2022 6th International Conference On Computing, Communication, Control And Automation (ICCUBEA)
Pimpri Chinchwad College of Engineering (PCCOE), Pune, India. Aug 26-27, 2022
Convolutional Neural Networks for Detection of
Crop Diseases and Weed
Bhavansh Gupta1 Saurabh Bomble2 Onkar Gaikar3
Department of Computer Engineering Department of Computer Engineering Department of Computer Engineering
Pimpri Chinchwad College of Pimpri Chinchwad College of Pimpri Chinchwad College of
2022 6th International Conference On Computing, Communication, Control And Automation (ICCUBEA) | 978-1-6654-8451-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICCUBEA54992.2022.10010772
Engineering Pune Engineering Pune Engineering Pune
Pune, India Pune, India Pune, India
[email protected] [email protected] [email protected] Shubham Chalekar4 Prof. S. R. Vispute5 Prof. Dr. K. Rajeswari6
Department of Computer Engineering Department of Computer Engineering Department of Computer Engineering
Pimpri Chinchwad College of Pimpri Chinchwad College of Pimpri Chinchwad College of
Engineering Pune Engineering Pune Engineering Pune
Pune, India Pune, India Pune, India
[email protected] [email protected] [email protected] Abstract - Agriculture is around $10 trillion industry very good, and accuracy with sample data is greater than 87
worldwide and has growth of more than 6% year on year. To percent. As a result, our study will be able to correctly
enhance crop output and modernize the system, an automated identify crop illness and predict the leaf disease. We have
and speedy disease detection mechanism is being developed. also used this in our application to provide detailed report to
Another area where these technologies play an important role user about methods of prevention and cure for the disease.
is the detection of Weed from farm using images of farmlands.
The images will be used to locate weed among the crops and Weed Detection – Weeds are the the unwanted plants
can be used along with aerial drones or robots for automated that grow in the farms and compete for the resources with
weed spraying operations. This paper compares proposes a the crops. In most of the developing countries like India
system for detection of crop/plant diseases using images of weeding is done by hand , however due to labour migration
plant and detection of weed using aerial images of farms. to cities most of these farmers resort to indiscrimiate use of
herbicide to kill the weeds. Weed growth that is unchecked
We compared MobileNetV2, InceptionV1, NASNetMobile can result in 100% yeild losses. It results in stunted growth
for crop disease detection on dataset of 3 different crops having in crops and can drastically impact the quality and
12 disease classes. Of these NASNetMobile performed the best. nutritional value of the crops.
Similarly we used ResNet50V2, MobileNetV2, InceptionV3,
NASNetMobile for weed detection on soyabean plant image Weed-infested areas can be identified using image
dataset and InceptionV3 had the best performance with detection. This is not only economically beneficial to the
81.23% Validation Accuracy for Bounding Boxes. Using the farmer but also an ecologically sustainable model as it will
result of this study as the basis we were able to develop an result in less use of harmful chemical herbicide on the crops.
application that can be used for weed detection and crop We have used bounding boxes for identify the weed affected
disease detection using images. areas in dataset of soyabean fields images for targeted
treatment of weeds using automated systems.
Keywords— Plant Disease, Image processing, Deep
Learning, Disease Detection, Weed Detection, Bounding Box, II. LITERATURE SURVEY
Agriculture
In [1] Image processing for weed detection was discussed
I. INTRODUCTION with 97% accuracy under ideal conditions. It discussed
several techniques like converting images to grayscale and
Agriculture has a direct influence on human survival on
increasing contrast of images to improve accuracy. In [3]
this planet, and while being an ancient profession, it still
discusses a method for creation of parallel weed detection
mostly adheres to mediaeval methods. We've identified two
system for creating a real-time CNN model that is scalable.
primary areas where precision farming may help farmers
In [5] the author discussed using semi supervised GCN-
and the agriculture industry.
