2024 4th International Conference on Pervasive Computing and Social Networking (ICPCSN)

2024 4th International Conference on Pervasive Computing and Social Networking (ICPCSN) | 979-8-3503-8634-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICPCSN62568.2024.00165

IMAGE-BASED PEST DETECTION AND

IDENTIFICATION SYSTEM FOR
AGRICULTURE
Sai Bhargav Vemulapalli Durga Prasada Rao Pillagolla Ganesh Vempala
Dept. of EIE Dept. of EIE Dept. of EIE
VR Siddhartha Engineering College VR Siddhartha Engineering College VR Siddhartha Engineering College
Vijayawada, India Vijayawada, India Vijayawada, India
[email protected] [email protected] [email protected]

Sri Mahalakshmi Ravuri

Dept. of EIE
VR Siddhartha Engineering College
Vijayawada, India
[email protected]

Abstract— For farmers, identifying pests is the most difficult

task in the agricultural industry. Farmers must use organic
insecticides in conjunction with other appropriate techniques to
combat pests. This project outlines a software prototype system
that uses pest identification to control pests. Farmers must use
the Android application to take an image of the pest. The pest
image must then be uploaded to the software. One of the most
difficult jobs for farmers and crop professionals in the
agricultural industry is identifying pests over crops. This harms
farmers as well as crops, resulting in a low output. The pests'
image database is also taken into account. A comparison is made
between a set of training and testing images. The convolutional
neural network classification (CNN) approach is adopted to
identify the class of Pests.
Fig. 1. Consumption of Pesticides in India(in%) [1]
Keywords—CNN approach, Farmers, Pests,
Fig. 1 depicts the decline of agriculture as a result of the
I. INTRODUCTION use of chemical pesticides. Using chemical insecticides on
crops without first identifying the pest results in decreased
India is an agricultural country. Farming is a source of crop growth. Identifying pests and using appropriate organic
income for many individuals. Agriculture is also vital to a pesticides in agriculture is critical to avoiding losses in farmer
country's economy. Farmers are rural residents. They rely yields and food quality. Because it is difficult to manually
only on agriculture for revenue. It accounts for 17.7% of gross identify individual pests and administer effective organic
domestic product. It will assist the nation by tackling pesticides, Digital Image Processing is used for early pest
unemployment. Crops are affected by some pests, including detection [2]. This approach also involves image acquisition,
bacteria, viruses, and fungi. As a result, agricultural yields pre-processing, segmentation, feature extraction, and
suffer from both quality and quantity issues. Crop quality and classification. Systems are complex computer programs that
quantity may therefore be regulated without using chemical may provide answers and guidance to specific problems in a
pesticides. This is accomplished by the use of organic given area by comparing photographs to databases.
pesticides. Organic insecticides are better since they kill pests
without damaging plants while increasing quality and quantity. Image-based pest detection and identification
technologies provide a long-term and efficient answer to these
difficulties. Pest detection and identification in agriculture
have traditionally depended mainly on physical labor, which
may be time-consuming and error-prone. However, with the
development of low-cost and accessible photography
equipment, such as smartphones and drones, as well as rapid
advances in artificial intelligence, the process may now be
automated and enhanced. Image-based pest detection and
identification technologies provide a more efficient, cost-
effective, and accurate way to monitor and manage pests in
agricultural settings[3].

