R SWATHI and K SWASTHIKA: DEEP LEARNING-BASED PLANT CROP DISEASE DETECTION USING CNN

DOI: 10.21917/ijdsml.2024.0136

DEEP LEARNING-BASED PLANT CROP DISEASE DETECTION USING CNN

R. Swathi and K. Swasthika
Department of Computer Science and Engineering, Amal Jyothi Engineering College, India

Abstract a large yield, despite the fact that tomatoes are among the crops
The smart farming is to deliver solutions that are revolutionary to the that are most important to human life. On the other hand, there are
question of how humankind can continue to exist in a sustainable diseases that affect tomatoes that can significantly reduce your
manner over the long stretch of time. Identification of the recorded harvest. The significance of correctly classifying the various
image is absolutely necessary in order to monitor the development of diseases is brought into focus by this phenomenon. Because of the
the plant and protect it from various diseases and pests. Currently, the abundance of tomatoes in Beijing, which is not generally the case
objective of automatic disease recognition is to conduct research on
in larger cities, the city was chosen to examine ailments that are
crop diseases through the use of deep learning. However, existing
classifiers have problems with a variety of challenges, including the typically associated with tomatoes. Investigations of tomato
identification of appropriate disease categories, among other things. illnesses that concentrate on particular regions within the same
This page is dedicated to the disease that specifically affects tomatoes city can provide helpful direction. This is because of the tight
as a crop, which is known as tomato disease. The purpose of this association that exists between the environment and the growth of
research is to improve the structure of tomato plant photographs for tomatoes.
the purpose of image identification. Because of this, the process of Farmers who have not had expertise in plant pathology have a
extracting features from photographs of plants is more effective and
precise than the approach that is typically taken in artificial
difficult time manually identifying the many different diseases
recognition. Using three separate sets of photographs recorded by a and pests that affect plants. As a means of addressing genuine
camera and a drone, the effectiveness of the proposed architecture was production difficulties brought about by the exponential increase
evaluated. These images were taken in three different environments of information and intellect, artificial intelligence technology has
where tomatoes are grown. Taking into consideration the statistics, this found extensive use in agriculture. This is because farming is one
method of counting articles achieves an accuracy rate of approximately of the most important industries in the world. Convolutional
96.30% on average. The decision-making process in precision neural networks and deep learning approaches have transformed
agriculture is aided by the scientific support and reference it receives. object recognition across the board as a result of their great
performance in resolving challenging computer vision
Keywords: difficulties. To be more specific, this revolution has occurred
Leaf Disease, CNN, MobileNets. when these techniques were applied. Studies that fit into the
machine learning (ML) and deep learning (DL) categories are the
1. INTRODUCTION two primary categories that precision agricultural research takes
into consideration. Prior to the examination of the sickness, the
machine learning-based technique [5,6,7] typically requires the
Precision and smart agriculture, which involves regulating
utilization of complex prediction and the extraction of disease-
crop planting acreage, is now the foundational agronomic
specific features. A significant amount of interest has been shown
technique for boosting food production, security, and
in deep learning picture identification systems over the course of
environmental preservation within the framework of sustainable
time, with a particular emphasis on the identification of
agriculture. Such an approach is a part of the framework of
agricultural diseases. Deep convolutional neural networks, also
sustainable agriculture. If experts and agriculturists use cutting-
known as DCNNs, function exceptionally well when it comes to
edge AI approaches rather than the labor-intensive traditional site
the diagnosis of diseases based on photographs. With the use of a
monitoring, they will be able to save time and effort [3].
machine vision system, it is possible to gain a better
Consequently, the development of intelligent applications and
understanding of plant diseases and crops can be safeguarded
efficient methods of data processing are essential to the success
more effectively. The problem of tomato disease picture
of precision agriculture applications. Precision agriculture, which
characteristics has been addressed by a number of deep
in turn relies on a wide range of technologies for things like
convolutional neural network (DCNN) models, such as NasNet
disease detection, data processing, data analysis, and sensor
[8, faster-RCNN [9], SSD method [10], mask-RCNN [11], and
deployment, is a fundamental component of precision agriculture.
EfficientNet [12]. These models have been applied for the
One of the most important aspects of precision agriculture is the
purpose of accurately identifying plant diseases. DCNN models
optimal use of agricultural inputs. Farmers have the ability to
perform exceptionally well when it comes to the extraction of
make progress toward smart agriculture, which is a form of
local characteristics, but they fail severely when it comes to the
agriculture that includes innovative technology to improve
capture of global features.
product quality, through the utilization of deep learning-based
intelligence technologies. Due to the fact that this limitation exists, vision transformer
(ViT) [13] is being utilized increasingly frequently. When
Recently, there have been advancements in visual algorithms
compared to older models, it performs better in terms of the
that have made it possible to utilize computer vision to evaluate
extraction of global features. By immediately applying an encoder
leaf disease independently of farmers. This has made it possible
unit to sequences of picture patches, ViT is able to successfully
to safeguard crops in a more precise and timely manner. It is
complete tasks involving picture classification [14]. The ability of
challenging to cultivate tomatoes in such a way that they produce
ViT to obtain global contextual information is the primary benefit

