Scientific African 19 (2023) e01550

Contents lists available at ScienceDirect

Scientific African
journal homepage: www.elsevier.com/locate/sciaf

Large scale survey for radio propagation in developing

machine learning model for path losses in communication
systems
Haruna Chiroma a,b, Ponman Nickolas c, Nasir Faruk d,e,∗, Emmanuel Alozie f,
Imam-Fulani Yusuf Olayinka f, Kayode S. Adewole g, Abubakar Abdulkarim h,
Abdulkarim A. Oloyede f, Olugbenga A. Sowande f, Salisu Garba i,
Aliyu D. Usman j, Lawan S. Taura k, Yinusa A. Adediran l
a
Computer Science Department Federal College of Education (Technical) Gombe, Nigeria
b
College of Computer Science and Engineering, University of Hafr Batin, Hafr Batin, Saudi Arabia
c
Department of Mathematical Sciences, Abubakar Tafawa Balewa University, Bauchi, Nigeria
d
Department of Information Technology, Sule Lamido University, Kafin Hausa, Jigawa State, Nigeria
e
Directorate of Information Technology, Sule Lamido University, Kafin Hausa, Jigawa State, Nigeria
f
Department of Telecommunication Science, University of Ilorin, Nigeria
g
Department of Computer Science, University of Ilorin, Nigeria
h
Department of Electrical Engineering, Ahmadu Bello University, Zaria, Nigeria
i
Department of Computer Science, Sule Lamido University, Kafin Hausa, Jigawa State, Nigeria
j
Department of Electronics and Telecommunication Engineering, Ahmadu Bello University, Zaria, Nigeria
k
Department of Physics, Sule Lamido University, Kafin Hausa, Jigawa State, Nigeria
l
Department of Electrical and Electronics Engineering, University of Ilorin, Nigeria

a r t i c l e i n f o a b s t r a c t

Article history: Several orthodox approaches, such as empirical methods and deterministic methods, had
Received 8 November 2022 earlier been used for the prediction of path loss in wireless communication systems. These
Revised 6 January 2023
approaches are either inefficient or complex. Robustness and performance motivated the
Accepted 9 January 2023
adoption of machine learning for modeling path loss in wireless communication systems
in place of traditional modeling schemes. Surveys on modeling path loss in communication
Editor: DR B Gyampoh systems exist in the literature; however, emerging deep learning architectures in-depth
analysis, machine learning taxonomies related to path loss, and in-depth analysis on fea-
ture engineering in modeling path loss are missing in the already published surveys. To

Abbreviations: SVM, Support Vector Machine; CNN, convolutional neural network; GA, Genetic Algorithm; ANN, artificial neural network; MLPNN, Multi-
layer Perception Neural Network; ReLU, Rectified linear tangent; FC, fully connected; ANFIS, adaptive-network-based fuzzy interference system; LOS, line of
sight; EIRP, Effective Isotropic Radiated Power; SINR, signal-to-interference-noise ratio; SNR, Signal to Noise Ratio; MDS, Minimum Detectable Signal; LAMS,
local area multi-line scanning; AE, attention-enhanced; DRNN, deep recurrent networks; LSTM, long short-term memory; PCA, principal component analy-
sis; RSRP, Reference Signal Receiving Power; RBFNN, Radial Basis Function Neural Network; TEMS, test mobile system; HTAF, hyperbolic tangent activation
function; LM, Levenberg-Marquardt; UWB, ultra-wide band; SVR, Support Vector Regression; UAV, unmanned aerial vehicles; KNN, K-nearest neighbour;
RT, regression tree; CI, Close-In; SDE, standard deviation error; MAE, mean absolute error; MSE, mean square error; CC, correlation coefficient; BPSK, binary
phase shift keying; MLFS, multi-layer fuzzy logic-based system; DE), Differential Evolution; CoDE, Composite Differential Evolution; PSO, particle swarm
optimization; ABC, artificial bee colony; GPS, global positioning system.
∗
Corresponding author at: Department of Information Technology, Sule Lamido University, Kafin Hausa, Jigawa State, Nigeria.
E-mail address: [email protected] (N. Faruk).

https://doi.org/10.1016/j.sciaf.2023.e01550
2468-2276/© 2023 The Authors. Published by Elsevier B.V. on behalf of African Institute of Mathematical Sciences / Next Einstein Initiative. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Keywords: fill this existing gap, a survey of machine learning modeling for path loss in wireless com-
Path loss modeling munication systems is conducted to resolve the outlined issues, hence making this survey
5G wireless communication systems unique. Synthesis and analysis of deep learning architectures to solve path loss problems
Deep learning
in communication systems are hereby presented. New taxonomy – deep learning, nature-
Artificial neural network
Feature engineering
inspired meta-heuristic algorithms, and shallow algorithms approach to path loss model-
Machine learning ing have been created. Analysis of feature engineering in path loss modeling is exploited.
Lastly, challenges militating against the full potential of modeling path loss based on ma-
chine learning are highlighted and discussed. Alternative approaches for resolving the out-
lined challenges are also presented in the survey to help in designing more practical ap-
plications in the future.
© 2023 The Authors. Published by Elsevier B.V. on behalf of African Institute of
Mathematical Sciences / Next Einstein Initiative.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

Introduction

The need to understand the technological advances in radio propagation and the study of how signals are transmitted
have made most research efforts in the area of mobile communication more interesting. Radio Propagation path loss and
its associated importance have been major considerations in the design of wireless networks. Signal propagation is the
movement of radio waves from the source (Transmitter) to the destination (Receiver). These waves are propagated from one
place to another, likened to light waves. Signals can propagate over considerable distances but can be affected by the choice
of the transmission medium and the objects that may appear on the transmission path. The transmission can be affected
by different phenomena such as signal reflection, diffraction, scattering, or absorption [79], possibly leading to signal fading.
The fading could occur within a short distance (small-scale fading) or over a long distance, which is commonly referred to
as propagation path loss.
Propagation path loss represents the reduction in signal strength of electromagnetic waves as it is being propagated
through a material medium, which could be space, vacuum, or built-up environment [48]. Path loss models are classified
as deterministic, probabilistic, and empirical models [85]. Determination of propagation path loss is very important in the
design of wireless communication systems as it is found to be one of the major constraints when deploying emerging
technologies such as 4G and 5G wireless networks [80]. Propagation path loss monitoring and evaluation have to be carried
out using certain models that have been developed over time [5]. However, inconsistencies in the models’ performance
were observed when the models are deployed in different environmental scenarios as reported in the works of Afolayan
et al. [108], Oloyede et al. [109], Faruk et al. [20,21], Faruk et al. [22], Surajudeen-Bakinde et al. [97], Jimoh et al. [110],
Greenberg and Klodzh [111], Faruk et al. [23], Popoola et al. [112] and Faruk et al. [24].
Path loss modeling is crucial in wireless communication system planning as it allows network designers to meet certain
expectations and design requirements [38]. According to Garg [32], path loss modeling plays a vital role in the design of
wireless networks as it affords the opportunity of specifying important system parameters such as transmission power,
frequency, and antenna heights. These models make predictions for a predetermined environment where their applications
are found suitable. Models have been developed to function in different environments which include indoor environments,
outdoor environments, urban, suburban, and rural environments. Propagation path loss models are also used to estimate the
expected number of cell sites that are needed for effective coverage for the designed network. Also, according to [32], the
coverage requirement and traffic requirements with every model, are taken care of by the design propagation model and
will determine the distribution of the network traffic.
Traditional models such as basic models have been used for decades for the modeling of propagation path loss. However,
the weakness of the traditional models, such as insufficient efficiency, lack of robustness, and ineffectiveness, motivated the
introduction of machine learning in modeling path loss in wireless communication systems. Path loss modeling employed
algorithms such as Artificial Neural Network (ANN), evolutionary ANN, Genetic Algorithm (GA), fuzzy systems, Support Vec-
tor Machine (SVM), K-nearest neighbor, etc. Deep learning architectures such as the Convolutional Neural Network (CNN) can
be applied for path loss modeling in communication systems. The machine learning-based path loss modeling has proven to
perform better than the traditional models [62,74,76,81,104,107,113,114]. Because of the large number of parameters typically
collected for path loss modeling, feature engineering has started gaining popularity in the practice of path loss modeling. As
a result, the research community tailored efforts towards the applications of machine learning in path loss modeling. This
makes it possible for a lot of research on modeling path loss based on machine learning to be conducted and published in
the literature. This motivated a number of surveys on path loss modeling in the literature.
Efforts have been made and gains recorded in machine learning-based path loss modeling. However, the emerging deep
learning architectures, machine learning taxonomies related to path loss, and in-depth analysis of feature engineering in
modeling path loss were missing in the already published surveys. To fill this existing gap, a survey of machine learning-

2
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 1
Summary of the difference between the current survey and the existing surveys.

Refs. Machine learning Shallow algorithms Hyperparameter settings Deep learning Feature engineering Environment
√ √
[70] 0 0 0 0
√ √
[115] 0 0 0 0
√ √
[107] 0 0 0 0
√ √
[9] 0 0 0 0
√ √ √ √ √ √
Current Survey
√ √
[1] 0 0 0 0
√ √ √
[86] 0 0 0

based path loss modeling in wireless communication systems is conducted to resolve the outlined issues, hence making the
survey unique.

