Received 2 July 2023, accepted 10 August 2023, date of publication 21 August 2023, date of current version 1 September 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3307311

An Agriprecision Decision Support System for
Weed Management in Pastures
HOSSEIN CHEGINI 1 , RANESH NAHA 2 , (Member, IEEE), ANIKET MAHANTI
MINGWEI GONG 3 , AND KALPDRUM PASSI 4

1,

1 School of Computer Science, University of Auckland, Auckland 1010, New Zealand

2 Centre for Smart Analytics, Federation University Australia, Churchill, VIC 3842, Australia
3 Department of Mathematics and Computing, Mount Royal University, Calgary, AB T3E 6K6, Canada
4 Bharti School of Engineering and Computer Science, Laurentian University, Sudbury, ON P3E 2C6, Canada

Corresponding author: Mingwei Gong (mgong@mtroyal.ca)

ABSTRACT Pastures are a vital source of dairy products and cattle nutrition, and as such, play a significant
role in New Zealand’s agricultural economy. However, weeds can be a major problem for pastures, making it
a challenge for dairy farmers to monitor and control them. Currently, most of the tasks for weed management
are done manually, and farmers lack persistent technology for weed control. This motivated us to design,
implement, and evaluate a Decision Support System (DSS) to detect weeds in pastures and provide decisions
for the cleanup of weeds. Our proposed system uses two primary inputs: weeds and bare patches. We created
a synthetic dataset to train a weed detection model and designed a fuzzy inference system to assess a pasture.
We also used a neuro-fuzzy system in our DSS to evaluate our fuzzy model and tune its parameters for better
functioning and accuracy. Our work aims to assist dairy farmers in better weed monitoring, as well as to
provide 2D maps of weed density and yield score, which can be of significant value when no digital and
meaningful images of pastures exist. The system can also support farmers in scheduling, recommending
prohibitive tasks, and storing historical data for pasture analysis, collaborated by stakeholders.
INDEX TERMS Fuzzy systems, object detection, pasture management, decision making, decision support
systems, fuzzy neural networks.
I. INTRODUCTION

Pastures provide the main source of nutrition for livestock,
with grass as the primary food source. Production of grass
for cattle significantly impacts several primary industries,
including dairy and meat production, and the milk industry.
According to New Zealand treasury [24], dairy is contributing up to 18.6 billion dollars to New Zealand economy in
2021 with a 5.3 % GDP and 23 % of total export values.
Any dairy farming methodology aims to increase pasture
production, considering the many existing problems limiting
this goal.
One of the most significant and long-lasting problems in
pastures is weeds. They limit grass’s space, nutrients, and
resources, leading to a loss of revenue and negative impact
on dairy production. Although dairy farmers do many tasks
The associate editor coordinating the review of this manuscript and
approving it for publication was Qi Zhou.
92660

to maintain pastures free of weeds, they need a digital tool or
software application to monitor, control, and evaluate their
pastures. They currently do most weed management tasks
manually, using no online and historical information on pastures, and no images, pastoral data, and automatic prohibitive
actions assisting them with weed management.
We have discussed in the section II that the existing
research on weed management mainly focuses on detecting
weeds in crops and not in pastures. Additionally, the majority
of these studies only focus on detection and do not provide further data processing or discussion on how to apply
them in weed management, which is crucial for a practical
application.
In the following, we have cited the studies which could be
categorised into two sections:
1) the studies which have extended their data processing
after weed detection to produce more informative and
practical outputs for farmers

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/

VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

2) the studies which have integrated their models into
other devices or hardware platforms to provide farmers
with more functional help in their fields
Our proposed system also considers the variable of bareness of a pasture, as patches of bare spots where neither grass
nor weeds are growing can shrink pasture productivity. Sowing grass seeds in these locations will contribute to increasing
productivity. We provide both the density of weeds and the
bareness to DSS for enhanced productivity of pastures.
It is important to note that while these studies may have
promising results, they may not be directly applicable to the
specific case of detecting weeds in pastures, as the conditions
and variables may be different. Further research and testing
would be needed to determine the effectiveness of these
methods in the specific context of pasture weed detection.
We can summarise the contributions of our work as
follows:
1) Introduction of quantification into the pastoral environment such as weed densities, bare patches, and yield
score for weed management and assessment
2) Use of fuzzy sets as an explainable and transparent
model in our DSS
3) Assessment and evaluation of our DSS
The paper is organised with a literature review in Section II,
methodology in Section III, results in Section IV, limitation
in Sections V and conclusion and future work in Section VI.
II. RELATED WORK
A. DECISION SUPPORT SYSTEMS (DSS) FOR
PASTURE MANAGEMENT

The advent of sensor technology within the agricultural sector
has fostered the development and execution of a multitude of
applications. Yet, the specific usage of these sensors within
pasture management is an aspect that warrants further investigative research and development.
We have classified our literature review into two categories: in-crop and in-pasture weed detection models. The
studies presented below showcase examples of deep learning models used for weed identification in agricultural crop
fields.
Nguyen et al. [1] designed an agriprecision application
with realistic weed images that were trained by a deeplearning model. The study shows the experiments in three
crop environments, and they have explored the prospect of
integrating a robot with a camera sensor to enable the detection of weeds. In order to evaluate the efficacy of the proposed
system, the authors have presented a set of performance
metrics.
Jogi et al. [40] studied in-crop weed identification with a
deep learning algorithm. The study focuses on the real-time
operations of weed detection to improve latency. They used
a spot-wise method to overcome latency issues arising from
uniform spraying. Kulkarni et al. [26] studied in-crop weed
detection using a Convolutional Neural Network (CNN) to
classify weeds in a crop environment. After detection, the
VOLUME 11, 2023

final task is sending detection images to farmers’ cell phones.
López-Granados et al.
Reference [36] examined features of weeds with an
object detection algorithm in an in-crop environment. Their
algorithm used a small airplane or drone for image collection and specified weed densities as output results.
Lottes et al. [37] studied another in-crop environment with a
weed detection model trained by a dataset of real images. The
research shows training a model on detecting weeds’ joint
stems instead of weed’s leaves or the entire weed. Once they
used the model on a spraying robot, they could detect joint
stems for better pasture monitoring. Irias Tejeda et al. [38]
showed another in-crop study of object detection. Although
they noted using their model in an agriprecision application,
it is not discussed how and where it can be applied in a system
to help farmers.
The following studies demonstrate the weed detection models in in-pasture environments. Chegini et al.
Zhang et al. [25] compared the accuracy of machine learning
and deep learning algorithms in an in-pasture environment
where similar visual characteristics exist between grass and
weeds. Reference [42] used the MaskRCNN model to detect
California thistle in New Zealand pastures and achieved high
accuracy with a synthetic dataset. Elakkiya et al.
The selected studies employed various techniques including object detection algorithms, convolutional neural networks, and joint stem detection models to accurately identify
and classify weeds in different environments. The primary
objective of these investigations was to develop effective
systems that could aid farmers in monitoring and managing
weed populations more efficiently. Some studies explored
the transmission of detection images to farmers’ mobile
devices, while others proposed the use of spraying robots
equipped with joint stem detection capabilities to enhance
pasture monitoring. However, it is important to note that not
all studies extensively discussed the practical implementation and impact of these models and systems in assisting
farmers.
Jin et al. [2] proposed an innovative approach by combining an object detection model with a genetic algorithm for
segmenting images based on color. Their research primarily
concentrated on crop fields and employed two distinct models: one for detecting easily identifiable weeds and another for
detecting unclear and blurry weed instances. This integration
of techniques resulted in improved detection accuracy. However, it is important to note that further discussion is required
regarding the integration of the entire model into a practical
automated system for real-world applications.
Yu et al. [39] conducted research on a deep learning
algorithm trained specifically for three types of weeds. The
study demonstrated high performance of their model in
detecting and classifying these weeds. However, there is a
gap in the discussion regarding the integration and practical
utilisation of this algorithm as a component within a comprehensive weed management system. Further exploration is
necessary to explore the potential integration of their model
92661