650
ISSN: 2583-9292 (ONLINE) ICTACT JOURNAL ON DATA SCIENCE AND MACHINE LEARNING, JUNE 2024, VOLUME: 05, ISSUE: 03

of using this technology. This capability enables the construction light, noise, and other distracting aspects. On the other hand,
of a long-distance dependency on the features that are being images that have a complex background contain elements such as
targeted. the sky, soil, countless leaves, and other backdrops.

2. RELATED WORKS
3. PROPOSED METHOD
After utilizing Kapur's thresholding to differentiate between
the damaged and healthy parts of leaves and doing parameter In order to address CNN's shortcomings in terms of extracting
searches with an emperor penguin optimizer approach, global picture aspects, a solution that is recommended is the use
Ashwinkumar et al. [9] were ultimately able to achieve the best of Transformer, which has the potential to generate relationships
model with a recognition accuracy of 98.5%. The minimally between distant qualities. This indicates that the processing
parameterized Reduced MobileNet that Kamal et al. [10] efficiency of the improved model as well as the maximum number
proposed was able to manage latency and performance in an of parameters will be fine-tuned. Concurrently, the inductive bias
effective manner. Ji et al. [11] combined the features of Inception- observed in CNN will prove to be beneficial to the global features
width V3 with the depth features of ResNet-50 in order to achieve that Transformer extracts. The work that was done in this area
a better representation of the target features. They put their resulted in the development of improved ways for efficiently
algorithm through its paces by testing it on grape datasets that recognizing sick leaves in highly complex scenarios.
contained four distinct diseases. Through the incorporation of
For the purpose of evaluating the performance of the various
batch normalization and global pooling into AlexNet, Sun et al.
models, we utilized the public dataset provided by Plant Village.
[12] were able to develop a novel model that demonstrated rapid
This dataset contains both healthy and damaged leaves from
convergence. As part of their research, Too et al. [13] improved
fourteen different types of crops. In addition, two additional
and evaluated six different models. Out of all of them, DenseNets-
datasets on crop leaf diseases with complex backgrounds were
121 had parameter quantities and running periods that were more
generated. These datasets took into consideration the problem of
realistic, and training prevented both performance degradation
severe environmental disturbance in the process of disease
and overfitting. Zhao et al. [14] used a pretrained model to train a
detection in the field. Leaf samples and Kaggle were the sources
new model utilizing cotton datasets, which helped them to
of the original images that were used for the dataset, which
completely avoid the issue of overfitting. This allowed them to
currently include images of apple, cassava, and cotton. There are
eliminate the problem of overfitting. The author of [15] selected
only 6,891 photographs included in the initial dataset, and their
VGG16 for training after analyzing the benefits and drawbacks of
distribution is not consistent. In addition to causing overfitting
CNN and traditional machine learning. They also investigated the
and poor generalization performance in the models, this would
impact that multiple-layer feature extraction and transfer learning
also lead categories with a high number of items to accumulate
had on recognition performance. When it comes to the results, the
more mistakes in consecutive iterative training. This would be the
model that they provided serves them exceptionally well. On the
case because of the combination of the two of these factors. Using
other hand, they did not take into account the association between
training data from the same source to identify crop illnesses in
the aspects of the disease and instead directed their attention only
different regions is troublesome since it could lead to poor
toward the ways in which the model structures influenced the
generalization performance of the final model. Identification of
results. Following the realization that the generalization
agricultural diseases in multiple regions is problematic. To add
performance of the model was of utmost importance, Mohanty et
insult to injury, we need to investigate the existing data in order
al. put it through its paces by utilizing other disease images that
to obtain the images that correspond to the diseases that affect
were comparable, and the results were to their satisfaction. On the
crops, which is difficult to do without the support of specialists.
other hand, [16] simply entered the images into the model and
As a result, we will need to make use of background replacement
obtained the results without conducting any more research on the
technologies in order to generate new data from existing data. The
characteristics of the condition. In addition, the recognition
backdrop replacement technique is able to simulate a wide variety
accuracy to parameter ratio in the study that Mohanty and his
of recognition scenarios in a large number of different
colleagues did was not adequate. Following the completion of
environments since it makes use of the existing data to replace the
exhaustive dataset preparation operations, [17] proposed a neural
image backgrounds. The experiment is carried out in a manner
architecture search network and achieved remarkable results with
that is both accurate and effective.
it. The adaptability of the network is insufficient, and it takes
some time to determine the parameters that are best. 3.1 MOBILENET-V2
Despite the fact that the research on crop disease recognition
using convolutional neural networks has shown encouraging In order to improve the depth of the model, convolutional
results, the datasets that were used in these studies only included neural networks (CNNs) were initially trained to learn target
one leaf that was damaged and a basic background [18]-[22]. properties at a variety of abstract levels. This was accomplished
CNN's inability to collect and learn sufficient disease information by continuously stacking convolution and pooling layers. By way
from these datasets has a substantial influence on the models' of illustration, the initial 18 layers of ResNet could be expanded
capacity to generalize in a satisfactory manner. As a consequence to fifty, one hundred and ten, or even 152 with the assistance of
of this, researchers have been gradually turning their focus to leaf residual connections. Although extending the receptive field size
disease data in complex situations and backgrounds for the through the use of stacked layer design can result in an increase
purpose of their subsequent investigation. Images are deemed to in the pixel range, doing so comes at the expense of an increase in
have difficult surroundings when they have different amounts of the computational cost and the quantity of model parameters,