979-8-3503-8634-9/24/$31.00 ©2024 IEEE 989

DOI 10.1109/ICPCSN62568.2024.00165
Authorized licensed use limited to: BLDEA's College of Eng & Tech - Vijayaura. Downloaded on March 10,2025 at 11:29:30 UTC from IEEE Xplore. Restrictions apply.
 II. LITERATURE SURVEY III. EXISTING METHODS
Mayank et al., 2021, [4] proposed a pest control system A. K-Means grouping technique
that uses IoT and Image processing technologies. Effective
An unsupervised machine-learning technique called K-
image classification method for smartphones based on neural
means clustering is used to extract the target area from
networks. Infrared sensors are used by the system to identify
images of plant leaves. Sort the data elements into K-groups
pests. Insects are kept out of the field by pests identified using
according to how similar they are. By computing the absolute
sensors and ultrasonic wave technology with the aid of image
difference between each pixel and the clustering center in the
processing.
color space, the K-means technique uses automatic data
Yaowei et al., 2020, [5] Transfer learning algorithms
grouping (clustering) to separate wheat ailments such as leaf
combined to provide a new method for image recognition and
rust, powdery mildew, and striped rust from images.
classification. The significance of TL-MobileNet for welding
flaw detection is suggested in this paper. This experiment,
B. Edge detection techniques
which classifies welding flaws using "Weld" datasets,
An edge is a barrier with unique characteristics that
confirms that the TL-MobileNet can accurately identify
separates two places. It signifies the change from one thing
particular defects within a restricted set of training sets.
to another. Since the intensity drastically changes at the
Numerous tests have been conducted with different image
boundaries of the area, this detection technique is used to
sizes and transfer learning models. Additionally, it validates
quantify the physical scale of items. Edge detection can be
the suggested TL-MobileNet approach, which is superior in
used to separate images into specific target portions
terms of recognition accuracy and requires less computation
according to criteria and is a crucial step in understanding
time and model size.
image function.
Brian et al.,2020 [6] titled Biological Control and
Integrated Pest Management in Organic and Conventional C. Support Vector Machines “SVM”
Systems. According to this research, IPM and organic The maximum distance between the examples is used to
farming are viable methods for increasing agricultural construct a judgment threshold by the SVM, a non-
productivity, and they both rely on biological control as a tool probabilistic linear classifier. By splitting the data into two
that farmers can employ. reduced risk, use of pesticides, and subspaces with the greatest distance between the data closest
effects on development. Working together, individuals to the hyperplane (referred to as a distinct hyperplane), an
utilizing two different sets of methods can advance the SVM looks for the hyperplane. The margin is the space that
adoption of solutions for production issues that involve exists between the data and the hyperplane. The data points
biological control. closest to the hyperplane are referred to as support vectors.
There are two different types of support vectors: those for one
G. Singh 2020 [7] Project titled Pest management in subset of the data and those for another since the data is
organic farming, becomes a challenging task without using divided into two subgroups. They are referred to as negative
any insecticides. It uses meticulous preparation ahead of time and positive support vectors, respectively, in everyday
together with little adjustments to cultural norms as the main speech. This technique allows for the creation of linear
means of managing pests. employing eco-friendly methods separators without requiring the original input space, but it
as a backup against insect pests, such as the use of biological does not allow for linear definition. Lastly, SVM may
control agents and other bio-based goods. These products can characterize complex functions while resisting overfitting
reduce insect pests and boost the profitable yield of organic because it is a nonparametric method that is not easily
farming whether used alone or in conjunction with other overfitted.
strategies like integrated pest management.
IV. PROPOSED METHODOLOGY
Basri et al. 2020 [8] To identify cocoa fruit pests and The pest datasets used in this study were gathered from
diseases at their earliest stages, this study suggests using the Kaggle website. In a web-based data science
image processing techniques based on mobile applications. environment, users may search and share data sets, as well as
The idea of competence in the field of cacao farming was evaluate and construct models using Kaggle. Various pest
implemented by the system. Pre-processing the image using classes, including grasshoppers, cutworms, Mormon
deep learning algorithms and image processing techniques is crickets, locusts, Japanese beetles, and true bugs, are taken
how it's done for machine learning systems and pattern into consideration. The process flow is shown in Fig. 2.
recognition. To identify the pixels in the photos of cacao
fruit, image processing techniques are included in the
application program. Image Image
Capture Image Pre- Augmentatio
Indhumathi et al., 2019, [9] According to this model, the processing n
testing images and their set of extracted images should be
compared with the geometrical properties of the training
images, which include the mean and variance characteristics Process and
and morphological features. The likelihood and probability give solution Pest order Image
of pests are taken into account. The class of pests is identified for pest prediction Classificatio
using the naive Bayes approach. control n

990

Authorized licensed use limited to: BLDEA's College of Eng & Tech - Vijayaura. Downloaded on March 10,2025 at 11:29:30 UTC from IEEE Xplore. Restrictions apply.
 Fig. 2 Block diagram of the proposed system. training data sets are chosen according to the user's expertise.

