A classification model based on depthwise separable convolutional neural network to identify rice plant diseases

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 4, August 2022, pp. 3642~3654
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i4.pp3642-3654  3642
Journal homepage: http://ijece.iaescore.com
A classification model based on depthwise separable
convolutional neural network to identify rice plant diseases
Md. Sazzadul Islam Prottasha, Sayed Mohsin Salim Reza
Department of Information and Communication Technology, Bangladesh University of Professionals, Dhaka, Bangladesh
Article Info ABSTRACT
Article history:
Received Jan 17, 2021
Revised Dec 19, 2021
Accepted Jan 25, 2022
Every year a number of rice diseases cause major damage to crop around the
world. Early and accurate prediction of various rice plant diseases has been a
major challenge for farmers and researchers. Recent developments in the
convolutional neural networks (CNNs) have made image processing
techniques more convenient and precise. Motivated from that in this
research, a depthwise separable convolutional neural network based
classification model has been proposed for identifying 12 types of rice plant
diseases. Also, 8 different state-of-the-art convolution neural network model
has been fine-tuned specifically for identifying the rice plant diseases and
their performance has been evaluated. The proposed model performs
considerably well in contrast to existing state-of-the-art CNN architectures.
The experimental analysis indicates that the proposed model can correctly
diagnose rice plant diseases with a validation and testing accuracy of 96.5%
and 95.3% respectively while having a substantially smaller model size.
Keywords:
Agriculture
Convolutional neural network
Deep learning
Image processing
Plant disease
Rice plant diseases
This is an open access article under the CC BY-SA license.
Corresponding Author:
Md. Sazzadul Islam Prottasha
Department of Information and Communication Technology, Bangladesh University of Professionals
Mirpur Cantonment, Dhaka-1216, Bangladesh
Email: 19541026@bup.edu.bd
1. INTRODUCTION
Agricultural science has an enormous effect on the food production system around the world, hence
this field is emerging day by day. Technologies have brought a new dimension to this field. Researchers are
implementing different methodologies and invented different types of seeds, treatments and weeds to
improve the overall crop production. Developments of recent deep learning based image processing methods
have improved the disease classification accuracy significantly. Inspired by that, our research is primarily
focused on the categorization of various rice plant diseases using a deep learning approach. Contributing to
this field has developed a profound interest in us.
There are more than 40 different types of rice plant diseases that can be fatal to the rice plants as
described by Ou [1]. Diseases like rice blast, smut and leaf blight can cause severe damage to rice
production. There exists some other diseases that can be lethal unless necessary measurements are taken
early. Researchers have come up with numerous methodologies and models for the detection of rice plant
diseases over the years. Different kinds of segmentation and feature extraction methods have been
implemented. A multistage convolutional neural network architecture has been presented by Lu et al. [2] that
can identify 10 different rice plant diseases. A total of 500 images has been considered including healthy and
diseased images and trained using the convolutional neural network (CNN) model. The result reported in the
paper shows an accuracy of 95.38% while diagnosing the rice plant diseases. The work was conducted on 10
types of rice plant diseases, however there are only 500 training images meaning only 50 images per disease
class. Based on the minimal quantity of training pictures, it seems doubtful that this model will be viable in

Int J Elec & Comp Eng ISSN: 2088-8708 
A classification model based on depthwise separable convolutional … (Md. Sazzadul Islam Prottasha)
3643
the real life scenarios. The work by Liu et al. [3] proposed a machine learning based along with a CNN based
model for identifying rice false smut images. The proposed CNN architecture inspired from AlexNet and
VGGNet-16 architecture comprises a total of 10 layers which includes 7 convolution layers and 3 fully
connected layers. Experimental result suggests that, the proposed method performed better than AlexNet in
diagnosing the rice false smut disease. Jagan et al. [4] developed a two-stage technique for detecting rice
plant diseases, in which they detected the disease-affected region using Haar-like characteristics and the
AdaBoost classifier in the first stage. After identifying the disease-affected area, they employed scale
invariant feature transform to extract features. Finally, support vector machine and k-nearest neighbor
classifiers were used for the detecting procedure. An optimal deep CNN architecture has been presented
at [5] for diagnosing rice plant diseases. A cloud based infrastructure has been developed where Inception-
ResNet V2 model is used for the feature extraction process while weighted extreme learning machine
performs the disease classification. The work by Ramesh and Vydeki [6] presents an optimized neural
network model to identify 4 types of paddy plant diseases. Initially, the red, green, blue (RGB) images were
converted into hue, saturation, value (HSV) images and analyzing the hue and saturation they extracted
binary images from the hue and saturation difference. This work integrated Jaya optimization algorithm with
the neural network. Jaya algorithm has been used to update the weights of the proposed neural network in
identifying rice diseases. The work by Krishnamoorthy et al. [7] used InceptionResNetV2 model for
detecting multiple paddy leaf diseases. By using transfer learning method and tuning the hyperparameters the
proposed model achieved an accuracy of 95.67%. The method provided by Shrivastava et al. [8] uses the
AlexNet model using transfer learning to detect 3 different rice plant diseases. For an 80-20 train and test
data split, the AlexNet model achieved a classification accuracy of 91.37%. A simple CNN architecture for
diagnosing various rice plant diseases from Bangladesh has been proposed at [9]. The work focuses on
providing a lightweight CNN model by hyper-parameter tuning which can provide admissible accuracy. The
statistical analysis shows that using Adam optimizer their proposed model achieved a validation accuracy of
95.4%. Similar CNN model has been proposed for mango [10], wheat [11], banana [12], apple [13], and
peach [14] classification.
Recent developments in various computer vision models faster R-CNN, YOLOv3, mask R-CNN,
and RetinaNet have significantly improved the accuracy of object detection modules. These models not only
detect the disease but also identifies the exact location of the disease occurrence. Sethy et al. [15] presented a
faster R-CNN object detection method for rice false smut disease. After collecting 50 images of false smut
they used image augmentation methods to increase the images in the dataset. After extracting the features
using the ResNet-50 model, the feature vectors are fed through the fully connected layer to predict the smuts
using bounding boxes. In case of numerous false smut are present, the model occasionally fails to detect all
smuts using bounding boxes. Bari et al. [16] proposed a faster R-CNN based real-time rice leaf disease
diagnosis system. Experimental analysis shows that, the proposed method can localize the disease-affected
portions with higher precision. However, the model requires multiple recurrences to extract all objects inside
a single image which makes the model slower. Considering these computer vision models, we developed a
CNN model based on depthwise separable convolutions to diagnose 12 types of rice plant diseases.
2. METHOD
In this section, the details of the dataset collection process and proposed models are discussed in
appropriate subheadings. Also, the model hyperparameters are discussed here. Figure 1 illustrates the details
of our method.
Figure 1. Method of rice plant disease detection process