651
R SWATHI and K SWASTHIKA: DEEP LEARNING-BASED PLANT CROP DISEASE DETECTION USING CNN

neither of which are characteristics that are conducive to disease and 99.22%, respectively, when compared to the methodologies
field detection. It is imperative that farmers continue to make use that were previously used, as shown in Table.1. Having said that,
of mobile devices in order to detect diseases as they emerge. This the graphic composition of Plant Village is unimpressive, and as
is because infections are sometimes difficult to detect and crop a result, it will not be a reliable reference for the diagnosis of
farming is a labor-intensive process. On the other hand, because actual illnesses. As a result, in the following section, we will study
mobile devices have a limited capacity for processing, it is and find solutions to the difficulties that we have faced in the task
difficult to adapt traditional large models. Intelligent disease of disease recognition by employing images of crop leaf diseases
recognition is built on the principle of lightweightedness, which that are set against complex backgrounds as our research object.
serves as its foundation. MobileNet-V2, a lightweight application It is possible for a leaf ailment to present itself with a wide
that can be deployed on mobile devices, was released in 2019. To range of symptoms, some of which are easily confused with those
get things started, MobileNet-V2 continued the depthwise of other diseases that have occurred at various times. Brown
separable convolution that was implemented in MobileNet-V1. streak disease of cassava is characterized by late symptoms that
This convolution was designed to speed up the operation of the are similar to those of cassava bacterial blight. These late signs
model by lowering the number of convolutional kernels. In the include tawny leaf markings and, in many instances, wilted
second place, the Inverted Residual Block (IRB, rising first and leaves. The indicators of mosaic virus infection, which include
then descending) was proposed as a solution. This was done by yellowing and curling, are so similar to those of the first two
taking a page out of the playbook of the traditional bottleneck diseases that damage cassava leaves that it is difficult to
layer. Not only did this architecture significantly reduce the differentiate between them. This is because the signs of mosaic
amount of memory that was used during the process of model virus infection include yellowing and curling. A misclassification
inference, but it also ensured that the Depthwise Convolution of CNN models is another consequence that will result from the
(DWConv) layer that was contained within the IRB was able to scenario described above. One of the reasons for this problem is
capture all of the rich feature information. Converting MobileNet- that diseases that affect leaves that belong to the same category
V2's nonlinear activation function ReLU6 into a linear function have a lot of similarities with one another in terms of color, shape,
was ultimately the solution to the problem of feature loss that and other characteristics. To put it another way, there are not
occurred during feature compression. This allows for the particularly many variances across the many groups of diseases;
preservation of the diversity of characteristics while nevertheless, there are numerous differences within each class. As
simultaneously boosting the ability of target features to express a result of the fact that, prior to optimization, the distribution of
themselves. disease features recovered by CNN is sparse and there is a
considerable intersection among samples of multiple categories,
4. RESULTS AND DISCUSSION it is highly likely that these samples will be misclassified in the
upcoming recognition task. After optimization, features that are a
With the intention of avoiding local optimization, we decided part of the same cluster migrate closer to the center of that cluster,
to set the attenuation coefficient of the learning rate to 0.8. When while the distance between different clusters grows. This is
this occurs, it indicates that the learning rate will decrease to 80% happening while the distance between the clusters is increasing. It
of its starting value once ten epochs have passed. A 64-bit system is evident, on the basis of the comparative results, that the
environment running Ubuntu 18.04 LTS was utilized for the incorporation of Centerloss has not only reduced the previously
execution of each and every one of the experiments. Since Pytorch dispersed distribution of disease symptoms, but it has also greatly
1.6 is compatible with GPU acceleration and can be utilized with increased the degree to which related characteristics can be
dynamic neural networks, it was selected as the Python distinguished from one another.
programming language. CUDA 9.1 was another tool that was
helpful in the training process. 32 gigabytes of random access 5. CONCLUSION
memory (RAM) and an NVIDIA GeForce GTX 2080Ti graphics
card equip the system. Based on the responsibilities of field agricultural disease
The recognition accuracies achieved by the revised model on recognition, the datasets that were utilized in this study were more
three different versions of Plant Village were 99.62%, 99.08%, in line with the actual production demands.

Table.1. Performance Analysis

Accuracy Detection False Response Run

Method Dataset F-measure
(%) Rate (%) Positive Rate (%) Time (ms) Time (s)
Training 92.5 0.915 3.2 35 250 0.915
ShuffleNet-V2 Testing 91.2 0.902 3.5 38 255 0.902
Validation 91 0.9 3.7 37 252 0.9
Training 93.8 0.93 2.8 32 220 0.93
DenseNets-121 Testing 92.6 0.921 3 34 225 0.921
Validation 92.4 0.919 3.2 33 223 0.919
Inception-V3 Training 94.2 0.935 2.5 30 200 0.935

