Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control

Journal homepage: http://kinetik.umm.ac.id

ISSN: 2503-2267
Vol. 4,
7, No. 3, August
2, May, Pp.2019, Pp. 277-288
187-196 187
Convolutional Neural Network (CNN) models for crop diseases
classification

Deni Sutaji*1,2, Harunur Rosyid1

Informatics Department, Universitas Muhammadiyah Gresik, Gresik, Indonesia*1
Computer Science, Informatics Institute, Gazi University, Ankara, Turkey2

Article Info Abstract

Keywords: Crop diseases have a significant impact on agricultural production. As a result,
CNN Models, Crop Disease, Classification,
early diagnosis of crop diseases is critical. Deep learning approaches are now
Unfreeze Layers
promising to improve disease detection. Convolutional Neural Network (CNN)
Article history: models can detect crop disease using images with automatic feature
Received: April 29, 2022 extraction. This study proposes crop disease classification considering ten pre-
Accepted: May 11, 2022
trained CNN models. Fine-tuning for each model was conducted on the Plant
Published: May 31, 2022
Village dataset. The experimental results show that fine-tuning improves the
Cite: model's performance with an average accuracy of 8.85%. The best CNN model
D. Sutaji and H. Rosyid, “Convolutional was DenseNet121, with 94.48% and 98.97% accuracy for freezing all layers
Neural Network (CNN) Models for Crop
and unfreezing last block convolution layers. Moreover, fine-tuning produces
Diseases Classification”, KINETIK, vol. 7, no.
2, May. 2022. less time-consuming with an average of 2.20 hours. VGG19 is the less time-
https://doi.org/10.22219/kinetik.v7i2.1443 consuming reduction by 8 hours. On the other hand, MobileNetV2 is the
second-best performance model with less time-consuming than DenseNet121
*Corresponding author.
and produces fewer parameters, which is affordable for embedding it to mobile
Deni Sutaji
E-mail address: devices.
[email protected]

1. Introduction
Crop diseases are one of the leading causes of the decline in agricultural production. In recent years, various
crop diseases have increased and affected the agricultural, economic, and health sectors. The disease generally
generates noticeable marks or lesions on the crop leaves. Therefore, it is crucial to detect the presence of the disease
at an early stage. However, traditional approaches are commonly used by most farms and plantations to recognize the
disease. Consequently, it is time-consuming and leads to high costs and misdetection [1].
Hence, researchers have designed the Convolutional Neural Network (CNN) models as a computer-aided
disease diagnosis. CNN models can overcome the problem of object recognition, and classification has significantly
improved in the last decade [2], [3]. In recent years, CNN models have emerged as the most extensively used model in
crop disease detection [3]–[10]. The CNN model can be applied from scratch or by using pre-trained designs.
Promisingly, pre-trained CNN models outperformed the scratch designs. AlexNet was the first proposed pre-trained
CNN model by [11] on the 1.2 million high-resolution images within the 1.000 different classes in the ImageNet LSVRC-
2010 contest. Furthermore, many CNN models were offered in the next contest ImageNet LSVRC.
Recently, several pre-trained CNN architectures have been suggested, for example, GoogleNet [12], VGGNet
[13], ResNet [14], InceptionV3 [15], DenseNet [16], Xception [17], MobileNet [18], InceptionResNetV2 [19], NasNet [20],
and EfficientNet [21]. In prior research, researchers used two methods in crop disease recognition using CNN models
[22]. Some use transfer learning to fine-tune well-known designs. Others use well-known architectures, such as
modifying the convolutional layers, changing the order of layers, adding attention layers before the classifier layer, and
ensembling more than two pre-trained CNN models. They aim to improve the model performance.
The most considered dataset is the Plant Village dataset. In this dataset, [23] corresponded with the results of
AlexNet and GoogLeNet, in detecting 26 crop diseases in laboratory leave images. They employed transfer learning to
increase categorization performance and achieve an accuracy of 99.35%. CaffeNet was proposed by [24] on 15 class
leaf images and achieved an accuracy of 96.0%. Cucumber leaves were considered by [25]–[27] using Deep
Convolution Neural Network (Scratch designs), modified Deep Convolution Neural Network, and combined
DeepLabV3+ and U-Net, respectively. Moreover, the detection of citrus diseases conducted by [22], [28], [29] used
modified MobileNetV2 and an Ensemble AlexNet, VGG16, ResNet50, and InceptionResNetV2. Both cucumber and
citrus crops achieved an accuracy of over 98%.
In addition to the entire categories of the Plant Village dataset, researchers considered multiclass classification
with 38 classes proposed by [30]–[34]. Pre-trained MobileNet, EfficientNet, DensNet, and ensemble MobileNet with
DenseNets were applied. All suggested model performances achieved an accuracy of over 92%, and 100% for the
ensemble model was proposed by [34].