A software prototype that helps farmers identify and

successfully manage pest infestations is the center of the
proposed methodology for the image-based pest detection
and identification system for agriculture. An image of a pest
is uploaded to the software system once it is taken with the
Android application. Convolutional neural network (CNN)
classification is one of the advanced image recognition
technologies that the system uses to accurately evaluate and
classify uploaded pest photos. The CNN model learns to
recognize different types of pests that are frequently
encountered in agricultural settings by comparing a
collection of training and testing images, making precise and Fig. 3 Samples images of pest in dataset
effective pest identification possible. E. Convolutional Neural Network (CNN)
A. Image acquisition and pre-processing A component of the Deep Learning technique is a
First, the cropped image was taken, and then it was Convolutional Neural Network (CNN), which can
processed. Pre-processing techniques are then used to distinguish one image from another by using an input image
improve the image by removing unwanted distortion and and assigning learnable weights and biases (importance) to
enhancing the necessary attributes for further processing the image's many elements. A CNN Architecture is
[10]. Image quality improvement techniques are applied in composed of multiple layers that use a differentiable function
image pre-processing to reduce noise and sharpen the image to turn the input volume into an output volume [15].
for greater precision. This process improves image quality to Commonly, a few different kinds of layers are employed. The
help with more accurate identification and classification of CNN models consist of three different types of layers:
insects, pests, and diseases. convolutional, pooling, and fully-connected (FC) layers. A
CNN architecture will be created once these levels are joined.
B. Segmentation
F. Layers in CNN
Segmentation is the following stage of the discovery process.
To extract the target object from the sheet, segmentation is The convolutional layer, pooling layer, and fully-
used for the colorized image. Segmentation methods and connected layer are the layers in the CNN.
Gaussian mixture models are typically used. An exponential
1) Convolutional Layer
increase in execution time is caused by the numerous
One of the most important parts of the construction of a
divisions, multiplications, and computation averages that
convolutional neural network (CNN) is the convolutional
these algorithms require.
layer, also called the CNN layer. CNNs are deep learning
C. Feature extraction models specifically designed for computer vision and image
In image processing, feature extraction is crucial. To recognition applications involving visual input. It is applied
extract characteristics helpful for image recognition and to the input photos to extract various kinds of features.
classification, feature extraction techniques are used [11]. Convolution maintains the spatial link between pixels by
Information that helps differentiate one object from another studying small squares of input images to learn image
is contained in a feature set [12]. Shape feature extraction and properties. Convolution of the input image is achieved by a
texture feature extraction are the two categories into which collection of learnable neurons. In the output image, this
features can be divided. Certain image qualities, like gray creates a feature map or activation map, which is
covariance and region attributes, are taken into account when subsequently used as input data by the following
extracting characteristics. convolutional layer.

D. Pest classification 2) Pooling Layer

All of the image's pixels are categorized into distinct classes The pooling layer often comes after the convolution layer.
based on how similar or different they are in this final This layer's main objective is to minimize computation costs
classification. Supervised and unsupervised classification are by shrinking the convolved feature map. They accomplish
the two categories of classification techniques. this by down-sampling the spatial dimensions of feature
Pixel clusters in unsupervised pest classification, on the maps produced by convolutional layers, which reduces
other hand, are created by software analysis of an image computational processing while preserving important
rather than training data sets supplied by the user [13]. The information. Pooling layers help the network become more
computer uses methods to classify the pixels based on their resilient to changes in input data and more computationally
relationships. The concept of supervised categorization relies efficient. There are two categories for pooling layers: average
on the user's ability to choose representative pixels from an pooling and maximum pooling.
image that serves as a lesson [14]. Following that, the image
processing program classifies every pixel in the image using Max pooling: In max pooling, each feature map region is
these training data sets as a guide. As seen in Fig. 3, the divided into tiny, non-overlapping tiles, often measuring 2 by
2 or 3 by 3. The pixel with the highest value is selected by
the filter as it moves through the input and is sent to the

991

Authorized licensed use limited to: BLDEA's College of Eng & Tech - Vijayaura. Downloaded on March 10,2025 at 11:29:30 UTC from IEEE Xplore. Restrictions apply.
 output array. Max pooling reduces the spatial dimensions V. RESULTS
while maintaining the most important features. The application utilized for this is called Spyder, and it is
an open-source, free scientific environment for Python that
Average pooling: Just as with max pooling, the feature map includes logic, compute modules, and other features.
is divided into small tiles.
The average (mean) value within each tile is calculated and A total of nine distinct classes of datasets, including
used for the final feature map rather than selecting the highest WHITE FLY, CUTWORM, APHID, BEET ARMYWORM,
value. CUTWORM, THRIPS, ACARI, GRASSHOPPER, and
BACTROCERA CUCURBITAETHE, were initially
Even though the pooling layer loses a lot of information, gathered for this project. To obtain a more precise outcome,
CNN nevertheless benefits from it in several ways. They help these datasets are trained with an Epoch value of 10. The
to simplify, increase effectiveness, and lessen the risk of CNN Algorithm provides 90% or higher accuracy.
overfitting.
Using an Android mobile device, the farmer must first take
3) Fully-Connected Layer an image of the pest in the crop. This image must then be
Neurons comprise the Fully Connected layer. in addition uploaded into the user interface, as illustrated in Figure 5.
to acknowledged weights and biases. The layer is employed This will be converted into 224 x 224 x 333 pictures, which
to establish connections between neurons in two distinct will then be fed into a CNN model for pest class prediction,
layers. In a CNN architecture, this layer is often positioned employing several image augmentation methods on the
before the output layer. The fully connected layer computes dataset.
the final classification or regression task using the input from
the preceding layer. A fully connected layer (FCL) is one in
which every filter in the layer preceding it is connected to
every filter in the one succeeding it. Convolutional, pooling,
and layer outputs represent high-level features of the input
image.

