Dynamic Location-based multi-level Otsu

Clustering protocol in wireless sensor networks

for precision agriculture.

Abstract:
Wireless Sensor Networks (WSNs) have increasingly woven themselves into the fabric of our
daily lives, finding applications in domains ranging from smart home systems to inventory
management. In the agricultural sector, WSNs are revolutionizing traditional practices by
providing advanced sensing and communication services, thereby enabling automation in
critical processes like irrigation and fertilization. However, the sensor nodes in WSNs are
typically constrained by their limited, non-rechargeable energy sources, necessitating
efficient management strategies to prolong their operational lifespan.
To address this challenge, a novel dynamic clustering protocol tailored for heterogeneous
networks has been developed, enhancing the strategic deployment of nodes, formation of
clusters, and selection of cluster heads. This protocol introduces a region-based approach for
node deployment, ensuring comprehensive field coverage. It innovatively employs an
expanded multi-threshold Otsu algorithm for cluster formation, which is adapted to
accommodate the diversity of sensor node types within the network. For the pivotal task of
cluster head selection, the protocol leverages Convolutional Neural Networks (CNN),
ensuring an intelligent and adaptive decision-making process. This strategic integration of
methodologies not only extends the lifespan of the sensor nodes but also amplifies the overall
efficacy and reliability of the WSN, marking a significant stride in the application of smart
technologies in agriculture.
 1. Introduction:
Wireless Sensor Networks (WSNs) have become a transformative force in various domains,
and agriculture is no exception. In the context of agriculture, WSNs bring about a paradigm
shift by introducing automation and data-driven decision-making to traditional farming
practices. These networks consist of autonomous sensor nodes equipped with sensing,
computing, and communication capabilities. The collaboration among these nodes allows for
the collection, processing, and transmission of data from the monitored agricultural
environment.
The primary objectives of implementing WSNs in agriculture include precision farming,
resource optimization, early detection of issues, and data-driven decision-making. Precision
agriculture, facilitated by WSNs, involves real-time monitoring of environmental parameters
such as soil moisture, temperature, humidity, and crop health. This enables farmers to
optimize resource usage, including water, fertilizers, and pesticides, based on accurate and
timely information.
A wireless sensor networks consists of hundreds or thousands of battery-powered sensor
nodes that are randomly deployed in a large geographical area. The deployed sensor nodes
are able to collect large amounts of data/information and then communicate it to a central
authority. Sensor nodes are not rechargeable and once their battery gets exhausted they are
considered to be dead nodes. Therefore, a reliable energy efficient communication protocol is
necessary in order to extend the lifetime of the sensor nodes in the WSN
The main objective of this paper is to provide an energy-efficient routing protocol for
automated precision agriculture. Therefore, we propose a novel dynamic energy-efficient
routing protocol for precision agriculture in this paper.
The rest of the paper is organized as follows: design issues and literature review have been
discussed in Section 2, and Section 3 introduces proposed protocol in detail, a comparative
study of simulation based results are presented in Section 4 and finally we provide conclusion
and future work related to our proposed protocol in Section 5.
 2. Related Work
With the inclusion of wireless sensor networks (WSNs) in precision agriculture becoming
more prevalent, major issues such as low-cost production and energy conservation have
to be tackled. Various energy-efficient WSNs have been proposed for the purpose for
precision agriculture. Some of the existing WSNs have been discussed below.
In 2016, Sonam Maurya et al. [1] developed a region-based clustering method which
ensures complete coverage of the agricultural land. A fuzzy-logic based cluster head
selection protocol is utilised, where the parameters considered are energy of the sensor
nodes and distance from BS. The WSN is heterogeneous in nature. A threshold-sensitive
hybrid routing protocol is utilised for data transmission. The simulation results of the
protocol show a significant improvement in network lifetime and energy efficiency over
several existing protocols such as EEHC,DEEC, etc.

In 2015, Singh et al. [2] Highlights the importance of WSN in different applications and
points out the challenge in energy conservation. Fuzzy Based Cluster Head Selection
(FBCHS) protocol, aims to improve the cluster head selections in WSN by selecting them
based on residual energy and distance to the base station which increases the energy
efficiency. Comparative analysis between FBCHS and SEP (stable election protocol)
proves FBCHS has better network lifetime, energy efficiency, and overall performance

In 2017, R. Yadav et al. [3] developed an Energy Efficient Probabilistic Clustering

Technique in WSN. The main objective of the protocol is to improve the energy
efficiency of the WSN. This is done by optimising the number of clusters and the
distribution of cluster heads among the various clusters. This is achieved by integrating
the proposed protocol with pre-existing probabilistic clustering protocols. Simulation
results suggest that when the protocol is applied to various existing algorithms such as
LEACH, HEED, etc. an improvement in energy efficiency and network lifetime is
observed in the WSN.

In 2020, Rajan Sharma et al. [4] proposed a flower pollination algorithm-based clustering
approach for WSNs. This method utilises the pre-existing flower pollination algorithm
(FPA) to optimise the selection of cluster heads (CHs) across the WSN. Apart from this
the protocol also optimises both inter-cluster and intra-cluster data transmissions and also
employs a multi-hop communication strategy, thereby, helping to conserve energy during
the data transmission process as well. Simulation results suggest that the FPA based
protocol significantly improves the stability period and network lifetime of the WSN.
In 2020, Deepika Agarwal et al. [5] introduced a grey wolf optimizer-based clustering
protocol. The GWO algorithm is based on the social hierarchy and hunting tactics of grey
wolves. The protocol’s main objective is to optimise the cluster head selection in WSNs.
The fitness function for the protocol is calculated by considering various parameters such
as intra-cluster distance, residual energy, CH balancing factor, etc. simulation results
suggest that the GWO-c shows a significant improvement over various existing protocols
in terms of energy efficiency, network lifetime and data throughput.