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 4, August 2022: 3642-3654
3644
Our experimental study starts with data acquisition followed by data pre-processing. Then the
processed images are fed through the CNN model. The CNN model generates feature maps and eventually
the classification is performed by the fully connected dense layers.
2.1. Dataset collection
We have taken some extensive measurements to collect 1,677 rice plant raw sample images from
different regions of Bangladesh, consisting 12 types of fungal and pest diseases along with healthy plants.
The data were collected under different weather conditions over a total 4-month span. In addition to our
accumulated sample images, we took some more sample images from the dataset stated by Rahman et al.
[17]. The images were taken on different heterogeneous backgrounds under different lighting conditions both
in sunny and overcast environments. The collected types of diseases we contemplated are bacterial blight,
brown plant hopper, brown spot, false smut, hispa, leaf blast, leaf scald, leaf smut, neck blast, sheath blight
rot, stackburn and stemborer as shown in Figure 2(a) to (l). Along with these 12 fungal and pest diseases, we
have also taken images of healthy leaves and panicles to differentiate them from the disease-affected plants.
The dataset can be found here at [18]. Figure 2 shows the collected sample images of our dataset.
The images were captured during different stages of the disease infection considering different
symptoms of the diseases; hence we get a fully representative dataset considering all aspects of the disease.
There are 10 classes in our dataset that contain multiple symptom variations. Table 1 shows the number of
sample images collected for different symptoms of diseases and pests.
Bacterial leaf blight caused by a bacteria called Xanthomonas oryzae pv. oryzae which attacks the
leaves of the rice plant causing it to dry out [19]. Brown spot is a fungal disease caused by Cochliobolus
miyabeanus. This disease mostly occurs at the protective sheath covering the leaf of the rice plants, leaves
and spikelets [20]. False smut is a rice grain disease caused by a plant pathogen named Ustilaginoidea virens
which transforms the individual rice grains into yellow fruiting bodies [21]. In some cases, the smuts turn
black in color. We have considered both types of false smut images in our dataset. Leaf blast, neck blast, leaf
smut, leaf scald, sheath blight and stackburn are similar types of disease considering the fact that they mostly
occur at the rice leaf. Each of the diseases causes small spots or lesions in the rice plants. In different cases,
there can be single or multiple spots with similar diameter in a single leaf. Neck blast also attacks the neck
calm of the rice plant making it vulnerable. Leaf blast and neck blast are more deadly than the other diseases
and in severe cases, the yield loss can be as high as 100% as indicated in [22].
In our dataset, we have also considered certain pests and bugs. The Brown planthopper is a disease
caused by a pest named Nilaparvata lugens. The pests mostly reside at the root of the plants then spreading
through the entire rice plant causing the plant to dry out and turn brown [23]. Hispa is another disease that is
caused by a pest named Dicladispa armigera where the armigera insect scrapes the upper surface of the
leaves leaving only the lower epidermis. Similar to hispa, stemborer is a pest that also attacks the leaves of
rice plants. The stem borer initially resides at the base of the rice plants and later the stem borer larva drill
through the upper nodes and feeds the tillers which cause the tillers to dry out [24]. Different environmental
factors along with excessive use of nitrogen fertilizers are the main reason for these disease occurrences.
(a) (b) (c) (d) (e) (f)
(g) (h) (i) (j) (k) (l)
Figure 2. Sample images of our collected rice plant disease dataset (a) bacterial blight, (b) brown
planthopper, (c) brown spot, (d) false smut, (e) hispa, (f) leaf blast, (g) leaf scald, (h) leaf smut, (i) neck blast,
(j) sheath blight, (k) stackburn, and (l) stemborer

