Detection and Classification Rice Plant Quality Through

UAV Imagery Using Yolo V5 Algorithm

Adi Suheryadu, A. Sumarudin, Alifia Puspaningrum and Renold N. K. Natasasmita

Politenik Negeri Indramayu, Jalan Raya Lohbener lama no 08, Indramayu, Jawa Barat, Indonesia

Keywords: Artificial Intelligence, Smart Farming, UAV Imagery, YOLO Algorithm, Rice Plants, Detection and

Classification Object.

Abstract: Smart farming is an important technology in supporting agriculture 4.0. Farmers can solve some problems by

utilizing this smart farm, including monitoring crops in agricultural areas. The large area of agricultural land

owned by farmers makes it difficult for farmers to monitor the quality of rice crops on their land. In

overcoming this problem, an intelligent system is needed to detect the quality level of rice crops and classify

them appropriately, and the scope of detection is broad. By detecting and classifying rice plants using one of

the artificial intelligence methods, namely YOLO, farmers can find out the quality of their rice plants through

images taken using UAVs. For this YOLO algorithm to detect the quality level of rice plants from each field,

datasets taken through aerial imagery or drone technology are needed, these datasets will be used to train

models to detect the quality level of rice plants. The YOLO algorithm can see the quality level of rice crops

based on the point of interest in the image uploaded by the farmer to the server. Images processed using the

YOLO algorithm produce output from bounding boxes and confidence scores for each detected object. The

yield of mAP@0.5 (mean average precision) was 93.69%.

1 INTRODUCTION

The industrial revolution is developing until now. The

technology is used to facilitate human activities to be

more effective and efficient by automating and

optimizing surveillance and production processes

(Hudson, P. 2014) The development of technology

also has an impact on the agricultural sector. The

result of agricultural technology is expected to help

increase agricultural output to meet the food demand

of around 10 billion people by 2050, as stated by the

World Resources Institutes (WRI) in 2018 (Ayaz, M.

Ammad-Uddin 2018).

Indonesia is one of the world's largest agricultural

countries, with a harvested area of agricultural land,

especially rice, reaching around 10.41 million

hectares, with rice production in 2021 of 54.42

million tons of GKG. If converted into rice, rice

production in 2021 will get about 31.36 million tons

(Badan Pusat Statistik, Luas Panen dan Produksi Padi

di Indonesia 2021). In addition, some Indonesians

have a livelihood as farmers or grow crops. The

agricultural sector is vital in improving the economy

and meeting food needs because paying attention to

its quality is necessary. Technology with the proper

methods to obtain product quality and build

sustainability in the agricultural sector is needed

(Lakshmi, V., & Corbett, J. 2020).

The increasing need for agricultural products with

better environmental quality protection has

encouraged the birth of smart farming with the term

Agriculture 4.0. The idea of Agriculture 4.0 attracted

the attention of agricultural actors in supporting the

development of modern agriculture combined with

digital information technology, mainly supported by

big data, mobile internet, and cloud computing. Some

of the technologies developed, including Unmanned

Aerial Vehicles (UAV) (Radoglou-Grammatikis,

2020), the use of Computer Vision (CV), and the

continuous improvement of Artificial Intelligence

(AI) (Kashyap, Anand. 2021), have undergone severe

improvements to new quality standards (Siniosoglou,

2020).

Combining UAV technology and deep learning

has provided information that has not been possible

so far, such as crop quality status, soil type, and

disease/pest attacks, among others (Bouguettaya,

2022). Precise and automated classification of plants

using UAV-based remote sensing imagery and deep

learning techniques have been widely used to perform

crop quality monitoring. Gao et al. (Gao, Z., Luo, Z.,

Suheryadu, A., Sumarudin, A., Puspaningrum, A. and Natasasmita, R.

Detection and Classiﬁcation Rice Plant Quality Through UAV Imagery Using Yolo V5 Algorithm.

DOI: 10.5220/0011843200003575

In Proceedings of the 5th International Conference on Applied Science and Technology on Engineering Science (iCAST-ES 2022), pages 577-581

ISBN: 978-989-758-619-4; ISSN: 2975-8246

577

2020) reviewed the latest deep learning-based

computer vision algorithms for detecting plant stress.