Published surveys on path loss modeling and motivation for the current survey

This section presents the already published survey on path loss modeling. In addition, the differences between the exist-
ing survey and the current survey are discussed and the summary is presented in Table 1.
A survey on the use of traditional methods for path loss modeling, such as the basic models, terrain models, supple-
mentary models, stochastic fading models, and many ray models for path loss prediction, is discussed. It covers studies on
traditional modeling of path loss from 1946 to 2010 [70]. Adebowale et al. [115] published a survey on the applications of
shallow machine learning algorithms, such as ANN, KNN, SVM, GA, etc., in modeling path loss in wireless communication
systems. Research trends and a general overview of the research area are discussed. In another survey, 5G wireless com-
munication systems motivated [107] to conduct a survey on the procedure for modeling path loss with high accuracy and
low complexity using machine learning techniques. The survey examines shallow algorithms for the prediction of path loss.
It was established in the survey that the machine learning approach to modeling path loss outperformed the traditional
modeling methods. Bkassiny et al. [9] present a survey on the adoption of shallow machine learning algorithms in cognitive
radio communication systems. The survey indicates the difference among the shallow algorithms and the criteria for adopt-
ing the algorithm for application in cognitive radio. Seretis & Sarris, [86] presents an overview of the applications of ANN for
modeling propagation losses in communication systems. The review points out the benefits of feature selection/extraction,
hybrid models, correlation among features and large scale data in modeling the propagation losses. It was found that syn-
thetic data can be used to increase the propagation loss dataset. Abdulkarim et al. [1] presents survey on the applications
of shallow learning algorithms in the prediction of path loss.
Surveys of modeling path loss in communication systems exist in the literature as shown in the preceding paragraph.
However, emerging deep learning architectures have not received the desired attention, only Seretis & Sarris, [86] showcase
some variants, thus, in the current survey in-depth and rigorous analysis on the deep learning is presented. Machine learning
taxonomies related to path loss, hyperparameter settings, and in-depth analysis on feature engineering for modeling path
loss in communication systems were missing in the already published surveys.

Machine learning algorithms

This section presents, only the commonly used of them for the prediction of path loss in communication systems. From
the shallow algorithms, ANN and fuzzy systems are presented while CNN is discussed from the domain of deep learning.
The survey presents the basic concept so that readers can understand how these algorithms operate, thereby making the
survey self-contained.

Artificial neural network

Hassoun [37] described ANN as a system that is generally motivated by distributed parallel computation and occurs in
the brain which makes it easier to carry out complex computation and recognition/classification tasks. Park et al. [63] de-
fined ANN as a non-linear regression system that was driven mainly by utilizing learning and a generalized relationship
that exists between system input and output by a weighted network of neurons. The ANN is inspired by the human ner-
vous system and allows learning by example for data representation, which models a physical phenomenon or a process
of decision-making (Sadiq et al., 2018). One unique feature of ANN is that it can establish an empirical relationship that
exists between dependent and independent variables and extract subtle information and complex knowledge from the rep-
resentative dataset. These relationships are established without necessarily, making assumptions about any mathematical
representation of a phenomenon. The effectiveness of the ANN model in estimating performance is very high when com-
pared to the polynomial regression model and also can handle more dimensions than the look-up table methodology. ANN
consists of output and input layers nodes, which are connected by one or more hidden nodes. Information is routed through
the input nodes to the hidden nodes by triggering activation functions, while the hidden layers remain static and depen-
dent on what was supplied. ANN can be classified into various types which are deployed for different purposes as designed.

3
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

These deployments have over time achieved some meaningful and considerably accurate results, but its flexibility to func-
tion in another environment other than that which it was designed for has been lacking. Commonly, the following are some
of the types of ANN in the literature: perception, radial basis functions, probabilistic neural network, and so on. Perception
is the oldest and most common neural network, which was developed in 1958 by Frank Rosenblatt and is considered to
have the simplest form, being a structure with a single neuron. Multilayer Perception Neural Network (MLPNN) is the most
popular and common type of ANN. They have multiple neurons which are arranged in layers, starting with an input layer,
then hidden layers, and finally the output layers (Bishop, 1995). The output at each node is the weighted sum of its inputs
over an activation function, given by the relation in Eq. (1) [63]:
D
k=1 Xn,k .w1k,m for l =1
M l−1
A1n,m = k=1 Zn,k .wk,m
1
for l = 2...L − 1 (1)
l−1 l−1
k=1 Zn,k .wk,m for l = 2...L − 1
1

and
l
 
Zn,m = H l Aln,m (2)
l
where Wk,m is the entry in the kth row and mth column of a given weight matrix of the L-th layer on an input Xn,k with the
number of features given as G = 2 (frequency and distance), H l is the activation function for the L-th layer and it tweaks
the weighted sum of linear output, Alm,n .
Common activation functions that are used when deploying ANN are rectifier, hyperbolic tangent function, and logistic
sigmoid. The Rectifier or Rectified linear Tangent (ReLU) makes it possible for models to obtain sparse representation easily.
It is represented as:

H (a ) = max (0, a ) (3)

The hyperbolic tangent function can be viewed as a differential, nonlinear activation function that maps negative input
as large negative values and maps zero inputs close to zero. It is expressed as:

ea − e−1
H (a ) = (4)
ea + e−a
Logistic Sigmoid activation function is nonlinear, with a smooth threshold curve obtained from an artificial neural net-
work and given by:
1
H (a ) = (5)
1 + e−a

Convolutional neural network

CNN is a class of ANN that is mostly applied to visual imagery. It came as a result of the limitation of traditional ANN in
carrying out complex computations on image data. It is a special variant of a multilayer neural network, which is also known
as deep learning feed-forward architecture and is inspired by the visualized system of human beings [28,39]. CNN is very
good when applied to different computer vision fields and can also accurately process natural language. Traced back to the
discoveries by Hubel and Wiesel (1959), where it was discovered that the cells of the animal visual cortex recognize light
in small receptive field, thus leading to some models and architecture of the CNN. Thereafter, inspired by the contribution
of Wiesel in 1959, the first theoretical model for CNN, Neocognitron was proposed [29].
LeCun et al. [50] proposed a modern framework for CNN which was called LeNet-5. It has the capability of recognizing
handwritten digits. LeNet-5 makes use of a backpropagation algorithm for training, making it possible for LeNet-5 to rec-
ognize visual patterns directly from raw image data without necessarily carrying out feature engineering. However, there
is still no solution to complex problems as regards the earlier discoveries, as the need for a more sufficient and innovative
approach needs to be developed. Going further, the first model to break the limelight and pave the way for the modern
CNN model was the AlexNet designed by Krizhevsky et al. (2012) AlexNet can be applied for computer vision and natural
language processing [33].
Traditionally, CNN consists of single or multiple blocks of convolutional and pooling layers. It is then followed by multiple
fully connected (FC) layers and an output layer. But the core building block of the CNN is the convolutional layer which
learns the feature representation of the inputs. The convolutional layer is composed of many learnable convolutional kernels
or filters that are used for computing different feature maps [33]. Generally, CNN can be viewed as having two basic parts:
the feature extraction part and the classification part.
In the architecture of CNN, the activation functions perform an activity of mapping inputs to output. It obtains the input
value by calculating the weighted sum of input neurons and adding bias where it exists. The activation function handles the
decision as to whether a neuron will fire or will not fire. Some of these activation functions that are used in CNN are as
follows:

4
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Sigmoid
Considering sigmoid as an activation function, the function takes in real numbers as input and then binds the output
within the range [0,1]. The mathematical model of the sigmoid activation function is as given in Eq. (6) [33]:
1
f (x )sigm = (6)
1 + e−x

Tanh
It is used to bind input values within the range [-1,1] and must be real numbers. Tanh can be modeled mathematically
as:
ex − e−x
f (x ) (7)
ex + e−x Tanh

ReLU
The ReLU is a very popular and currently used activation function. In CNN. It carries out its operation by first converting
its inputs to positive numbers; and, it requires minimal computation load. ReLU is expressed as [33]:

f (x )ReLU = max(0, x ) (8)

Leaky ReLU
In this activation function, it does consider negative values in its operations, compared to the usual ReLU, where it
converts input values to positive values.

x if x > 0
f (x )Leaky = (9)
ReLU xm i f x ≤ 0

where m is the learnable parameter that controls the leakage of negative values.

Noisy ReLU
This activation function uses Gaussian distribution to create noise in ReLU. It can be represented mathematically as:

f (x )NoisyReLU = max(x + Y ), with Y ∼ N (0, σ (x ) ) (10)

Parametric linear units

This activation function is very similar to Leaky ReLU except that the leak factor here is tuned when modeling the
training process. It can be represented mathematically as:

x if x > 0
f (x ) parametric = (11)
LU ax i f x ≤ 0

Adaptive network-based fuzzy interference system

Adaptive-network-based fuzzy interference system (ANFIS) is a variant of ANN that combines the fuzzy logic and neural
network features thereby utilizing their benefits in a single framework. This interference system works correspondingly to
a set of fuzzy, IF-Then rules that can approximate non-linear functions. The ANFIS is popularly viewed to be a universal
estimator [41]. The basic architecture of the ANFIS is considered in two parts: the premise part and the consequence part.
Composing five layers, the first layer takes the input values and determines its membership function (the Fuzzification layer);
the second layer generates the fire strength for the rules (the Rule layer); the third layer normalizes the firing strength
through the division of total firing strength; the fourth layer takes the normalized value as input and passes the results to
the fifth layer which returns as the output [47].
ANFIS is predicated upon the fuzzy if-then-rules, which are conditional statements of expressions of the form IF A THEN
B, where A and B are fuzzy sets of labels characterized by the right membership function. In an environment of uncertainty,
fuzzy if-then-rules are employed to capture precise choices of decision that reflect the imprecise mode of reasoning in hu-
man decision problems. A fuzzy system is composed of basically a rule. The rule is a number of if-then rules and a database
that defines the membership functions of the fuzzy sets. The fuzzy set contained fuzzy rules, a decision-making unit that
performs the interference operation on the rules and Fuzzification and defuzzification. Fuzzification interface transforms the
crisp input to a matching degree of linguistic values, and defuzzification interface transforms the fuzzy results to a crisp
output [41].