Cite: D. Sutaji and H. Rosyid, “Convolutional Neural Network (CNN) Models for Crop Diseases Classification”, KINETIK, vol. 7, no. 2, May.
2022. https://doi.org/10.22219/kinetik.v7i2.1443
188 Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control
Other datasets such as iBean, Citrus, Rice, and Turk-Plants were also used to detect crop diseases. Several pre-
trained models were used, such as modified and ensembled by researchers, such as [35]–[40]. The model's average
performance is above 99%.
However, various datasets have been used in the literature, making it difficult to compare directly. Several
detection systems work only on two classes, while the others detect crop diseases in over 38 classes. Considering this
term, we suggested investigating the classification of crop disease by using ten pre-trained CNN algorithms to consider
the entire Plant Village dataset. This dataset consists of 54.305 RGB leaf images with 256 x 256 pixels. The dataset
subsists of 38 classes from 14 crop species: strawberries, apples, grapes, cherries, corn, oranges, blueberries,
potatoes, soybeans, pumpkins, peppers, raspberries, peaches, and tomatoes.
Another problematic aspect of comparing which CNN models can perform better is the various configurations
and modifications of the pre-trained CNN models to detect crop disease. As a simple investigation to compare, we
proposed regular and unfroze the final convolutional blocks aforementioned pre-trained CNN models. It aims to
generate high-dimensional extracted features to provide a more helpful characteristic of an object. Our main
contributions to this research are:
1. Comparing the performance of 10 pre-trained CNN models to detect crop disease.
2. Improving the performance of 10 pre-trained CNN models by unfroze the final convolution blocks.
This research is organized as follows: Section 1 introduces crop diseases, state-of-the-art methods for classifying
crop diseases, explains the problem of statements, and gives solutions as a contribution to the research. Section 2
discusses research methods. Section 3 presents the experimental results and performance comparison. Eventually,
Section 4 explains the conclusion of this study.

2. Research Method
2.1 Method
This study considered 10 pre-trained CNN architectures, such as AlexNet, GoogleNet, VGG16, VGG19,
ResNet50V2, InceptionV3, Xception, MobileNetV2, DenseNet121, and InceptionResnetV2 to classify crop diseases.
These CNN architectures were developed with a transfer learning approach and implemented fine-tuning by unfreezing
the last block of convolutional layers. It aims to enhance the high-dimensional extracted features. Updating the weight
parameter impacts capture more details of high-dimensional extracted features. Figure 1 illustrates our research
method, and Figure 2 describes our suggested transfer learning and fine-tuning design.

Figure 1. Our Research Method

Figure 2. Our Proposed Transfer Learning and Fine-Tuning CNNs Design

2.2 Dataset
Plant Village dataset consists of 38 classes and includes 54,305 images with 256 x 256 size in the RGB channel.
The dataset consists of 14 plant species: apples, blueberries, cherries, corn, grapes, oranges, peppers, raspberries,
potatoes, pumpkins, peaches, raspberries, soybeans, strawberries, and tomatoes. The diseases from the 14 plant
© 2022 The Authors. Published by Universitas Muhammadiyah Malang
This is an open access article under the CC BY SA license. (https://creativecommons.org/licenses/by-sa/4.0/)
Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control 189
species are distributed as 17 fungal diseases, 4 bacterial diseases, 2 fungal diseases, 2 viral diseases, and the last
disease caused by mites has 1 class. Meanwhile, there are 12 plant species in the healthy category of plant leaves [23].
The considered dataset was collected from https://www.kaggle.com/abdallahalidev/plantvillage-dataset. The example
of each class image is shown in Figure 3.
We pre-processed the collected dataset in four steps. The first step is resizing all images into 224 x 224 pixels.
The second step is to split and distribute the dataset into a training set, a validation set, and a test set with a composition
of 70%, 10%, and 20%, respectively. The detail of distribution is described in Figure 3. The third step is enforced image
normalization, dividing the intensity value of each pixel by 255. The reason is to reduce computational costs because
the intensity value of each pixel is between 0 and 1 [41].
Furthermore, the last step is image augmentation, completing the image in the data series more varied since the
acquired image is subject to different conditions (angle, illumination, and background). Image augmentation is only
performed on the training data because this model can accept different conditions of the leaf image under test. This
step's configuration includes image rotation with a rotation angle of 30 degrees, zooming image with a percentage of
0.3, cropping with width_shift_range, height_shift_range, shear_range, and zoom_range settings with values of 0.3.
The last configuration is horizontal_flip with the setting True.