In 2018, N. Mittal. [6] suggested a moth flame optimization (MFO) routing protocol for
clustering the WSN in an energy efficient manner. The major objectives of the MFO
algorithm are improving cluster head selection mechanism and data transmission routes.
These changes assist in improving the energy efficiency, the network lifetime and the
network stability of the WSN. Simulation results show that the application of moth flame
optimization algorithm to a WSN, enables it to significantly outperform several existing
protocols such as LEACH, HEED, SEP, etc.

In 2019, Jingxia Zhang et al. [7] created an Adaptive Clustering algorithm for Dynamic
Heterogeneous Wireless Sensor Networks (ACDH). The ACDH algorithm is used to
optimise the cluster head selection process in a dynamic heterogeneous wireless sensor
network (DHWSNs). The ACDH algorithm adapts to any changes occurring in the
network. The ACDH algorithm uses a weighted probability for cluster head selection.
This probability changes if there any changes in the network topology. Simulation results
suggest that the ACDH algorithm improves various performance metrics such as network
lifetime, data transmission efficiency, etc.

In 2022, Yin-Di Yao et al. [8] proposed an Energy-Efficient Routing Protocol for
Precision Agriculture in Wireless Sensor Networks (WSNs). It uses multi-threshold
image segmentation to address node clustering, combining the Otsu algorithm with node
angle and number for improved Cluster Heads' (CHs) distribution uniformity and load
balance. It identifies communication scenarios to mitigate energy consumption induced
by multipath fading. Closed-form expressions for optimal CH location and number form
an objective function for CH selection based on node residual energy. The integrated
node clustering and CH selection aim to reduce energy consumption in WSNs for
Precision Agriculture.
In 2020, Piyush Rawat et al. [9] introduced a Particle Swarm Optimization-based Energy
Efficient Clustering Protocol (PSO-EEC) for wireless sensor networks. PSO-EEC utilizes
the Particle Swarm Optimization (PSO) technique to select optimal Cluster Heads (CHs)
and relay nodes, aiming to improve network lifetime and minimize energy expenditure.
The protocol is designed for heterogeneous networks with normal and power nodes,
differing in initial energy levels. Parameters such as energy ratio, distance, and node
degree are considered in a fitness function for CH selection using PSO. Multi-hop
communication is employed for data transfer to the Base Station (BS), using relay nodes
selected through another fitness function employing PSO.

In 2019, Rodrigue K. Domga et al. [10] proposed an energy-efficient sensor deployment

strategy for wireless sensor networks (WSNs) with linear topology, focusing on multi-hop
Low-power and Lossy Sensor Networks (LWSN) with convergecast traffic. It addresses
the challenge of uneven energy distribution and potential network failure near the sink
due to high traffic load and overhearing. The proposed approach divides the LWSN into
virtual nodes, allowing for wireless connectivity between consecutive virtual nodes. A
greedy algorithm calculates the number of physical sensors needed for each virtual node,
considering traffic load. Unlike existing solutions, it doesn't require precise inter-sensor
distances and accounts for both message transmissions and receptions, improving
network lifetime by up to 40%.
In 2015, Stefanos A. Nikolidakis et al. [11] proposed an energy-efficient automated
irrigation management in agriculture which utilizes wireless sensor networks (WSNs) to
monitor climate parameters such as humidity, temperature, and wind. It employs
historical data and compares past and current states to make efficient irrigation decisions,
considering factors like temperature changes. The complexity of the model is 2n + n, with
n as the number of nodes in the cluster. The system comprises two subsystems: WSN for
data collection and a decision-making system. Sensors placed in parcels monitor soil and
air conditions, sending data wirelessly to a base station. The system adjusts irrigation
based on crop types, soil components, and seasonal variations, providing real-time
monitoring for growers. The routing protocol assumes a base station and homogeneous
sensor nodes organized into clusters.

In 2013, Jing Li et al. [12] addressed energy and fault tolerance challenges in Wireless
Sensor Networks (WSNs) for black pepper monitoring in tropical Precision Agriculture.
It optimizes antenna height and introduces the Energy-efficient Sensor Protocol for
Information via Negotiation (ESPIN). ESPIN reduces redundant data transmission and
enhances success rates compared to traditional protocols. A proposed Symmetrical
Double-Chain (SDC) topology surpasses the traditional tree topology in network lifetime.
The system aids farmers in optimizing resource utilization for improved black pepper
productivity in tropical agriculture.
In 2017, Korhan Cengiz et al. [13] proposed an Energy Aware Multi-Hop Routing
(EAMR) protocol that utilizes fixed clusters to establish communication between sensor
nodes and the base station. Once a sensor node joins a cluster, it remains a member for
the entire network lifetime, reducing the energy overhead associated with frequent cluster
formation. EAMR employs fixed clusters near the base station for direct data
transmission, while other cluster heads forward data to relay nodes for multi-hop
transmission, enhancing scalability and reducing communication energy. The protocol
consists of set-up and steady-state phases, where fixed clusters are formed, and data
collection and transmission occur, with the possibility of dynamic cluster head and relay
node changes during the steady-state phase.

In 2018, Xiaohan Lai et al. [14] introduced predicted remaining deliveries (PRD), which
is a combination of link quality, distance factors etc. to improve network performance and
energy efficiency. PRD balances energy consumption across the network by focusing on
various factors such as the node’s residual energy. PRD significantly improved network
performance, especially in harsh environmental conditions, over traditional routing
protocols.

In 2022, Premkumar Chithaluru et al. [15] presented ARFOR which is a routing protocol
that helps to extend the network lifetime. This is done by addressing the various energy
constraints in the WSN. Two types of roles are given to the sensor nodes by the protocol,
one is being a parent node and the other is being a volunteer node. The roles are assigned
based on various criteria such as residual energy, proximity, etc. ARFOR also
significantly improved the network efficiency in situations where the volunteer nodes are
quite less.