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

TABLE 1. Key findings of peer-reviewed works on weed detection.

into real-world applications for effective weed detection and
management.
Chegini et al. [41] presented a system that utilizes weed
density and bare patches as input to calculate a yield score
for pasture areas. The researchers extracted two fuzzy variables from visual data of pastures and integrated them into a
Decision Support System (DSS) model.
Table 1 presents a summary of the reviewed papers, categorizing them based on the environmental study they focus on,
the employed method or algorithm, and whether they discuss
the automation aspect of the system.
Despite the existence of automated Decision Support Systems (DSS) software for various agricultural contexts, such as
soil monitoring and animal behaviour, there is still a need for
more research and attention in the field of pastures. According to literature on precision agriculture, there is a strong
need for software implementation that can effectively process
pastoral environments with a high degree of automation. Such
software could aid dairy farmers in monitoring and implementing preventative actions, as well as assisting with weed
management and destruction.
The DSSs have been designed based on farmers’ behaviour
and actions. Macé et al. [10] depict a simple behavioural DSS
for weed control in three stages:
1) pasture observation
2) choosing proper actions such as spraying or mowing
3) evaluating the pasture
Additionally, incorporating a stage of re-evaluation into
the process can make it iterative, continuous, and consistent.
The aforementioned stages can serve as key considerations
when designing and implementing any DSS model for weed
management.
Sønderskov et al. [31] present a DSS that utilises a
behavioural model from [10] to recommend the appropriate
dose of sprays based on the duration of observation. The
model utilises data from the crop environment, but there is
no explanation on the specific model used in the DSS. Additionally, the study does not address the practical application
of the system in pastures.
Colas et al. [32] conducted a survey among farmers to
gain insights on how to design and implement a DSS for
92662

weed control. The study focuses on the creation of a decision
tree model to present survey data. The survey results indicate
that farmers desire a synthetic tool with rule-based decisions
for a DSS. The study is considered as a prototype for a DSS,
emphasising the need for a practical and functional system
among farmers. However, the study only focuses on the crop
environment and does not provide any modelling or insight
for pastures.
Kanatas et al. [33] evaluated a weed DSS for its accuracy
and effectiveness. According to the study, a DSS should
provide useful information on fields to aid in management.
The DSS should also be designed to be interactive, allowing
for validation and improvement of experiments. The paper
highlights the use of advanced technology and AI models in
DSS to assist farmers in estimating weed growth, assessing
yield, and recommending preventative actions. The paper also
suggests that a DSS should be able to quantify the level of
weeds and evaluate the yield. These suggestions and recommendations from the paper have been applied in practical use
in our current study.
Vishwajith et al. [34] developed a DSS that utilises a single
computer-aided platform on crop environment to assist farmers in acquiring basic information on soil, water, and weed
conditions. However, the study lacks clarity on the specific
model used in their DSS, as well as the methods and data
processing techniques employed. Additionally, the study only
focuses on crops and does not address pasture environments.
While the study provides insight on DSS for crop management, we employed a unique methodology for our study in
pastures using a coded model that can be validated and tested
using pastoral images.
Masin et al. [35] investigated the positive effects of monitoring and recording weed densities in crops. They proposed
that incorporating weed densities as useful crop data can
enhance the reliability, precision, and accuracy of a DSS.
Their DSS model aims to estimate the impact of environmental factors such as temperature, rainfall, and soil temperature
on weed growth. However, the study solely focuses on crops
and does not mention pastures. Similarly to other studies,
we employed an in-pasture analysis of weed densities in
our DSS.
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

When it comes to DSS, it is important for farmers and
dairy farmers to have easy-to-use and transparent systems.
One reason why dairy farmers may not be inclined to use
software platforms and technologies for weed management
is that they can be complex and lack transparency in their
internal operations for data processing. This motivated us to
use a fuzzy inference system for data processing.
Fuzzy sets and fuzzy inference systems are useful models
for handling real systems and environments. They follow
human linguistics and provide transparency in data processing and internal operations, in contrast to Neural Networks
(NN). Some studies have employed fuzzy systems and fuzzy
sets in their DSS.
S. Sivamani et al. [16] utilised a fuzzy system for animal
control. The system takes two inputs of age and weight, and
provides four outputs of change-diet, change-diet schedule,
need health check-up, and be-ready-for-a-sale. The fuzzy
inference system suggests actions for the farmers based on
two fuzzy membership functions. For example, the output of
change-diet would indicate that the farmers should change the
animals’ diets.
Nguyen-Anh and Le-Trung [17] address the problem
of adaptive programming in an IoT environment, using a
fuzzy inference system for controlling complex contexts.
Khanum et al. [18] applied a fuzzy system for understanding
the conditions of leaves and detecting fungal diseases, using
five inputs.
Pandey et al. [19] designed a fuzzy system for agricultural
data processing, using crop input data for disease detection to
aid decision-making on sprays. They used two inputs, wind
and temperature. These examples demonstrate the various
ways fuzzy systems can be employed for different agricultural problems to aid farmers in improving the accuracy and
timing of routine crop, stock and pasture management tasks.
Few studies on modelling agricultural environments with
fuzzy systems motivated us to use fuzzy inference systems
in our DSS.
Table 2 summarises the studies on fuzzy inference systems
in agriculture. The papers presented in this section demonstrate how fuzzy systems can be used for agricultural analysis.
As they are efficient in data processing, we employed fuzzy
systems in our DSS for processing pastoral images. These
studies typically produce outputs that can aid farmers in
decision-making, whether through classifying tasks or recommending the best course of action.
Table 3 summarises the key findings and directions from
our literature review. We found that a DSS that produces weed
density and yield scores is highly desired by dairy farmers.
We also incorporated a fuzzy inference system and added a
quantification module to convert visual data into variables of
weed density and bare patch, which was recommended by
our review of DSS papers. Our study context is pasture, which
has not been extensively studied in the reviewed papers. After
the design and implementation of our DSS, we also applied
an Adaptive Neuro-Fuzzy Inference System to evaluate and