(a1) (a2) (a3) (a4) (b1) (ce1) (ce2) (co1) (co2) (co3)

(co4) (g1) (g2) (g3) (g4) (o1) (p1) (p2) (pp1) (pp2)

(po1) (po2) (po3) (r1) (s1) (sq1) (st1) (st2) (t1) (t2)

(t3) (t4) (t5) (t6) (t7) (t8) (t9) (t10)

Figure 3. Randomly image examples: Apples (a1) Healthy, (a2) Scap, (a3) Balck rot, (a4) Cedar rust; Blueberry (b1)
Healthy; Cherry (ce1) Healthy, (ce2) Powdery mildew; Corn, (co1) Healthy, (co2) Cercospora, (co3) Rust, (co4)
Northern blight; Grape (g1) Healthy, (g2) Blak rot, (g3) Black measles; Orange (o1) Healthy; Peach (p1) Healthy, (p2)
Bacterial spot; Pepper (pp1) Healthy, (pp2) Bacterial spot; Potato (po1) Healthy, (po2) Early blight; Raspberry (r1)
Healthy; Soybean (s1) Healthy; Squash (sq1) Powdery mildew; Strawberry (st1) Scorch; Tomato (t1) Healthy, (t2)
Bacterial spot, (t3) Early blight, (t4) Late blight, (t5) Mold, (t6) Septoria spot, (t7) Spider mites, (t8) Target spot, (t9)
Mosaic virus, (t10) Yellow curl virus

2.3 Software Tools for Experimental

Our experimental research was conducted on the Google Colab Pro with the following specifications: Intel(R)
Xeon(R) 2 CPU @ 2.20GHz Processor, 13GB RAM, and GPU Tesla P100-PCIE-16GB. This platform runs Python and
Jupyter notebooks with Tensorflow and Keras libraries as a backend for VGG16, VGG19, ResNet50, InceptionV3,
Xception, MobileNetV2, and InceptionResnetV2 models. Moreover, we used the PyTorch framework to train and test
pre-trained AlexNet and GoogleNet models.

2.4 Test Scenarios

In this research, we considered two test scenarios. We initially considered regular transfer learning for ten pre-
trained CNN architectures without fine-tuning. Then, transfer learning and fine-tuning as a second scenario. We were
unfreezing the final convolutional block layers to enhance the high-dimensional extracted features.
This research considers Accuracy (Acc), Sensitivity (Sen), Specificity (Spe), and Precision (Prec) metrics to
calculate the performance of tested CNN models. The correctly classified positive ratio of all True Positives is
distinguished as Sensitivity. The correctly classified negative ratio of all False Positives is marked as Specificity. Then,
Accuracy directs to the ratio of correctly classified samples to all samples. Precision guides to the correctly classified
positive ratio of all accurate recognitions.

Cite: D. Sutaji and H. Rosyid, “Convolutional Neural Network (CNN) Models for Crop Diseases Classification”, KINETIK, vol. 7, no. 2, May.
2022. https://doi.org/10.22219/kinetik.v7i2.1443
 190 Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control
3. Results and Discussion
Determining the hyperparameters for model training is critical work. We accept Adam as the optimizer, including
a learning rate of 0.0001 as the best value. We initially tried different learning rate values starting from 0.1 to 0.0001 in
0.1 additions. We also tested other optimizers, such as SGD and RMSprop. Therefore, Adam optimizer produces
excellent training, validation, and testing performances. Another parameter is an epoch, which we set at 100 and
consider to have an early stop value of 10. Training can be more efficient, and the process does not waste time in 100
epochs. We also employed cross-entropy to track each epoch's decreasing error rate and accuracy. Finally, a batch
size of 25, 18, and 18 for training, validation, and testing data were presented in Table 1.

Table 1. Dataset Split Distribution (70:10:20) of Training, Validation, and Testing

Crop Species Total Train Val Test Crop Species Total Train Val Test
Apple Scab 630 441 63 126 Pepper Healthy 1478 1035 148 295
Apple Black root 621 435 62 124 Potato Ealy Blight 1000 700 100 200
Apple Cedar 275 192 28 55 Potato Late Blight 1000 700 100 200
apple
Apple rust
Healthy 1645 1151 165 329 Potato Healthy 152 106 15 31
Blueberry 1502 1051 150 301 Raspberry 371 260 37 74
Healthy
Cherry Powdery 1052 736 105 211 Soybean 5090 3563 509 1018
mildew
Cherry Healthy 854 598 85 171 Squash Powdery Mildew 1835 1285 183 367
Corn Cercospora 513 359 51 103 Strawberry scorch 1109 776 111 222
Corn rust 1192 834 119 239 Strawberry Healthy 456 319 46 91
Corn Northern 985 690 98 197 Tomato Bacterial spot 2127 1489 213 425
CornBlight
Healthy 1162 813 116 233 Tomato Early Blight 1000 700 100 200
Grape Black rot 1180 826 118 236 Tomato Late Blight 1909 1336 191 382
Grape Esca 1383 968 138 277 Tomato Leaf Mold 952 666 95 191
Black
GrapeMeasles
Blight 1076 753 108 215 Tomato Septoria spot 1771 1240 177 354
Grape Healthy 423 296 42 85 Tomato Spider mites 1676 1173 168 335
Orange 5507 3855 551 110 Tomato Target Spot 1404 983 140 281
Haunglongbing
Peach Bacterial 2297 1608 230 1
459 Tomato Yellow curs virus 5357 3750 536 1071
spot
Peach Healthy 360 252 36 72 Tomato mosaic virus 373 261 37 75
Pepper Bacterial 997 698 100 199 Tomato Healthy 1591 1114 159 318
Spot Total 54305 38012 5430 10863

