also be of heteroge-
spots instead of image of whole leaf were used by Barbedo neous background. This research addresses all these issues
(2019) for identifying 79 diseases of 14 plant species. Few while collecting data. This paper focuses on recognizing eight
researches have also been conducted on rice disease clas- different rice plant diseases and pests that occur at different
sification (Atole & Park, 2018; Lu, Yi, Zeng, Liu, & Zhang, times of the year at Bangladesh Rice Research Institute (BRRI).
2017). Lu et al. (2017) conducted a study on detecting 10 This work also includes a ninth class for non-diseased rice
different rice plant diseases using a small handmade CNN plant recognition.
architecture, inspired by older deep learning frameworks In summary, this paper makes two important contribu-
such as LeNet-5 and AlexNet, using 500 images. Atole and tions in rice disease and pest detection. First, state-of-the-art
Park (2018) used AlexNet (large architecture) to distinguish large scale deep learning frameworks have been tested to
among three classes - normal rice plant, diseased rice plant investigate the effectiveness of these architectures in rice
and snail infested rice plant using 227 images. plant disease and pest identification from images collected
The research mentioned above mainly focused on accurate from real-life environments. Second, a novel two-stage
plant disease recognition and classification. For this purpose, training based light-weight CNN has been proposed that is
they implemented various types of CNN architectures such as highly effective for mobile device based rice plant disease and
AlexNet, GoogLeNet, LeNet-5 and so on. In some studies, pest detection. This can be an effective tool for farmers in
ensemble of multiple neural network architectures have been remote environment.
114 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0

collected in real life scenario with heterogeneous backgrounds

2. Materials and methods from December, 2017 to June, 2018 for a total of seven months.
The image collection has been performed in a range of weather
2.1. Data collection conditions - in winter, in summer and in overcast condition in
order to get as fully representative a set of images as possible.
Rice diseases and pests occur in different parts of the rice Four different types of camera have been used in capturing the
plant. Their occurrence depends on many factors such as images. These steps increase the robustness of our model. This
temperature, humidity, rainfall, variety of rice plants, season, work encompasses a total of five classes of diseases, three
nutrition, etc. An extensive exercise was undertaken to collect classes of pests and one class of healthy plant and others - a
total 1426 images of rice diseases and pests from paddy fields total of nine classes. The class names along with the number of
of Bangladesh Rice Research Institute (BRRI). Images have been images collected for each class are shown in Table 1. Note that
Sheath Blight, Sheath Rot and their simultaneous occurrence
have been considered in the same class, because their treat-
ment method and place of occurrence are the same.
Table 1 e Image collection of different classes.
Symptoms of different diseases and pests are seen in
Class Name No. of Collected Images different parts such as leaf, stem and grain of the rice plant.
False Smut 93 Bacterial Leaf Blight disease, Brown Spot disease, Brown Plant
Brown Plant Hopper (BPH) 71 Hopper pest (late stage) and Hispa pest occur on the rice leaf.
Bacterial Leaf Blight (BLB) 138
Sheath Blight disease, Sheath Rot disease and Brown Plant
Neck Blast 286
Hopper pest (early stage) occur on the rice stem. Neck Blast
Stemborer 201
Hispa 73 disease and False Smut disease occur on rice grain. Stemborer
Sheath Blight and/or Sheath Rot 219 pest occurs on both rice stem and rice grain. All these aspects
Brown Spot 111 have been considered while capturing images. To prevent
Others 234 classification models from being confused between dead parts

Fig. 1 e A sample image of each detected class.