VOLUME 11, 2023

adjust the internal parameters of the DSS, resulting in a more
accurate and functional system.
B. WEED DETECTION MODELS FOR PASTURE
MONITORING

Our DSS and fuzzy inference system work with visual data
produced by our object detection model. We have used the
Mask Region Convolutional Neural Network (MaskRCNN)
to process pastoral images and detect weeds. MaskRCNN
is a state-of-the-art model in object detection. There are
studies that have used object detection models in agricultural
applications.
Thanh V. Le et al. [1] employed a FasterRCNN model
trained with realistic weed images for improved latency
in testing mode. Jin et al. [2] used Mobile net, VGG, and
a CNN model with 15000 training images for feature
extraction of weeds and detecting 10 weed types in crops.
Abdulsalam et al. [4] used You Only Look Once (YOLO) and
a ResNet model for classifying four types of weeds. However,
these studies focus on weed detection in crop environments
and do not discuss their application in pastures.

FIGURE 1. Three main components of our literature review in our system
design and implementation.

Figure 1 illustrates the three main components of the literature that we studied for designing and implementing our DSS.
These are fuzzy inference systems, weed detection models,
and Adaptive Neuro-Fuzzy Inference Systems (ANFIS). The
combination of these research components influenced our
concept of having a fuzzy inference system that can be trained
and enhanced by image output.
III. THE PROPOSED DSS WEED MANAGEMENT SYSTEM

This section describes the design and implementation of our
DSS weed management system and its components. Our DSS
consists of three main components:
92663

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

TABLE 2. Examples of fuzzy inference systems on agricultural applications.

TABLE 3. Key findings of reviewed papers on DSS.

1) MaskRCNN model capable of receiving pastoral
dataset images and training for weed detection
2) Fuzzy inference system for processing weed density
and bareness
3) A Neuro-Fuzzy system called ANFIS for evaluating
the DSS
We considered the importance of weed information
obtained from images in decision-making. As our survey
study revealed that seeding can help control weed invasion
and growth, we defined a second model based on empty areas
of a pasture to identify bareness in pastures.
We utilised a fuzzy inference system to process the output
of the MaskRCNN model. Figure 2 illustrates our DSS design
and implementation flowchart. The weed detection component includes the stages of image preparation, synthetic
dataset creation, and model training. The next component,
below the weed detection model, is a fuzzy system, which
includes a fuzzy inference system for 2D map creation and
yield score calculation. An Adaptive Neuro-Fuzzy Inference System (ANFIS) is a component that improves the
system.
For the first component, we created two synthetic datasets
for weed and bare patch detection. We then developed a
weed detection model and trained it with these datasets. The
accuracy of the model was enhanced by fine-tuning its hyperparameters. We then stored the masks and image outputs in
an array. The quantification component converts the weed
output masks into clear ratios and links the weed detection
software to a fuzzy inference system. The fuzzy inference
system receives the weed and bareness values and evaluates
the pasture.
A. MASKRCNN

To detect weeds from images captured in the field, our system
employs MaskRCNN. As an enhancement of FasterRCNN,
which was used for object detection such as weeds, MaskRCNN generates masks of detected objects. These masks are
used to calculate the density of weeds.
92664

The model processes pastoral images as input and detects
the weeds and bare spaces. A synthetic dataset was used to
train the models. The steps for creating a synthetic dataset
are as follows:
1) object extraction from pastoral images
2) background creation by erasing weeds from the images
3) setting up the maximum number of weeds in every
image
4) setting up the image resolution
5) defining the required weed orientation such as transformation, rotation, and scale
6) setting up many images
7) attaching weed objects to the background
After creating the synthetic dataset, the object detection
model can be trained. Figure 4 illustrates a schematic process
of creating our synthetic dataset. We extracted weed and
bare patch objects from the images in the first step. Then,
we transformed and placed them in the background images.
We repeated this process for the number of images required
for our synthetic dataset.
A few examples of the MaskRCNN model are shown in
Figure 3. The images were collected in the field under various
conditions for different types of weeds. Despite the variety
of weed types and growth patterns, the model was able to
accurately and reliably detect the density of weeds.
Figure 5 displays the detected weeds and empty spaces by
our model. The top image shows two weeds in the middle and
several scattered empty spots. The bottom images depict the
detected weeds and empty spots in colorful masks.
Once the model was trained, we proceeded to carefully
select the critical hyperparameters that would significantly
influence the accuracy of the model. In this experiment,
we fine-tuned a dependable range for each hyperparameter
and trained the model multiple times, considering various
values within those ranges. The following hyperparameters
were specifically chosen for this section:
1) learning rate
2) RoI
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 2. Flowchart of our DSS: weed detection, quantification, fuzzy inference, and fuzzy evaluation.

TABLE 4. Precision, Recall, and F1 score values for weed and empty
models.

3) Maximum instance
4) RESNET backbone
Epochs and image resolutions were two main training
parameters studied in-depth. The study revealed that the scale
of 640 X 480 had the best accuracy. Table 4 presents the
final results of our experiments on the number of images and
epochs in the training set. Precision, recall, and F1-score are
the evaluation metrics in our study.
B. QUANTIFICATION

This section describes the module that we have designed and
coded to calculate the ratio of weeds and bareness in an image
based on the number of detected weeds and bare patches.
We calculate the ratio by using the number of output masks
and their areas. For example, an array with the shape of (480,
640, 7) represents the detection of seven weeds. The output
masks of weeds have a value of 1.
Tweeds = if

Tbareness = if

#weeds
X
i=1
#empty
X

r[640, 480, i] > 0

(1)

r[640, 480, i] > 0

(2)

i=1

P
Total weed pixels
Tweeds
Weed ratio =
=
Total image pixels
640 ∗ 480
P
Total bare pixels
Tbare patch
Bare patch ratio =
=
Total image pixels
640 ∗ 480
VOLUME 11, 2023

(3)
(4)

We have two approaches for the quantification process:
• Calculating from bounding box
• Calculating from mask
Equations 1-4 depict the quantification stages for calculating the weed-to-grass ratio and the bareness-to-grass
ratio. Equation 1 calculates the area of detected weeds, and
equation 2 calculates the area of detected bareness. Equation
3 calculates the weed-to-grass ratio, and equation 4 calculates
the bareness-to-grass ratio. The results are two scalars, which
are used as crisp input for the fuzzy inference system.
C. FUZZY INFERENCE SYSTEM

This section illustrates the design and implementation of our
fuzzy inference system. Fuzzy systems are particularly useful
for measuring the complexity and uncertainty of a process
by defining linguistic variables and fuzzy rules, as discussed
in [11] and [12]. The following reasons summarise why a
fuzzy inference system is powerful for modelling complex
and uncertain processes:
1) Fuzzy systems are best to manage uncertainties in an
environment.
2) Fuzzy systems are explainable.
3) Fuzzy rules are flexible and can be adjusted according
to process changes.
4) Fuzzy systems are one of the best models for decision
making.
We constructed our fuzzy logic model with two variables:
weed density and bareness. We then defined our fuzzy rules
based on different combinations of variable conditions. The
following are the three stages of a fuzzy inference system:
1) Fuzzification: converting crisp values of weed density
and bareness into fuzzy membership functions
2) Fuzzy rules’ excitement: execution of fuzzy rules to
drive a fuzzy output
92665

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 3. The MaskRCNN output on several images.