In 2020, Zongshan Wang et al. [16] presented a novel protocol to improve the energy
efficiency in wireless sensor networks (WSN). A cluster head selection method based on
the artificial bee colony (ABC) algorithm is used to select cluster heads by considering
several factors such as node energy levels, distance, etc. further, a combination of fuzzy
c-means and ant colony optimization is used for clustering and routing in the WSN. This
method facilitated in improving the energy efficiency of the WSN.

In 2021, Akhilesh Panchal et al. [17] created Energy Efficient Optimum Number of
Cluster Head and Grid Head Selection (EOCGS) which is a protocol used to optimize the
number of cluster heads and grid heads in a WSN. This is done by taking various factors
such as residual energy, node distance, etc. into consideration. The implementation of
EOCGS in a WSN shows significant improvement in network lifetime, energy
consumption, etc. it is also dynamic in nature, thereby, quickly adapting to any changes in
the WSN.
In 2021, Trupti Mayee Behera et al. [18] created an adaptive fuzzy-based cluster head
selection protocol, whose main objective was to improve the energy efficiency of the
network. This protocol considers the residual energy of the nodes and the distance as the
primary parameters to select a cluster head. The protocol is also dynamic in nature and
quickly adapts to any changes in the network. Simulation results have shown significant
improvements in both energy efficiency and throughput of the WSN when compared to
other existing protocol.

In 2021, Akhilesh Panchal et al. [19] introduced an Energy Efficient Hybrid Clustering
and Hierarchical Routing (EEHCHR) protocol. The major objective of the protocol is to
optimize energy consumption among cluster heads and cluster members. EEHCR uses an
adaptive and hybrid clustering protocol which considers various factors such as base
station location, residual energy of nodes, etc. clustering only occurs in selected rounds
and CH selection is optimized with fitness function, designed to conserve energy. Further,
a hierarchial routing method is used with the help of direct cluster heads (DCH) and
central cluster heads (CCH) to manage the packet routing. EEHCHR has shown
significant improvement in various performance metrics over several existing protocols.

In 2021, Jia Yanfei et al. [20] presented an Energy-Efficient Routing Protocol Based on
Zone for Heterogeneous Wireless Sensor Networks. This protocol uses a zone-based
clustering process where different types of sensor nodes are deployed in distinct zones of
the heterogeneous wireless sensor network (WSN). A new and improved threshold for
cluster head selection is introduced and this ensures that nodes with higher energy levels
are given prioritised for cluster head selection. A mixture of single-hop and multi-hop
data transmission methods are used to improve the network performance. Simulation
results show a significant improvement in network efficiency when compared to other
existing protocol.

In 2020, Yu Han et al. [21] developed a Meta-Heuristic Approach to cluster a wireless

sensor network. The clustering protocol based on Meta-Heuristic approach (CPMA)
works by integrating two separate meta-heuristic algorithms. These are the harmony
search (HS) algorithm and artificial bee colony (ABC) algorithm. The HS algorithm is
used for dynamic cluster head selection and the ABC algorithm is used to optimize the
protocol’s parameters. CPMA algorithm works in two different phases: online and
offline. When online, HS algorithm is used to select cluster heads by considering various
factors such as energy usage. When offline, ABC algorithm is used to adapt to the
specific needs of the network. Simulation results suggest that a significant improvement
in network lifetime is observed.
In 2022, Shankar Madkar et al. [22] proposed a machine learning-based efficient routing
protocol. This protocol combines machine learning techniques with cyclic redundancy
check (CRC) methods. This improves data recovery and routing efficiency in the wireless
sensor network. Channel impulse response (CIR) prediction based on sensor node
location is used to improve the packet delivery ratio and reduce the data loss in the WSN.
Simulation results show that the machine learning-based efficient routing protocol (ML-
ERP) notably enhance the packet delivery ratios and improve the energy efficiency over
other conventional protocols.

In 2022, Hemant Kumar Vijayvergia et al. [23] created a multi-objective Beetle Swarm
Optimization (BSO) algorithm. The main objective of the algorithm is to balance the load
around the gateways and create optimal routing paths. The BSO algorithm improves
clustering and routing by using objective fitness functions for both clustering and routing.
This leads to a decrease in power consumption and an increase in energy efficiency.
Simulation results show that the BSO algorithm outperforms several existing algorithms
in various performance metrics such as energy consumption, throughput, network
lifetime, etc.

In 2022, Manar Ahmed Hamza et al. [24] introduced Grey Wolf Improved Particle
Swarm Optimization with Tabu Search Techniques (GW-IPSO-TS). The main objectives
of the protocol are to optimize cluster head selection and create efficient routing paths.
The GW-IPSO-TS protocol uses a unique combination of grey wolf optimization
(GWO) and particle swarm optimization (PSO). This is combined with Tabu search
techniques to improve the energy efficiency of the WSN. Simulation results suggest that
the GW-IPSO-TS protocol outperforms various existing protocols such as LEACH, PSO,
etc. in various performance metrics such as network lifetime, energy efficiency, etc.

Based on the things we have learned from our literature survey, the key components of
our proposed protocol are:

 In the network, heterogeneous types of sensor nodes are considered where some
nodes have less energy, sensing range and communication range as compared to other
nodes. These nodes are randomly deployed in certain regions which are pre-defined
by the user.
 The multi-threshold Otsu algorithm, used to segment images based on the intensity of
the pixels, has been extended to segment multiple different images. A relationship has
been constructed between the cluster formation problem and the multi threshold Otsu
algorithm. a cluster forming method based on the variance of node angle and number
of different types of nodes between clusters is proposed
  A CNN based cluster head election technique has been adapted to reduce the load on
the sensor nodes and to increase the lifetime of the network.