5
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

The basic concept of radio propagation

Radio wave propagation, by concept, mostly pertains to that thing that happens between the transmitter and the receiver.
At every link end, there exists a receiver attached to the antenna of a particular geometry that produces signals which
are then modulated onto the carrier signal. As the modulated signal is sent out to the receiver, mostly at the speed of
light, it will be demodulated. On an account of obstacles in the environment, the so transmitted signal will be forced to
undergo absorption, reflection, refraction, diffraction, or a combination of these, which attenuates and/or distorts the signal.
The obstacles that are closer to the Line of Sight (LOS) path can obstruct the Fresnel zone and can cause fading of the
received signal level which can cause problems. Complication exists if some antennas emit their signals at the same time
and frequency in different directions. These signals could pick a variety of paths to the receiver thus causing every path to
communicate with its surroundings differently. When such signals arrive at the receiver, they are delayed by a particular
amount. If the delays lead to signals that are in phase with each other, then constructive interference will occur. However,
when they are not in phase with each other, they will produce an interference called destructive interference. The results
of this interference are known as multipath fading. The attenuation caused by large static obstacles, like mountains and
buildings is called slow fading or shadowing or large-scale fading. If the attenuation is a result of small but transient objects
with a variation of time, it is referred to as fast-fading or small-scale fading. For mobile receivers, a frequency shift may
occur as a result of Doppler spreading. Small-scale fading occurs as a result of frequency shifts and delay [70].

Path loss in wireless mobile communication systems

The geometry of the transmitter or receiver antenna determines the nature of the arriving signals from different di-
rections. Multi-directional antennas (i.e., antennas that receive signals from all directions) reflect electromagnetic waves
uniformly in a specific plane, often the azimuthal plane. Therefore, its radiation pattern will then take the form of a donut.
A directional antenna, such as a sectoral, parabolic, or patch antenna will succinctly signify signals that arrive basically from
one direction as indicated by the beamwidth. The radiation pattern of such an antenna is similar to the shape of the bat of
a baseball. However, it is impossible to isolate perfectly, and its geometries represent one-direction gain when considering
other directions. The gain of an antenna is typically measured in dBi, that is, the gain in decibels is relative to that of the
spherical isotropic antenna.
If the transmitting power Ptx (dBm) of the transmitter’s radio with antenna gain, Gtx (dBi), PL is the overall losses, the
Effective Isotropic Radiated Power (EIRP) is given as:

EIRPdBW = 10log10 Ptx [dBm] + 10log10 Gtx [dBi] − 10log10 P L[dB] (12)
with Prx the received power by the receiver and Gtx the gain of the transmitting antenna which must be in the transmitter’s
direction, Grx is the receiver’s antenna gain where (EIRP=10log10 Prx [dBm] + 10log10 Grx [dBi]). The PL is the overall path
loss, including attenuation. It is assumed that the radio link is kept away from external pollution and other signal sources
within the environment (e.g., Interference from other transmitters and thermal noise). Relatively, the quality of the signal
at any point is taken as the signal-to-noise ratio, SNR = Prx − N (dB), where N is the noise power. Interchangeably, with
consideration to interferences, the signal-to-interference-noise ratio (SINR) is given by
n
SINR = PrX − (N + Ij ) [dB] (13)
j

where I j is interference power due to source j.

Signal to Noise Ratio (SNR) and bit error rate have an established relationship for a given receiver design and modulation
system. Using this relationship, the Minimum Detectable Signal (MDS) for a given radio can be calculated as a function of
the acceptable error rate: MDS(Pe ). Where, Pe satisfies the probability of bit error. The set of receiver locations that satisfy
the inequality is then used to determine the covered points as shown in (14):

Ptx + Gtx + Grx − P L ≥ MDS(Pe ) [dB] (14)

Knowing that P and G in the equation satisfy a particular link, problems may arise when attempting to predict the exact
value of P L, There may be the need to interpolate the P L value from measured data.
The task of the model as stated here is the prediction of certain values for Lt + Ls (dB) where Lt is the inconsequential
free space path loss, and Ls is the loss that is caused by shadowing or slow fading as a result of a very mighty static object
such as buildings or mountains.

P L = Lt + L f (t ) [dB] (15)
L f (t ) represents the moderate fast fading which is caused by destructive hindrances from the effects of multipath and
minor scatters (where there is a variation with time t). The inconsequential fading comprises frequencies with time. Not
having greater knowledge of the surroundings, models cannot predict a certain quantity represented as L f (t ) The applica-
tion can be mostly stochastically computed with the aid of probability distribution (e.g., Raleigh, m-Nakagami, and Ricean
are popular). Doing this, simulation of frequency- and time-selective fades is possible, and most often, can be predicted
accurately, giving room for the examination of its effects on the modulation scheme [36,96].

6
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

2.5

Number of publications
2

1.5

0.5

0
1957

1991
1946

1968
1969
1977
1982
1988

1993
1994
1997
1998
1999
2000
2001
2002
2003
2006
2007
2008
2009
2010
Year of publication
Fig. 1. Sample of studies that applied the traditional models for path loss prediction for the period between 1946 and 2010.

Traditional methods of modeling path loss

The investigation of dynamic spectrum access network by researchers requires a precise map of the radio environment
to enable proper and timely allocations of choice of frequency; but performance-wise, the systems are tied firmly to their
capabilities of making informed and clear predictions on the present and future possession of the radio channel. Tracing
back to the 1940s, efforts by engineers and researchers have all been channeled towards finding the needed solution to this
problem. This has led to the development of countless schemes which seek to predict the exact value and/or distribution
of path loss (i.e., signal attenuation) with variation in environments and frequencies. Models that are presented here are
a priori, that is, their predictions are only based on the availability of prior understanding and will not make use of ex-
plicit measurement in making predictions. This makes these models most suitable for predictions where situations make it
succinctly difficult to obtain desired measurements [70].

Theoretical/Foundational models
Models, worthy of note that are first considered here are purely analytical models which are derivatives of idealized
electromagnetic propagation theory [70].

Basic models
The basic models are numerically plenteous compared to other model types. Here, computation of path loss is done
along one particular path and, most times takes corrections based on the already made measurement which are carried out
in the concerned environment(s). The inputs to these models are distance, carrier frequency, transmitter, and transmitter
heights. Examples of the basic models include Egli, Okumura- Hata, COST-231, etc.

Terrain models
Similarly, terrain models are not too dissimilar to the basic models. They attempt computation from diffracted losses that
occur along the LOS path because of obstructions that result from the terrain or buildings. An example of the terrain model
is the ITU Terrain

Supplementary models
Another classification of models to be presented here are the supplementary models which, as the name implies, cannot
stand independently but are used to effect changes/corrections to existing models that are complete

Stochastic fading models

Stochastic Fading models are augmented with some random variables to account for further fading in wireless channels.
They include fading that is a result of scattering and effects of multipath which are not correlated in measurement on a
small distance (i.e., less than 1 wavelength).

Many-Ray models
The Many-Ray Models try to compute the path loss by taking the summation of the loss along distinct paths, not on
a single LOS path. An accurate and precise understanding of the environment is very important to the operation of these
models. The commonly used data are the two- and three-dimensions vector models of interfering structures and buildings.
Fig. 1 presents the number of papers that have applied the traditional modeling approach for path loss modeling from
1946 to 2010 based on the data extracted from Phillips et al. [70]. This indicates that modeling path loss has been well
established in the literature.

7
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

20
18

Number of publications
16
14
12
10
8
6
4
2
0
Fundational Basic Basic/Terrain Terrain Supplementary
Traditional models

Fig. 2. Sample of studies that used the traditional models for path loss prediction.

Fig. 2 depicts the frequency of the different traditional models for modeling path loss in communication systems based
on data extracted from Phillips et al. [70]. The bars indicate that the basic model is the most frequently used model for path
loss prediction compared to other types of traditional models.

Machine learning modeling of path loss in wireless communication systems

Fig. 3 depicts a machine learning procedure for path loss prediction in wireless communication systems. The machine
learning concept can be used in finding estimated functions for path loss. Path loss prediction is typically a supervised
learning paradigm. As such, input features and output are required for modeling to predict path loss. The function esti-
mated by the machine learning procedure is mapped on to the output (path loss). Fig. 1 shows the first stage of data
collection where samples of measurement, in which the value of path loss and the features corresponding to the path
loss values, are collected. The features fall into two categories namely, system and environmental parameters. The system-
dependent parameter includes carrier frequency, antenna-separation distance, height, transmitter and receiver positions, the
angle between the line of sight and horizontal plane, etc. These parameters are independent of the environment propaga-
tion. The environment-dependent parameters include terrain, vegetation, and building condition that are determined by the
geographical environment and weather conditions [107].
The parameters are collected mostly from 3D map topographic databases as well as the database for the land cover. The
parameters obtained from the weather include humidity, temperature, precipitation rate, etc. The size of the sample data is
highly related to the performance of the path loss prediction model. In a practical sense, features collected from the mea-
surement contain a large number of features, running into hundreds, including unwanted and irrelevant parameters. The
irrelevant parameters typically have a negative impact on the algorithm by reducing its performance for modeling the path
loss for prediction. Therefore, feature selection is applied to reduce the number of features while maintaining the quality of
the features. The feature selection selects the most relevant subset features that contain the information required to predict
the path loss. Different methods exist for the feature selection such as embedded, filter, and wrapper. After the feature se-
lection, a suitable algorithm is chosen or modified for modeling the path loss for prediction. The factors considered in the
selection of the appropriate algorithm include the accuracy and complexity of the algorithm. The hyperparameter values of
the algorithms are set before the beginning of the learning process. Different approaches exist for the selection of the opti-
mum parameters setting, such as random search, grid search, and Bayesian optimization. The hyperparameters play a critical
role in the performance of the algorithms in path loss prediction. The performance is measured to ascertain the effectiveness
and efficiency of the algorithm. Some of the performance metrics are: root mean square error, correlation factor, maximum
prediction error, error standard deviation, mean absolute error, and mean absolute percentage error [107]. However, there
are no universally acceptable performance metrics; hence, different studies use different performance metrics.