3) Defuzzification: Converting fuzzy output into crisp
values
1) FUZZY VARIABLES

A fuzzy variable, also known as a linguistic variable, is a function that represents the membership of a variable to a certain
92666

phenomenon. It ranges from 0 to 1, with 0 indicating minimal membership and 1 indicating maximum membership.
Before using any fuzzy system, the input must be converted
into a fuzzy variable. For example, a fuzzy variable might
describe a car’s speed with measurements such as fast, slow,
and medium, instead of using exact numerical values (km/h).
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 4. The schematic diagram of synthetic creation.

FIGURE 6. Five fuzzy membership functions for pasture productivity.

2) FUZZY RULES AND SURVEY DATA

FIGURE 5. MaskRCNN model’s output on detecting weeds (left-bottom)
and detecting empty spaces (right-bottom).

Therefore, in the first stage, we need to convert crisp data into
linguistic variables that can be understood by a fuzzy system.
In a fuzzy system, several types of fuzzy membership
functions or fuzzy linguistic variables can be defined and
designed for a particular problem. The triangular function is
the most common type used. Other types include trapezoidal,
gaussian, pending, linear, and bell.
For our decision support system (DSS), we designed
and coded two fuzzy variables: weed density and bareness.
We categorised each with three fuzzy functions: low density,
medium density, and high density for weed density, and three
fuzzy functions for bare patches.
For the fuzzy output variable, we defined a variable to show
pasture productivity with a qualifying degree, named the yield
score, and categorised it into five conditions, each represented
by a fuzzy triangular function: excellent yield, good yield,
average yield, poor yield, and very poor yield.
VOLUME 11, 2023

Fuzzy rules are the core of a fuzzy inference system and
are used for reasoning. They process the fuzzy inputs and
produce the output as fuzzy membership functions. They are
written in the form of ‘‘if-then’’ statements. The rules can also
be more complex, such as ‘‘if input1 and input2 then.’’
There are two types of fuzzy output systems: Mamdani and
Takagi, Sugeno, and Kang (TSK) [13]. The Mamdani fuzzy
system uses a fuzzy membership function for its output, while
the TSK fuzzy system uses a linear proposition to represent
the fuzzy output. In our study, we used the Mamdani type for
our fuzzy inference system.
Equation 5 shows the fundamental Mamdani formula in a
fuzzy inference system. Equation 6 shows the same formula
for adjusting the fuzzy membership and rule numbers. The
dividend of the equation has an outer sigma, which sums up
the output of the nine rules we have defined in our fuzzy
inference system. Each proposition in the sigma contains
a y and internal production of two inputs. The production
calculates the membership functions µ, of weed density and
bareness of each rule. x is a quantised value of weed density
and bareness. In each proposition, the membership function
of x is multiplied by y inverse, which results from the output
of each rule. In the divisor, there is no y inverse, and the final
yield score is achieved after the dividing operation.
We needed 3 × 3=9 rules to define all the conditions based
on our fuzzy rules, as we had two input variables (weed
density and bareness) each with three conditions.
Table 5 shows the fuzzy rules of our decision support
system (DSS). Each rule checks a condition of weed density
and bareness, leading to an Adaptive Network-based Fuzzy
Inference System (ANFIS) calculation to determine the yield
score. If more conditions and situations of the pasture need
to be considered, for example by adding more inputs, they
should be included in the fuzzy rules.
Figure 6 displays the five membership functions representing different quality states of a typical pasture, ranging from
‘‘very poor’’ to ‘‘excellent.’’
P#Rules −1 Q#Inputs l
µi (xi )
l=1 y
i=1
Yield score = P
(5)
#Rules Q#Inputs l
µi (xi )
l=1
i=1
P#9 −1 Q#2 l
l=1 y
i=1 µi (xi )
Yield score = P
(6)
#9 Q#2
l
l=1
i=1 µi (xi )
92667

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

We can extend the fuzzy rules by adding more inputs,
such as wind, humidity, and temperature. Following is an
example of a rule that includes the new inputs and their
conditions:
if(weed growth is high) and if(temperature is low) and
if(wind is becoming high) and if(bareness is low) and
if(weed density is high) and if(tiny weeds are above 35%)
then (very low score) and (need a high amount of spray)
3) FUZZY OUTPUT

Fuzzy outputs are the third and final part of a fuzzy inference
system. They represent the result of data processing and rule
handling in a fuzzy inference system. In our case, as we had
used the Mamdani system, we chose a triangular membership
function for our fuzzy output.
Figure 7 illustrates the network diagram of our fuzzy
system, which assesses a pasture. It showcases the flow of
data from the fuzzy inputs to the nine fuzzy rules and their
integration to generate the fuzzy output.

A. FUZZY INFERENCE SYSTEM ENHANCEMENT

This section presents the experiments on our fuzzy inference
system to evaluate its accuracy in predicting yield scores.
We set up our fuzzy structure in the Matrix Laboratory
(MATLAB) software and designed and implemented an
Adaptive Network-based Fuzzy Inference System (ANFIS)
to train our pastoral data for yield score predictions.
ANFIS is a combination of a neural network model and
a fuzzy system. The neural components are used to train
the membership functions of the fuzzy system to reach the
desired level of accuracy. It has been used in various control
systems. By training our fuzzy inference system, we can
obtain metrics for evaluating the accuracy of yield score
assessment. In this case, we used paired pastoral data points
as training data, consisting of weed density and bareness as
inputs and yield scores as output. Therefore, our training
dataset for ANFIS had two inputs and one output, a total of
three columns. The training process allows us to evaluate our
fuzzy inference system and improve its accuracy by making
adjustments.
v
u
u 1 Data
X
(y((xi ) − y0 (xi ))
(7)
RMSE = t
Data
i=0

FIGURE 7. Our fuzzy network for yield scoring based on weed density
and pastoral bareness.