Nguyen Cao Tri et al. (Tri, N. C., Hoai, 2016)

proposed using CNN and unmanned aerial vehicles

(UAVs) to assess the quality of rice fields. The UAV

is responsible for taking pictures of rice fields at low

or very low altitudes. Then, the neural network will

process these high-resolution images to produce low-

quality and high-quality classifications of the rice

fields. The accuracy result they got was about 72%.

Then in the study (i, N. C., Duong, H. N., 2017), we

proposed an approach using DNN, a UAV, to assess

rice field yields. They used UAVs to collect high-

resolution imagery of rice fields from low altitudes.

Then, DNN is used to classify images for evaluating

rice field yields.

In this study, we will create a classification model

to obtain rice quality information as an effort to

monitor using YOLO v5 Deep learning.

2 COMPARISON METHOD

The classification algorithm used in this study is

YOLO V5 by includes inputs of Mosaic data,

enhancement, adaptive anchor, calculation, and

adaptive image scaling. The technology used in the

'Backbone' consists of the Focus structure and CSP

(Common Spatial Pattern) structure with techniques in

the 'Neck' section, including the FPN + PAN structure;

YOLO v5 is less capable than its predecessor YOLOv4

in terms of performance, but it is much more flexible

and faster than YOLOv4, so it has an advantage in the

use of its model. Yang et al. (Yang, D., Cui, Y., 2021).

YOLO accepts an input image divided into a grid

of S x S and sent to a neural network to create a

bounding box and prediction class. Each grid cell

predicts the bounding box and confidence score of

each box. This confidence score reflects how

confident the model has that the object in the box is

predicted. YOLO rates confidence as Pr (Object) *

IOU truth (Intersection of Union). Most previous

detection systems used a classifier or localizer to

perform the detection by applying the model to the

image in multiple locations and assigning confidence

to the image as a material for detection. YOLO uses

a different approach to the previous algorithm:

applying a single neural network to the entire image.

This network will divide the image into regions and

then predict the bounding box and probability, for

each bounding region box weighed its probability of

classifying as an object or not.

YOLOv5 has three important parts, namely

Backbone, Neck, and Head. The Backbone model on

YOLOv5 includes CSP (Cross Stage Partial

Networks), which extract informative features from

input images. CSP showed a significant improvement

in processing time with deeper networks. In the

Backbone process, the image will be extracted using

CSPNet, storing important information from the image

while reducing complex models. The next layer is the

Neck model used to generate the feature pyramid.

Pyramid features help the model to identify the

same object with different sizes and scales in a way

that is well generalized in scaling its objects. Feature

pyramids are very useful, and help models work on

invisible data, YOLO v5 uses PANet to obtain or

generate feature pyramids. In the Neck process of

information generated by the backbone, the Neck is

used to create helpful information at each feature

level that will spread directly to the sub-network. The

Head of the Output model is used to perform a part of

the final detection; this model applies boxes,

accuracy, and classes to the objects detected in the

image. Model Head will produce three different layer

sizes to achieve multi-scale detection; multi-scale

detection will ensure that the model can keep up with

the size changes. Next, the layer will generate

predictions such as class, accuracy and bounding box

(Muhammad, M. A., & Mulyani, Y., 2021). The

architecture of YOLOv5s is shown in figure 1.

Figure 1: YOLOv5s Architecture.

iCAST-ES 2022 - International Conference on Applied Science and Technology on Engineering Science

578

Figure 2: SSDMobileNetV2 Architecture.

This study was also tested using the

SSDMobileNetV2 architecture designated in figure 2.

This architecture consists of an SSD that acts as a base

model and MobileNet as a Network Model. The SSD

will set up object detection by creating a bounding

box. MobileNet will work to extract features that will

later be classified. The merging of SSD and

Mobilenet will help in the process of creating an

object detection application. An object detection

application takes an SSD to create an image

localization to determine the object’s position.

Meanwhile, Mobilenet will be needed to help classify

objects contained in an image. The classification will

result in categories for each object, namely Good and

Bad.

3 RESULT AND DISCUSSION

The rice plant data used as a training dataset consists

of two labels Healthy and Unhealthy label contains

3.610 images with a size of 640 x 640 pixels, and the

Unhealthy label contains 4.210 images with a size of

640 x 640 pixels. The data is an image of a rice plant

taken through aerial or drone imagery (Anuar

Marzhar. 2021).