3645
Table 1. Total collected sample images with different symptom variations
Name Sample image variation No. of collected images Total images
Bacterial blight Early stage yellowish brown 46 126
Later stage full brown 80
Brown plant hopper Early stage of attack 50 75
Later stage of attack 25
Brown spot Small size spots 102 102
False smut Black smut 78 112
Brown smut 34
Hispa Visible pests on the leaves 56 86
Visible spots on the leaves without pests 30
Leaf blast Deep brown spot 110 136
Yellowish grey spot 26
Neck blast Blast spot on neck 225 225
Leaf scald Dark brown lesion 23 73
Greyish white lesion 50
Leaf smut Spots on the whole leaf 103 103
Sheath blight Black stem 112 228
White stem 116
Stemborer Grain symptom 171 187
Stem symptom 16
Stackburn Symptom on whole leaf 49 71
Symptom on a part 22
Healthy Healthy leaf 90 153
Healthy grain 63
2.2. Image preprocessing
After collecting the rice plant sample images, we processed the sample images before inputting
them to the convolutional neural network model. At first, we removed out-of-focus and blurry images from
our collected dataset. Once the faulty images were discarded, we move on to the image augmentation process
to increase the size of our dataset. Since we have small number of images for each of the class indicated by
Table 1, hence for better interaction and considering all possible dependencies at the feature level, we
perform image augmentation. The first augmentation method we carried out is the contrast stretching method.
Due to the gloomy weather and mobile camera lens, few sample pictures had very low contrast, hence
making it difficult to differentiate between the foreground subject and the background. This sort of low
contrasting pictures will prompt false positive and false negative rates in the network resulting in low
accuracy. Therefore, we utilised the contrast limited adaptive histogram equalization method described by
Pizer et al. [25] to enhance the low contrast sample images. We performed local histogram enhancement
using the contrast limited adaptive histogram equalization (CLAHE) method. CLAHE restricts the
amplification level by clipping the histogram at a predefined value before calculating the cumulative
distribution function. The clipping value depends upon the normalization of the histogram hence on the size
of the local area. The result of the histogram equalization process is illustrated in Figure 3, where Figure 3(a)
shows input image with distribution range in Figures 3(b) and (c). Figure 3(d) shows histogram equalized
image with distribution range in Figure 3(e) and (f).
It is evident that the intensity values of histogram equalized images are spread through the full
dynamic range hence the details are considerably more discernible. After performing the histogram
equalization method, we performed various image augmentation methods on our dataset. We performed
flipping, skewing, random rotation, random zoom, shear transformation, noise and distortion as distinct
image augmentation processes. Using these procedures, 10 augmented images have been created for each
example image and thus our dataset increased to a total of 16,770 training images. We used 80-20 data split
on our dataset by which we have 13,415 and 3,355 images for training and validating the model respectively.
2.3. State-of-the-art CNN architectures
Eight different state-of-the-art CNN architectures have been employed on our rice disease dataset
based on different criteria. We used VGG net which has 3×3 filters throughout the CNN architecture and they
are very deep CNN architectures with a large number of parameters [26]. The Inception-v3 architecture by
Szegedy et al. [27] has an inception module. Despite being a very deep architecture, the inception module
greatly reduces the amount of the parameters. By the use of asymmetric convolution, factorization is
performed which reduces the complexity of this model. ResNet architecture has residual networks with skip
connections and overcomes the problem of overfitting in ImageNet dataset [28]. The Xception [29]
introduced depthwise separable convolutions along with residual connections. Mobilenet v2 architecture also
uses depthwise separable convolutions having inverted shortcut network with linear bottleneck [30]. NasNet
Mobile architecture is similar to Mobilenet v2 architecture and it is used primarily for mobile devices. We

 ISSN: 2088-8708
3646
intended to see how our dataset performs on these models and later compared the results of these state-of-the-
art CNN models with our proposed CNN model.
(a) (b)
(c) (d)
(e) (f)
Figure 3. Contrast limited adaptive histogram equalization method (a) low contrast input image,
(b) histogram of input image, (c) RGB channel of input image, (d) high contrast output image,
(e) histogram of output image, and (f) RGB channel of output image
2.4. Proposed depthwise CNN model
We proposed a deep CNN architecture based on depthwise separable convolution. In the 1st layer of
our proposed CNN model, we use a 3×3 2D convolution followed by batch normalization and 2×2
max-pooling. In the 2nd layer, we used a depthwise separable convolution layer which consists of a 3×3
depthwise convolution followed by a 1×1 point-wise convolution. Unlike the normal convolution, depthwise
separable convolution deals with not only the spatial dimension but also the depth dimension.
For normal convolution considering a kernel K having spatial height 𝐾ℎ and spatial width of 𝐾𝑤 and
image tensor with input and output channels of 𝐼𝑖 and 𝐼𝑜 respectively. The convolution layer is represented
by, 𝐶 ∈ ℝ𝐼𝑖×𝐼𝑜×𝐾ℎ×𝐾𝑤. After applying the filter to a 3D tensor x, with size 𝐾ℎ × 𝐾𝑤 × 𝐼𝑖, we obtain a
response vector 𝑦 ∈ ℝ𝐼𝑜.