IV. WEED KNOWLEDGE: PASTURE ASSESSMENT

This section explains how we implemented the fuzzy inference system to produce the desired output. We used the
python packages of skfuzzy and skfuzzy control for coding and implementation. After designing and implementing
the system, we then used Adaptive Network-based Fuzzy
Inference System (ANFIS) to process the fuzzy output and
membership functions for system enhancement and evaluation. We also proposed a Graphical User Interface (GUI) to
display the fundamental parameters and results. A subsection
is included to discuss how the system could be used for
predicting the yield scores. Figure 8 shows the five pastoral
images used for our experiments. Table 6 shows the results
of scoring on each image, which is presented in Figure 8.
Figure 9 shows a simulation of a land, with the colorful maps
showing the density of membership functions and the output
of the yield score. The produced score for the studied land
is 84.27.
92668

We have used the Root Mean Squared Error (RMSE) as our
metric to evaluate these experiments. (Equation 7). y is the
yield score, and ‘‘y0 ’’ is the ANFIS output. Data represents
the number of yield Scores collected over time as fuzzy score
outputs.
We examine two cases of our system: static and dynamic
fuzzy systems. In the static fuzzy system, the parameters of
the membership functions (such as means and deviations) are
fixed and cannot be trained. In contrast, the dynamic fuzzy
system has the capability to train and adjust these parameters
based on the input data. We have used 100 data points as the
training dataset and set up 100 epochs for training. Figure 10
shows the errors of static and dynamic fuzzy systems while
trained with 100 epochs. The DSS without training and with
no parameter enhancement has an error of 0.25 as the red dot
in Figure 10.
In the next stage, we performed hyperparameter tuning to
improve the accuracy of our Adaptive Network-based Fuzzy
Inference System (ANFIS). The number of membership functions and the type of membership functions are two main
parameters that are crucial for the fuzzy model’s accuracy
and performance. To find the best value of fuzzy membership functions, we changed the configuration of our ANFIS
model, trained it and observed the accuracy. To experiment
with the shape or type of membership functions, we considered four main types of membership functions. After
configuring the ANFIS with each type, we trained our model
and recorded the accuracy.
Figure 11 shows the changes in RMSE metric based on
increasing numbers of the two mentioned parameters. The
best values of accuracy are in the centre of the radar plot.
As values close to the centre of the most internal circle
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

TABLE 5. Fuzzy rules of our weed system framework.

FIGURE 8. Five sample images of our examined pasture with their fuzzy output.

represent the best experimentation, the bell-shape of the
membership function and ten membership functions would
have the best accuracy. In other words, using ten membership
functions with a bell shape will result in the best accuracy of
our ANFIS model.
Algorithm 1 shows the sequence of pastoral image processing. The first step is to train a weed detection model
using several pastoral images. Next, we quantify the images
by producing crisp numbers representing weed density and
bareness. These crisp numbers are the input data for the
fuzzy inference system. The rules and defuzzification components then produce the yield score. We then use an Adaptive
Network-based Fuzzy Inference System (ANFIS) model to
train fuzzy membership parameters. We conduct our training and record the Root Mean Squared Error (RMSE), our
evaluation metric, according to different parameters and configurations.
In this section we tried to illustrate a prototype for the
proposed DSS model for pasture forecasting. The fuzzy inference system can produce yield scores on a specific date and
time. Collecting yield scores at different time intervals can
lead to a time series of yield scores. Having historical data on
a pasture in the form of a time series can help dairy farmers
VOLUME 11, 2023

predict their pasture yield. This way, they can have a better
insight into the productivity of their pastures and organize
their tasks in a proactive manner rather than a reactive one.
Figure 12 illustrates the process of sequencing the three
variables: weed density, bareness, and yield score. Each variable can represent a recording point of a time series, suitable
for any predictive forecasting model. Incorporating the forecasting values into the yield score of the fuzzy inference
system can enhance the decision support system (DSS) for
better and more accurate services for dairy farmers.

B. A COMPARISON TO SEGMENT ANYTHING
MODEL (SAM)

When it comes to the functionalities of MaskRCNN and
SAM, both operate on pixel-wise images and generate masks
for the objects they can recognize within an image. However,
there is a significant distinction between the two. MaskRCNN
is trained by targeting a specific object within an image,
whereas SAM can detect any object present in an image.
Another key difference is the size of their trained models. MaskRCNN’s model size is 200MB, while SAM’s is
approximately 2.5GB, which could potentially cause latency
92669

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 9. The produced 2D maps of our weed system framework. (a) shows a pasture divided into
100 individual images. (b) shows the 2D map of weed density. (c) shows the bareness 2D map, and
(d) shows the 2D map of the yield score. For this studied pasture, the yield score is 84.27.

TABLE 6. Fuzzy rules of our weed system framework.

FIGURE 10. The errors of our static fuzzy system with no training (upper
section) and with training (lower section).

issues during model transfer and deployment in real-world
applications.
The training time of MaskrRCNN is much lower than
SAM. This is also true about the inference time. For high
resolution images a RAM crash message may be received.
92670

For inference time, images were scaled to different resolutions all in 3:4 ratio and the inference time was recorded.
Figure 13 shows a snapshot of the image scales in various
resolutions.
Figure 14 shows the inference time of the same image
with different resolutions and scales. The analysis shows how
costly it is to detect objects on a high-resolution image with
more than 4 minutes as compared to a small resolution image
which takes less than a minute. With this experiment, for
a practical application the resolution analysis is a crucial
stage which should determine the best image scale in the
inference time.
For the SAM analysis we demonstrate the impact of Union
over Intersection (UoI) on the number of detected masks.
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 11. RMSE error of the number and type of membership functions.

Algorithm 1 The Algorithm of Calculating a Yield
Score for a Pasture and the Fuzzy System Evaluation
Input: N: Number of images for a pasture
Image processing section
for 1 to N do
Calculate the weed density of image(i)
Calculate the bareness of image(i)
end
Quantification section
For each image produce the crisp numbers of weed
density and bareness
Fuzzy inference section
Fuzzify the weed density and bareness values
Incite the rules
Calculate the fuzzy score for each image
Defuzzify
Average the scores of all images of pasture
Result: Scoring section
ANFIS: The Neuro Fuzzy evaluation section
Prepare the yield score data as paired data for training
set
Configure the fuzzy structure
for 1 to Epochs do
Train the weights(j) of neurons of ANFIS
Apply the changes to weights
Record the RMSE
end
Output: The last RMSE

By setting a low UoI prediction ratio of 0.74, the model
detects around 500 masks. In contrast, setting a higher prediction ratio of 0.92 results in the detection of only 100 masks.
This highlights the importance of studying and properly
setting parameters such as UoI in determining the model’s
behaviour.
C. A PREDEFINED DASHBOARD FOR DAIRY FARMERS

FIGURE 12. The time-series creation of fuzzy inference system.