3. Proposed Work

3.1 Assumptions

1. The agricultural field is circular in shape with a radius of R.

2. Three types of heterogeneous sensor nodes are considered for sensing and
transmitting data to BS.

3. BS is located at the centre and has no energy constraints.

4. BS is within the range of each sensor node.

5. All the sensor nodes are aware of their own locations. The location of the nodes can
be achieved by satellite-based GPS system (Global Positioning System) or other
positioning algorithms.

3.2 System Composition

Three classes of heterogeneous sensor nodes as class-1, class-2 and class-3 nodes are
considered in the network for sensing and transmitting information to BS. In the proposed
protocol, the energy level, sensing range and transmitting range for each class-1, class-2 and
class-3 sensor nodes are different. We have considered that the higher energy sensor nodes
have higher sensing and transmitting range. The sensor nodes used here are similar to the
sensor nodes used in [1].
class-1 node’s energy < class-2 node’s energy < class-3 node’s
energy.
class-1 sensing range < class-2 sensing range < class-3 sensing
range.
class-1 node’s transmission range < class-2 node’s transmission
range < class-3 node’s transmission range.
3.3 Energy consumption model
The energy consumption model use here follows the one given in [25]. The first-order radio
energy model for 3-level heterogeneous wireless sensor network is used for calculating
energy dissipation of each CH and CM node. The energy dissipated in a CH node during a
particular round is calculated as:
en ch= ¿ (1)

where N is the packet size and Eelec is the per bit energy dissipation in running the transmitter
or receiver circuit of a sensor device. E fs is the energy consumed by the amplifier to transmit
over a shorter distance, E amp is the energy consumed by the amplifier to transmit at a longer
distance and d th =√ (E fs ¿ ¿ Eamp) ¿ . E DA is the data aggregation energy and dist BS is the average
distance between a CH node and BS. The energy dissipated by CM node during a particular
round is equal to:
en cm= ¿ (2)

where dist CH is the average distance between a CM node and its CH

node. The total initial energy of entire heterogeneous network (E) is
defined as:
E=n × EO ×1+ mo ×(α +m1 × β) (3)

where EO is the initial energy of class-1 nodes. In the network, α and

β are the energy factor i.e. class-3 nodes have β times more energy
than class-1 nodes, and class-2 nodes haveα times more energy
than class-1 nodes.

3.4 Node deployment

The circular agricultural field is split into 3 concentric circular locations where the three
classes of nodes are deployed. The nodes are deployed in the following manner:
 class-1 nodes are deployed in the inner circular region, which is the closest one to the
base station, because they have the least energy and least sensing range.
  class-3 nodes are deployed in the middle circular region and cover the most area out
of all three types of sensor nodes because they have the most energy and the greatest
sensing range out of all three types of nodes.
 class-2 nodes are deployed in the outer circular region.
All three types of sensor nodes are deployed randomly within the defined location. The
reason for deploying class-3 nodes in the middle region instead of the class-2 nodes will be
explained in the following sections.

Algorithm 1. Location-based node deployment algorithm

Input: R; n; r; x
Output: Nodeclass , deployed sensor node in π R 2 area

Initialization: Field area = π R 2, BS position = (0,0)

Node LOC 1= π r 2

Node LOC 2 = π x2−π r 2

Node LOC 3 = π R 2−π x 2

Step 1:// For class-1 node deployment in location-1 (i.e. LOC-1)

Step 2: for i = 1 to Node LOC 1 do

Step 3: Nodeclass (i)= ‘class-1’

Step 5: end for
Step 6:// For class-2 node deployment in location-2 (i.e. LOC-2)
Step 7: for i = Node LOC 1 + 1 to Node LOC 2 do

Step 8: Nodeclass (i) = ‘class-3’

Step 9: class-3 sensor nodes are randomly deployed in LOC-2 which has an area π x2−π r 2
Step 10: end for
Step 11:// For class-2 node deployment in location-3 (i.e. LOC-3)
Step 12: for i = Node LOC 1 + Node LOC 2 + 1 to Node LOC 3 do

Step 13: Nodeclass (i) = ‘class-2’

Step 14: class-2 sensor nodes are randomly deployed in REG-3 which has an area π R 2−π x 2
Step 15: end for
3.5 Cluster formation

3.5.1 Multi-Threshold Otsu algorithm

The cluster formation is based on the multi-threshold Otsu algorithm which is used to find the
optimal threshold of an image to divide the image into many classes according to the
maximization of variance between classes. This has been expanded to segment three different
images which can be correlated to the three different sensors used in the WSN.
If you have three different images, let's call them types 'A', 'B', and 'C', each with its own
colour spectrum (e.g., 'X', 'Y', 'Z'), you can extend the equations as follows:
Probability Distribution (pi,c):
Calculate the joint probability distribution for each colour channel c in the combined image
('A', 'B', and 'C'). The probability of intensity i (pi,c) now considers all three types of pixels
and colour channels:
hi , A ,c +hi , B , c +h i ,Cc
pi,c = (4)
N h , combined, c

where ℎi,A,c, ℎi,B,c and ℎi,C,c, are the number of pixels with intensity i in types 'A', 'B', and 'C'
images for colour channel c, respectively, and Nh,combined,c is the total number of pixels in the
combined image for colour channel c.
Cumulative Probability (βk,c):
Calculate cumulative probability distributions for each threshold thk in each colour channel c
in the combined image.
th k
βk,c = Σ ⅈ=0 pi,c (5)

Average Intensity Levels (μk,c):

Find the average intensity levels for each class Ck and each colour channel c considering all
three types of pixels in the combined image.

thk pi ,c
​Σ ⅈ=0 i
μk,c = β k ,c (6)
β k ,c

Between-Class Variance (σ 2B , c):

Calculate the between-class variance based on the average intensity levels and cumulative
probabilities for each colour channel c in the combined image.
k
σ B , c = ∑ β k,c(μk,c - μT,c)2
2
(7)
k =1

where μT,c is the total gray mean of the combined image for colour channel c:
k
μT,c = ∑ β k,cμk,c (8)
k =1

Objective Function (FOtsu,c):

Define the objective function to maximize the between-class variance for each colour channel
c in the combined image.