 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0 115

and diseased parts of rice plant, images of dead leaf, dead stem of parameters such as VGG16 and InceptionV3. Later the
and dead grain of rice plants have been incorporated into the proposed light-weight two-stage Simple CNN has been tested
dataset. For example, diseases like BLB, Neck Blast and Sheath and compared with three state-of-the-art memory efficient
Blight have similarity with dead leaf, dead grain and dead stem CNN architectures such as MobileNetv2, NasNet Mobile and
of rice plant respectively. Thus images of dead leaf, dead stem SqueezeNet. VGG16 (Simonyan & Zisserman, 2014) is a
and dead grain along with images of healthy rice plant have sequential CNN architecture using 3
3 convolution filters.
been considered in a class that has been named others. Sample After each maxpooling layer, the number of convolution filters
images of each class have been depicted in Fig. 1. gets doubled in VGG16. InceptionV3 (Szegedy et al., 2015) is a
False Smut, Stemborer, Healthy Plant class, Sheath Blight non-sequential CNN architecture consisted of inception
and/or Sheath Rot class show multiple types of symptoms. blocks. In each inception block, convolution filters of various
Early stage symptoms of Hispa and Brown Plant Hopper are dimensions and pooling are used on the input in parallel. The
different from their later stage symptoms. All symptom vari- number of parameters of these five architectures along with
ations of these classes found in the paddy fields of BRRI have simple CNN architecture have been given in Table 3. Three
been covered in this work. These intra-class variations have
been described in Table 2. BLB, Brown Spot and Neck Blast
disease show no considerable intra-class variation around Table 3 e State-of-the-art CNN architectures and their
parameter no.
BRRI area. An illustrative example for Hispa pest has been
given in Fig. 2. CNN Architecture No. of Parameters
VGG16 138 million
2.2. Experimental setup InceptionV3 23.8 million
MobileNetv2 2.3 million
Keras framework with tensorflow back-end has been used to NasNet Mobile 4.3 million
SqueezeNet 0.7 million
train the models. Experiments have been conducted with two
Simple CNN 0.8 million
state-of-the-art CNN architectures containing large numbers

Table 2 e Intra-class variation in some diseases and pests.

Class Name Symptom Variation Sample No.
BPH Early stage of Brown Plant Hopper attack 50
Late stage of Brown Plant Hopper attack 21
False Smut Brown symptom 66
Black symptom 27
Others Healthy green leaf and stem 96
Healthy yellow grain 71
Dead leaf and stem 67
Hispa Visible black pests and white spot on plant leaf 53
No visible pest, intense spot on leaves 20
Stemborer Symptom on grain 180
Symptom on stem 21
Sheath Blight and/or Sheath Rot Black Stem 70
White spots 77
black and white symptom mixed 72

Fig. 2 e Hispa Variations: Image on the left has visible black pests and white spots on the plant leaf which occur during early
stage of Hispa attack. Image on the right has intense spots on leaves with no visible pest occurring during later stage of
Hispa attack.
116 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0

different types of training methods have been implemented 10-fold cross-validation accuracy along with standard
on each of these five architectures. deviation have been used as model performance metric
Baseline training: All randomly initialised architecture since the dataset used in this work does not have any major
layers are trained from scratch. This method of training takes imbalance. Categorical Crossentropy has been used as loss
time to converge. function for all CNN architectures since this work deals
Fine Tuning: The convolution layers of the CNN architectures with multi-class classification. All intermediate layers of
are trained from their pre-trained ImageNet weights, while the the CNN architectures used in this work have relu as acti-
dense layers are trained from randomly initialised weights. vation function while the activation function used in the
Transfer Learning: In this method, the convolution layers last layer is softmax. The hyperparameters used are as
of the CNN architectures are not trained at all. Rather pre- follows: dropout rate of 0.3, learning rate of 0.0001, mini
trained ImageNet weights are kept intact. Only the dense batch size of 64 and number of epochs 100. These values
layers are trained from their randomly initialised weights. have been obtained through hyperparameter tuning using
10-fold cross-validation. Adaptive Moment Estimation
(Adam) optimiser has been used for updating the model
weights.
All the images have been resized to the default image
size of each architecture before working with that archi-
tecture. For example, InceptionV3 requires 299
299
3
pixel size image while VGG16 requires image of pixel size
224
224
3. Random rotation from 15 to 15 , rotations of
multiple of 90 at random, random distortion, shear trans-
formation, vertical flip, horizontal flip, skewing and in-
tensity transformation have been used as part of the data
augmentation process. Every augmented image is the result
of a particular subset of all these transformations, where
rotation type transformations have been assigned high
probability. It is because CNN models in general are not
rotation invariant. In this way, 10 augmented images from
every original image have been created. Random choice of
the subset of the transformations helps augment an original
image in a heterogeneous way.
A remote Red Hat Enterprise Linux server of RMIT Uni-
versity has been used for carrying out the experiments. The
configuration of the server includes 56 CPUs, 503 GB RAM, 1
petabyte of user specific storage and two NVIDIA Tesla P100-
PCIE GPUs each of 16 GB.