Comparison of traditional and machine learning methods

The comparison between these two approaches to propagation path loss prediction in wireless communication signal
transmission can be reduced to the fact that all the approaches aim at predicting propagation path loss, though in different
ways. For instance, traditional approaches are widely viewed as a priori, which means that they act on prior knowledge of
the environment to be able to predict future occurrences in the transmission through the use of explicit measurements [70].
Thus, machine learning methods, on the other hand, involve data inputs and outputs in building or developing the models.
Models are developed to accept transmission data inputs and features of path loss data are extracted/selected from the input
data [107]. Machine learning models carry out predictions based on the features they can select/extract from the input data

8
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Fig. 3. Typical stages in path loss prediction based on machine learning approach [107].

to make an informed decision on the propagation path loss. The machine learning approach is limited to carrying out such
functions whenever it becomes impossible or difficult to carry out measurements [70]. The basic advantage of the machine
learning approach, however, is its ability to build a model and train the model based on the need of the required output,
and test the model against a real-world situation. These enhance the accuracy of the machine learning models compared
to traditional models. Furthermore, where there is no prior knowledge of the environment, machine learning can still make
predictions based on the features available in the environment and come up with result more accurate than the traditional
approach [70].
To measure strength in the machine learning approach in path loss prediction, models that are built must be supplied
with the right parameter set as there is always a variation in the propagation environment. Therefore, for effective path loss
prediction results, measured against performance, proper feature selection/extraction must be ensured to achieve consider-
able results which form the basic advantage of the machine learning approach over the traditional approach [107]. Machine
learning approaches can predict propagation path loss at different locations compared to the traditional propagation path
loss approach which makes a prediction based on a single path or similar paths. The machine learning approach can make
predictions based on labeled and unlabeled data and establish a relationship between input data and output data [103].
One of the major advantages of the traditional approach in propagation path loss prediction over the machine learning
approach is that they are suitable to predict propagation path loss when it is fervently impossible or difficult to carry out
the measurement [70]. Though the accuracy of the traditional model approach can be contested or questionable, they are
however very easy and simple to understand and implement which makes it easy for adoption into several network sim-
ulators. Furthermore, as a limitation though, traditional approaches make propagation path loss prediction along a single
path only and can make corrections only based on the measurement available to it based on one or more environments.
Traditional models take transmitter and receiver heights, carrier frequency, and distance as inputs. The traditional approach
allows complementary models which may not be able to stand on their own but can step in to complement other underper-

9
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Fig. 4. Taxonomy for machine learning including deep learning for path loss prediction.

forming models, thus making them attractive [70]. Fig. 4 shows the taxonomy of the machine learning algorithms commonly
used for the prediction of path loss.

Deep learning predictive model for path loss

In this section, the adoption of deep learning to predict path loss in communication systems is presented. The commonly
used deep learning architectures include CNN and hybrid deep learning architecture where two or more individual deep
learning algorithms (e.g., shallow algorithm and deep learning, or traditional method and deep learning), are hybridized to
predict path loss.

Prediction of path loss based on convolutional neural network

The CNN and its variants have been applied for the prediction of path loss in wireless communication systems. For exam-
ple, Ahmadien et al. [4] proposed a transfer learning VGG-16 network model for the prediction of path loss distribution in
an area using 2D satellite images. The VGG-16 network architecture makes suitable modifications to pre-train data through
several trainable parameters. Feature extraction was performed by the convolutional layers while prediction was done by the
regression layers. Research results showed that VGG-16 can predict path loss distribution of an area more accurately from 2D
satellite images than ray tracing. Wang & Lee [101] proposed a meta-learning-configured CNN for the prediction of indoor
path loss for 5G communications. In the model, the reptile algorithm was adopted as the training algorithm for the CNN
which was trained to extract features automatically. The experimental result suggested that the meta-learning-configured
CNN model outperformed the conventional CNN and empirical models.
Cheng et al. [13] proposed a CNN-based path loss modeling method for 28 GHz mmWave in suburban scenarios. The
proposed model automatically extracted environmental features and mmWave propagation features with no much pre-
processing. The model, while extracting a 2D google map image, also extracted relevant features to form the geometric
information on the environment. The model made use of the Enhanced Local Area Multi-scanning (E-LAMS) algorithm to
obtain input images for the CNN. Experimental results suggested that the proposed model outperformed empirical and de-
terministic models in terms of accuracy and complexity. Levie et al. [53] proposed the RadioUNet model for the prediction of
path loss in cellular optimization and efficient device-to-device link scheduling. The RadioUNet returns one output channel
which is the desired radio map and is denoted by RadioUNetc where c stands for Clean maps and the RadioUNet receives
as an input three channels: version of the environment, measurement of radio map, and transmitter location. The experi-
mental result showed that RadioUNet model outperformed models based on radial basis functional interpolation and tensor
completion.

10
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Nobuaki et al. [59] proposed an AlexNet model of CNN for the prediction of path loss considering diffraction and re-
flection in an open environment. The model considered transmitter height, the height of the receiver, and distance to the
receiver in the nearest line of sight when it is non-line-of-sight. The depth of diffraction is represented by the relationship
between the heights of buildings and antennae while the depth of reflection is represented by the angle formed between
transmitter and receiver. Results suggested that the proposed model improved estimation accuracy compared to ray tracing.
Ma et al. [55] proposed a CNN-based model for the prediction of indoor Wi-Fi path loss. The model adopted the local area
multi-line scanning (LAMS) algorithm to generate input for the CNN. The model was trained to learn features from indoor
environments with a further approximation of the functions of propagated Wi-Fi signals. Feature extraction was carried out
by CNN. Research results showed that the proposed deep learning model CNN performed better than the 3D ray tracing
methods.
Kuno et al. [49] proposes CNN for the prediction of path loss in communication systems. The propose model applied
building side view image and top view image as the typical convention in modeling the path loss. The experiment was
also conducted with the model using only top-view images. The CNN is found to perform better than the conventional
models and the CNN model with only top-view images in predicting path loss. Qiu et al. [77] proposes the use of path
loss prediction network (PP-Net) for the prediction of path loss in 5G scenario directly. The algorithm uses 3D ray tracing
generated data for offline training. The propose PP-Net accuracy performance is comparable with the state-of-the-art 3D ray
tracing simulator. However, the propose approach is faster than the state-of-the-art 3D ray tracing simulator 30 times. Bal
et al. [11] applied the VGG-16 for the extraction of image features to model path loss. The VGG-16 is model using region’s
satellite image and height map. The result indicated that the region’s satellite image or height map can be used to predict
path loss.
Sotiroudis et al. [91] proposed a CNN model for the prediction of path loss in urban areas using image-driven regression.
The proposed model assigned an image to each of the path loss measurements because path loss depends largely on build-
ings along the line of sight. Feature extraction was carried out by the convolutional layers of the proposed model. Research
results showed that the proposed model though generated similar results with input based on tabular data, adapted swiftly
to the new environment with less computational complexity. Table 3 presents a summary of the studies that adopted CNN
and its variant for path loss prediction.

Hybrid deep learning algorithms

This section presents the use of hybridized deep learning algorithms to predict path loss in wireless communication sys-
tems. For example, Thrane et al. [99] propose the hybrid of ANN and CNN (ANN-CNN) for path loss prediction in mobile
communication systems. Experimental data, including satellite images, were collected for modeling the path loss. In the ar-
chitecture, the CNN processes the satellite images while the ANN computes the numerical features. It was found that the
image improved the prediction accuracy of the path loss compared to the stochastic models and ray-tracing methods [99].
Sotiroudis et al. [92] proposed a CNN-XGBoost hybrid approach for predicting path loss in 5G and IoT communication sys-
tems. It is performed by fusing different input modalities of both tabular data and images of the target area. XGBoost was
used to process tabular data; and, CNN to process the images. Feature extraction was carried out by a series of convolutional
and pooling layers, while predictions were carried out by a set of fully connected networks. Experimental result shows that
CNN-XGBoost comfortably predicts path loss with higher accuracy compared to stacked generalization and feature concate-
nation. Lee et al. [52] proposed a 3D ray tracing-CNN deep learning approach for path loss exponent prediction for outdoor
millimeter wave channels. The 3D ray tracing was used to generate the outdoor millimeter wave band channel and path loss
exponent. The proposed model uses CNN to learn the characteristics of data abstractions. Feature extraction was carried out
by CNN. Results showed that the proposed model enhanced the prediction of path loss within the shortest possible time
compared to Deterministic Channel Models. Sotiroudis et al. [95] propose a hybrid of CNN and natural gradient boost (CNN-
NGBoost) for probabilistic path loss prediction. The footprint of the urban built-up area is the study environment. The CNN
served as feature extractor whereas the NGBoost perform the path loss probabilistic prediction. The CNN-NGBoost it found
to predict the path loss with an improved performance. Juang [45] proposed the hybrid of autoencoder and a generative
adversarial network (AE-GAN) to translate images from street map for the purpose of modeling path loss. The AE-GAN is
applied to predict the path loss. It was found that the propose model perform better compared to the conventional models
operating at the frequency of 3.5 GHz band.
Ates et al. [7] proposed a VGG-16-ResNet-50 model for the prediction of path loss exponent and shadowing factor in
wireless channel systems. This modeled the relationship that exists between model parameters such as path loss expo-
nent/shadowing and 2D satellite images of an area. Experimental results show that VGG-16 and ResNet-50 network model
predicts path loss exponent and large-scale shadowing factor with 88% and 77% accuracy compared to ray tracing. Sani et al.
[82] also proposed a hybrid of ResNet50V2 and regression algorithm for multiple environments- and multiple parameter-
path loss prediction. Extraction of features from 2D satellite images in the hybrid model was carried out by the ResNet50V2.
Experimental result shows that the proposed model can accommodate multiple parameters and multiple environment path
loss predictions with high accuracy. Cheng et al. [14] considered an attention-enhanced CNN (AE-CNN) hybrid model for the
prediction of millimeter wave path loss in a 5G communication network. The model made use of distance-embedded local
area multi-scanning for generating input for the model. Feature extraction was carried out by CNN, and is composed of four
dilated convolutional layers and three general convolution blocks. Experimental results suggest that the proposed model

11
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 2
The summary of the path loss prediction based on hybrid deep learning architecture.