The goal of employing the FCL is to use these features to

classify the input image according to the training dataset. The
accuracy vs. epochs and loss vs. epochs graphs from the
training procedure are displayed in Fig. 4.

Fig. 5 Initial page to upload an image

Fig. 5 illustrates the UI for uploading a pest image to the

backend, where the user can view the selected image as the
original. Upon uploading the pest image to the user interface,
the saved CNN model will be used to identify the pest and
Fig. 4 Accuracy and loss of pest control management systems.
provide the farmer with the pest class name. Additionally,
instead of employing chemical pesticides, it advises farmers
to apply organic insecticides when necessary to minimize
pest populations.

992

Authorized licensed use limited to: BLDEA's College of Eng & Tech - Vijayaura. Downloaded on March 10,2025 at 11:29:30 UTC from IEEE Xplore. Restrictions apply.
 As demonstrated in Fig. 6, upon uploading the input image, [7] G. Singh, Pest management in organic farming is a challenging task
the system recognizes the type of pest and provides the without the use of insecticides, ICICICT, 1, August 2020.
appropriate organic pesticides required to minimize the [8] B. Rosmawati Tamin, H. Karim, I. Bay, I.S. Arani, Mobile Image
Processing Applications For Cacao’s Fruits Pest And Disease Attack
infestation in agricultural areas. Using Deep Learning Algorithms, IEEE, 2020.
[9] S.K Praveen Kumar, Computerized System for Identification of Pests
Using Machine Vision, 8, IJRTE, Indhumathi, Coimbatore, 2020.
[10] D. Patel, N. Bhatt, Improved accuracy of pest detection using
augmentation approach with Faster R-CNN To cite this article, 2021
IOP Conf. Ser.: Mater. Sci. Eng. 1042 012020, ICMLSC, 2020.
[11] Simonyan, K., & Zisserman, A. (2014). "Very Deep Convolutional
Networks for Large-Scale Image Recognition". arXiv preprint
arXiv:1409.1556.
[12] Nagesh, N., Jayanthi, V., & Umesh, G. (2017). "Pest Detection and
Classification using Image Processing Techniques: A Survey".
International Journal of Computer Applications, 159(1), 1-5.
[13] Kavitha, C., & Sreelekshmi, K. (2020). "Pest Detection and
Classification using Convolutional Neural Networks in Agriculture".
International Journal of Scientific & Technology Research, 9(1), 442-
446.
[14] Anupama, K. R., & Ravi, K. N. (2019). "A Review on Crop Pest
Detection using Image Processing and Machine Learning
Techniques". International Journal of Engineering and Advanced
Technology, 8(5), 1-6.
[15] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). "ImageNet
Classification with Deep Convolutional Neural Networks". Advances
in Neural Information Processing Systems, 25, 1097-1105.

Fig. 6 Identification and recommendation of organic pesticides

VI. CONCLUSION
Cutting-edge agricultural innovation involves CNNs into
image-based pest monitoring and identification systems.
These technologies provide a more precise, data-driven,
long-lasting answer to the problems associated with pest
management in agriculture. CNN-based systems are
positioned to play a critical role ensuring the crop health,
boosting agricultural productivity, and contributing to a more
sustainable and food-secure future as technology advances
and their use grows. By using more datasets and image
classes during model training, accuracy can be raised.
VII. REFERENCES

[1] Jasheena. C.J, Ashutosh Kumar, Future Prospect of Packaged

Sustainable Food in India, International Journal of Research and
Analytical Reviews, vol. 7, no.1, pp. 101-121.
[2] L. Sun, K. Liang, Y. Song, Y. Wang, An Improved CNN-Based Apple
Appearance Quality Classification Method with Small Samples,
IEEE, August 2021.
[3] The Future of Organic Insect Pest Management: Be a Better
Entomologist or Pay for Someone Who is David Headrick MDPI
Horticulture and Crop Science Department, California Polytechnic
State University,
[4] Mayank Mishra, Tanupriya Choudhury, Tanmay Sarkar, CNN Based
Efficient Image Classification System for Smartphone Device,
Universitat Autònoma de Barcelona, Barcelona, Spain, 2021
ELCVIA ISSN:1577-5097.
[5] H. Pan, Z. Pang, Y. Wang, Y. Wang A new image recognition and
classification method combining transfer learning algorithm and
mobile net model for welding defects, June 29, 2020.
[6] B.P. Baker, Thomas A. Green, A.J. Loker Biological control and
integrated pest management in organic and conventional systems,
2020.

993

Authorized licensed use limited to: BLDEA's College of Eng & Tech - Vijayaura. Downloaded on March 10,2025 at 11:29:30 UTC from IEEE Xplore. Restrictions apply.