3.1 Experimental Results Scenario 1

The CNN model weight parameters were unchanged in this scenario because layers were configured as trainable
equal false. We added one layer before the classifier layer, GlobalAveragePooling2D. It aims to provide features in one
channel, as the classifier layer needs. The experimental results show that GoogleNet, DenseNet121, ResNet50, and
MobileNetV2 achieved better accuracy than the seven pre-trained CNN models. It is similar to previous studies done
by [23], [38], [42], and [29], respectively, that model performances were outperformed.
On the other hand, the less time-consuming was provided by the AlexNet model. Moreover, MobileNetV2, both
model performance and time-consuming outperformed. Table 2 presents the model performances in the first scenario
design.

Table 2. Model Performance for the First Scenario

Acc Prec Rec F1 Epoch
No Model Computation Time
Train Val Test Test data Stop
1 AlexNet 95.63 89.47 89.47 91.42 89.47 89.54 3:10:55.722126 26
2 GoogleNet 95.53 96.32 96.22 96.55 96.22 96.23 10:07:40.969602 22
3 VGG16 81.04 86.00 84.18 84.91 84.18 83.95 8:27:10.904154 85
4 VGG19 78.42 84.33 81.80 81.81 81.80 81.25 12:49:57.617564 97
5 InceptionV3 88.39 89.28 88.14 88.51 88.14 87.96 4:36:43.910856 37
6 ResNet50V2 92.67 94.62 93.26 93.56 93.26 93.21 4:07:58.418349 33
7 InceptionResnetV2 89.32 91.58 90.64 90.85 90.64 90.46 5:51:53.510285 39
© 2022 The Authors. Published by Universitas Muhammadiyah Malang
This is an open access article under the CC BY SA license. (https://creativecommons.org/licenses/by-sa/4.0/)
Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control 191
8 DenseNet121 93.40 94.99 94.48 94.62 94.48 94.45 4:55:42.577147 39
9 MobileNetV2 93.08 93.17 93.11 93.35 93.11 93.11 4:03:49.903792 44
10 Xception 91.81 92.30 84.57 86.45 84.57 84.65 7:44:02.629347 47

3.2 Experimental Results Scenario 2

As mentioned in the previous subsection, this scenario considered unfreezing the last convolution blocks of the
CNN models. The model performances significantly improved. Each architecture achieved accuracy above 97%.
Moreover, the time-consuming of each model decreasing about 50%. Table 3 describes the model performances of the
second scenario design.

Table 3. Model Performance for the Second Scenario

Acc Prec Rec F1 Epoch
No Model Computation Time
Train Val Test Test data Stop
1 AlexNet 99.29 98.78 98.12 98.30 98.25 98.25 1:29:02.212469 31
2 GoogleNet 99.28 99.02 98.34 98.34 98.42 98.42 1:32:59.543974 27
3 VGG16 98.11 98.51 98.27 98.28 98.27 98.26 2:15:17.964204 16
4 VGG19 96.64 97.29 97.19 97.31 97.19 97.19 4:13:36.753806 27
5 InceptionV3 97.73 98.38 97.30 97.47 97.30 97.31 2:01:08.610317 21
6 ResNet50V2 98.22 98.91 98.43 98.45 98.43 98.43 2:19:37.840994 18
7 InceptionResnetV2 98.70 99.04 98.76 98.78 98.76 98.75 3:13:28.997381 22
8 DenseNet121 99.13 99.15 98.97 99.00 98.97 98.97 3:38:19.336488 27
9 MobileNetV2 98.74 99.10 98.95 98.96 98.95 98.94 2:07:51.701822 22
10 Xception 99.53 99.45 97.58 97.68 97.58 97.58 6:12:13.524522 25

3.3 Research Analysis and Limitations

Implementing fine-tuning by unfreezing the last convolution block layers significantly affects the system's
performance. As shown in Table 4, the accuracy value for each CNN model has improved with 4.49, 15.39, and 8.85
values of lower, higher, and average, respectively. The model can perform better when the weights of high-dimensional
features are updated.