FOtsu,c = max(σ 2B , c(TH)), (9)

0≤thi≤L-1,
i=[1,2,….,K-1]
This optimization problem aims to find the intensity levels thi that maximize the between-
class variance for each colour channel c in the combined image.
In summary, these modified equations consider all three types 'A', 'B', and 'C' pixels in each
colour channel. The calculations are done separately for each colour channel, and the
optimization problem is solved independently for each channel in the combined image.

3.5.2 Node clustering based on angular variance and number of nodes

the node clustering problem has been correlated with the multi-threshold Otsu algorithm as
done in [8], where the angular threshold and the number of nodes in each cluster is chosen as
the main deciding factors to decide the clusters.
The node clustering using multi-threshold Otsu segmentation is done as follows:
 The joint probability distribution for each sensor class ‘t’ in the network is calculated.
The probability of angular separation ‘Ɵ’ (pi,t) considers all three classes of sensors in
the network.

hθ , A ,t +hθ , B , t+ hθ ,C ,t
pi, t= (10)
N h , combined ,t

Where hθ , A , t , hθ , B ,t , hθ , C , t are the number of sensor nodes with angular

separation Ɵ in types ‘A’ , ‘B’ and ‘C’ sensors for class ‘t’ respectively and
N h , combined ,t is the total number of sensors of class ‘t’ in the combined network.

 Cumulative probability distribution for each angular thresholdT k in each class ‘t’ in
the combined sensor network is calculated.
 Tk
β k ,t = ∑ pi , t (11)
Ɵ=0

 Average angular separation level for each cluster C k and each class ‘t’ considering all
three classes of sensors in the combined network is to be found
Tk

∑ θ . pi, t (12)
Ɵ=0
μk , t =
β k ,t

 The between-class angular variance based on the average angular separation levels
and cumulative probabilities for each class ‘t’ in the combined sensor network is
calculated.

K
σ 2β ,t =∑ β k ,t (μ k ,t −μT ,t )2 (13)
k=1

Where μT ,t is the total mean angular separation of sensors for class ‘t’:

K
μT ,t =∑ β k ,t . μk , t (14)
k=1

 The objective function to maximize the between-class angular variance for each class
‘t’ in the combined sensor network is defined.

2
FOtsu , t=max ⁡( σ β ,t ( T h ) ) (15)

Where 0 ≤ T h ≤ L−1∧h ∈ [ 1 , 2 , … … , K −1 ] .

3.5.3 finding optimal clusters using a hybridized PSO-WCA algorithm

In order to find the optimal clusters, various swarm intelligence algorithms such as bat
algorithm, artificial bee colony, etc. are widely used. In this scenario we have chosen
a hybridized particle swarm optimisation (PSO)- water cycle algorithm (WCA)
algorithm has been chosen to attain optimal clusters.
 To attain the optimal angular threshold using PSO-WCA algorithm the following
steps are to be taken:

 N particles are initialized with positions X i and velocities V i for i=1,2,…,N.

 Each X i has three components: X i 1, X i 2, X i 3 , corresponding to the three types of sensor

nodes in the network
Similarly, each V i has three components V i 1 ,V i 2,V i 3

 Similarly, N particles are initialized with water level W i , precipitation level Pi and
evaporation coefficient E vap

 α ∧β are the weighing factors for the fitness function

 The fitness of each particle is evaluated using the objective function which is the
FOtsu objective.

Fitness ( X ij ) =¿ α .Overall Angular Variance+ β . Number Variance (16)

Hybrid PSO-WCA update:

 Update particle swarm velocity:

(t +1 ) t t t t
V ij =λ . V ij + (1−λ ) . min ⁡( K water , W i + Pi −E vap .W i ) (17)

 Update particle swarm position:

(t +1) t (t+1 )
X ij =X ij +V ij (18)

 Evaluate fitness for each particle using the FOtsu objective function

 Update personal best and global best if fitness improves

 Update water levels:

(t +1) t t t
Wi =W i + Pi −Evap . W i (19)

 Update precipitation levels based on the FOtsu objective function

 (t +1) ( t +1)
Pi =FOtsuObjective ( X ij ) (20)

Algorithm 2: Cluster Formation Algorithm

Require:
 N sensor nodes
 Number of clusters K ¿
 Maximum iterations t max
Ensure:
 Angular threshold set T ¿
 Optimal clusters set C opt

Step 1: Initialize particle swarm positions X i and velocities V i

Step 2: Define inertia weight ω , cognitive coefficient c 1, social coefficient c 2

Step 3: repeat
Step 4: for each particle i=1 to N do
Step 5: Evaluate fitness for particle i using FOtsu function
Step 6: Update personal best position if fitness improves
Step 7: Update global best position if fitness improves
Step 8: Update particle velocity and position using PSO equations
V (ijt+1 )=λ . V tij + ( 1−λ ) . min ( K water ,W ti + Pti −Evap . W ti )
(t +1) t (t+1 )
X ij =X ij +V ij

Step 9: end for

Step 10: until convergence or maximum iterations reached
¿
Step 11: Let T = Global best position
 ¿
Step 12: Acquire optimal clusters C opt based on T
Function: FOtsu
Step 13: Calculate joint probability distribution pi , t for each type 't '
hθ , A ,t + hθ , B ,t + hθ ,C , t
pi, t=
N combined ,t