ResNet-101 model architecture for weed detection by
Crop Disease Detection –Potato, Bell Pepper, Tomato leveraging feature associations. [6] discusses weed detection
are among the most commonly consumed crops in India. on soyabean crop images using segmentation tool. Dataset
The production of these crops is mostly affected by a variety had over 15,000 images and was trained on CNN with 98%
of pests and illnesses which results in lower yeild. Early accuracy over all classes. In paper [2], algorithms such as
blight, late blight, Yellow Leaf Curl virus, bacterial spots, SVM, ANN and CNN were compared with each other to
Mosaic Virus and other devastating crop diseases have classify weed and crops plants. This paper concluded that
occurred in India in the past. Every year, the a large CNN performs better than ANN and SVM. SVM gives
majority of crops in India are hampered by this illness. accuracy of 87% and ANN gives maximum accuracy of
93%. CNN gives accuracy more than 98% for classifying
FAO(Food and Agriculture Organization) estimates that weeds and crops plants. The confusion matrix below
Plant diseases cost the world economy some $220 billion provides crop and weed classification validation for several
each year. We used image processing to diagnose the crop classifiers.
leaf diseases. Accuracy with training data in the project is
978-1-6654-8451-0/22/$31.00 ©2022 IEEE 1
Authorized licensed use limited to: Bahria University. Downloaded on December 18,2023 at 19:02:32 UTC from IEEE Xplore. Restrictions apply.
In paper [4], weeds were detected in wheat fields.
Background subtraction algorithms to detect weeds and
wheat crops. In background subtraction method, first images
of barren land were captured. Then, images of field with
wheat crops are taken. Using background subtraction newly
grown plants are detected from the frame. After that Hue,
Saturation and value (HSV) is applied to differentiate
between crops and weeds. Here colour of wheat and weed
plants were different in every stage of its growth so it was
easy to classify weed plants in image by applying threshold
value for HSV.
In [9] dataset of more than 50,000 images was used with
a 60:40 split to train and validate for 35 different diseases for
16 different categories of plants. The CNN model trained had
accuracy of 93%. In [14] disease detection was performed on
soyabean plant leaf. The author used dataset consisting of
2200 synthetic images. In [15] CNN model was used to
detect 9 different tomato diseases on 14,529 training and
3,631 validation images. After training it over 5000 epochs
model gave accuracy of 98.4%. [11] compares different
CNN architectures. The images were taken in lab and the
accuracy in field dropped dramatically. In paper [13], Pre-
processing, feature extraction, and classification are all
stages of image processing. An input picture was scaled
during processing. Then, from an input for network training,
colour and texture characteristics were retrieved and
identified. The technology was used to automatically identify Fig. 1. Proposed System
the test photographs in order to determine whether the plant
was abnormal or not. CNN classifier was used. In paper [12]
Image capture was the initial phase, followed by Image pre-
processing, image segmentation, and feature extraction,
which took colour, shape, and size into account. The dataset
was split in 65:35 for training and testing. They used two
classification techniques SVM and Neural Network with
accuracy 89% and 80% respectively. They further claim that
as compared to unhealthy wheat, the healthy wheat
histogram has the highest peak incidence.
III. METHODOLOGY
Two different CNN based models are used for the
purpose of crop disease detection and weed detection. The
models were trained on the different CNN architectures and
their results were compared. The best models were chosen to Fig. 2. Block Diagram
be used in the application developed for farmers to identify
various crop diseases and weed growth using images. A. Dataset
As a result of this study an application is created with the Dataset for Crop Disease detection -The dataset we used
following flow. The application is able to identify weed is plant village dataset available freely from Kaggle. We
growth and return image with bounding boxes around weed were able to collect data on about 12 types of diseases of
infested areas and returns the crop disease details and potato, tomato and bell pepper. Of these 12 types of disease,
suggests possible methods for treatments using the CNN 7 unique types of diseases are listed below. There is total
models created during this study. 2952 images in this plant disease dataset.
The block diagram shows the structure of the application x Bacterial Spot
developed. The application consists of two different models
x Early Blight
trained for crop disease detection and weed detection which
are hosted on a flask server. A web interface is provided to x Late Blight
the user to upload image over which the results are required.
The request is then sent to the server which based on the x Leaf Mold
image sends the report with confidence percentage of x Septoria Leaf Spot
whether the crop is healthy or unhealthy and if unhealthy the
disease by which it is affected in case of crop disease x Yellow Leaf Curl Virus
detection. For weed detection the server sends back the
image with bounding box around the weed identified in the x Mosaic Virus
image and also the confidence percentage. x Spider Mite
2
Authorized licensed use limited to: Bahria University. Downloaded on December 18,2023 at 19:02:32 UTC from IEEE Xplore. Restrictions apply.