Figure 3: Process Diagram.

The flow diagram is processed to create a deep

learning model, as shown in figure 3. Starting with

labelling, the dataset is given a bounding box and

annotations or labels at this stage. The following

process is splitting data or dividing data into train

folders and validation with a ratio of 9 to 1 or 90% to

10% as shown in table 1. Then Image Augmentation

or known as image augmentation, image

augmentation is a process of modifying data to

become another shape, such as the size that was

originally 640 x 640 to 512 x 512. The training

process aims to train the model to be used to detect;

the training process produces two models, namely,

last and best weights. The last stage is testing; testing

is carried out to see how high the accuracy is, the

model that has been trained against the new data.

Figure 4: Training Implementation.

Table 1: Training Data.

Data Training

90% 7038

Data Validation

10% 782

Total Data

100% 7820

The measurement matrices used in this study are

Precision, Recall, mAP@0.5, mAP@.5:.95, detection

time and file size of each model. The confusion

matrix can be used to determine precision and recall.

To determine the precision, you can use the equation

below (Hidayatullah, P., 2021).

precision  











(1)

recall  











 (2)

AP  





xAP



0  AP



0,1. . . AP



1,0 (3)

Detection and Classiﬁcation Rice Plant Quality Through UAV Imagery Using Yolo V5 Algorithm

579

mAP  

∑





(4)

where TP is true positive, FP is false positive, FN is

false negative, AP is Average Precision and mAP is

Mean Average Precision.

The training process on the first model was

completed within 0.457 hours with quite good results,

with a Precision value of about 0.991, Recall around

0.973, mAP@0.5 around 0.979 and mAP@0.5:0.95

around 0.944. And the second model using SSD was

completed within 0,25 hours, with a Precision of

about 0,850, Recall of about 0,907, mAP@0,5 about

0,946, and mAP@0,5:0.95 of about 0,851. The

hyperparameter settings used are shown in table 2.

Table 2: Hyperparameter setting.

Parameter Value

lr0 0.01

lrf 0.01

Momentum 0.937

weight_decay 0.0005

warmup_epochs 3.0

warmup_momentum 0.8

warmup_bias_lr 0.1

Box 0.05

Cls 0.5

cls_pw 1.0

Obj 1.0

obj_pw 1.0

iou_t 0.2

anchor_t 4.0

hsv_h 0.015

hsv_s 0.7

hsv_v 0.4

Translate 0.1

Scale 0.5

Fliplr 0.5

mosaic 1.0

The dataset for testing was taken directly from

farms located in Indonesia part of the districts of

Majalengka, Cirebon and Indramayu, there are about

100 data with 111 bounding boxes in the picture,

including 40 pictures of rice fields with healthy

quality labels with 47 bounding boxes and 60 images

of rice fields with unhealthy quality labels with 64

bounding boxes taken using drone technology. This

test aims to select the best model to be used or

implemented. The tests in this study will compare two

models, namely the results of yolov5s training and

SSDMobileNetV2. The resulting output is Precision,

Recall, mAP@0.5, mAP@.5:.95. The result of

comparison testing is shown in table 3 and figure 5.

Table 3: Result Testing.

Metrics YOLOv5s SSDMobileNetV2

F1-Score

0,90 0,750

Precision

0,936 0,683

Recall

0,87 0,832

mAP@0.5

0,93 0,790

mAP@.5:.95

0,916 0,683

time

0,063s 0,144s

Size model

14,4Mb 18,4Mb

Figure 5: YOLOv5s vs SSDMobileNetV2 Metrics.

From the results of the test above, YOLOv5s

is relatively better and can work faster than

the SSDMobileNetV2 and the detection time and file

size of the model are also smaller. YOLOv5s have

advantages in F1-Score, Precision, Recall and mAP

(mean average precision) values. This means that

the resulting FPS will be much higher, and

the computation time will also be relatively faster.

The example result of the classification is shown in

figure 6.

Figure 6: Result of Classification.

iCAST-ES 2022 - International Conference on Applied Science and Technology on Engineering Science

580

4 CONCLUSIONS

The detection system for the quality level of rice

plants built using the YOLOv5s Algorithm when

compared to SSDs gets relatively quite good results.