Step 14: Calculate cumulative angular probability β k ,t

Tk
β k ,t = ∑ pi , t
Ɵ=0

Step 15: Calculate average angular separation μk , t for each cluster C k and each class ‘t ’
Tk

∑ θ . pi, t
Ɵ=0
μk , t =
β k ,t

K
Step 16: Calculate between-class angular variance σ 2
β ,t σ 2
β ,t =∑ β k ,t (μ k ,t −μT ,t )2
k=1

2
Step 17: Define objective function FOtsu , t as max ⁡(σ β , t ( T h ))

Step 18: return FOtsu , t

Function: PSO
Step 19: Initialize particle positions and velocities
Step 20: repeat
Step 21: for each particle i=1 to N do
Step 22: Evaluate fitness for particle i using FOtsu function
Step 23: Update personal best position if fitness improves
Step 24: Update global best position if fitness improves
Step 25: Update particle velocity and position using PSO equations
V (ijt+1 )=λ . V tij + ( 1−λ ) . min ( K water ,W ti + Pti −Evap . W ti )
(t +1) t (t+1 )
X ij =X ij +V ij

Step 26: end for

Step 27: until convergence or maximum iterations reached
Step 28: return Global best position
Step 29: Call PSO to find optimal angular threshold T ¿
 ¿
Step 30: Acquire optimal clusters C opt based on T

3.6 Cluster head selection

3.6.1 Cluster head selection using convolutional neural network (CNN)
The BS first asks every sensor node to advertise itself while broadcasting a TDMA schedule.
Subsequently, every node sends its ID, starting energy, and position data to BS. Only class-3
sensor nodes are which are present in the middle region of the field are considered for CH
selection. The BS chooses the optimal CH node from the middle region using CNN. Once the
CH node has been selected, the BS broadcasts the IDs and position data of the new CHs. This
data is kept in internal memory by each sensor node in the network. The assumption made in
this network is that BS is reachable by every sensor node, allowing the nodes to communicate
field conditions and their own initial data, such as IDs, initial energy, and location.
Class-1 sensor nodes in each cluster will directly transfer the sensed data to the BS. All the
class-2 and class-3 sensor nodes will transfer the sensed data to the cluster head which will
aggregate all the collected data and send it to the BS. . Since the CH node in the middle
region has to collect data from both class-2 and class-3 CM nodes present in the middle
region and the outer region respectively, the CH node has to have a large amount of energy in
order to manage the huge amount of data aggregation tasks entrusted to it. This is why class-3
nodes are deployed in the middle region instead of the outer region.
The utilization of CNN enhances the appropriate distribution of energy dissipation between
the sensor nodes in each cluster. Every round, BS considers two factors while choosing the
CH node, these two factors are: the class-3 node's distance from BS and the class-3 node's
residual energy. During the data pre-processing phase of cluster head selection using a
Convolutional Neural Network (CNN), each class-3 sensor node's distance and residual
energy are calculated and normalized. The nodes transmit their positional data and residual
energy to the base station, where these values are processed to form the input for the CNN
model.
The CNN model consists of several layers, including convolutional layers with ReLU
activation, pooling layers, and fully connected layers, culminating in an output layer that uses
a sigmoid function. The training phase utilizes an Adam optimizer and a binary cross-entropy
loss function. For selecting cluster heads, the model predicts probabilities for each node,
ranking them accordingly. A selection algorithm then iteratively adds nodes to the cluster
heads set based on their rank and the clusters they belong to, avoiding duplicates until the
desired number of cluster heads is formed. This method ensures an efficient selection
process, leveraging the CNN's predictive capabilities to identify optimal nodes as cluster
heads.
Algorithm 3: CH Selection Algorithm

Require: N sensor nodes, Distance threshold TH distance, Residual energy threshold TH energy

Ensure: Selected CH nodes for the current round

Step 1: for each Class-3 sensor node:
Step 2: Transmit positional data and energy to the BS.

Step 3: distance= √ x 2 + y 2

distance−min distance
Step 4: Normalized distance=
max distance−min distance
residual energy−min energy
Step 5: Normalized energy=
max energy−min energy
Step 6: end for
Step 7: X =[ Normalized distance , Normalized energy ]
Step 8: CNN Model Architecture

Step 9: Conv i=ReLU (Convolution ( X , filter i ) )

Step 10: Pool i=MaxPooling(Conv i)

Step 11: FC i=ReLU (FullyConnected ( Pool i ))

Step 12: Output =Sigmoid (FullyConnected ( FC n ) )

Step 13: Training

Step 14: Optimizer= Adam ( learning rate )

Step 15: Loss=BinaryCrossEntropy (Output , Labels)

Step 18: Predictions=Predict( X)

Step 19: Rank nodes based on predictions

Step 20: Initialize an empty set Cluster heads
Step 21: Initialize an empty set Selected Clusters
Step 22: for each node i in Rank:

Step 23: If i' s cluster c i is not in Selected Clusters

Step 24: Add node i to Cluster heads

Step 25: Add cluster c i to Selected Clusters

Step 26: If the length of Cluster heads reaches N

Step 27: end for

3.7 Data transmission

Once CHs are selected and the transmission scheduled is made the following protocol is
followed:
 Class-1 CMs will periodically collect environmental data and transmit the data
directly to the BS in their TDMA timeslot.
 Class-2 and Class-3 CMs collect environmental data and transmit it to the CH in their
TDMA timeslot.
 CHs, which belong to Class-3, receive and aggregate the data collected from both the
CMs present in the middle region and the outer region and forwards it to BS.