652
ISSN: 2583-9292 (ONLINE) ICTACT JOURNAL ON DATA SCIENCE AND MACHINE LEARNING, JUNE 2024, VOLUME: 05, ISSUE: 03

Testing 93.1 0.925 2.7 32 205 0.925

Validation 92.9 0.923 2.9 31 203 0.923
Training 95.5 0.95 2 28 180 0.95
Proposed
Testing 94.6 0.94 2.3 29 185 0.94
Classification
Validation 94.4 0.938 2.5 28 182 0.938

Since crop disease characteristics tend to manifest in complex [3] R. Kohavi and G.H. John, “Wrappers for Feature Subset
environments with characteristics such as uneven distribution and Selection”, Artificial Intelligence, Vol. 97, No. 1-2, pp. 273-
wide distribution regions, we investigated the shortcomings of 324, 1997.
MobileNet-V2 and improved the model in such a way that it [4] A. Jain and D. Zongker, “Feature Selection: Evaluation,
achieves a good balance between the recognition accuracy and the Application, and Small Sample Performance”, IEEE
parameter quantity. As a result of the incorporation of Transactions on Pattern Analysis and Machine Intelligence,
Transformer Encoder, the improved model was able to more Vol. 19, No. 2, pp. 153-158, 1997.
effectively extract global sickness characteristics and focus its [5] Shivnath Ghosh and Santanu Koley, “Machine Learning for
attention on the regions that were specifically impacted. Through Soil Fertility and Plant Nutrient Management using Back
the utilization of centerloss, which is founded on the concept of Propagation Neural Networks”, International Journal on
cross-entropy loss, the automatic clustering of sample Recent and Innovation Trends in Computing and
characteristics toward the feature center of their respective Communication, Vol. 2, No. 2, pp. 292-297,2014.
categories was made possible, hence enhancing the separability of [6] R.C. Gonzalez, R. E. Woods and S. L. Eddins, “Digital
specific sickness characteristics. On Plant Village, the improved Image Processing using Matlab”, Prentice-Hall, 2003.
model achieved a recognition accuracy score of 99.62%. In spite [7] X. Yang and T. Guo, “Machine Learning in Plant Disease
of the fact that it was presented with complex background Research”, European Journal of Biomedical Research, Vol.
interference in Dataset 1, it surpassed other models with an 3, No. 1, pp. 6-9, 2017.
accuracy of 96.58%. The updated model used in this study [8] Eibe Frank, Mark Hall, Len Trigg, Geoffrey Holmes and Ian
demonstrated a high degree of generalizability, as evidenced by H. Witten, “Data Mining in Bioinformatics using Weka”,
its recognition accuracy of 95.03% in Dataset2, 94.03% in Bioinformatics, Vol. 20, No. 15, pp. 2479-2481, 2004.
Dataset1, and 96.39% in Dataset2. In terms of recognition [9] H. Benjamin Fredrick David and S. Antony Belcy, “Heart
accuracy, the improved model surpasses rival models while Disease Prediction using Data Mining Techniques”,
simultaneously utilizing a smaller number of parameters. In ICTACT Journal on Soft Computing, Vol. 9, No. 1, pp. 1817-