Figure 15 shows that increasing the prediction ratio of
UoI leads to a more conservative model, resulting in fewer
detected masks. On the contrary, a lower prediction ratio will
lead to more masks being detected, as the model becomes less
strict in its criteria.
Figure 16 illustrates this trend, the two images display the model’s output for different UoI prediction ratios.
VOLUME 11, 2023

This section presents a recommended Graphical User Interface (GUI) for dairy farmers, which can be used for data
upload and model training as services for pasture management. Figure 17 shows a predefined dashboard that can
be used during data entry and model training for pasture
management. To avoid complexity, the GUI includes simple
controlling components for image upload and model type.
The user can also define the parameters of a fuzzy inference
system, such as input type and the number of fuzzy rules. The
trained models and resulting 2D maps can be downloaded
and used.
D. A COLLABORATIVE FRAMEWORK FOR DAIRY FARMERS

The power of our fuzzy inference system is not limited
to assessing pastures and recommending actions for weed
monitoring, but also in generating pasture data as historical knowledge. Other potential contributions of our system
could be:
92671

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 16. The mask output of two images by SAM.

FIGURE 13. The snapshot of image scales.

FIGURE 14. The inference time analysis of SAM.
FIGURE 17. The Dashboard of DSS model for pasture management.

FIGURE 15. The mask analysis of SAM.

1) A pasture dataset that comes from pasture monitoring
of weed density, bareness and yield score
2) A federated model that comes from aggregation of local
models of dairy farmers
A pasture dataset can be created when the system is producing yield scores according to weed density and bareness.
Each time of monitoring can result in a new row of a dataset
92672

and by adding more data we can have a more sophisticated
knowledge of a particular pasture. We can even include environmental variables such as temperature, wind, humidity,
moisture, rain, pressure, etc in the dataset and calculate cross
correlations among them to gain new knowledge. If any weed
action is conducted, the monitoring tasks can provide the
impact of the action, which can be used in the dataset for more
accurate knowledge.
In federated learning, a number of local models (any type
of object detection) are aggregated in a server for a more
accurate model. In this method, each local model is labelled
and saved. It is then transferred to an enterprise storage server
to be processed and aggregated. The global model from the
aggregated models can then be sent back to the local users,
improving their weed detection and pasture assessment.
With these two concepts we can define an innovative architecture for a collaborative weed management system that
helps farmers to use the service of monitoring weeds but
in the meanwhile collaborating to the knowledge of weed
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

FIGURE 18. A weed management architecture based on weed dataset and federated mdoels.

management at the same time. Figure 18 shows the proposed
system architecture.
V. LIMITATION

For this study, we faced limitations in accessing a pasture
to take photos and collect images. We initially planned to
conduct our experimentation on a real pasture using tiling
images, but due to COVID-19 restrictions, we experimented
with a simulation map. Additionally, Python does not have
a package or library for coding Adaptive Network-based
Fuzzy Inference System (ANFIS) and training the fuzzy
model, so we performed our training experimentation using
the ANFIS model in MATLAB.

dairy farmers, helping them to have much better pasture
monitoring.
As future work, one can study the influences of other
variables such as temperature, humidity, and wind on pasture
productivity. A longitudinal study on the improved productivity of using the Decision Support System (DSS) would
provide insight into how weed invasion starts, spreads, and is
controlled, and how effectively seeding covers bare patches.
Satellite images can be another useful source for weed controlling and covering bare patches of pastures. Adopting more
advanced object detection algorithms for measuring weed
density will also improve pasture productivity.
REFERENCES

VI. CONCLUSION AND FUTURE WORK

This research integrates several Artificial Intelligence (AI)
models and topics: an object detection model, a fuzzy inference system, Adaptive Network-based Fuzzy Inference System (ANFIS), and a time-series analysis. Our work has
contributed to scoring and recommending actions for dairy
farmers, which has not been previously conducted. As dairy
farmers in New Zealand lack technological tools for managing their pastures, the proposed system can help them
understand their pastures and manage them systematically.
Processing two random variables of weed density and the
bareness of a pasture can provide sufficient knowledge for
VOLUME 11, 2023

[1] V. N. T. Le, G. Truong, and K. Alameh, ‘‘Detecting weeds from crops under
complex field environments based on faster RCNN,’’ in Proc. IEEE 8th Int.
Conf. Commun. Electron. (ICCE), Jan. 2021, pp. 350–355.
[2] X. Jin, J. Che, and Y. Chen, ‘‘Weed identification using deep learning
and image processing in vegetable plantation,’’ IEEE Access, vol. 9,
pp. 10940–10950, 2021.
[3] P. Bir, R. Kumar, and G. Singh, ‘‘Transfer learning based tomato leaf disease detection for mobile applications,’’ in Proc. IEEE Int. Conf. Comput.,
Power Commun. Technol. (GUCON), Oct. 2020, pp. 34–39.
[4] M. Abdulsalam and N. Aouf, ‘‘Deep weed detector/classifier network for
precision agriculture,’’ in Proc. 28th Medit. Conf. Control Autom. (MED),
Sep. 2020, pp. 1087–1092.
[5] C. L. Susilawati, ‘‘Rainwater management model development for agriculture in the Savu Island semi-arid region,’’ Civil Eng. Dimension, vol. 14,
no. 1, pp. 36–41, Mar. 2012.
92673