4. Simulation and analysis

4.1 simulation results

For the simulation purposes, we have taken a circular of radius 100 metres. The base station
is located at the centre of the circular field. The field is divided into three circular regions
where the three classes of sensor nodes are randomly deployed within the specified regions.
Class-1 sensor nodes are deployed within 0 to 20 metres of the base station. Class-3 sensor
nodes are deployed within 20 to 70 metres of the base station and Class-2 sensor nodes are
deployed within 70 to 100 metres of the base station as specified in Algorithm-1. The
simulation parameters are mentioned below.
Table 1
Simulation parameters
parameters Values
Total number of sensor nodes(n) 85
 Total number of class-1 sensor nodes 25
Total number of class-2 sensor nodes 30
Total number of class-3 sensor nodes 30
Initial energy of class-1 sensor node 0.3J
Initial energy of class-2 sensor node 0.6J
Initial energy of class-3 sensor node 0.9J
E amp 0.0013 pJ/bit/m^4
E fs 10 pJ/bit/m^2
Eelec 50 nJ/bit
E DA 5 nJ/bit/report
packet size (N) 500 bytes

Table 2
Simulation results
Performance metrics LEACH EEHC DEEC HEED PROPOSED
PROTOCOL
Number of rounds after which 1st sensor 34 2205 851 1391 217
node is dead
Number of rounds after which all sensor 152 3824 3161 3763 4442
nodes are dead
Number of rounds after which 1st class-1 57 843 851 1393 1390
sensor node is dead
Number of rounds after which all class-1 59 2238 1217 1454 1498
sensor nodes are dead
Number of rounds after which 1st class-2 34 1130 1643 1451 1416
sensor node is dead
Number of rounds after which all class-2 54 1938 2170 2043 2596
sensor nodes are dead
Number of rounds after which 1st class-3 69 2205 2362 2590 217
sensor node is dead
Number of rounds after which all class-3 152 3824 3161 3763 4442
sensor nods are dead
Number of packets transmitted to the 2108 6465 103619 94814 123469
BS(throughput)
Fig.1 overall dead sensor nodes Fig.2 class-1 dead sensor node
at each round at each round

Fig.3 class-2 dead sensor nodes Fig.4 class-3 dead sensor nodes
at each round at each round
Fig.5 Residual energy of the WSN at Fig.6 number of packets transmitted
rounds 1000,1500,2000 to BS at each round

we have done a comparative study of the proposed protocol against LEACH, HEED, EECH
and DEEC protocols.
4.2 Analysis of Results
In this section, we’ll be doing a complete analysis of the simulation we have done and the
results we have achieved.
Fig.1 shows the network lifetime of all LEACH, HEED, EECH, DEEC, and the proposed
protocols. In LEACH a randomised rotation of the role of cluster head among sensor nodes is
used to distribute the energy consumption in the network. All three classes of sensor nodes
have the possibility of becoming a cluster head irrespective of their energy level and any
other parameters. In HEED a secondary parameter along with the residual energy of the
nodes is used to select a cluster head. Once again all classes of sensor nodes are considered
for the role of cluster head. In EEHC the cluster head is selected randomly after a few rounds,
once again all classes of sensor nodes are considered to be selected as cluster head and
because of this, the network lifetime is significantly reduced. In DEEC the cluster head
selection is based on a probabilistic ratio that considers the average energy of the network
and the current energy level of each node.

The proposed protocol outperforms LEACH, HEED, EECH, and DEEC by 2822%,
18%,16%, and 40% respectively as seen in Fig.1. EEHC outperforms all the other protocols
but the difference is not as significant as seen between the proposed protocol and all the other
protocols. In terms of the network lifetime of class-1 nodes, EEHC outperforms all the other
protocols including the proposed protocol as seen in Fig.2. For the class-2 nodes, the
proposed protocol outperforms LEACH, HEED, EECH, and DEEC by 4707%,27%,34%, and
19% respectively as seen in fig.3. Similarly, the proposed protocol also outperforms all the
other protocols in terms of network lifetime of class-3 nodes. The network lifetime of class-3
nodes is similar to the overall network lifetime of the WSN as seen in Fig.4. The proposed
protocol has the highest amount of residual energy at each round mentioned when compared
to LEACH, HEED, EECH, and DEEC as seen in Fig.5.
The location-based deployment of the different classes of sensor nodes, used in the proposed
protocol, ensures complete coverage of the agricultural land, whereas, this may not be
observed in the other protocols due to the random deployment of nodes. The location-based
deployment also ensures improved energy efficiency among class-2 and class-3 sensor nodes.
Class-3 sensor nodes cover a majority of the agricultural land and have more data aggregation
tasks than class-2 and class-1 sensor nodes. The cluster head selection algorithm only
considers class-3 nodes, which have the highest energy and the greatest communication
range, to be selected as cluster heads. Therefore, the load on class-1 and class-2 sensor nodes
is significantly reduced and the lifetime of the network is improved.

Fig.6 shows the throughput of each protocol till the end of their respective network lifetimes.
The proposed protocol outperforms all the other protocols with a throughput of 123469. It
shows an improvement of 5757%,1810%,19%, and 30% over LEACH, EEHC, DEEC, and
HEED respectively. This is because the class-1 sensor nodes directly send the aggregated
data to the BS instead of sending it to the cluster head.

5. Conclusion and future work

In this dynamic location-based clustering protocol, the location-based deployment ensures
complete coverage of the field and the clustering protocol ensures that the WSN remains
dynamic, thereby making it easier to adapt to any changes occurring in and around the
network. When we compare the protocol with LEACH, HEED, EECH, and DEEC significant
improvements are seen in the network lifetime, throughput, and all other performance
metrics.

For future work, the development of a more robust data transmission protocol could help to
further improve the energy efficiency of the network. Some examples include threshold-
sensitive data transmission protocol, Power-Efficient Gathering in Sensor Information
Systems (PEGASIS), Sensor Protocols for Information via Negotiation (SPIN) etc.