Refs. Hybrid deep learning Baseline algorithms Indoor/outdoor Generation Result/contributions

[99] ANN–CNN Stochastic models and Not Stated 5G The satellite image
ray-tracing improves the prediction
accuracy of the model
[92] XGBoost-CNN Stacked generalization and Not Stated 5G XGBoost-CNN provides
feature concatenation better prediction which
enhanced path loss
prediction
[52] 3D Ray Tracing-CNN Deterministic Channel Outdoor 5G The proposed model
Models enhanced the prediction of
path loss within the
shortest possible time.
[7] CGG-16-ResNet-50 Ray Tracing Not Stated Not Stated CGG-16-ResNet-50
improves the prediction of
path loss exponent and
shadowing of an area
[82] ResNet50V2-Regressor MLP Deep learning Models Not Stated 5G ResNet50V2-Regressor
improves path loss
prediction with respect to
multiple environments and
multiple parameters.
[14] AE-CNN CI, ABG, and 3D Ray tracing Outdoor, 5G The model enhances the
outdoor-to-indoor prediction of path loss
[95] CNN–NGBoost CNN Not stated Not stated CNN–NGBoost improved
prediction accuracy of
path loss
[45] AE-GAN Close-In model and Outdoor 5G The AE-GAN improve the
alpha-beta-gamma performance of the path
loss prediction

outperforms other state-of-the-art empirical and deterministic methods. Table 2 presents a summary of path loss prediction
methods using hybrid deep learning algorithms.

Deep recurrent neural network

A few deep recurrent networks (DRNN) were found to be used in the prediction of path loss. For example, Sasaki et al.
[84] proposed a deep learning method for predicting path loss using the long short-term memory (LSTM) model. The study
considered the prediction at frequencies of 2.4, 4.7, and 26.4 GHz. The LSTM overcomes the problem of gradient vanishing
of RNN, thus making long-term memory possible. The prediction is performed using 100 fast-fading data as input. The
proposed model was compared with the CNN method, with results showing that the proposed model improved prediction
accuracy for any frequency band measured including higher frequency bands. Liu et al. [54] proposed the DeepLora deep
learning model based on Bi-LSTM for predicting path loss for long-distance links in low-power wide area networks. The
model also depended on remote sensing to automatically recognize land cover types along a Lora link. The proposed model
was compared with free-space, Bor, StateLoc, and Path/Intersection models, thus showing that DeepLora reduced estimation
errors to less than 4 dB, which is twice smaller than for the stated models. A summary of the studies is presented in Table 4.

Shallow algorithms in predicting path loss

Although there is growing interest in the prediction of path loss in communication systems using deep learning algo-
rithms as shown in Section 6, shallow machine learning algorithms are still active in path loss prediction. It is found in the
literature that ANN, SVM, fuzzy systems, GA, etc. have been applied recently. The survey presents the applications of shallow
algorithms in the prediction of path loss.

Artificial neural network adoption for path loss prediction

This section presents studies that utilize ANN to predict path losses in wireless communication systems. Wu et al.
[104] proposed a multilayer perception ANN for the prediction of path loss. The model predicted accurately the path loss for
a wireless network. Low-dimensional environmental features were extracted by using Principal Component Analysis (PCA),
with the environmental features combined with information about the base station and receiver to construct a dataset. The
experimental result showed that the proposed model outperformed the Close-In Path loss model and α − β path loss model.
Wang et al. [102] proposed a Back Propagation-ANN which is a simplified path loss model for radio wave propagation and
path loss prediction. The model predicted radio wave propagation through minimization of drive test and operated on a
large amount of the minimization of drive test data generated from antenna adjustment. It also divided the cells into grids.

12
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 3
Summary of path loss prediction based on convolutional neural network.

Refs. Deep learning Architecture Baseline Algorithms Indoor/outdoor Generation Result/contributions

[4] VGG-16 Ray Tracing Outdoor Not revealed The VGG-16 improved
accuracy of the path
loss distribution
prediction
[101] Meta-learning configured CNN, FI Indoor 5G The proposed model
CNN outperforms the
conventional CNN.
[13] CNN CI, ABG Not Stated 5G The proposed method
enhanced the
prediction accuracy of
path loss
[53] RadioUNet tensor completion Not Stated Not available The proposed model
predicted path loss
that is close to
estimation.
[59] AlexNet Ray Tracing Not Stated Not available The proposed model
improved estimation
accuracy.
[55] CNN Ray Tracing Indoor Not available Deep learning-based
CNN for indoor path
loss modeling
improved path loss
prediction
[91] CNN Tabular-data-Driven CNN Not Stated Not available Image-driven CNN
path loss prediction
enhanced path loss
prediction with less
computational
complexity
[49] CNN Conventional model and Not available Not available Perform better than
CNN model with only the conventional
top-view images models
[77] PPNet 3D ray tracing and UNet Outdoor 5G Improve the speed of
the path loss
prediction 30 times
[11] VGG-16 ————- Not available Not available Predict path loss using
satellite images with
good accuracy

Table 4
The summary of the path loss prediction based on a Deep recurrent network.

Refs. Deep learning Baseline algorithms Indoor/outdoor Generation Result/contributions

architecture

[84] LSTM Conv Method Not Stated 5 G, 6G The model improved the
prediction accuracy of the
system
[54] Bi-LSTM Free space, Bor Model,Outdoor Not available The proposed model
StateLoc model, and DeepLora based on
Path/Intersection Bi-LSTM outperformed the
Model. state-of-the-art models.

The Back Propagation ANN was trained by using the means of Reference Signal Receiving Power (RSRP) in grids with suffi-
cient samples before the antenna is adjusted and was used for the modeling. Research findings suggested that the proposed
model made an accurate prediction with minimal cost. Ojo et al. [61] considered the use of Radial Basis Function Neural
Network (RBFNN) and MLPNN for the prediction of path loss for LTE networks in the multi-transmitter signal propagation
environment. An input dataset of the same scenario was used on the two models. The RBFNN and MLPNN as a compo-
sition both have three layers. The researchers used a smartphone that was preinstalled with a test mobile system (TEMS)
for testing and measurement. MLPNN and Results showed that RBFNN outperformed MLPNN. Jo et al. [43] proposed a tri-
pod machine learning framework of ANN-based multi-dimensional regression, Gaussian process-based variance analysis, and
principal component analysis-based (PCA-based) feature extraction for modeling path loss. In the proposed techniques, the
ANN learned path loss structure from a dataset with reduced dimension; Gaussian process learned the shadowing effects,
while PCA did the feature extraction. Experimental results showed that the combined path loss and shadowing models per-
formed accurately with more flexibility than the conventional linear path loss and log-normal shadowing path loss models.