Table 4 e Quantitative performance of different state-of-

the-art CNN architectures obtained from 10 fold cross
validation (best accuracy of each architecture has been
mentioned in bold character).
CNN Training Method Mean Standard
Architecture Used Validation Deviation
Accuracy
VGG16 Baseline training 89.19% 10.28
Transfer Learning 86.52% 5.37
Fine Tuning 97.12% 2.23
InceptionV3 Baseline training 91.17% 3.96
Transfer Learning 72.09% 7.96
Fine Tuning 96.37% 3.9
MobileNetv2 Baseline training 78.84% 7.38
Transfer Learning 77.52% 8.56
Fine Tuning 96.12% 3.08
NasNet Mobile Baseline training 79.98% 6.96
Transfer Learning 78.21% 8.09
Fine Tuning 96.95% 3.35
SqueezeNet v1.1 Baseline training 74.88% 8.18
Transfer Learning 42.76% 9.12
Fine Tuning 92.5% 3.75
Fig. 3 e Simple CNN architecture. Simple CNN Two Stage Training 94.33% 0.96
 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0 117

2.3. Proposed simple CNN model into 17 classes by keeping all intra-class variations in separate
classes. These variations have been shown in detail in Table 2.
Apart from adapting state-of-the-art CNN models, a memory For example, others class is divided into three separate classes.
efficient two-stage small CNN architecture, namely Simple Thus, the model is trained with this 17 class dataset. As a result,
CNN shown in Fig. 3 has been constructed from scratch the final dense layer of the model has 17 nodes with softmax
inspired by the sequential nature of VGG16. Fine tuned VGG16 activation function. In stage two, the original dataset of nine
provides excellent result on rice dataset. This Simple CNN ar- classes is used. All layer weights of simple CNN architecture
chitecture has only 0.8 million parameters compared to 138 obtained from stage one are kept intact except for the topmost
million parameters of VGG16. All five of the state-of-the-art layer. This dense layer consisting of 17 nodes is replaced with a
CNN architectures trained and tested in this work have dense layer consisting of nine nodes with softmax activation
shown the best result when fine tuning has been used (see function. Such measures are taken, because stage two training
Section 3). Two stage training is inspired from fine tuning. In data are divided into the nine original classes. Now all the
stage one, the entire image dataset of nine classes are divided layers of the Simple CNN architecture are trained using this

Fig. 4 e Confusion matrix generated using simple CNN on entire dataset.

Fig. 5 e First convolution layer output of simple CNN.

118 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0

Fig. 6 e Last convolution layer output of simple CNN.