conclusion, the improved model reported in this research not only 1823, 2018.
provides superior crop leaf disease classification even when [10] Vaneesbeer Singh and Abid Sarwar, “Analysis of Soil and
presented with complex backgrounds, but it also proposes Prediction of Crop Yield (Rice) using Machine Learning
approaches to transfer deep learning models to mobile devices for Approach”, International Journal of Advanced Research in
the purpose of disease detection. Computer Science, Vol. 8, No. 5, pp. 1254-1259, 2017.
Despite the fact that it is more significant and difficult to [11] K.P. Satamraju, K. Shaik and N. Vellanki, “Rural Bridge: A
identify illnesses at an earlier stage, the majority of the approaches Novel System for Smart and Co-Operative Farming using
that are now used to identify crop diseases concentrate on sick IoT Architecture”, Proceedings of International Conference
leaves. Although the patches are thinner and less apparent when on Multimedia, Signal Processing and Communication
the disease is in its early stages, they are still present. Prior to the Technologies, pp. 22-26, 2017.
creation of sickness spots, the RGB-based image recognition [12] H.B.F. David and A. Suruliandi, “Empirical Study of
algorithm is unable to distinguish this particular type of Ensemble Classifications on Benchmark Datasets”, Journal
unwellness image. In further research, it is conceivable to of Analysis and Computing, Vol. 12, No. 2, pp. 1-14, 2018.
incorporate multimodal images of crop diseases into deep [13] Ginne M James, S.C. Punitha, “Tomato Disease
learning in order to accomplish early disease identification. This Classification using Ensemble Learning Approach”,
would allow for effective disease detection. International Journal of Research in Engineering and
Technology, Vol. 5, No. 10, pp. 104-108, 2016.
[14] U. Mokhtar, M.A.S Ali and A.E. Hassanien, “Identifying
REFERENCES
Two of Tomatoes Leaf Viruses Using Support Vector
Machine”, Advances in Intelligent Systems and Computing,
[1] Kirsi Juottonen, Mikko P. Laakso, Kaarina Partanen and
Vol 339, pp. 1-19, 2015.
Hilkka Soininen, “Comparative MR Analysis of the
[15] S. Jeyalakshmi, R. Radha, “A Novel Approach to Segment
Entorhinal Cortex and Hippocampus in Diagnosing
Leaf Region from Plant Leaf Image using Automatic
Alzheimer Disease”, American Journal of Neuroradiology,
Enhanced Grabcut Algorithm”, International Journal of
Vol. 20, No. 1, pp. 139-144, 1999.
Advanced Computer Technology, Vol. 8, No. 11, pp. 3485-
[2] R.L. Simoes, “Towards Earlier Detection of Alzheimer’s
3493, 2018.
Disease using Magnetic Resonance Images”, PhD
[16] H. Pourazar, F. Samadzadegan and F.D. Javan, “Aerial
Dissertation, Department of Computer Science, University
Multispectral Imagery for Plant Disease Detection:
of Twente, pp. 1-125, 2013.