From the training time, the SSD is almost 50% faster,

but when the model has been formed and the testing

process is carried out, the results obtained by YOLO

are more detectable. The system can detect two

quality labels based on training results using the

dataset that has been collected. The accuracy results

(mAP@0.5) obtained using the YOLOv5 model were

0.9369.

REFERENCES

T. S. Ashton, P. Hudson, The industrial revolution, 1760-

1830, Oxford University Press, Oxford, 1997

Ayaz, M.; Ammad-Uddin, M.; Sharif, Z.; Mansour, A.;

Aggoune, E.H.M. Internet-of-Things (IoT)-Based

Smart Agriculture: Toward Making the Fields Talk.

IEEE Access 2019, 7, 129551–129583.

Lakshmi, V., & Corbett, J. (2020). How artificial

intelligence improves agricultural productivity and

sustainability: A global thematic analysis.

Radoglou-Grammatikis, P.; Sarigiannidis, P.; Lagkas, T.;

Moscholios, I. A compilation of UAV applications for

precision agriculture. Comput. Netw. 2020, 172,

107148.

Kashyap, P.K.; Kumar, S.; Jaiswal, A.; Prasad, M.;

Gandomi, A.H. Towards Precision Agriculture: IoT-

Enabled Intelligent Irrigation Systems Using Deep

Learning Neural Network. IEEE Sens. J. 2021, 21,

17479–17491.

Anand, T.; Sinha, S.; Mandal, M.; Chamola, V.; Yu, F.R.

AgriSegNet: Deep Aerial Semantic Segmentation

Framework for IoT-Assisted Precision Agriculture.

IEEE Sens. J. 2021, 21, 17581–17590.

Siniosoglou, I.; Argyriou, V.; Bibi, S.; Lagkas, T.;

Sarigiannidis, P. Unsupervised Ethical Equity

Evaluation of Adversarial Federated Networks. In

Proceedings of the 16th International Conference on

Availability, Reliability and Security, Vienna, Austria,

17–20 August 2021; Association for Computing

Machinery: New York, NY, USA, 2021.

Badan Pusat Statistik, Luas Panen dan Produksi Padi di

Indonesia 2021. 05100.2203. 2022-07-12

Bouguettaya, A., Zarzour, H., Kechida, A., & Taberkit, A.

M. (2022). Deep learning techniques to classify

agricultural crops through UAV imagery: a review.

Neural Computing and Applications, 1-26.

Gao, Z., Luo, Z., Zhang, W., Lv, Z., & Xu, Y. (2020). Deep

learning application in plant stress imaging: a review.

AgriEngineering, 2(3), 29.

Tri, N. C., Hoai, T. V., Duong, H. N., Trong, N. T., Vinh,

V. V., & Snasel, V. (2016, December). A novel

framework based on deep learning and unmanned aerial

vehicles to assess the quality of rice fields. In

International Conference on Advances in Information

and Communication Technology (pp. 84-93). Springer,

Cham.

Tri, N. C., Duong, H. N., Van Hoai, T., Van Hoa, T.,

Nguyen, V. H., Toan, N. T., & Snasel, V. (2017,

October). A novel approach based on deep learning

techniques and UAVs to yield assessment of paddy

fields. In 2017 9th International Conference on

Knowledge and Systems Engineering (KSE) (pp. 257-

262). IEEE.

Yang, D., Cui, Y., Yu, Z., & Yuan, H. (2021). Deep

learning based steel pipe weld defect detection. Applied

Artificial Intelligence, 35(15), 1237-1249.

Muhammad, M. A., & Mulyani, Y. (2021, October).

Library Attendance System using YOLOv5 Faces

Recognition. In 2021 International Conference on

Converging Technology in Electrical and Information

Engineering (ICCTEIE) (pp. 68-72). IEEE.

Anuar Marzhar, “Paddy Field Health”. kaggle.com.

https://www.kaggle.com/datasets/marzharanuar/paddy

-field-health. (accessed oct 3, 2022).

Hidayatullah, P (2021). Buku Sakti Deep Learning

Computer Vision Menggunakan YOLO untuk Pemula.

Detection and Classiﬁcation Rice Plant Quality Through UAV Imagery Using Yolo V5 Algorithm

581