3647
𝑦 = 𝐶 ∗ 𝑥 (1)
where 𝑦𝑜 = ∑ 𝐶𝑜,𝑖 ∗ 𝑥𝑖 , 𝑜 ∈ 𝐼𝑜, 𝑖 ∈ 𝐼𝑖,
𝐼𝑖
𝑖=1 Here ∗ is the convolution operator. 𝐶𝑜,𝑖 = 𝐶[𝑜, 𝑖, : , : ] is a tensor
patch along the i-th input channel and o-th output channel. 𝑥𝑖 = 𝑥[𝑖, : , : ] is a tensor patch along the i-th input
channel of 3D tensor x.
For depthwise separable convolution, the convolutions are performed at each depth by which the
output is split for each dimension. Then pointwise convolution is performed for projecting the outputs into a
single channel space. Pointwise convolution is done by performing 1×1 cross-channel convolutions. Since the
depthwise convolution is performed at each depth at a time hence the kernel tensor is 𝐶𝑑 ∈ ℝ𝐼𝑖×1×𝐾ℎ×𝐾𝑤 and
for pointwise convolution the kernel tensor is 𝐶𝑝 ∈ ℝ𝐼𝑖×𝐼𝑜×1×1
. After applying this to a 3D tensor x, we
obtain a response vector y' as,
𝑦′ = (𝐶𝑝 ∘ 𝐶𝑑) ∗ 𝑥 (2)
where 𝑦𝑜′ = ∑ 𝐶𝑝𝑜,𝑖
(𝐶𝑑𝑜,𝑖
∗ 𝑥𝑖),
𝐼𝑖
𝑖=1 ∘ is a element wise operation. 𝐶𝑝(𝑜,𝑖)
= 𝐶𝑝[𝑜, 𝑖, : , : ] is a tensor patch
along the i-th input channel and o-th output channel. 𝐶𝑑(𝑖)
= 𝐶𝑝[𝑖, : , : ] is a tensor patch along the i-th input
channel of 3D tensor x.
As we can see, rather than performing the convolution in spatial dimension in RGB images,
depthwise separable convolutions are performed for each color channel separately. Due to single-channel
calculation, the overall complexity of the depthwise separable convolution is less than the normal
convolution on spatial dimension which accelerates the overall classification process. Because of the small
parameter size and faster execution, we specifically used depthwise separable convolution in our proposed
CNN model.
Our proposed CNN network starts with a normal 3×3 convolution followed by 2 depthwise
separable convolutions and then again a normal 3×3 convolution. Finally, we used 2 fully connected layers
having 128 neurons each and a 0.3 dropout layer in between the dense layer to reduce overfitting. 2×2 max
pooling and batch normalization were performed after every convolution. The rectified linear unit (ReLU)
activation function has been used for every layer except for the final dense layer which uses the softmax
activation function for the classification task. For updating the weights, several loss functions such as Adam,
RmsProp, Stochastic gradient descent, Hingeloss were utilized, and their performance was evaluated. Our
proposed depthwise CNN architecture is demonstrated in Figure 4.
Figure 4. Proposed depthwise separable CNN model
Our proposed CNN model's hyperparameters are listed below.
− Fold: 10-fold cross validation − Total Data: 16770
− MiniBatch: 64 − Data Split: 80-20
− Epochs: 100 − Training Data: 13415
− Activation Function: ReLU − Validation Data: 3355
− Momentum: 0.9 − Testing Data: 1600
− Dropout: 0.3 − Image Size: 224×224