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

[6] A. Bonfante, E. Monaco, P. Manna, R. De Mascellis, A. Basile,
M. Buonanno, G. Cantilena, A. Esposito, A. Tedeschi, C. De Michele,
O. Belfiore, I. Catapano, G. Ludeno, K. Salinas, and A. Brook, ‘‘LCIS
DSS—An irrigation supporting system for water use efficiency improvement in precision agriculture: A maize case study,’’ Agricult. Syst., vol. 176,
Nov. 2019, Art. no. 102646.
[7] M. Rinaldi and Z. He, ‘‘Decision support systems to manage irrigation in
agriculture,’’ J. Adv. Agronomy, vol. 123, pp. 229–279, Jan. 2014.
[8] S. Fountas, D. Wulfsohn, B. S. Blackmore, H. L. Jacobsen, and
S. M. Pedersen, ‘‘A model of decision-making and information flows
for information-intensive agriculture,’’ Agricult. Syst., vol. 87, no. 2,
pp. 192–210, Feb. 2006.
[9] P. G. Cox, ‘‘Some issues in the design of agricultural decision support
systems,’’ J. Agricult. Syst., vol. 52, nos. 2–3, pp. 355–381, 1996.
[10] K. Macé, P. Morlon, N. Munier-Jolain, and L. Quéré, ‘‘Time scales as
a factor in decision-making by French farmers on weed management in
annual crops,’’ Agricult. Syst., vol. 93, nos. 1–3, pp. 115–142, Mar. 2007.
[11] L. A. Zadeh, ‘‘Toward a theory of fuzzy information granulation and its
centrality in human reasoning and fuzzy logic,’’ Fuzzy Sets Syst., vol. 90,
no. 2, pp. 111–127, 1997.
[12] H. J. Zimmermann, Fuzzy Set Theory and Its Applications. Dordrecht, The Netherlands: Springer, 2021. [Online]. Available: https://link.
springer.com/book/10.1007/978-94-010-0646-0
[13] G. Chen and T. Pham, Fuzzy Sets, Fuzzy Logic, and Fuzzy Control Systems.
Boca Raton, FL, USA: CRC Press, 2000, p. 328,
[14] R. J. Wieringa, ‘‘Design science methodology: For information systems
and software engineering,’’ J. Des. Sci. Methodol., Inf. Syst. Softw. Eng.,
vol. 90, no. 2, pp. 1–332, 2014.
[15] A. R. Hevner, ‘‘A three cycle view of design science research,’’ Scandin.
J. Inf. Syst., vol. 19, no. 2, pp. 87–92, 2007.
[16] S. Sivamani, H. G. Kim, J. Park, and Y. Cho, ‘‘A study on decision support
system based on the fuzzy logic approach for the livestock service management,’’ Int. J. Services Technol. Manag., vol. 23, nos. 1–2, pp. 83–100,
2017.
[17] T. Nguyen-Anh and Q. Le-Trung, ‘‘RFL-IoT: An IoT reconfiguration
framework applied fuzzy logic for context management,’’ in Proc. IEEERIVF Int. Conf. Comput. Commun. Technol. (RIVF), Mar. 2019, pp. 1–6.
[18] A. Khanum, A. Alvi, and R. Mehmood, ‘‘Towards a semantically enriched
computational intelligence (SECI) framework for smart farming,’’ in Proc.
Int. Conf. Smart Cities, Infrastruct., Technol. Appl., in Lecture Notes of the
Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, vol. 224, 2018, pp. 247–257.
[19] P. Pandey, R. Litoriya, and A. Tiwari, ‘‘A framework for fuzzy modelling in
agricultural diagnostics,’’ J. Européen des Systèmes Automatisés, vol. 51,
nos. 4–6, pp. 203–223, Dec. 2018.
[20] A. Kumar, A. Sharma, and A. Nayyar, ‘‘Fuzzy logic based hybrid model
for automatic extractive text summarisation,’’ in Proc. 5th Int. Conf. Intell.
Inf. Technol., 2020, pp. 7–15.
[21] Y. Hafeez, S. Ali, N. Jhanjhi, M. Humayun, A. Nayyar, and M. Masud,
‘‘Role of fuzzy approach towards fault detection for distributed components,’’ Comput., Mater. Continua, vol. 67, no. 2, pp. 1979–1996, 2021.
[22] G. W. Bourdôt, S. V. Fowler, G. R. Edwards, D. J. Kriticos, J. M. Kean,
A. Rahman, and A. J. Parsons, ‘‘Pastoral weeds in new zealand: Status
and potential solutions,’’ New Zealand J. Agricult. Res., vol. 50, no. 2,
pp. 139–161, Jun. 2007.
[23] J. T. Saunders, G. Greer, G. Bourdôt, C. Saunders, T. James, C. Rolando,
J. Monge, and M. S. Watt, ‘‘The economic costs of weeds on productive
land in New Zealand,’’ Int. J. Agricult. Sustainability, vol. 15, no. 4,
pp. 380–392, Jul. 2017.
[24] New Zealand Treasury. (2021). FEU Special Topic—Medium-Term
Outlook for Dairy Exports. Accessed: Apr. 24, 2023. [Online].
Available:
https://www.treasury.govt.nz/publications/research-andcommentary/rangitaki-blog/feu-special-topic-medium-term-outlookdairy-exports
[25] W. Zhang, M. F. Hansen, T. N. Volonakis, M. Smith, L. Smith, J. Wilson,
G. Ralston, L. Broadbent, and G. Wright, ‘‘Broad-leaf weed detection
in pasture,’’ in Proc. IEEE 3rd Int. Conf. Image, Vis. Comput. (ICIVC),
Jun. 2018, pp. 101–105.
[26] S. Kulkarni and S. A. Angadi, ‘‘IoT based weed detection using image
processing and CNN,’’ Int. J. Eng. Appl. Sci. Technol., vol. 4, no. 3,
pp. 606–609, Jul. 2019.
[27] K. Simiński, ‘‘Neuro-fuzzy system with weighted attributes,’’ Soft Comput., vol. 18, no. 2, pp. 285–297, Feb. 2014.
92674

