Weed Identification Technique in Basil Crops using Computer Vision

RICARDO YAURI1,2, BRYAN GUZMAN2, ALAN HINOSTROZA2, VANESSA GAMERO3

BRAZIL

Abstract: - The promotion of organic and ecological production seeks the sustainable and competitive growth

article is to demonstrate how an artificial intelligence model based on computer vision can contribute to the

identification of weeds in basil plots. An iterative and incremental development methodology is used to build

the system. In addition, this is complemented by a Cross Industry Standard Process for Data Mining

methodology for the evaluation of computer vision models using tools such as YOLO and Python language for

weed identification in basil crops. As a result of the work, various Artificial Intelligence algorithms based on

neural networks have been identified considering the use of the YOLO tool, where the trained models have

shown an efficiency of 69.70%, with 3 hours of training, observing that, if used longer training time, the neural

network will get better results.

Key-Words: - Crops, farm, agricultural, basil, machine learning, Classification algorithms, yolo, Weed

of organic crops in Peru, [1], [2]. During 2019,

agricultural exports have been characterized by

products such as fruit and vegetables, with the

United States, Germany, the Netherlands, and Spain

as main clients. These markets need to comply with

organic certification regulations to enter the US, it is

necessary to be accredited under the USDA Organic

Standards 7 CFR 205, [3]. Its Operational Standard

for Pests and Weeds indicates that the problem of

weed control must be carried out with coverage of

product in great demand for local and foreign

consumption, [1]. These products are affected by

weeds due to competition for water, light, nutrients,

CO2, and space. This is why it is important to

recognize the weeds of the crop, as it helps to

increase the yield per hectare to solve this problem.

crops can be solved from the technological field

using solutions such as performing a mechanical

removal of the weeds, using a robot that can move

necessary to differentiate between the crop and the

weed whose most used technique is image

processing. This causes the staff to be replaced by

robots in the near future using Deep Learning image

recognition techniques in blueberries during the

rooting stage, for efficient use of resources, [4], [5],

[6].

processes is related to identifying diseases in plant

leaves through image processing, using artificial

contribute with a solution is Edge Computing, [8],

which allows for reducing latency and improving

computing power closer to the client together with

cloud computing.

In the case of this research, the area of study is

basil fields with weed presence problems in rural

areas and limited access to cable internet. Based on

what was previously described, the problem has

been identified and this paper poses the following

research question: How is it possible to identify

identification algorithm for the classification of

weeds in a basil plot; Implement the weed

identification algorithm in a basil plot; Evaluate the

effectiveness of the identification algorithm. As a

first step, the use of a research methodology is

proposed through literature review techniques to

evaluate the most appropriate technologies.

Subsequently, a methodology based on CRISP

(Cross Industry Standard Process for Data Mining)

is selected to implement computer vision

product per hectare. In addition, the efficiency of the

artificial intelligence tool to differentiate weeds

from basil crops is described, which can then be

applied to a mechatronic device or to identify areas

with many weeds.

The development of this paper is written as

follows. Section 2 shows the most relevant papers.

The most important concepts about the technologies

used are shown in section 3. Subsequently, section 4

describes the proposed system and section 5

to the use of identification and classification

algorithms through deep learning techniques

through image processing in crop fields and the

importance of the development of a country.

In, [9], the author describes how globalization

has driven development in many third-world

countries. It is mentioned that there is a growing

demand for vegetables and fruits since 1986, having

a great leap since 2000, thanks to agro-export

agriculture. This is why many companies, related to

this area, manage better resources to optimize

production, among which we have: technical

irrigation, and access to water projects, in products

(Quinua, Grapes, Asparagus, Avocados, etc). The

results of the study indicate that the export sectors

use valid development strategies for low incomes.