 ISSN: 2088-8708
3648
All the images were resized to 224×224 pixels before inputting them in the CNN model. The first
layer of our CNN model takes an input (224, 224, 3) and produces an output tensor of (222, 222, 32). After
every convolution, batch normalization and 2×2 max-pooling are performed. Table 2 shows the kernel
dimension of each layer along with the generated feature maps of our proposed CNN model.
The output tensor (12, 12, 28) is fed through the flatten layer which maps the input matrix into a 1-D
vector of 18432 values. This vector is passed through two fully connected layers each of them consisting of
128 neurons and finally, we get 13 probabilistic values and thus the classification is performed. Figure 5
demonstrates the input images along with their feature maps generated by the last convolution layer.
Table 2. Proposed model generated feature vector output of each layer
Layer name Kernel dimension Stride Input feature vector Output feature vector
Conv_2d 3×3 - 224×224×3 224×224×32
Maxpooling1_2d 2×2 2 224×224×32 112×112×32
SeparableConv1_2d 3×3 - 112×112×32 109×109×32
Maxpooling2_2d 2×2 2 109×109×32 54×54×32
SeparableConv2_2d 3×3 - 54×54×32 52×52×64
Maxpooling3_2d 2×2 2 52×52×64 26×26×64
Conv3_2d 3×3 - 26×26×64 24×24×128
Maxpooling4_2d 2×2 2 24×24×128 12×12×128
Flatten - - 12×12×128 18432
Figure 5. Final convolution layer feature map output for each class
3. RESULTS AND DISCUSSION
In this section, the detailed analysis of various state-of-the-art CNN models performance along with
our proposed model are discussed. Also, the findings are reported extensively. The complete result analysis
has been discussed in appropriate subheadings.
3.1. State-of-the-art CNN models performance
For experimental work, Keras framework with TensorFlow back-end consisting Tesla P-100 GPU
has been used to train the models. For the training process, we used two different methodology transfer
learning and fine-tuning. In the transfer learning method, every model had pre-trained ImageNet weights
associated with them. Randomly initialized weights are only used to train the final dense layer. The
fine-tuning technique however unfreezes the final dense layer of the CNN and freezes all other layer weights.
Then the dense layers are trained using our rice plant dataset and different weights are initialized to the
nodes. The networks later adjusted the weights using optimizers. 10-fold cross-validation accuracy has been
considered as the model performance metric. We used different train, validation and testing set for multiple
instances and recorded the mean results. Table 3 shows the mean training, validation and testing result after
100 iterations of the rice plant disease dataset on different state-of-the-art CNN architectures.

3649
The result reported at Table 3 indicates that, fine-tuning produced a better result than the transfer
learning method for all of the CNN models. The performance curve of 8 state of the art CNN architectures
are shown in Figure 6 with transfer learning validation accuracy in Figure 6(a) and fine tuning validation
accuracy in Figure 6(b). Among the state-of-the-art CNN architectures, MobileNet v2 architecture performed
best in terms of accuracy. VGG-16 architecture provided high mean validation and testing accuracy. The
other deep CNN architectures (e.g. Inception, ResNet-50) with big parameters size mostly failed to identify
the test and validation rice plant images correctly. In the transfer learning method, we noticed that most of the
CNN architectures showed overfitting problem thus resulting in low testing accuracy. Even during the
fine-tuning method, Inception v3 and ResNet-50 model showed overfitting characteristics. The architectures
performed better in training data but failed in validation and test data.
Table 3. Statistical analysis of different CNN architectures on rice plant disease dataset. Here bold numbers
indicate best result
Architecture Training method Mean train accuracy Mean validation accuracy Mean test accuracy Total parameters
VGG 16 Transfer learning 94.2±1.8% 80.3±1.5% 78.2±1.0% 134 million
Fine tuning 98.6±0.9% 97.1±0.4% 95.7±0.5%
VGG 19 Transfer learning 89.5±1.5% 81.3±3.5% 79.1±2.0% 143 million
Fine tuning 97.3±0.7% 93.6±0.3% 86.2±1.5%
Inception v3 Transfer learning 84.8±3.0% 65.2±2.5% 61.7±3.0% 24 million
Fine tuning 98.6±0.4% 84.1±1.5% 83.2±2.0%
Xception Transfer learning 90.2±2.5% 84.2±1.5% 79.4±2.0% 20 million
Fine tuning 99.3±0.6% 95.1±0.8% 93.1±0.6%
MobileNet v2 Transfer learning 88.5±3.0% 83.2±2.0% 78.6±1.4% 3.5 million
Fine tuning 98.9±0.5% 97.3±0.4% 95.7±0.3%
ResNet-50 Transfer learning 39.2±6.0% 24.1±4.5% 20.2±5.0% 25 million
Fine tuning 67.7±1.5% 58.5±3.0% 52.1±2.5%
DenseNet-121 Transfer learning 94.2±2.8% 84.1±0.9% 78.0±1.5% 8 million
Fine tuning 98.1±0.9% 95.7±0.5% 92.3±1.0%
NasNet Mobile Transfer learning 94.1±1.0% 85.0±1.0% 78.4±1.8% 4.3 million
Fine tuning 97.6±0.5% 93.2±1.0% 91.0±1.2%
Proposed CNN Train from scratch 98.7±0.3% 96.5±0.3% 95.3±0.5% 2.4 million
(a)
(b)
Figure 6. Performance curve of different CNN models on transfer learning and fine tuning method,
(a) transfer learning validation accuracy and (b) fine tuning validation accuracy