Table 4. Model Performances Comparison

Test Accuracy % Computational Time
No Models
Scenario 1 Scenario 2 Improve Scenario 1 Scenario 2 Gap (hour)
1 AlexNet 89.18 98.12 8.94 01:03:50 01:29:02 0.00
2 GoogleNet 94.00 98.34 4.34 02:02:11 01:33:00 0.00
3 VGG16 84.18 98.27 14.09 08:27:11 02:15:18 6.00
4 VGG19 81.80 97.19 15.39 12:49:58 04:13:37 8.00
5 InceptionV3 88.14 97.30 9.16 04:36:44 02:01:09 2.00
6 ResNet50V2 93.26 98.43 5.17 04:07:58 02:19:38 1.00
7 InceptionResnetV2 90.64 98.76 8.12 05:51:54 03:13:29 2.00
8 DenseNet121 94.48 98.97 4.49 04:55:43 03:38:19 1.00
9 MobileNetV2 93.11 98.95 5.84 04:03:50 02:07:52 1.00
10 Xception 84.57 97.58 13.01 07:44:03 06:12:14 1.00
Average 89.34 98.19 8.85 - - 2.20

The highest accuracy obtained by the DenseNet121 model of 98.97% is quite promising for crop disease
detection. It will significantly assist farmers or agricultural experts in detecting crop diseases.

Cite: D. Sutaji and H. Rosyid, “Convolutional Neural Network (CNN) Models for Crop Diseases Classification”, KINETIK, vol. 7, no. 2, May.
2022. https://doi.org/10.22219/kinetik.v7i2.1443
192 Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control
AlexNet GoogleNet

VGG16 VGG19

Inception Resnet V2 MobileNetV2

InceptionV3 Xception

DenseNet121 ResNet50

Figure 4. Training Accuracy and Loss Curve of the Second Scenario Design

In addition to improving accuracy, unfreezing the final block of the convolution layer speeds up the training
process. The average reduction in time consumption in each model is almost 50%, or 2.20 hours, with the most
significant reduction being 8 hours.
However, our research did not ultimately outperform the earlier research models. We accomplished just simple
fine-tuning by unfreezing the final convolution block layers. For example, the CNN AlexNet and GoogleNet models
proposed by [23] get an accuracy value of 99.35%, while our proposed model is only 98.34%. If we refer to the model
performance graph in Figure 4, our proposed model is overfitting, and the VGG19 model followed this condition. While
the DenseNet121 model, which is our best model, does not outperform the model proposed by [31]. Our DenseNet121
model only gets 98.97% accuracy and 99.75% for [31] model.
On the other hand, our model outperforms the model proposed by [43], which was only implemented on citrus
crops. Their models get 97%, 96.2%, and 97% accuracy with the VGG16, VGG19, and InceptionV3 models. Meanwhile,
our proposed model achieved an accuracy of 98.27%, 97.19%, and 97.30%. This positive result also outperforms the

© 2022 The Authors. Published by Universitas Muhammadiyah Malang

This is an open access article under the CC BY SA license. (https://creativecommons.org/licenses/by-sa/4.0/)
Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control 193
Table 5. Performance Comparison with Prior Studies
Authors Datasets Model Proposed Accuracy
Mohanty [23] Entire Plant Village AlexNet, GoogleNet 99.35%
15 class custom leaf
Sladojevic [24] CaffeNet 96.30%
images
Gandhi [30] Entire Plant Village MobileNetV1 and GAN 92%
Cucumber Plant Deep Convolution Neural
Ma [25] 93.40%
Village Network (Scratch)
Too [31] Entire Plant Village DenseNets 99.75%
Deep Convolution Neural
Geetharamani [32] Entire Plant Village 96.46%
Network (Scratch)
Citrus Modified MobileNetV2 92% MobileNetV2, 99%
Barman [29]
SSCNN SSCNN
EfficientNet 99.91% on B5,
Atila [44] Entire Plant Village
(B0 – B7) 99.97% on B4.
Custom Tomato Leaf 99.60% on lab data
I. Ahmad [35] Inception V3
(lab & field) 93.70% on field data
Apple, Grape, Potato
(Plant Village), 99.94% on plant village
Chen [36] MobileNet-Beta
Custom Maize and 99.85% on custom
Rice
Plant Village Maize,
92% on plant village
Chen [37] Custom rice and INC-VGG
80.38% custom
maize
User-defined (3
specieses of Plant DenseNet with 5-fold Cross-
Tiwari [45] 99.19%
village), iBean , Validation
Citrus, and Rice
Custom Plant Village
(Apple, Corn Potato, CNN adoption from
Gajjar [39] 96.88%
Tomato) and field Inception Design (Scratch)
dataset
VGG16, VGG19, Inception
Sujatha [43] Private Citrus leaf 97%, 96.2%, 97%
V3
Custom Plant Village Modified MobileNetV2 + Soft 99.71% on Plant Village
Chen [40]
and local dataset Attention 99.13% on local
Ensemble AlexNet, VGG16,
99.04% on augmented
Khanramaki [22] Private Citrus leaf ResNet50, and
98.64%
InceptionResNetV2
Wang [27] Cucumber leaf DeeplabV3+MobileNet 92.85%
Ensemble six CNN models
Turk-Plants 97.56% MV Model,
Turkoglu [1] (feature extractor) and SVM
96,83% EF
(classifier).
Ensemble weighted
MobileNet (0.3451),
Vallabhajosyula [34] Entire Plant Village 100%
Densenet121 (0.6541),
DenseNet201 (0.0009)
DenseNet121 (scenario 2) 98.97%
Our Proposed Entire Plant Village
MobileNetV2 (scenario 2) 98.95%