13
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Sotiroudis et al. [90] compared ANN with random forest for path loss modeling in NB-IoT networks, the two models
tested. The ANN acquired knowledge from the samples and generalized predictions based on the training while the ran-
dom forest is an ensemble learning method with multiple regression trees. Experimental results show that the two models
performed at the same level. Popoola et al. [74] investigated the suitable ANN parameters for path loss prediction in very
high-frequency wireless channels. They trained various ANN architectures with varying parameters for path loss prediction.
The performance of all neural networks was determined, and ANN produced the best result. Comparison made with Hata,
COST 231, ECC-33, and Egli models still showed that ANN-based path loss prediction had better prediction accuracy. Zaarour
et al. [106] carried out a comparative study of 60 GHz path loss channel modeling in a mining environment. In the study,
ANN and RBF were considered. They modeled the path loss attenuation as a function of distance and frequency. Experi-
mental results suggested that although both types of ANN have the same potential to predict path loss MLP had less error
compared to RBF. Ostlin et al. [62] proposed an ANN model for macro cell path loss prediction. ANN was used for radio wave
path loss predictions in the proposed model, which was composed of a simple neuron model, a feed-forward model with
different numbers of hidden neurons. The inputs to the model were chosen such that it considers the distance to the base
station, parameters from terrain path profiles, and vegetation density/land used near the receiving antenna. Results obtained
in the proposed model were evaluated against other traditional models with respect to prediction accuracy, complexity, and
prediction time. The proposed model was found to have performed very well in comparison to the Okumura-Hata model
and ITU-R P.1546. Popoola et al. [74] investigated the performances and applications of various types of ANN learning algo-
rithms to model path loss prediction in the VHF band (30–300 GHz). In the research, prediction results of ECC-33, Egli, Hata,
and COST 231 models at varying distances are fed into the ANN as input. Experimental results suggested that the ANN mod-
els, which were trained with hyperbolic tangent activation function (HTAF) and Levenberg-Marquardt (LM) algorithm with
80 neurons in the hidden layer produced the best results in terms of the Mean Error, Root Mean Square Error, Standard
Deviation, and coefficient of determination.
Eichi et al. [27] performed a comparative study between the basic models and ANN-based models for path loss predic-
tion. The study considered path loss parameters such as distance, terrain environment, and transmitting power, which was
then used as inputs to the ANN for path loss prediction. In the research, MLPNN parameters were varied in carrying out
performance evaluation, thereby choosing the weight and bias values of the best performed MLPNN. Results suggested that
the proposed model outperformed the basic empirical path loss models such as Hata, Egli, COST-231, Ericsson model, and
the modified path loss model. Park et al. [63] proposed ANN modeling for path loss prediction in an urban environment.
They proposed an ANN-based multi-dimensional regression framework for path loss modeling in an urban environment in
the 3 - 6 GHz frequency band. They implemented a three-layer MLPNN which learned features from the multi-dimensional
path loss and adopted the use of activation functions such as rectifier, hyperbolic tangent, and logistic sigmoid. Research
Results showed that the proposed model was more accurate and flexible compared to conventional linear models. Popescu
et al. [72] carried out a comparison of ANN models for path loss prediction, particularly on the applications of ANN in an ur-
ban and suburban environment. The researchers investigated an error correction model based on the combination between
the theoretical model and ANN. Experimental results suggested that the ANN model produced an enhanced accuracy when
compared to COST-231-Walfisch-Ikegami model.
Popescu et al. [73] proposed a multilayer perception and a generalized regression neural network (Radial Basis Function)
to undertake a comparison of ANN-based models for path loss prediction in an indoor environment. The performance of
the proposed network was investigated for prediction error, standard deviation, and root mean square error. The model was
trained with measured data collected in the 1890 MHz band. Experimental results Findings showed that the model produced
high accuracy when compared with the empirical model with high accuracy. Kalakh et al. [46] proposed an ultra-wide band
(UWB) propagation channel model for propagation path loss in a mining environment. The model focused on the change
in path loss attenuation as a function of distance and frequency. The model was trained by performance sufficiency and
several experiments were carried out to verify its validity. The model was tested by estimation of the model’s accuracy in
predicting correctly the received power values. Experimental findings as compared to empirical models, the proposed model
produced high prediction accuracy in the mine environment. Jo et al. [43] proposed a machine learning framework to model
path loss using three techniques: ANN-based multi-dimensional regression (ANN), Gaussian process-based variance analysis,
and PCA-aided feature extraction. The researchers used PCA for feature extraction to reduce the dimension of the dataset.
ANN, on the other hand, learned the path loss structure from the dataset with reduced dimension, while Gaussian process
was adopted to learn the shadowing effects. Research results showed that the combined method of path loss model and
shadowing model produced more accurate results with higher flexibility in comparison to the conventional linear path loss
models and log-normal shadowing models.
Wen et al. [103] proposed a four-model technique based on machine learning path loss prediction for aircraft cabin envi-
ronments. They investigated the parameters of path loss to ensure the reliability and stability of the wireless communication
in the cabin. The study adopted the BPNN, Support Vector Regression (SVR), random forest, and AdaBoost. The BPNN model
the linear relationship between inputs and outputs, while the SVR deals with non-linear regression problems. The random
forest handles two random processes of feature and sample selection based on bagging. AdaBoost algorithms, also known
as boosting, can be used to superimpose the base learner layer. The data expansion method was found to have enhanced
prediction performance with few measurement samples at varying frequencies. Tahat et al. [98] presented a simplified ma-
chine learning-based air-to-ground path loss modeling for an urban environment. The study suggested certain strategies for
constructing an empirical path loss model for air-to-ground radio frequency signals. They proposed regression algorithms,

14
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

including the KNN, Regression tree (RT), and ANN. The research used the GPS coordinates of unmanned aerial vehicles (UAV)
transmitters and ground receivers. The verification of the validity of the proposed model was carried out against numerical
results. Research results show that the proposed method improved path loss prediction. Duangsuwan et al. [19] proposed
path loss characterization machine learning models for a ground sensor for UAV-enabled communication in smart farming
scenarios. They adopted the SVR and ANN to measure data in different scenarios. The performance of the proposed method
was compared with the GUT-R model, with results showing that the proposed model of SVR and ANN can optimally char-
acterize path loss in a smart farming scenario with the corresponding accuracy of 95% for SVR and 97% for ANN.
Egi & Otero [26] proposed a 3-dimensional point cloud-based signal power path loss model. They proposed a smart strat-
egy for deploying wireless communication systems to reduce the need for predefined estimation parameters. They used 2D
image color classification to extract features from 3D point-cloud and ANN models. They made use of light detection and
ranging with machine learning to predict signal power loss. The results of the proposed model were compared with the
empirical models such as the log-distance and Cost 231-Hata model, with The results showing that the proposed method
enhanced signal power path loss determination with a received signal level error of 4.26%. Piacentini & Rinaldi [71] proposed
a machine learning and dimensionality reduction technique for path loss prediction in an urban environment. The SVM and
ANN were proposed as a new approach for path loss prediction. The SVM and ANN which are learning machines were
trained to predict propagation path loss. The result obtained after comparison of the SVM and ANN using no dimensionality
reduction as well as using PCA and nonlinear PCA over different intervals, showed that the PCA-based prediction model
produce a better result than the NPCA hence improved the accuracy of path loss prediction. Moraitis et al. [57] suggested
an evaluation strategy to explore machine learning methods for path loss prediction in a rural environment at 3.7 GHz. The
study investigated machine learning models such as ANN, SVM, random forest, and KNN with bagging learners. The models
were evaluated against empirical models and were deduced that the proposed methods outperformed the empirical models
like the Close-In (CI) and extended-Hata. The performance of the models was also evaluated against each other and research
results show that, with increased hidden layers of ANN (which produced better prediction results), the ANN was highly rec-
ommended for path loss prediction modeling. Singh et al. [88] proposed a machine learning approach to predict path loss
for a smart campus environment. Machine learning models, such as ANN and random forest, were trained and evaluated.
The empirical COST-231 Hata model was compared with the performance of the proposed models to determine their su-
periority. Results showed that the proposed method of path loss prediction for a smart campus environment improved the
prediction accuracy, thus making machine learning models very attractive. Fu et al. [25] compared the performance of ANN
and Statistical model the prediction of path loss. Results indicated that the ANN perform better than the mixture models
but the mixture models matches well with the path loss probability density function.
Ayadi et al. [8] proposed an ultra-high frequency path loss model using a learning machine for heterogeneous networks.
They projected an MLPNN consisting of several elementary neurons located at different layers. The proposed model is based
on MLPNN which utilized the backpropagation algorithm and obtains inputs from standard propagation models. The pro-
posed model was compared with the ITU-R P.1812-4 and Standard Propagation Model (SPM), with results showing that the
proposed model performed better than the other models and accurately predict the path loss for heterogeneous networks.
A summary of the studies that adopted ANN for the prediction of path loss in mobile wireless communication systems is
presented in Table 5.

Adaptive neuro-fuzzy inference system for path loss prediction

Fuzzy neural networks have been adopted to predict path loss in wireless communication systems. This section presents
studies that have adopted ANFIS to predict path loss, Adeyemo et al. [2] proposed a path loss prediction model using ANFIS.
The data for this model were collected from measured Received Signal Strength (RSS) and employed to determine the path
loss. The model was compared with the free-space, Okumura-Hata, and COST-231 models. Research findings showed that
ANFIS gave a lower error which is indicative of its stability for path loss prediction. Bhupuak & Tooprakai [12] proposed a
fuzzy logic and k-means clustering-based algorithmic scheme to reduce the error in the prediction of path loss. The model
simultaneously employs area topographical variability and several path loss prediction models to reduce the prediction error
inherent in the independent usage of a traditional path loss model. Comparison showed that the errors in prediction of the
scheme (k-means fuzzy scheme) was 2.67% compared to that from drive-test measurement data. Also, in relation to other
conventional models (free space loss, Walfisch–Ikegami, HATA, ECC-33, Stanford University Interim, and ERICSSON models).
the prediction error is smaller in the k-means clustering and fuzzy logic model.
Cruz et al. [16] used ANN and Neuro-Fuzzy logic to propose methodologies for the prediction of path loss in LTE-1.8 GHz
networks. They predicted propagation path loss for 4G signals through a static empirical approach, with the study combin-
ing the common characteristics of ANN and Fuzzy Logic. They measured the transmitted mobile signals through a Samsung
smartphone equipped with GNETtrack that incorporated mobile GPS. The RMSEs of the model proposed were compared
with the RMSE of some notable models for predicting path loss such as COST 231-Hata, Ikegami-Walfisch, and the ITU-R
P1546-4, and were found to be more accurate. Dalkilic et al. (2010) used ANFIS to tune the model and solve function ap-
proximation problems. The proposed model performance was compared with the Bertoni-Walfisch model and was adjudged
to be better for analyzing propagation involving buildings and structures, particularly in urban areas. Their findings showed
that the proposed model error was less compared to that of the Bertoni-Walfisch model. Danladi & Vasira [18] used spline
interpolation and fuzzy logic techniques to modify the Hata model. Fuzzy logic undergoes offline training 65 times with

15
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 5
Summary of path loss prediction based on artificial neural network.