IV. MODEL PERFORMANCE
A. Training, Testing and the Validation
Crop Disease Detection Model – We used transfer
learning to build the following models using the
MobileNetV2, InceptionV1, NASNetMobile architectures as
our base from Keras Library. Over these architectures dense
layers with ReLU and SoftMax activation functions was
used. For these models we used Adam optimizer with
learning rate as 0.001 and trained them for 15 epochs with
image size of 224x224. We used binary cross entropy as loss
function for training these models.
Weed Detection Model – We used transfer learning to
build the following models using the ResNet50V2,
MobileNetV2, InceptionV3, NASNetMobile models as our
base layer from the Keras Library that provides pretrained
weights. The CNN model was trained using imagenet
weights and with the activation function of ReLU for the
dense layers. The model had two output layers with
activation function as SoftMax and Sigmoid. Adam
Optimizer was used with learning rate of 0.00005 for training
the models. All these models were initially trained for 40
epochs to identify the best performing models and the best
performing model was trained for 100 epochs before using it
in the application. The loss function was calculated using
mean squared error for the bounding box accuracy.
B. Model performance
Crop Disease Detection Model – The models were
initially trained for 15 epochs. The model with MobileNetV2
architecture had validation accuracy of 84.60%. The model
Fig. 3. Crop Disease Dataset Example took total time of 31min 10s for training with around 2min
average time per epoch. The InceptionV1 model had
Dataset for Weed detection - It contains 1300 images of validation accuracy of 84.26% and took total time of 49min
sesame crops and different types of weeds with each image 11s with average time of 3min 16s per epoch. The
labels. The dataset was created using 546 images and then NASNetMobile model had validation accuracy of 82.57%
data augmentation was used to generate rest of the images. and took total time of 17min 18s with average time of 1min
Then it was manually labelled to draw the bounding boxes. 9s per epoch which is the fastest of all the models and with
highest accuracy. Thus, we decided to use to NASNetMobile
used for our application. The performance of NASNetMobile
model is shown below.
Fig. 5. Comparison of Accuracy of the trained models
Fig. 4. Weed Dataset Example
3
Authorized licensed use limited to: Bahria University. Downloaded on December 18,2023 at 19:02:32 UTC from IEEE Xplore. Restrictions apply.
Fig. 7. Bounding Box accuracy and loss and for InceptionV3 Weed
Detection Model
Fig. 6. Accuracy and Loss for Disease Detection Model
Weed Detection Model – The models were initially
trained for 40 epochs. The model with ResNet50V2
architecture had bounding box validation accuracy of
72.31%. The model took total time of 23min 55s for training
with around 36s average time per epoch. The model with
MobileNetV2 architecture is lightweight model specifically
designed for mobiles. It had bounding box validation
accuracy of 68.31 %, the model took total time of 15min 28s
for training with around 23s average time per epoch which is Fig. 8. Weed Detection Performed by Model
significantly faster than previous model. The model with
InceptionV3 architecture had the best bounding box V. CONCLUSION
validation accuracy of 77.23%. The model took total time of
12min 34s for training with around 19s average time per This study focuses on disease detection and weed
epoch. The model with NASNetMobile architecture had detection using CNN. The scope of the study is limited to 3
bounding box validation accuracy of 73.85%. The model was different crops and 12 different classes of diseases for those
fastest in terms of training and took total time of 11min 49s plants for disease detection due to limited availability of
for training with around 18s average time per epoch. Of the datasets. It focuses on identifying best CNN model that can
above models The model with InceptionV3 performed be used with transfer learning for providing best results with
significantly better than the other models and was chosen to least possible training time and over datasets with very
be used for our application. The performance of InceptionV3 limited datasets. Similarly, the weed dataset used contained
model is given below. manually labelled images since the size of dataset was small.
As the larger datasets will become more accessible the scope
of this study can be expanded.
Crop Disease Detection of three crops namely Bell
Pepper, Tomato and Potato for 12 classes of diseases was
performed using three different models and the accuracy of
those models was compared. The model with the best
accuracy was trained on the image dataset for 100 epochs
and the model gives 87% of validation accuracy on the
validation dataset.
The other use case for this application is weed detection.
It uses bounding boxes to identify the weed growth in the
farms. This can be further used to provide input to robotic
systems where the robotic systems or drones can capture
4
Authorized licensed use limited to: Bahria University. Downloaded on December 18,2023 at 19:02:32 UTC from IEEE Xplore. Restrictions apply.
images from farm and these will be processed in Realtime to [6] Alessandro dos Santos Ferreira, Daniel Matte Freitas, Gercina
provide the robots with the marked areas where they will Gonçalves da Silva, Hemerson Pistori, Marcelo Theophilo Folhes,
Weed detection in soybean crops using ConvNets, Computers and
have to use herbicides. Electronics in Agriculture, Volume 143, 2017.