References
[1] S. Maurya and V. K. Jain, "Threshold Sensitive Region-Based Hybrid Routing protocol
for precision agriculture," 2016 IEEE Wireless Communications and Networking Conference
Workshops (WCNCW), pp. 169-174
[2] K. Singh and A. K. Daniel, "Load Balancing in Region Based Clustering for
Heterogeneous Environment in WSNs Using AI Techniques," 2015 Fifth International
Conference on Advanced Computing & Communication Technologies, pp. 641-646
[3] Yadav, R.K., Gupta, D. & Lobiyal, D.K. Energy Efficient Probabilistic Clustering
Technique for Data Aggregation in Wireless Sensor Network. Wireless Pers Commun 96,
4099–4113 (2017).
[4] Sharma R, Mittal N, Sohi BS. Flower pollination algorithm-based energy-efficient stable
clustering approach for WSNs. Int J Commun Syst. 2020;e4337.
[5] Agrawal D, Wasim Qureshi MH, Pincha P, et al. GWO-C: Grey wolf optimizer-based
clustering scheme for WSNs. Int J Commun Syst. 2020;e4344.
[6] Mittal, N. Moth Flame Optimization Based Energy Efficient Stable Clustered Routing
Approach for Wireless Sensor Networks. Wireless Pers Commun 104, 677–694 (2019).
[7] Zhang J., Chen J. An adaptive clustering algorithm for dynamic heterogeneous wireless
sensor networks. Wireless Netw 25, 455–470 (2019)
[8] Y. -D. Yao, X. Li, Y. -P. Cui, J. -J. Wang and C. Wang, "Energy-Efficient Routing
Protocol Based on Multi-Threshold Segmentation in Wireless Sensors Networks for Precision
Agriculture," in IEEE Sensors Journal, vol. 22, no. 7, pp. 6216-6231
[9] Rawat, P., Chauhan, S. Particle swarm optimization-based energy efficient clustering
protocol in wireless sensor network. Neural Comput & Applic 33, 14147–14165 (2021).
[10] Domga, R.K., Stanica, R., Tchuente, M. et al. Sensor deployment in wireless sensor
networks with linear topology using virtual node concept. Wireless Netw 25, 4947–4962
(2019).
[11] Stefanos A. Nikolidakis, Dionisis Kandris, Dimitrios D. Vergados, and Christos
Douligeris. 2015. Energy efficient automated control of irrigation in agriculture by using
wireless sensor networks. Comput. Electron. Agric. 113, C (April 2015), 154–163.
[12] Li, J.; Shen, C. An Energy Conservative Wireless Sensor Networks Approach for
Precision Agriculture. Electronics 2013, 2, 387-399
[13] K. Cengiz and T. Dag, "Energy Aware Multi-Hop Routing Protocol for WSNs," in IEEE
Access, vol. 6, pp. 2622-2633, 2018
[14] X. Lai, X. Ji, X. Zhou and L. Chen, "Energy Efficient Link-Delay Aware Routing in
Wireless Sensor Networks," in IEEE Sensors Journal, vol. 18, no. 2, pp. 837-848
[15] P. Chithaluru, S. Kumar, A. Singh, A. Benslimane and S. K. Jangir, "An Energy-
Efficient Routing Scheduling Based on Fuzzy Ranking Scheme for Internet of Things," in
IEEE Internet of Things Journal, vol. 9, no. 10, pp. 7251-7260
[16] Z. Wang, H. Ding, B. Li, L. Bao and Z. Yang, "An Energy Efficient Routing Protocol
Based on Improved Artificial Bee Colony Algorithm for Wireless Sensor Networks," in IEEE
Access, vol. 8, pp. 133577-133596
[17] Panchal, A., Singh, R.K. EOCGS: energy efficient optimum number of cluster head and
grid head selection in wireless sensor networks. Telecommun Syst 78, 1–13 (2021).
[18] T. M. Behera, S. Nanda, S. K. Mohapatra, U. C. Samal, M. S. Khan and A. H. Gandomi,
"CH Selection via Adaptive Threshold Design Aligned on Network Energy," in IEEE
Sensors Journal, vol. 21, no. 6, pp. 8491-8500
[19] Akhilesh Panchal, Rajat Kumar Singh, EEHCHR: Energy Efficient Hybrid Clustering
and Hierarchical Routing for Wireless Sensor Networks, Ad Hoc Networks, Volume 123,
2021, 102692, ISSN 1570-8705
[20] Jia Yanfei, Chen Guangda, Zhao Liquan, "Energy-Efficient Routing Protocol Based on
Zone for Heterogeneous Wireless Sensor Networks", Journal of Electrical and Computer
Engineering, vol. 2021, Article ID 5557756
[21] Y. Han, G. Li, R. Xu, J. Su, J. Li and G. Wen, "Clustering the Wireless Sensor
Networks: A Meta-Heuristic Approach," in IEEE Access, vol. 8, pp. 214551-214564
[22] Madkar S, Pardeshi S, Kumbhar MS. Machine learning based efficient routing protocol
in wireless sensor network. Salud Cienc. Tecnol2022;2(S2):195
[23] H. K. Vijayvergia and U. S. Modani, "Energy efficient load balancing and routing using
multi-objective based algorithm in wsn," Intelligent Automation & Soft Computing, vol. 35,
no.3, pp. 3227–3239
[24] M. A. Hamza, A. H. Abdalla Hashim, D. H. Elkamchouchi, N. Nemri, J. S. Alzahrani et
al., "Energy-efficient routing using novel optimization with tabu techniques for wireless
sensor network," Computer Systems Science and Engineering, vol. 45, no.2, pp. 1711–1726
[25] W. B. Heinzelman, A. P. Chandrakasan, and H. Balakrishnan, “An application-specific
protocol architecture for wireless microsensor networks,” IEEE Trans. Wireless Commun.,
vol. 1, no. 4, pp. 660–670