Model proposed by [37], which obtained an accuracy of 92% on rice plants. Likewise, the InceptionResNetV2
model we propose has an accuracy of 98.76%, higher than the model proposed by [22] of 98.64%. This positive result
is followed by the Xception model we propose to get almost the same results as the Xception model proposed by [46],
which are 97.58% and 97.60%, respectively. Unfortunately, our MobileNetV2 model did not follow these positive results.
Our MobileNetV2 model is inferior to some of the previously proposed models, which applied modification to an
ensemble of MobileNet proposed by [29], [30], [36] and obtains more than 99% accuracy. Their models extract features
better than our model. Moreover, their models added several convolutional layers before the classification layer. Our

Cite: D. Sutaji and H. Rosyid, “Convolutional Neural Network (CNN) Models for Crop Diseases Classification”, KINETIK, vol. 7, no. 2, May.
2022. https://doi.org/10.22219/kinetik.v7i2.1443
 194 Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control
MobileNetV2 model provides 98.95% of accuracy in the second scenario. Table 5 describes the performance
comparison of our model with earlier studies.
In completing the discussion of this section, the fine-tuning model that we propose is relatively simple compared
to the model proposed by previous studies. Some of them applied architectural modifications, ensembles, and other
advance works. Fortunately, our model is not entirely inferior to the previously proposed model. Some models like
VGG16, VGG19, InceptionV3, and InceptionResNetV2 outperform models by [43]. The best models we
proposeDenseNet121 and MobileNetV2, promise to be improved in terms of accuracy and time consumption. Moreover,
MobileNetV2 has fewer parameters than DenseNet121, which means more affordable to embed in the mobile device
for live detection.
Although the proposed models in this article have achieved good results for the entire Plant Village dataset, there
are several works to improve the performance:
1. Adding several convolution layers before the classifier layer.
2. Model modifications by unfreezing the initial convolution layers to update low-dimensional weight or ensembling
models.

4. Conclusion
Fine-tuning by unfreezing the final convolution layer blocks can significantly improve the performance of pre-
trained 10 CNN models. From the results of our experiments, the best performance was obtained by DenseNet121 and
MobileNetV2 models of 98.97% and 98.95%, respectively. This result increases the average performance of CNN
models by 8.85%. Meanwhile, there was a drastic decrease of 50% for each model in time consumption, or 2.20 hours
on average. Further research can be developed with architectural modifications by adding several convolution layers
before the classifier layer, unfreezing the initial convolution layers to update low-dimensional weight, and ensembling
models.

Acknowledgment
Thank you to Lembaga Penelitian dan Pengabdian Masyarakat. (LPPM) University of Muhammadiyah Gresik
(UMG). This reseach article supported by LPM UMG Nomor: 066/MoU.In/II.3.UMG/DPPM/F/2021 for Internal Research
Funded.