 ISSN: 2088-8708
3650
3.2. Proposed CNN model performance
Our proposed model was trained from scratch and the weights were updated accordingly. We used
different hyperparameter combination and evaluated the performance thus selected the most suitable
parameters for our proposed model. By hyperparameter tuning, our model achieved a mean validation and
mean testing accuracy of 96.5% and 95.3% respectively. Considering the small parameter size, our model
performed significantly well in diagnosing rice plant diseases. We used different optimizers in our proposed
model and achieved the best accuracy using the Adam optimizer. Figure 7 demonstrates the best loss and
accuracy curve achieved with our proposed model using Adam optimizer with a 0.001 learning rate.
From the loss and accuracy curve, it is evident that our proposed model slowly converged to the
without showing any overfitting characteristics. We employed a variety of optimizers, including Adam,
stochastic gradient descent (SGD), Rmsprop, Hingeloss along with variable learning rates on our model.
Among the optimizers, adaptive moment estimation (Adam) [31] outperformed all the other optimizers in
terms of accuracy. Rmsprop optimizer performed similarly to Adam optimizer. Table 4 shows the
performance of using different optimizers with variable learning rates on our rice plant disease dataset.
Figure 7. Performance curve of the proposed CNN model using Adam optimizer with 0.001 learning rate
Table 4. Different optimizers performance with variable learning rate on rice plant disease dataset
Optimizer Learning rate Mean validation accuracy
Adam 0.01 95.4±0.4%
0.001 96.5±0.3%
0.0001 95.1±0.4%
RmsProp 0.01 94.1±0.3%
0.001 95.01±0.3%
0.0001 94.4±0.4%
HingeLoss 0.01 78.2±1.8%
0.001 82.7±1.2%
0.0001 73.5±1.5%
SGD 0.01 89.2±1.0%
0.001 88.5±1.5%
0.0001 82.1±1.5%

3651
From Table 4 it is evident that, stochastic gradient descent (SGD) and the support vector machine
SVM optimizer Hingeloss didn’t perform as well as Adam or Rmsprop. Adam and Rmsporp performed quite
similarly. The accuracy curve of the Adam optimizer with variable learning rates on our dataset is illustrated
in Figure 8.
From Figure 8 we can observe that the learning rate of 0.01 converged very quickly to our rice
disease dataset. On the other hand, the learning rate of 0.001 converged slowly to our dataset but achieved
overall maximum accuracy in the validation dataset. The learning rate of 0.0001 also performed well and
provided good accuracy. Hence, the optimal learning rate for our proposed CNN model is 0.001. However,
the difference is subtle and learning rate of 0.01 and 0.0001 can be used on our proposed model. The
normalized confusion matrix for the testing data of our proposed model is shown in Figure 9.
Figure 8. Adam optimizer with variable learning rate
Figure 9. Normalized confusion matrix

 ISSN: 2088-8708
3652
Our proposed model achieved a mean testing accuracy of 95.3%. By analyzing the normalized
confusion matrix, we observe that, apart from the brown spot disease all of the other disease classes have
high detection accuracy which clearly indicates that our proposed model can recognize the disease classes
correctly. The comparison of our proposed model with existing related works is represented in Table 5. We
observe from Table 5 that the majority of the research has focused on a small number of disease types. And
their proposed method comprising lower accuracy than us. Taking a modest parameter size into account, it is
evident that our proposed model performed significantly well in detecting rice plant diseases.
Table 5. Comparison of our proposed model with existing works
Authors Proposed Method Dataset Iteration Performance
Lu et al. [2] A simple 5-layer CNN architecture 500 sample images of 10 classes
collected from fields
10 Simple CNN: 95.38%
Liu et al. [3] A custom 6-layer CNN architecture 693 sample images of 2 classes
10,000 Custom CNN: 99.5%
AlexNet-8: 98%
VGGNet-11: 96.7%
Jagan et al. [4] SIFT combined with different
classifiers
120 sample images of 3 classes - SVM classifier: 91.1%
KNN classifier: 93.33%
Sowmyalakshmi
et al. [5]
Inception ResNet v2 integrated with
OWELM
115 sample images of 3 classes - Inception ResNet v2: 94.2%
Krishnamoorthy
et al. [7]
Inception ResNet v2 with transfer
learning method
3 types of disease classes along
with healthy plants
- Inception ResNet v2: 95.67%
Simple CNN: 84.75%
Sethy et al. [15] Faster R-CNN to localize disease
spots
50 sample images of false smut 5 In most cases the model can
identify the false smuts
Rahman et al. [17] A simple CNN architecture 1,426 sample images of 9 classes
100 Simple CNN: 94.33%
VGG-16: 97.12%
Proposed CNN A CNN model based on Depthwise
separable convolutions
1,677 sample images of 13 classes
100 Proposed CNN: 96.5%
3.3. Discussion
From our experimental analysis we have obtained some interesting findings regarding our rice
disease dataset. We identified some interesting facts about how different model parameter metrics improve
the performance of classification model. The findings are reported as follows: i) using transfer learning
method with pre-trained ImageNet weights on state-of-the-art CNN architectures cannot provide higher
classification accuracy. In all eight state-of-the-art CNN architectures, the fine-tuning method outperforms
the transfer learning method in terms of accuracy; ii) the lightweight CNN architectures (MobileNet v2,
NasNet Mobile, DenseNet) converge quickly to our dataset and provides a greater degree of precision;
iii) the deep CNN architectures with pre-trained ImageNet weights cannot converge quickly to our rice plant
disease dataset and mostly shows overfitting characteristics; iv) categorical cross-entropy with Adam
optimizer performs better in updating the weights for the dense layer on proposed model. However, different
learning rate provides almost similar performance which indicates that learning rate has very limited effect
on updating the weights, and v) on our proposed model, a learning rate of 0.001 and momentum of
0.9 performs well and produces better accuracy in classifying rice plant diseases.
4. CONCLUSION
In this paper, a depthwise separable convolution based neural network model has been proposed that
can effectively identify 12 distinct rice plant diseases along with healthy rice plants. Initially, the raw sample
images were collected from different regions of Bangladesh and during the pre-processing method contrast
limited adaptive histogram equalization algorithm (CLAHE) has been implemented to enhance the contrast
of the images and to reduce the noise. Along with histogram equalization different image augmentation
method has been performed by which the sample images of the dataset are increased to 16770 images.
Besides our proposed model, 8 different state-of-the-art CNN architectures have been used on the rice disease
dataset. Among all CNN architectures in terms of accuracy, MobileNet v2 was found to have the best
validation and testing accuracy of 97.3% and 95.7% respectively. The proposed CNN model has been
evaluated with different loss functions along with variable learning rates. Adam optimizer with a learning
rate of 0.001 provided the best mean validation and mean test accuracy of 96.5% and 95.3% respectively for
our proposed CNN model. Based on our model's performance, it is evident that the proposed model was able
to properly diagnose a variety of rice plant diseases. Considering a small parameter size of 2.4 million we
conclude that, the proposed model is a substantial improvement over the traditional convolutional neural
network architectures in rice plant disease detection.