 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0 119

nine class dataset which are initialised with the pre-trained The confusion matrix generated from the application of
weights obtained from stage one training. Experiments show Simple CNN on the entire dataset (training and validation set
the effectiveness of applying this method. combined) has been shown in Fig. 4. 4.3% of the False Smut
images existing in the dataset have been misclassified, which
is the highest among all present classes of this work. False
3. Results and discussion Smut symptoms cover a small portion of the entire image
(captured with heterogeneous backgrounds) compared to
Experimental results obtained from 10-fold cross-validation other existing pest and disease images.
for the five state-of-the-art CNN architectures along with The first convolution layer outputs of Simple CNN have
Simple CNN have been shown in Table 4. Transfer learning been shown in Fig. 5. The three rows from top to bottom
gives the worst result in all five of the models. For the smallest represent output for Fig. 1a, d and i respectively, while the
architecture SqueezeNet, it is below 50%. Rice disease and pest left and right columns represent output for stage one and
images are different from images of ImageNet dataset. Hence, stage two of the Simple CNN model respectively. Each of the
the freezing of convolution layer weights disrupts learning of six images contains 16 two dimensional mini images of size
the CNN architectures. Although baseline training, also 222
222 (first convolution layer outputs a matrix of size
known as training from scratch, does better than transfer 222
222
16). The last convolution layer outputs of Simple
learning, the results are still not satisfactory. For the three CNN have been shown in Fig. 6 in a similar setting. Each of
small models, the accuracy is less than 80%. The standard the six images of Fig. 6 contains 64 two dimensional mini
deviation of validation accuracy is also large which denotes images of size 10
10 (last convolution layer outputs a
low precision. This shows that the models are not able to learn matrix of size 10
10
64). The first layer maintains the
the distinguishing features of the classes when trained from regional features of the input image, although some of the
randomly initialised weights. More training data may solve filters are blank (not activated). The activations retain
this problem. Fine tuning gives the best result in all cases. It almost all of the information present in the input image.
also ensures high precision (lowest standard deviation). It The last convolution layer outputs are visually less under-
means that for the state-of-the-art CNN architectures to standable. This representation depicts less information
achieve good accuracy on rice dataset, training on the large about the visual contents of the input image. Rather this
ImageNet dataset is necessary prior to training on the rice layer attempts to present information related to the class of
dataset. Fine-tuned VGG16 achieves the best accuracy of the image. The intermediate outputs for different classes
97.12%. The Simple CNN architecture utilising two stage are visually different for different classes. An interesting
training achieves comparable accuracy and the highest pre- aspect can be observed in Fig. 6. The last convolution layer
cision without any prior training on ImageNet dataset. Rather, output for stage one model carries considerably fewer blank
this model is trained from scratch. two dimensional mini images than does the stage two
From Table 3, it is evident that the Simple CNN model model. This shows the capability of the stage two model in
has a small number of parameters even when compared to terms of learning with less features. This helps Simple CNN
small state-of-the-art CNN architectures such as MobileNet achieve good accuracy and high precision after stage two
and NasNet Mobile. The number of parameters of Squee- training.
zeNet (the smallest of the five state-of-the-art CNN archi-
tectures in terms of parameter number) is comparable to the
parameter number of the Simple CNN. Sequential models 4. Conclusion
like VGG16 need depth in order to achieve good perfor-
mance, hence they have large numbers of parameters. This work has the following contributions:
Although Simple CNN is a sequential model with a low
number of parameters, its high accuracy (comparable to the  A dataset of rice diseases and pests consisting of 1426
other state-of-the-art CNN architectures) proves the effec- images has been collected in real life scenario which
tiveness of two stage training. Future research should aim at cover eight classes of rice disease and pest. This dataset
building miniature version of memory efficient non- is expected to facilitate further research on rice diseases
sequential state-of-the-art CNN architectures such as and pests. The dataset is available here (https://drive.
InceptionV3, DenseNet and Xception. These architectures google.com/open?id¼1ewBesJcguriVTX8sRJseCDbXAF_
should be able to achieve similar excellent results with even T4akK). The details of the dataset have been described
smaller number of parameters. in Subsection 2.1.
A major limitation of two stage training is that the entire  Three different training methods have been imple-
dataset has to be divided manually into symptom classes. In mented on two state-of-the-art large CNN architectures
a large dataset, detecting all the major intra-class variations and three state-of-the-art small CNN architectures
is a labour intensive process. There is a great chance of (targeted towards mobile applications) on the rice
missing some symptom variations. Minor variety within a dataset. Fine tuning from pre-trained ImageNet weight
particular class maybe misinterpreted as separate symptoms. always provided the best result for all five architectures.
One possible solution could be to use high dimensional  A new concept of two stage training derived from the
clustering algorithms on each class-specific image set sepa- concept of fine tuning has been introduced which en-
rately in order to automate this process of identifying intra- ables proposed Simple CNN architecture of this work to
class variations. perform well in real life scenario.
120 b i o s y s t e m s e n g i n e e r i n g 1 9 4 ( 2 0 2 0 ) 1 1 2 e1 2 0