[28] M. Hassanzadeh and Z. Rahmani, ‘‘An intelligent predictive controller for power and battery management in plug-in hybrid electric
vehicles,’’ J. Energy Resour. Technol., vol. 143, no. 11, Nov. 2021,
Art. no. 112105.
[29] Z. Pourtousi, S. Khalijian, A. Ghanizadeh, M. Babanezhad, A. T. Nakhjiri,
A. Marjani, and S. Shirazian, ‘‘Ability of neural network cells in learning
teacher motivation scale and prediction of motivation with fuzzy logic
system,’’ Sci. Rep., vol. 11, no. 1, May 2021, Art. no. 9721.
[30] J. B. Mohapatra, P. Jha, M. K. Jha, and S. Biswal, ‘‘Efficacy of
machine learning techniques in predicting groundwater fluctuations in
agro-ecological zones of India,’’ Sci. Total Environ., vol. 785, Sep. 2021,
Art. no. 147319.
[31] M. Sønderskov, R. Fritzsche, F. de Mol, B. Gerowitt, S. Goltermann,
R. Kierzek, R. Krawczyk, O. M. Bøjer, and P. Rydahl, ‘‘DSSHerbicide:
Weed control in winter wheat with a decision support system in three south
Baltic regions—Field experimental results,’’ Crop Protection, vol. 76,
pp. 15–23, Oct. 2015, doi: 10.1016/j.cropro.2015.06.009.
[32] F. Colas, S. Cordeau, S. Granger, M.-H. Jeuffroy, O. Pointurier, W. Queyrel,
A. Rodriguez, J. Villerd, and N. Colbach, ‘‘Co-development of a decision support system for integrated weed management: Contribution from
future users,’’ Eur. J. Agronomy, vol. 114, Mar. 2020, Art. no. 126010, doi:
10.1016/j.eja.2020.126010.
[33] P. Kanatas, I. S. Travlos, I. Gazoulis, A. Tataridas, A. Tsekoura and
N. Antonopoulos, ‘‘Benefits and limitations of decision support systems
(DSS) with a special emphasis on weeds,’’ Agronomy, vol. 10, no. 4,
pp. 1–10, 2020.
[34] K. P. Vishwajith, P. K. Sahu, B. S. Dhekale, P. Mishra, and C. Fatih,
‘‘Decision support system (DSS) on pulses in India,’’ Legume Res., vol. 43,
no. 4, pp. 530–538, 2020.
[35] R. Masin, V. P. Vasileiadis, D. Loddo, S. Otto, and G. Zanin, ‘‘A single-time
survey method to predict the daily weed density for weed control decisionmaking,’’ Weed Sci., vol. 59, no. 2, pp. 270–275, Jun. 2011.
[36] F. López-Granados, J. Torres-Sánchez, A. Serrano-Pérez, A. I. de Castro,
F.-J. Mesas-Carrascosa, and J.-M. Peña, ‘‘Early season weed mapping in
sunflower using UAV technology: Variability of herbicide treatment maps
against weed thresholds,’’ Precis. Agricult., vol. 17, no. 2, pp. 183–199,
Apr. 2016.
[37] P. Lottes, J. Behley, N. Chebrolu, A. Milioto, and C. Stachniss, ‘‘Robust
joint stem detection and crop-weed classification using image sequences
for plant-specific treatment in precision farming,’’ J. Field Robot., vol. 37,
no. 1, pp. 20–34, Jan. 2020.
[38] A. J. I. Tejeda and R. C. Castro, ‘‘Algorithm of weed detection in crops
by computational vision,’’ in Proc. Int. Conf. Electron., Commun. Comput.
(CONIELECOMP), Feb. 2019, pp. 124–128.
[39] J. Yu, S. M. Sharpe, A. W. Schumann, and N. S. Boyd, ‘‘Deep learning
for image-based weed detection in turfgrass,’’ Eur. J. Agronomy, vol. 104,
pp. 78–84, Mar. 2019, doi: 10.1016/j.eja.2019.01.004.
[40] Y. Jogi, P. N. Rao, Raksha, S. Shetty, and Shreekari, ‘‘CNN based synchronal recognition of weeds in farm crops,’’ in Proc. 4th Int. Conf. Electron., Commun. Aerosp. Technol. (ICECA), Nov. 2020, pp. 1373–1378.
[41] H. Chegini, F. Beltran, and A. Mahanti, ‘‘Fuzzy logic based pasture assessment using weed and bare patch detection,’’ in Proc. SSA Conf., Jun. 2021,
pp. 1–18, doi: 10.1007/978-3-030-88259-4_1.
[42] H. Chegini, F. Beltran, and A. Mahanti, ‘‘Designing and developing a weed
detection model for California thistle,’’ ACM Trans. Internet Technol.,
vol. 23, no. 3, pp. 1–29, Jul. 2022, Art. no. 48, doi: 10.1145/3544491.
[43] A. Sial, A. Singh, and A. Mahanti, ‘‘Detecting anomalous energy consumption using contextual analysis of smart meter data,’’ Wireless Netw., vol. 27,
no. 6, pp. 4275–4292, Jul. 2019, doi: 10.1007/s11276-019-02074-8.
[44] A. Sial, A. Singh, A. Mahanti, and M. Gong, ‘‘Heuristics-based detection of abnormal energy consumption,’’ in Proc. Int. Conf. Smart Grid
Inspired Future Technol., Jul. 2018, pp. 21–31, doi: 10.1007/978-3-31994965-9_3.
[45] R. Elakkiya, V. Subramaniyaswamy, V. Vijayakumar, and A. Mahanti,
‘‘Cervical cancer diagnostics healthcare system using hybrid object detection adversarial networks,’’ IEEE J. Biomed. Health Informat., vol. 26,
no. 4, pp. 1464–1471, Apr. 2022.
[46] R. Elakkiya, D. K. Jain, K. Kotecha, S. Pandya, S. S. Reddy,
E. Rajalakshmi, V. Varadarajan, A. Mahanti, and V. Subramaniyaswamy,
‘‘Hybrid deep neural network for handling data imbalance in precursor
MicroRNA,’’ Frontiers Public Health, vol. 9, pp. 1–12, Dec. 2021, doi:
10.3389/fpubh.2021.821410.
VOLUME 11, 2023

H. Chegini et al.: Agriprecision DSS for Weed Management in Pastures

[47] D. K. Jain, A. Mahanti, P. Shamsolmoali, and R. Manikandan, ‘‘Deep
neural learning techniques with long short-term memory for gesture
recognition,’’ Neural Comput. Appl., vol. 32, no. 20, pp. 16073–16089,
Oct. 2020, doi: 10.1007/s00521-020-04742-9.
[48] H. Chegini, R. K. Naha, A. Mahanti, and P. Thulasiraman, ‘‘Process
automation in an IoT–fog–cloud ecosystem: A survey and taxonomy,’’ IoT,
vol. 2, no. 1, pp. 92–118, Feb. 2021, doi: 10.3390/iot2010006.
[49] H. Chegini and A. Mahanti, ‘‘A framework of automation on contextaware Internet of Things (IoT) systems,’’ in Proc. 12th IEEE/ACM Int.
Conf. Utility Cloud Comput. Companion, Dec. 2019, Art. no. 157162, doi:
10.1145/3368235.3368848.

HOSSEIN CHEGINI received the bachelor’s
degree in software engineering and the master’s
degree in AI from the University of Tehran, Iran,
and the Ph.D. degree in information systems from
the University of Auckland. He is currently a Data
Scientist experienced in AI and ML pipelines for
software production. His research interests include
computer vision, evolutionary computation, and
LLM.

RANESH NAHA (Member, IEEE) received the
M.Sc. degree from Universiti Putra Malaysia and
the Ph.D. degree from the University of Tasmania, Australia. He is currently a Research Fellow with Federation University Australia. Prior
to this, he was a Grant-Funded Researcher with
The University of Adelaide. He has authored more
than 30 peer-reviewed scientific research articles. His research interests include the Internet of
Things (IoT), AI and ML, software-defined networking (SDN), distributed computing (fog/edge/cloud), cybersecurity, and
blockchain.

VOLUME 11, 2023

ANIKET MAHANTI received the B.Sc. degree
(Hons.) in computer science from the University of
New Brunswick, Canada, and the M.Sc. and Ph.D.
degrees in computer science from the University
of Calgary, Canada. He is currently a Senior Lecturer (Associate Professor) of computer science
with the University of Auckland, New Zealand.
His research interests include network science, distributed systems, and internet measurements.

MINGWEI GONG received the B.Eng. degree
in computer engineering from Tianjin University,
Tianjin, China, in 2001, and the M.Sc. and Ph.D.
degrees in computer science from the University of
Calgary, in 2003 and 2009, respectively. He is currently an Associate Professor with the Department
of Mathematics and Computing, Mount Royal
University. His research interests include computer networking, resource allocation, and network security.
KALPDRUM PASSI received the Ph.D. degree
in parallel numerical algorithms from the Indian
Institute of Technology, Delhi, India, in 1993.
He is currently a Full Professor with the School
of Engineering and Computer Science, Laurentian
University, Sudbury, ON, Canada. He has collaborative work with faculty in Canada and U.S. and
the work was tested on the CRAY XMP’s and
CRAY YMP’s. He transitioned the research to web
technology and more recently has been involved
in machine learning and data mining applications in bioinformatics, social
media, and other data science areas. He received funding from NSERC
and Laurentian University for the research. He has published many papers
on parallel numerical algorithms in international journals and conferences.
He has also published several papers related to machine learning applications
in medical image processing, natural language processing, sports analytics,
and social media platforms. His research interest includes bioinformatics,
which has been on improving the accuracy of predicting diseases, such as
different types of cancer using microarray data. He is a member of the ACM
and the IEEE Computer Society.

92675