3653
ACKNOWLEDGEMENTS
We gratefully acknowledge the ICT division of the Government of the People’s Republic of
Bangladesh for funding this research work.
REFERENCES
[1] S. H. Ou, Rice diseases, 2nd ed. Commonwealth Mycological Institute, 1985.
[2] Y. Lu, S. Yi, N. Zeng, Y. Liu, and Y. Zhang, “Identification of rice diseases using deep convolutional neural networks,”
Neurocomputing, vol. 267, pp. 378–384, Dec. 2017, doi: 10.1016/j.neucom.2017.06.023.
[3] C. Liu, C. Xu, S. Liu, D. Xu, and X. Yu, “Study on identification of rice false smut based on CNN in natural environment,” in
2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI),
Oct. 2017, pp. 1–5, doi: 10.1109/CISP-BMEI.2017.8302016.
[4] K. Jagan, M. Balasubramanian, and S. Palanivel, “Detection and recognition of diseases from paddy plant leaf images,”
International Journal of Computer Applications, vol. 144, no. 12, pp. 34–41, Jun. 2016, doi: 10.5120/ijca2016910505.
[5] R. Sowmyalakshmi et al., “An optimal classification model for rice plant disease detection,” Computers, Materials and Continua,
vol. 68, no. 2, pp. 1751–1767, 2021, doi: 10.32604/cmc.2021.016825.
[6] S. Ramesh and D. Vydeki, “and classification of paddy leaf diseases using optimized deep neural network with jaya algorithm,”
Information Processing in Agriculture, vol. 7, no. 2, pp. 249–260, Jun. 2020, doi: 10.1016/j.inpa.2019.09.002.
[7] N. Krishnamoorthy, L. V N. Prasad, C. S. P. Kumar, B. Subedi, H. B. Abraha, and S. V E, “Rice leaf diseases prediction using
deep neural networks with transfer learning,” Environmental Research, vol. 198, Jul. 2021, doi: 10.1016/j.envres.2021.111275.
[8] V. K. Shrivastava, M. K. Pradhan, S. Minz, and M. P. Thakur, “Rice plant disease classification using transfer learning of deep
convolutional neural network,” The International Archives of the Photogrammetry, Remote Sensing and Spatial Information
Sciences, pp. 631–635, Jul. 2019, doi: 10.5194/isprs-archives-XLII-3-W6-631-2019.
[9] M. S. I. Prottasha, Z. Tasnim, S. M. S. Reza, and D. A. Hossain, “A lightweight CNN architecture to identify various rice plant
diseases in Bangladesh,” in 2021 International Conference on Information and Communication Technology for Sustainable
Development (ICICT4SD), Feb. 2021, pp. 316–320, doi: 10.1109/ICICT4SD50815.2021.9396927.
[10] A. S. A. Mettleq, I. M. Dheir, A. A. Elsharif, and S. S. Abu-Naser, “Mango classification using deep learning,” International
Journal of Academic Engineering Research (IJAER), vol. 3, no. 12, pp. 22–29, 2019.
[11] M. M. Hasan, J. P. Chopin, H. Laga, and S. J. Miklavcic, “Detection and analysis of wheat spikes using convolutional neural
networks,” Plant Methods, vol. 14, no. 1, Dec. 2018, doi: 10.1186/s13007-018-0366-8.
[12] N. Saranya, K. Srinivasan, and S. K. P. Kumar, “Banana ripeness stage identification: a deep learning approach,” Journal of
Ambient Intelligence and Humanized Computing, May 2021, doi: 10.1007/s12652-021-03267-w.
[13] L. Huang and D. He, “Ripe fuji apple detection model analysis in natural tree canopy,” TELKOMNIKA Indonesian Journal of
Electrical Engineering, vol. 10, no. 7, pp. 1771–1778, Nov. 2012, doi: 10.11591/telkomnika.v10i7.1574.
[14] S. Yadav, N. Sengar, A. Singh, A. Singh, and M. K. Dutta, “Identification of disease using deep learning and evaluation of
bacteriosis in peach leaf,” Ecological Informatics, vol. 61, Mar. 2021, doi: 10.1016/j.ecoinf.2021.101247.
[15] P. K. Sethy, N. K. Barpanda, A. K. Rath, and S. K. Behera, “Rice false smut detection based on faster R-CNN,” Indonesian
Journal of Electrical Engineering and Computer Science, vol. 19, no. 3, pp. 1590–1595, Sep. 2020, doi:
10.11591/ijeecs.v19.i3.pp1590-1595.
[16] B. S. Bari et al., “A real-time approach of diagnosing rice leaf disease using deep learning-based faster R-CNN framework,”
PeerJ Computer Science, vol. 7, Apr. 2021, doi: 10.7717/peerj-cs.432.
[17] C. R. Rahman et al., “Identification and recognition of rice diseases and pests using convolutional neural networks,” Biosystems
Engineering, vol. 194, pp. 112–120, Jun. 2020, doi: 10.1016/j.biosystemseng.2020.03.020.
[18] M. S. I. Prottasha, “Rice plant disease dataset,” archive.org. [Online]. Available: http://archive.org/details/ricediseasedataset
[19] Y. Tagami, “Epidemiology of the bacterial leaf blight of rice,” Japanese Journal of Phytopathology, vol. 31, no. 1, pp. 145–151,
1965, doi: 10.3186/jjphytopath.31.Special1_145.
[20] S. Sunder, R. Singh, and R. Agarwal, “Brown spot of rice : an overview,” Indian Phytopathology, vol. 67, no. 3, pp. 201–215,
2014.
[21] H. Ikegami, “Studies on the false smut of rice,” Japanese Journal of Phytopathology, vol. 27, no. 1, pp. 16–23, 1962, doi:
10.3186/jjphytopath.27.16.
[22] E. C. Roumen and W. S. de Boef, “Latent period to leaf blast in rice and its importance as a component of partial resistance,”
Euphytica, vol. 69, no. 3, pp. 185–190, Jan. 1993, doi: 10.1007/BF00022364.
[23] J. Xue et al., “Transcriptome analysis of the brown planthopper nilaparvata lugens,” PLoS ONE, vol. 5, no. 12, Dec. 2010, doi:
10.1371/journal.pone.0014233.
[24] M. D. Pathak and Z. R. Khan, Insect pests of rice. Los Banos: International Rice Research Institute, 1994.
[25] S. M. Pizer et al., “Adaptive histogram equalization and its variations,” Computer Vision, Graphics, and Image Processing,
vol. 39, no. 3, pp. 355–368, Sep. 1987, doi: 10.1016/S0734-189X(87)80186-X.
[26] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in 3rd International
Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, Sep. 2014.
[27] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” in
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2016, pp. 2818–2826, doi:
10.1109/CVPR.2016.308.
[28] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), Jun. 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90.
[29] F. Chollet, “Xception: deep learning with depthwise separable convolutions,” in 2017 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Jul. 2017, pp. 1800–1807, doi: 10.1109/CVPR.2017.195.
[30] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: inverted residuals and linear bottlenecks,” in
2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun. 2018, pp. 4510–4520, doi:
10.1109/CVPR.2018.00474.
[31] J. L. Ba and D. P. Kingma, “Adam: a method for stochastic optimization,” in 3rd International Conference on Learning
Representations, ICLR 2015 - Conference Track Proceedings, 2015, pp. 1–15.