We have developed a web application that can detect the [7] Fenil Dankhara, Kartik Patel, Nishant Doshi, Analysis of robust weed
disease of the above-mentioned plants and will suggest detection techniques based on the Internet of Things (IoT), Procedia
Computer Science, Volume 160, 2019.
possible treatment for each such diseases. The application
[8] Muhammad Hamza Asad, Abdul Bais, Weed detection in canola
can also be used along with automated weed removing fields using maximum likelihood classification and deep
systems to identify patches of weed for effective spraying of convolutional neural network, Information Processing in Agriculture,
herbicides. We will be able to improve accuracy in the future Volume 7, Issue 4, 2020.
by expanding our image dataset and training of model.Using [9] Parul Sharma, Yash Paul Singh Berwal, Viqas Ghai, “Krishimitr
our project Indian farmers can receive immediate support (Farmer Friend): Using Machine Learning to Identify Diseases in
and advise on the condition of their crops Plants” , IEEE International Conference on Internet of Things and
Intelligence System (IoTaIS), 2018.
ACKNOWLEDGMENT [10] S. Ramesh, Hebbar Ram, P V Vinod, Niveditha M, Pooja R, Prasad
Bhat N, Shashank N, “Plant Disease Detection Using Machine
We are grateful to Prof. Sushma Vispute, our Project Learning”, International Conference on Design Innovations for 3Cs
Guide, for her aid and guidance during our Final Year Compute Communicate Control, 2018.
Project, particularly for the important advice she provided [11] Konstantinos Ferentinos, “Deep learning models for plant disease
throughout the Project and for laying the groundwork for the detection and diagnosis”, Computers and Electronics in Agriculture,
ELSEVIER, Feb 2018.
success of our study.
[12] Varsha P. Gaikwad, Dr. Vijaya Musande, “Wheat Disease Detection
REFERENCES Using Image Processing”, 1st International Conference on Intelligent
Systems and Information Management (ICISIM), IEEE 2017.
[1] J. Irías Tejeda and R. Castro Castro, "Algorithm of Weed Detection in [13] Sudarshan Barure, Bhushan Mahadik, Monali Thorat, Akhil Kalal ,
Crops by Computational Vision,” 2019 International Conference on “Disease Detection in Plant using Machine Learning”, International
Electronics, Communications and Computers (CONIELECOMP). Research Journal of Engineering and Technology (IRJET).
[2] S. T., S. T., S. G. G.S., S. S. and R. Kumaraswamy, "Performance [14] Keke Zhang, Qiufeng Wu, Yiping Chen, Detecting soybean leaf
Comparison of Weed Detection Algorithms,“2019 International disease from synthetic image using multi-feature fusion faster R-
Conference on Communication and Signal Processing (ICCSP). CNN, Computers and Electronics in Agriculture, Volume 183, 2021.
[3] S. Umamaheswari, R. Arjun and D. Meganathan, "Weed Detection in [15] Mohit Agarwal, Suneet Kr. Gupta, K.K. Biswas, Development of
Farm Crops using Parallel Image Processing,“IEEE 2018 Conference Efficient CNN model for Tomato crop disease identification,
on Information and Communication Technology (CICT). Sustainable Computing: Informatics and Systems, Volume 28, 2020.
[4] S. Hameed and I. Amin, "Detection of Weed and Wheat Using Image [16] Bhavansh Gupta, Saurabh Bomble, Onkar Gaikar, Shubham
Processing," 2018 5th International Conference on Engineering Chalekar, Sushma Vispute, Kannan Rajeswari, “An Overview of
Technologies and Applied Sciences (ICETAS). Decision Support Systems for Precision Farming using Deep
[5] Honghua Jiang, Chuanyin Zhang, Yongliang Qiao, Zhao Zhang, Learning” 21st Century Innovations in Management, Science and
Wenjing Zhang, Changqing Song, CNN feature based graph Technology, Education and Social Sciences
convolutional network for weed and crop recognition in smart
farming, Computers and Electronics in Agriculture, Volume 174,
2020.
5
Authorized licensed use limited to: Bahria University. Downloaded on December 18,2023 at 19:02:32 UTC from IEEE Xplore. Restrictions apply.