653
R SWATHI and K SWASTHIKA: DEEP LEARNING-BASED PLANT CROP DISEASE DETECTION USING CNN

Radiometric Calibration Necessity Assessment”, European [20] Rice Leaf Disease, Available at
Journal of Remote Sensing, Vol. 52, No. 3, pp. 17-31, 2019. https://www.kaggle.com/vbookshelf/rice-leaf-diseases,
[17] Canny Edge Detection, Available at Accessed in 2021.
https://www.cse.iitd.ac.in/~pkalra/col783-2017/canny.pdf, [21] S. Iniyan, R. Jebakumar, P. Mangalraj, M. Mohit and A.
Accessed in 2009. Nanda, “Plant Disease Identification and Detection using
[18] P. Sharma, Y.P.S. Berwal and W. Ghai, “Performance Support Vector Machines and Artificial Neural Networks”,
Analysis of Deep Learning CNN Models for Disease Proceedings of International Conference on Advances in
Detection in Plants using Image Segmentation”, Intelligent Systems and Computing, pp. 15-27, 2020.
Information Processing in Agriculture, Vol. 6, pp. 2214- [22] G. Owomugisha, E. Nuwamanya, J.A. Quinn, M. Biehl and
3173, 2019. E. Mwebaze, “Early Detection of Plant Diseases using
[19] Plant Village, Available at Spectral Data”, Proceedings of International Conference on
https://www.kaggle.com/abdallahalidev/plantvillagedataset Electrical and Computer Engineering, pp. 1-13, 2020.
/version/1, Accessed in 2021.

654