(Vertical Dimensions) and DSE-H (Horizontal)

modules, [11], [12]. The results in [10], show an

effectiveness of 85% in the detection of the crop, in

a natural environment. In addition, YOLO version 3

was used, which has the Binary Cross Entropy

(BCE) function, allowing to efficiently estimate the

loss of classification and loss of confidence using

mean square error (MSE).

identify weeds in crops, [13], in real-time. The

irrelevant features, improving the performance of

the network. A Jetson AGX was used for the

training and deployment of the solution, verifying a

performance improvement from 0.90 to 0.92.

and integrates methods of artificial intelligence,

[14], or the use of drones, [15]. This makes it

possible to contribute to protecting crops from

aspects such as: climate change, population growth,

and food security. This paper describes weeding,

Computer Vision are used.

of Things technologies in agriculture enables early

recognition. For the training and tests with images

of the rice paddy, software technologies such as

Keras and Tensorflow are used, [17], [18].

Some of the most used tools for the development of

artificial intelligence include Python, C++, R, Java,

Prolog, and Matlab. In the case of Python, this is a

programming language, with great development for

different platforms. Among its advantages we have

Treatment of a large volume of data, ideal for Big

Data, Simple and easy to learn, and Dynamic

pass over the entire image, instead of using sliding

windows or multi-stage detection methods. This

makes YOLO faster and more efficient compared to

other approaches.

of lasers, which eliminate unwanted plants.

Another alternative is the application of

radiation, X-rays, gamma, etc.

crops, using object detection algorithms, and

integrating artificial intelligence techniques and the

YOLO software package. In Fig. 3, we show a

flowchart of our project, with the following stages:

classifiers. These classifiers are combined to form a

strong classifier that can detect specific objects. On

the other hand, YOLO is based on applying

convolutional neural networks by dividing the

image into a grid and bounding boxes for object

classification. In the inference stage, the neural

network processes the image and predicts the

classes of the detected objects. Both approaches

require a training stage using labeled data sets to

apply the inference stage.

following steps are performed:

 The training and validation procedure begins

with the preparation of the images, considering

different types of weeds and they are used in the

training and validation stage.

 Subsequently, the images are divided into

training and validation sets, which allow us to

evaluate and adjust the performance of the

model during training.

 The Google Colab platform is used, and the

 During training, the model learns to recognize

and classify weeds in the basil images by

performing multiple iterations to gradually

improve model performance in the Validation.

resulting images with the "weed" class mark.

In the field of object detection, there are different

approaches and algorithms used. These include two-

stage algorithms such as Region-based

Convolutional Neural Networks (R-CNN) and their

variants such as Fast R-CNN, Faster R-CNN, and R-

single-stage approach that uses a 24-layer

convolutional neural network based on GoogleNet.

The choice of YOLO V4 was based on its ability

to run on different operating systems, such as Linux

and Windows, and its ability to work with both

static and real-time images. By using YOLO V4,

accurate object identification was achieved in the

context of this project.

that will allow renaming each photo as: weed_0,

weed_1, and so on. To facilitate the training of the

neural network, these renamed files were stored in

the “imagenesMaleza_rename” folder. Then the

Labeling of these images was carried out, to

manually create and locate the name of the class:

"Weed".

subfolders: Train (80%), Validation (15%) and Test

(5%). This was done with another script called

“Split Files” (Fig. 8).

Fig. 6: Rectbox creation and weed class

Fig. 8: Training, testing, and validation folders

Analyzing the results described above, it was

considered that the data preparation and labeling

steps were crucial to ensure the quality and accuracy

of neural network training in weed recognition in

basil crops. As a discussion of the results, it can be

stated that the successful implementation of these

processes contributed to obtaining an effectively

To deploy the algorithm, the following steps are

performed in Google Colab:

 The Darknet Framework is installed to create a

virtual workspace.

 Create a MakeFile file, which will give

instructions to the Make function, to perform all

tasks.

 Download the weights of the algorithm that

process.

 Execution of the configuration script

 Training process.

 It will proceed from inference, using the images

from the test folder.

weights obtained are used, through the "darknet"

command, to which a series of options will be

passed as parameters to display the results (Fig. 9).

[1] J. I. Sanca Mendoza, “Informe por servicios

profesionales: ‘Manejo del cultivo de

albahaca (ocimum basilicum) Var. Genovessa

para la planta procesadora agroindustrial La

Joya S.A.C. - Arequipa,’” Universidad

Nacional de San Agustín de Arequipa, 2018.