 ISSN: 2088-8708
3654
BIOGRAPHIES OF AUTHORS
Md. Sazzadul Islam Prottasha obtained his bachelor’s and master’s degree in
Information and Communication Engineering from Bangladesh University of Professionals
(BUP), Dhaka, Bangladesh. Currently he is working in multiple projects based on image
processing and computer vision. His research interests include machine learning, computer
vision, image processing, big data analytics, artificial intelligence, and IoT. He can be
contacted at email: 19541026@bup.edu.bd.
Sayed Mohsin Salim Reza is an Assistant Professor at the Department of
Information and Communication Technology, Bangladesh University of Professionals (BUP),
Dhaka, Bangladesh. He is a member of a variety of professional organizations. He is an
effective planner and operator, possessing expertise in overseeing academic and administrative
operations including organizational standards, and computer science training for faculty. He
has published over 50 research articles in ISI/Scopus-indexed journals and several renowned
international conferences. He can be contacted at email: salim.reza@bup.edu.bd.

A classification model based on depthwise separable convolutional neural network to identify rice plant diseases

More Related Content

Similar to A classification model based on depthwise separable convolutional neural network to identify rice plant diseases (20)

More from IJECEIAES (20)

Recently uploaded (20)

A classification model based on depthwise separable convolutional neural network to identify rice plant diseases