References
[1] M. Turkoglu, B. Yanikoğlu, and D. Hanbay, “PlantDiseaseNet: convolutional neural network ensemble for plant disease and pest detection,”
Signal, Image and Video Processing, 2021. https://doi.org/10.1007/s11760-021-01909-2
[2] N. Ahmad, S. Asghar, and S. A. Gillani, “Transfer learning-assisted multi-resolution breast cancer histopathological images classification,”
Visual Computer, 2021. https://doi.org/10.1007/s00371-021-02153-y
[3] S. Kaur, S. Pandey, and S. Goel, “Plants Disease Identification and Classification Through Leaf Images: A Survey,” Archives of Computational
Methods in Engineering, vol. 26, no. 2, pp. 507–530, Apr. 2019. https://doi.org/10.1007/s11831-018-9255-6
[4] S. Sachar and A. Kumar, “Survey of feature extraction and classification techniques to identify plant through leaves,” Expert Systems with
Applications, vol. 167. Elsevier Ltd, Apr. 01, 2021. https://doi.org/10.1016/j.eswa.2020.114181
[5] V. Singh, N. Sharma, and S. Singh, “A review of imaging techniques for plant disease detection,” Artificial Intelligence in Agriculture, vol. 4, pp.
229–242, 2020. https://doi.org/10.1016/j.aiia.2020.10.002
[6] J. Lu, L. Tan, and H. Jiang, “Review on convolutional neural network (CNN) applied to plant leaf disease classification,” Agriculture
(Switzerland), vol. 11, no. 8. MDPI AG, Aug. 01, 2021. https://doi.org/10.3390/agriculture11080707
[7] A. Abade, P. A. Ferreira, and F. de Barros Vidal, “Plant diseases recognition on images using convolutional neural networks: A systematic
review,” Computers and Electronics in Agriculture, vol. 185. Elsevier B.V., Jun. 01, 2021. https://doi.org/10.1016/j.compag.2021.106125
[8] R. Manavalan, “Automatic identification of diseases in grains crops through computational approaches: A review,” Computers and Electronics
in Agriculture, vol. 178. Elsevier B.V., Nov. 01, 2020. https://doi.org/10.1016/j.compag.2020.105802
[9] Z. Iqbal, M. A. Khan, M. Sharif, J. H. Shah, M. H. ur Rehman, and K. Javed, “An automated detection and classification of citrus plant diseases
using image processing techniques: A review,” Computers and Electronics in Agriculture, vol. 153. Elsevier B.V., pp. 12–32, Oct. 01, 2018.
https://doi.org/10.1016/j.compag.2018.07.032
[10] L. C. Ngugi, M. Abelwahab, and M. Abo-Zahhad, “Recent advances in image processing techniques for automated leaf pest and disease
recognition – A review,” Information Processing in Agriculture, vol. 8, no. 1. China Agricultural University, pp. 27–51, Mar. 01, 2021.
https://doi.org/10.1016/j.inpa.2020.04.004
[11] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” in Advances in Neural
Information Processing Systems 25, 2012, pp. 1–9.
[12] C. Szegedy et al., “Going Deeper with Convolutions,” Sep. 2014.
[13] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” Sep. 2014.
[14] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” Dec. 2015.
[15] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,” Dec. 2015.
[16] G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely Connected Convolutional Networks,” Aug. 2016.
[17] F. Chollet, “Xception: Deep Learning with Depthwise Separable Convolutions,” 2017.
[18] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” Jan. 2018.
[19] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi, “Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,” Feb.
2016.
[20] B. Zoph, V. Vasudevan, J. Shlens, and Q. v. Le, “Learning Transferable Architectures for Scalable Image Recognition,” Jul. 2017.
[21] M. Tan and Q. v Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” 2019.
© 2022 The Authors. Published by Universitas Muhammadiyah Malang
This is an open access article under the CC BY SA license. (https://creativecommons.org/licenses/by-sa/4.0/)
Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control 195
[22] M. Khanramaki, E. Askari Asli-Ardeh, and E. Kozegar, “Citrus pests classification using an ensemble of deep learning models,” Computers and
Electronics in Agriculture, vol. 186, Jul. 202. https://doi.org/10.1016/j.compag.2021.106192
[23] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep learning for image-based plant disease detection,” Frontiers in Plant Science, vol.
7, no. September, Sep. 2016. https://doi.org/10.3389/fpls.2016.01419
[24] S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D. Stefanovic, “Deep Neural Networks Based Recognition of Plant Diseases by Leaf
Image Classification,” Computational Intelligence and Neuroscience, vol. 2016, 2016. https://doi.org/10.1155/2016/3289801
[25] J. Ma, K. Du, F. Zheng, L. Zhang, Z. Gong, and Z. Sun, “A recognition method for cucumber diseases using leaf symptom images based on
deep convolutional neural network,” Computers and Electronics in Agriculture, vol. 154, pp. 18–24, Nov. 2018.
https://doi.org/10.1016/j.compag.2018.08.048
[26] S. Zhang, S. Zhang, C. Zhang, X. Wang, and Y. Shi, “Cucumber leaf disease identification with global pooling dilated convolutional neural
network,” Computers and Electronics in Agriculture, vol. 162, pp. 422–430, Jul. 2019. https://doi.org/10.1016/j.compag.2019.03.012
[27] C. Wang, P. Du, H. Wu, J. Li, C. Zhao, and H. Zhu, “A cucumber leaf disease severity classification method based on the fusion of DeepLabV3+