[2] PromPerú, “Departamento de Inteligencia de

Mercados - PromPerú,” 2019. Accessed: Jan.

23, 2023. [Online]. Available:

[4] R. Moreira, L. F. Rodrigues Moreira, P. L. A.

Munhoz, E. A. Lopes, and R. A. A. Ruas,

“AgroLens: A low-cost and green-friendly

Smart Farm Architecture to support real-time

leaf disease diagnostics,” Internet of Things,

vol. 19, p. 100570, Aug. 2022, doi:

10.1016/J.IOT.2022.100570.

[5] S. Kallapur, M. Hegde, A. D. Sanil, R. Pai,

hyperspectral images of agricultural and food

products,” Chemom. Intell. Lab. Syst., vol.

198, p. 103936, Mar. 2020,

doi: 10.1016/J.CHEMOLAB.2020.103936.

[7] I. A. Quiroz and G. H. Alférez, “Image

recognition of Legacy blueberries in a Chilean

smart farm through deep learning,” Comput.

Electron. Agric., vol. 168, p. 105044, Jan.

2020, doi: 10.1016/J.COMPAG.2019.105044.

[8] R. Yauri and R. Espino, “Edge device for

[9] J. Schwarz, E. Mathijs, and M. Maertens, “A

dynamic view on agricultural trade patterns

and virtual water flows in Peru,” Sci. Total

doi: 10.1016/J.COMPAG.2022.107057.

[11] R. Punithavathi et al., “Computer Vision and

Deep Learning-enabled Weed Detection

Model for Precision Agriculture,” Comput.

Syst. Sci. Eng., vol. 44, no. 3, pp. 2759-2774,

2023,

doi: 10.32604/CSSE.2023.027647.

[12] A. Wang, T. Peng, H. Cao, Y. Xu, X. Wei,

and B. Cui, “TIA-YOLOv5: An improved

approach incorporating YOLO v5 and

attention mechanisms for field real-time

detection of the invasive weed Solanum

rostratum Dunal seedlings,” Comput.

Electron. Agric., vol. 199, Aug. 2022,

doi: 10.1016/J.COMPAG.2022.107194.

[14] T. Talaviya, D. Shah, N. Patel, H. Yagnik, and

M. Shah, “Implementation of artificial

intelligence in agriculture for optimisation of

irrigation and application of pesticides and

segmentation using motion blurred UAV

images of sorghum fields,” Comput. Electron.

10.1016/J.COMPAG.2022.107388.

[16] O. Debnath and H. N. Saha, “An IoT-based

intelligent farming using CNN for early

disease detection in rice paddy,”

Microprocess. Microsyst., vol. 94, p. 104631,

Oct. 2022,

doi: 10.1016/J.MICPRO.2022.104631.

doi: 10.3390/ELECTRONICS11244100.

[18] A. A. Albraikan, M. Aljebreen, J. S.

doi: 10.3390/APP122412828.

[19] R. Johansson, Numerical python: Scientific

computing and data science applications with

numpy, SciPy and matplotlib, Second edition.

Apress Media LLC, 2018.

doi: 10.1007/978-1-4842-4246-9/COVER.

[20] N. Ketkar and J. Moolayil, Deep Learning

with Python. Apress, 2021,

doi: 10.1007/978-1-4842-5364-9.

[21] S. Russel and P. Norvig, Artificial

[23] O. Masurekar, O. Jadhav, P. Kulkarni, and S.

Patil, “Real Time Object Detection Using

YOLOv3,” Int. Res. J. Eng. Technol., 2020,

Accessed: Jan. 21, 2023. [Online]. Available:

[24] M. Bustamante, R. Martínez, R. Suazo N., and

D. Sharma, Manual de control de malezas.

Escuela Agricola Panamericana, 2014, 2014.

Accessed: Jan. 21, 2023. [Online]. Available:

All authors have contributed equally to the creation

of this article.

No funding was received for conducting this study.

Creative Commons Attribution License 4.0