In future, incorporating location, weather and soil data Brahimi, M., Boukhalfa, K., & Moussaoui, A. (2017). Deep learning
along with the image of the diseased part of the plant can be for tomato diseases: Classification and symptoms
investigated for a comprehensive and automated plant dis- visualization. Applied Artificial Intelligence, 31(4), 299e315.
Calpe, C. (2002). Rice in world trade, part ii. status of the world rice
ease detection mechanism. Segmentation or object detection-
market. In Proceedings of the 20 th session of the international rice
based algorithm can be implemented with a view to detecting commission.
and classifying rice diseases and pests more effectively in the Coelli, T., Rahman, S., & Thirtle, C. (2002). Technical, allocative,
presence of heterogeneous background. cost and scale efficiencies in Bangladesh rice cultivation: A
non-parametric approach. Journal of Agricultural Economics,
53(3), 607e626.
Declaration of Competing Interest Cruz, A. C., Luvisi, A., De Bellis, L., & Ampatzidis, Y. (2017). X-fido:
An effective application for detecting olive quick decline
syndrome with deep learning and data fusion. Frontiers of Plant
None declared. Science, 8, 1741.
DeChant, C., Wiesner-Hanks, T., Chen, S., Stewart, E. L.,
Yosinski, J., Gore, M. A., et al. (2017). Automated identification
of northern leaf blight-infected maize plants from field
Acknowledgments imagery using deep learning. Phytopathology, 107(11),
1426e1432.
We thank Bangladesh ICT division for funding this research. Ferentinos, K. P. (2018). Deep learning models for plant disease
We also thank the authority of Bangladesh Rice Research detection and diagnosis. Computers and Electronics in
Institute for supporting the research by providing us with the Agriculture, 145, 311e318.
opportunity of collecting images of rice plant diseases in real Fuentes, A., Yoon, S., Kim, S. C., & Park, D. S. (2017). A robust
deep-learning-based detector for real-time tomato plant
life scenario. We also acknowledge the help of RMIT Univer-
diseases and pests recognition. Sensors, 17(9), 2022.
sity who gave us the opportunity to use their GPU server. The Karmokar, B. C., Ullah, M. S., Siddiquee, M. K., & Alam, K. M. R.
authors would like to thank anonymous reviewers for (2015). Tea leaf diseases recognition using neural network
providing suggestions which improved the quality of this ensemble. International Journal of Computer Application, 114(17).
article. Liu, B., Zhang, Y., He, D., & Li, Y. (2018). Identification of apple leaf
diseases based on deep convolutional neural networks.
Symmetry, 10(1), 11.
references Lu, Y., Yi, S., Zeng, N., Liu, Y., & Zhang, Y. (2017). Identification of
rice diseases using deep convolutional neural networks.
Neurocomputing, 267, 378e384.
Mahmud, H., Hossain, I., & Ahmad, M. (2016). Efficacy of bau-
Atole, R. R., & Park, D. (2018). A multiclass deep convolutional biofungicide on brown spot and bacterial leaf blight disease and
neural network classifier for detection of common rice plant vigour index of rice.
anomalies. International Journal of Advanced Computer Science Mohanty, S. P., Hughes, D. P., & Salathe , M. (2016). Using deep
and Applications, 9(1), 6770. learning for image-based plant disease detection. Frontiers of
Babu, M. P., & Rao, B. S. (2007). Leaves recognition using back Plant Science, 7.
propagation neural network- advice for pest and disease Simonyan, K., & Zisserman, A. (2014). Very deep convolutional
control on crops. IndiaKisan. Net: Expert Advisory System. networks for large-scale image recognition. arXiv preprint
Barbedo, J. G. A. (2018). Impact of dataset size and variety on the arXiv:1409.1556.
effectiveness of deep learning and transfer learning for plant Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., &
disease classification. Computers and Electronics in Agriculture, Stefanovic, D. (2016). Deep neural networks based recognition
153, 46e53. of plant diseases by leaf image classification. Computational
Barbedo, J. G. A. (2019). Plant disease identification from Intelligence and Neuroscience.
individual lesions and spots using deep learning. Biosystems Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D.,
Engineering, 180, 96e107. et al. (2015). Going deeper with convolutions. In Proceedings of
Bhagawati, R., Bhagawati, K., Singh, A., Nongthombam, R., the ieee conference on computer vision and pattern recognition (p.
Sarmah, R., & Bhagawati, G. (2015). Artificial neural network 19).
assisted weather based plant disease forecasting system. Wang, G., Sun, Y., & Wang, J. (2017). Automatic image-based plant
International Journal on Recent and Innovation Trends in Computing disease severity estimation using deep learning. Computational
and Communication, 3(6), 4168e4173. Intelligence and Neuroscience.

View publication stats