Refs. Algorithm Baseline algorithms Indoor /outdoor Generation Result/contributions

[104] ANN CI, A-B Not Stated Not available The MLP ANN improved
path loss prediction
[102] BPNN Not compared Not Stated Not available BP NN and simplified path
loss model make an
accurate prediction for
RSRP mean of grids with
minimal cost thereby
enhancing path loss
prediction
[61] RBFNN COST-321 Hata, Indoor/Outdoor 4G RBFNN improved path loss
ECC-33, Ericsson and prediction with minimal
MLPNN root mean square error
[43] ANN conventional linear Outdoor Not available The combine techniques
path loss and enhance the accuracy of
log-normal shadowing path loss prediction.
path loss model
[90] ANN Random Forest Not available NB-IoT The comparison produced
a similar performance rate
and accuracy
[74] ANN COST 231, Hata, Not available Not available The proposed model and
ECC-33 method enhanced the
prediction of path loss
with more accurate
results.
[106] ANN Not compared Mining site Not available The model enhanced the
prediction of path loss in a
mining environment
[62] ANN Okumura-HATA, Not available Not available ANN improved path loss
ITU-RP.1546 prediction as compared to
other traditional models.
[74] ANN Empirical Models Not available Not available The ANN models that are
trained with HTAF
improved path loss
prediction
[27] ANN Hata, Egli, COST-231, Not available Not available The adopted path loss
Ericson model prediction model
performed better in path
loss prediction compared
to other models
[63] ANN Conventional Linear Not available Not available The proposed model
model (HATA, produced more accurate
COST-231) and flexible prediction
results.
[72] ANN COST-231-Walfisch- Not available Not available The proposed model
Ikegami enhanced the accuracy of
path loss prediction in an
urban environment
[73] RBF Empirical models Indoor Not available The proposed model
predicted path loss with
high accuracy.
[46] ANN Empirical models Mines Not available The model predicted with
high accuracy path loss
prediction in a mine
environment
[43] ANN + PCA Linear conventional Not available Not available The proposed combine
models technique enhanced the
accuracy and flexibility of
path loss prediction
[103] BPNN, Random Forest, N/A Not available 5G The proposed method
AdaBoost, and Support enhanced path loss
Vector Regression prediction with few data
samples at different
frequencies.
[98] KNN RT and ANN Outdoor Not available The KNN improved path
loss prediction in terms of
RMSE
(continued on next page)

16
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 5 (continued)

Refs. Algorithm Baseline algorithms Indoor /outdoor Generation Result/contributions

[19] ANN SVR Not available Not available The proposed models of
SVR and ANN improved
path loss prediction for
smart farm scenarios with
an accuracy of 95% and
97% respectively
[26] ANN log-distance and Cost Not available Not available The prediction of received
231-Hata signal level path loss was
enhanced by the model
[57] ANN, SVR, RF, KNN CI, and Extended Hata Not available Not available ANN in the evaluated
methods enhances path
loss prediction.
[88] ANN, RF COST-231 Hata Not available 5G The proposed method
improved the accuracy of
path loss prediction
[71] ANN +SVM with ANN +SVM Not available Not available The proposed method
dimension reduction enhanced the accuracy of
path loss prediction
[8] ANN ITU-R P.1812, SPM Not available Not available The method enhanced the
prediction of path loss in a
heterogeneous network
[25] ANN Statistics model Not available 5G The ANN performs better
in the path loss prediction

many fuzzified inputs while spline interpolation enhanced the model sensitivity. They recommended fuzzy logic and spline
interpolation models for path loss prediction because the models reduced path loss by 1.94 dB for the study area.
Gupta et al. [34] used the ANFIS model to predict path loss for multi-transmitter radio wave propagation. The model was
trained using the received signal strength and evaluated for performance using the RMSE, standard deviation error (SDE),
mean absolute error (MAE), mean square error (MSE), and correlation coefficient (CC). The model considered fringe areas for
path loss prediction. The proposed model was likened to empirical models such as Hata, WI, SUI, Ericsson, COST-231, Egli,
and ECC-33. Results showed that the developed model produced the required qualities for the planning of the radio network.
Gupta & Sharma [35] considered fringe areas and proposed a fuzzy-logic approach for protected path loss prediction. The
method predicted correctly the path loss for the location of the base station in cellular mobile communication networks.
Fuzzy-logic manipulates vague concepts and produces a rational uncertain event. They classified the propagation medium
into many well-established propagation environments, called fuzzy set, such as free space, flat area terrain, light structures
terrain, heavy structures terrain, and village terrain. The model was compared with the HATA model and the results showed
a significant difference (0.536) in path loss rate over the HATA model and the proposed model having a higher path loss
rate. Jafri et al. [40] presented a combined modulated ANN and fuzzy logic model that used measured Interference path
loss data, with expert knowledge used to model coupling patterns in A320 and A319 airplanes. The neural network learned
interference patterns according to the locations of doors, windows, and aircraft antenna locations. It also learned interference
patterns from one aircraft to another. Fuzzy logic handled and incorporated expert knowledge when modeling. The Neuro-
fuzzy model was compared to their previous model of ANN and with results showing that Neuro-Fuzzy model enhanced
Interference path loss modeling.
Mathew et al. [56] employed a modulated signal-based binary phase shift keying (BPSK) method, in conjunction with
Hata’s empirical formula, to determine path loss in wireless communication systems. A triangular membership function was
used to assign identified values. Defuzzification was done by a center-of-sum method which is flexible and faster. Fuzzy
Logic was used to represent the data obtained from the path loss model in linear regression. Results showed that the model
was efficient and could be used to predict path loss in all kinds of signal modulation regimes. Phaiboon et al. [69] proposed
a fuzzy logic method for the prediction of path loss in cellular mobile communication systems. The study classified the
propagation medium into two propagation environment densities defined as an input fuzzy set, which was then classified
by the fuzzifier and inference using the fuzzy linguistic rule. The proposed model was compared with the conventional
HATA model, and findings indicated that the proposed model and method were more accurate in the predictions of path
loss compared to the HATA model. Phaiboon et al. [66] determined the characteristics of path loss on urban streets in the
microwave band. Fuzzy logic was used to determine unknown path loss out of a set of known path loss values. The result
obtained from the proposed method was compared with conventional models such as multi-ray models. The upper and
lower bounds of the model represent the output of the fuzzy system and have a formula with values from the upper bound.
Results indicated that fuzzy logic provided an enhanced prediction.
Phaiboon et al. [67] developed a multi-layer fuzzy logic-based system (MLFS) to predict mobile path loss in forest areas.
The proposed model has a supervisory layer, called the tree density decision layer, that selects the next layer using fuzzy
decision. Crisp inputs to the model are categorized by the fuzzifier into fuzzy sets and inference through linguistic rules.
Measurements were performed in various forest densities at 1.8 GHz. Results indicated that the approach of fuzzy logic

17
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

produced desired results for slopes of path loss in comparison to a conventional empirical mathematical model. Phaiboon
& Phokharatkul [65] proposed an empirical-based methodology with the effects of tree density for a 1.8 GHz mobile com-
munication system by fuzzy regression. They presented a path loss model for the upper and lower bounds in the forests
using fuzzy linear regression. The boundary line spread was dependent on the minimum and maximum values of certain
data. The proposed model was compared with the conventional regression models of previous research. Results suggested
that the fuzzy regression model aligns with measured data, thereby enhancing path loss prediction in forest areas. Salman
et al. [81] proposed an ANFIS- based model to predict path loss in the VHF band. They used an optimized ANFIS of five
layers which are predicated on the algorithm of backpropagation gradient descent and the least square error estimates. The
proposed model was compared with popular conventional models which include Hata, COST 231, Egli, and ECC-33 mod-
els. Results showed ANFIS producing the lowest Root Mean Square Error and Mean Error, hence its suitability for path loss
prediction.
Shoewu et al. [87] presented a methodology for path loss prediction in the metropolitan environment using fuzzy-logic
models. They conducted a drive test for the collection of data input for model training. Fuzzy logic makes use of the fuzzy
sets theory where one variable belongs to one or more sets with some degree of membership. The performance of the model
was compared to the free space propagation model and Hata model. Results show that the model has better efficiency,
is faster, and produced more accurate results compared to the physical and empirical models. Surajudeen-Bakinde et al.
[97] proposed an ANFIS-based method to predict path loss in the very high-frequency band for multi-transmitter radio
propagation. The study trained an optimized five-network structure of ANFIS based on the backpropagation gradient descent
algorithm. The researchers adopted the pi membership function to maintain stability and fastness in convergence with a
minimum number of epochs. The model’s performance was compared to empirical models like the HATA, COST 231, Egli,
and ECC-33. Experimental result showed an enhanced path loss prediction. Table 6 presents a summary of the studies that
adopted ANFIS for the prediction of path loss.

Evolutionary neural network

This section presents the studies that applied an evolutionary algorithm to optimize the ANN to improve its performance
for path loss prediction. Sotiroudis et al. [89] proposed an ANN for the prediction of propagation path loss in an urban
environment. They made use of an algorithm called Differential Evolution (DE), which is also called composite Differential
Evolution (CoDE), to optimally model the ANN to predict path loss. The adopted algorithm uses three different strategies of
trial-vector generation and has three preset control parameter settings. The performance of CoDE was compared to some
variants of the DE (jDE, JADE, and SaDE). Results suggested that the proposed model (CoDE-ANN) yielded satisfactory ac-
curacy in path loss prediction compared to the ray tracing model. Table 7 presents the studies that used the evolutionary
ANN.