and U-Net,” Computers and Electronics in Agriculture, vol. 189, Oct. 2021. https://doi.org/10.1016/j.compag.2021.106373
[28] H. T. Rauf, B. A. Saleem, M. I. U. Lali, M. A. Khan, M. Sharif, and S. A. C. Bukhari, “A citrus fruits and leaves dataset for detection and
classification of citrus diseases through machine learning,” Data in Brief, vol. 26, Oct. 2019. https://doi.org/10.1016/j.dib.2019.104340
[29] U. Barman, R. D. Choudhury, D. Sahu, and G. G. Barman, “Comparison of convolution neural networks for smartphone image based real time
classification of citrus leaf disease,” Computers and Electronics in Agriculture, vol. 177, Oct. 2020.
https://doi.org/10.1016/j.compag.2020.105661
[30] R. Gandhi, S. Nimbalkar, N. Yelamanchili, and S. Ponkshe, “Plant Disease Detection Using CNNs and GANs as an Augmenting Approach,”
May 2018.
[31] E. C. Too, L. Yujian, S. Njuki, and L. Yingchun, “A comparative study of fine-tuning deep learning models for plant disease identification,”
Computers and Electronics in Agriculture, vol. 161, pp. 272–279, Jun. 2019. https://doi.org/10.1016/j.compag.2018.03.032
[32] G. Geetharamani, P. J. Arun, M. Agarwalc, and S. K. Gupta, “Identification of plant leaf diseases using a nine-layer deep convolutional neural
network,” Computers and Electrical Engineering, vol. 76, pp. 323–338, Jun. 2019. https://doi.org/10.1016/j.compeleceng.2019.04.011
[33] Ü. Atila, M. Uçar, K. Akyol, and E. Uçar, “Plant leaf disease classification using EfficientNet deep learning model,” Ecological Informatics, vol.
61, Mar. 2021. https://doi.org/10.1016/j.ecoinf.2020.101182
[34] S. Vallabhajosyula, V. Sistla, and V. K. K. Kolli, “Transfer learning-based deep ensemble neural network for plant leaf disease detection,”
Journal of Plant Diseases and Protection, 2021. https://doi.org/10.1007/s41348-021-00465-8
[35] I. Ahmad, M. Hamid, S. Yousaf, S. T. Shah, and M. O. Ahmad, “Optimizing pretrained convolutional neural networks for tomato leaf disease
detection,” Complexity, vol. 2020, 2020. https://doi.org/10.1155/2020/8812019
[36] J. Chen, D. Zhang, and Y. A. Nanehkaran, “Identifying plant diseases using deep transfer learning and enhanced lightweight network,”
Multimedia Tools and Applications, vol. 79, no. 41–42, pp. 31497–31515, Nov. 2020. https://doi.org/10.1007/s11042-020-09669-w
[37] J. Chen, J. Chen, D. Zhang, Y. Sun, and Y. A. Nanehkaran, “Using deep transfer learning for image-based plant disease identification,”
Computers and Electronics in Agriculture, vol. 173, Jun. 2020. https://doi.org/10.1016/j.compag.2020.105393
[38] V. Tiwari, R. C. Joshi, and M. K. Dutta, “Dense convolutional neural networks based multiclass plant disease detection and classification using
leaf images,” Ecological Informatics, vol. 63, Jul. 2021. https://doi.org/10.1016/j.ecoinf.2021.101289
[39] R. Gajjar, N. Gajjar, V. J. Thakor, N. P. Patel, and S. Ruparelia, “Real-time detection and identification of plant leaf diseases using convolutional
neural networks on an embedded platform,” Visual Computer, 2021. https://doi.org/10.1007/s00371-021-02164-9
[40] J. Chen, D. Zhang, M. Suzauddola, and A. Zeb, “Identifying crop diseases using attention embedded MobileNet-V2 model,” Applied Soft
Computing, vol. 113, Dec. 2021. https://doi.org/10.1016/j.asoc.2021.107901
[41] E. Prasetyo, N. Suciati, and C. Fatichah, “Multi-level residual network VGGNet for fish species classification,” Journal of King Saud University
- Computer and Information Sciences, 2021. https://doi.org/10.1016/j.jksuci.2021.05.015
[42] V. Kumar, H. Arora, Harsh, and J. Sisodia, “ResNet-based approach for detection classification of plant leaf diseases,” in Proceedings of the
International Conference on Electronics and Sustainable Communication Systems (ICESC 2020), 2020, pp. 495–502.
https://doi.org/10.1109/ICESC48915.2020.9155585
[43] R. Sujatha, J. M. Chatterjee, N. Z. Jhanjhi, and S. N. Brohi, “Performance of deep learning vs machine learning in plant leaf disease detection,”
Microprocessors and Microsystems, vol. 80, Feb. 2021. https://doi.org/10.1016/j.micpro.2020.103615
[44] Ü. Atila, M. Uçar, K. Akyol, and E. Uçar, “Plant leaf disease classification using EfficientNet deep learning model,” Ecological Informatics, vol.
61, Mar. 2021. https://doi.org/10.1016/j.ecoinf.2020.101182
[45] V. Tiwari, R. C. Joshi, and M. K. Dutta, “Dense convolutional neural networks based multiclass plant disease detection and classification using
leaf images,” Ecological Informatics, vol. 63, Jul. 2021. https://doi.org/10.1016/j.ecoinf.2021.101289
[46] M. A. Moid and M. A. Chaurasia, “Transfer Learning-based Plant Disease Detection and Diagnosis System using Xception,” in Proceedings of
the 5th International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud), I-SMAC 2021, 2021, pp. 451–455.
https://doi.org/10.1109/I-SMAC52330.2021.9640694

Cite: D. Sutaji and H. Rosyid, “Convolutional Neural Network (CNN) Models for Crop Diseases Classification”, KINETIK, vol. 7, no. 2, May.
2022. https://doi.org/10.22219/kinetik.v7i2.1443
196 Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control

© 2022 The Authors. Published by Universitas Muhammadiyah Malang

This is an open access article under the CC BY SA license. (https://creativecommons.org/licenses/by-sa/4.0/)