Path loss prediction based on the nature-inspired meta-heuristic algorithm

This section presents the studies that applied nature-inspired meta-heuristic algorithms to enhance path loss prediction
in communication systems. For example, Nafea & Hamza [58] proposed a path loss optimization approach in WiMAX net-
works using the GA. The researchers made use of the COST-231 Hata model optimized using the GA for the prediction of
path loss. The study area was divided into urban and suburban to determine the minimum path loss value. The GA used
fitness functions to calculate the average of the difference that existed between the evaluated and measured path loss rates.
Performance comparison was carried out by varying the distances of the path and was found through the optimization of
the COST-231 Hata model that shorter distance optimization produced better results because of its minimum value in path
loss. Jawad et al. [42] considered smart agriculture and proposed an empirical model that was based on optimized particle
swarm optimization (PSO) for path loss determination in wireless sensor networks. The researchers formulated two path
loss models which were based on polynomial and exponential functions before hybridizing with the PSO. They asserted that
the hybridization of polynomial and exponential functions with PSO each improved the coefficient of determination of the
regression line. Results also showed that the proposed method had mean absolute error of 1.6 and 2.7 dB for exponential
PSO and polynomial PSO respectively. Further investigation showed that they outperformed previous state-of-the-art models
(Log Distance, LNSM, and FSPL).
Anireh & Osegi [6] proposed artificial bee colony (ABC)-PLOSS, which is a tool for minimization of path loss in GSM
telecommunication networks. The ABC-PLOSS software is an open-source tool that is established for use in the processes of
optimization. The architecture of this tool uses sequential processing, which is based on the algorithm of ABC that possesses
swarm intelligence, possible. The proposed model was compared to the GA and PSO. Results revealed that the ABC-PLOSS
approach gave a lower estimated path loss compared with the conventional approaches. Garah et al. [31], in an attempt to
reduce path loss in suburban and rural areas, proposed a PSO. The optimization method for tuning parameters is precisely
that of the COST231 model which improves prediction accuracy. A comparison was made between optimized COST231 and
other path loss prediction models such as the SUI, Hata, and Egli models in terms of the root mean square error, mean error,
standard deviation and Relative Error Percentage. Experimental results show that the optimized COST231 outperforms other
models.

18
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 6
Summary of path loss prediction based on fuzzy-based neural network.

Refs. Algorithm Baseline algorithms Indoor /outdoor Generation Result/contributions

[2] ANFIS Okumura-Hata, Not available Not available The model was stable and
COST-231, and Free enhance path loss
space prediction
[12] K-Means fuzzy logic free space loss, Not available Not available There was an
Walfisch–Ikegami, enhancement in the path
HATA, ECC-33 loss prediction accuracy
compared to conventional
models
[16] Neuro-Fuzzy-ANN COST 231-Hata, Outdoor 4G The methods enhanced
Ikegami-Walfisch, and path loss prediction for
ITU-R P1546–4 the 4 G network
[17] ANFIS Bertoni-Walfisch Not available Not available The proposed model
reduced the error of path
loss prediction
[18] Fuzzy logic and spline Hata Model Not available Not available the proposed model
interpolation reduced path loss by
1.94dB
[34] ANFIS Hata, WI, SUI, Ericsson, Not available Not available The Developed model
COST-231, Egli, and shows desirable qualities
ECC-33 in path loss prediction
results
[35] Fuzzy-Logic HATA model Not available Not available The proposed model has
better path loss prediction
for an unknown
environment
[40] Neuro-Fuzzy ANN Not available Not available The model enhanced
interference path loss
prediction in airplanes
[56] BPSK Not available Not available The model enhanced the
prediction of path loss for
different kinds of
modulated signals
[69] Fuzzy-Logic HATA Not available Not available The model outperforms
the conventional HATA
model in terms of
accuracy of path loss
prediction
[66] Fuzzy Logic Multi-ray models Not available Not available Fuzzy logic models
provided a better
prediction of path loss
[66] Fuzzy Logic N/A Not available Not available The fuzzy logic models
provided better prediction
accuracy of path loss
[67] Fuzzy Logic Conventional empirical Not available Not available The fuzzy logic approach
mathematical models produced an accurate
result of path loss
compared to conventional
empirical mathematical
models
[65] Fuzzy Regression Conventional models Not available Not available The fuzzy regression
model aligns with
measured data thereby
enhanced path loss
prediction in forests area
[81] ANFIS Hata, COST 231, Egli, Not available Not available Experimental results
and ECC-33 models indicate the suitability of
the proposed model for
path loss prediction
[87] Fuzzy Logic Hata, Free space Not available Not available The proposed model was
efficient, faster, and
accurate in the prediction
of propagation path loss
[97] ANFIS Hata, Egli, ECC-33, Not available Not available The proposed model
COST-234 shows higher accuracy and
efficiency in propagation
path loss prediction

19
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 7
Summary of path loss prediction based on evolutionary algorithms.

Ref. Neural Network architecture Baseline algorithms Indoor /outdoor Generation Result/contributions

[89] CoDE-ANN Ray Tracing Not available Not available The proposed ANN model improved path loss prediction

Table 8
Summary of path loss prediction based on Nature Inspired Meta-Heuristic Algorithm.

Refs. Algorithm Baseline algorithms Indoor /outdoor Generation Result/contributions

[58] Optimized Hata Model Traditional Not available Not available Optimized Hata Model
performs better for path
loss prediction
[42] PSO Log Distance, LNSM Not available Not available The proposed optimization
and FSPL method enhanced path
loss prediction in an
agricultural area.
[6] ABC-PLOSS PSO, GA Not available Not available The proposed approach
efficiently improves path
loss prediction
[31] Optimized COST231 Egli, Hata, SUI Not available Not available The Optimized COST 231
model enhanced the
accuracy of path loss
prediction
[3] Optimized Conventional approach Not available Not available The optimized model
Okumura-Hata model enhanced path loss
prediction
[15] PSO GA Outdoor Not available The experimental results
enhanced the design of
wireless networks as it
reduces path loss
[100] GA N/A Outdoor Not available The proposed method
improves path loss design
[51] GA Baseline method Indoor Not available The proposed method
produces a performance
improvement of 31.09%

Adeyemo et al. [3] proposed a GA-based optimization process for path loss determination in long-term evolution net-
works. The study investigated the appropriateness of the use of the Okumura-Hata model for the determination of long-term
evolution at 2.3 GHz. Data was acquired by undertaking a drive test around four LTE stations which are located in Lagos
State Nigeria. The Okumura-Hata model was optimized by GA and evaluation of the Okumura-Hata model was taken against
the optimized model through the path loss. Results showed that the values of the optimized model were above those of
the data measured. Chiu et al. [15] carried out a comparative analysis of PSO and GA to reduce path loss in urban areas.
They used the shooting and bouncing ray/image to make the computation of path loss for several outdoor environments.
GA and PSO were applied to a higher-order nonlinear optimization problem. From the investigated antenna rays, it was de-
duced that PSO had a better performance reduction rate of path loss than GA. Tu & Chiu [100] suggested a methodology
that used the SBR/Image to compute the path loss in different outdoor environments to reduce it. The study used the GA
for optimization of the exciting phases and voltages and antenna arrays to form clear antenna patterns. The antenna pat-
terns were used to deduce the path that had lower path loss. Antenna arrays by the GA were used to decrease path loss in
the outdoor environment. Lee et al. [51] employed GA to model path loss selection in signal strength that is dependent on
indoor localization. They considered multiple candidates by developing an overlapping multi-state model of path loss, in-
cluding those in the line-of-sight path and those not in the line of sight. The GA was applied to take care of computational
complexity to allow the proposed method to execute in real-time. Experimental validation was carried out by ray tracing
simulation and the received signal strength measurement at a real site as it showed an improvement in accuracy for Wi-Fi
in indoor environments. Results also showed that the proposed method produced performance improvements of 31.09% on
the benchmark scenario. Table 8 presents the summary of the studies based on the network-inspired algorithms.

Other algorithms used for the prediction of path loss

There are other algorithms used in predicting path loss but do not fall into any of the classes discussed in the preced-
ing sections. Therefore, a survey of such studies is presented in this section. Sotiroudis et al. [93] proposed an enhanced
model to predict path loss using image texture methodology. In the work, they used images that represented the footprint
of an urban setup which had features revealing information about the heights of the buildings in the study area. Feature
extraction is carried out by the segmentation-based fractal texture analysis. They made use of three machine learning mod-
els - XGBoost, Random Forest, and LightGBM - the determination performance of the proposed model. Results proved that

20
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 9
Summary of path loss prediction based on other machine learning algorithms.

Refs. Algorithm Baseline algorithms Indoor /outdoor Generation Result/contributions

[92] Algorithms with Image Algorithms without Not available Not available The use of image
texture Image texture texture-based features
enhanced the accuracy of
the prediction
[76] Extreme learning ANN-BP, COST-231, Outdoor Not available The proposed method
machine Okumura-Hata, enhanced path loss
prediction in outdoor
environments
[105] RF, KNN COST231-Hata, Not available Not available The proposed method
Okumura-Hata enhanced path loss
prediction with low RMSE
[83] Ensemble method KNN, XGBoost, Not available Not available It was found that the
extreme learning trees, ensemble approach
RF and gradient improve accuracy
boosting
[94] GBoost RF Not available 5G GBoost predict path loss
better than the RF

the summation of the texture-based features enhanced the prediction of path loss for varying transmitter heights. Popoola
et al. [76] proposed an algorithm of extreme learning machine for outdoor path loss prediction. Extreme learning machines
were considered to develop an optimal prediction of path loss for an outdoor environment. Here, a ANN, which has a single
hidden layer was trained and tested with the data of path loss which was computed based on received signal strength. An
analysis between ANN backpropagation and extreme learning machines was carried out to determine their effectiveness and
generalization. Results showed that extreme learning machine models outperformed ANN backpropagation, Okumura-Hata,
and COST-231. It then showed that extreme learning machines guaranteed optimally and converged, and with a high level
of accuracy. Yang et al. [105] proposed a machine learning-based methodology for path loss determination and delay spread
in air–to–ground channels of millimeter waves. The algorithms of random forest and KNN were employed for the prediction
of path loss. To possess a generalization ability, the researchers also introduced a feature engineering scheme of the subset
having the best information among all its sizes with the values of evaluation measured. The proposed prediction model was
compared to the Okumura-Hata model and COST231-Hata models, with result showing that the proposed method had the
smallest root mean square error than other models, hence an improvement on the prediction. Sani et al. [83] proposed an
ensemble – KNN, extreme learning trees, RF, gradient boosting and XGBoost for the prediction of path loss in communi-
cation systems. Features of the multiple environment were extracted from satellite images. The categorical clutter height
values were also used. The propose ensemble approach is used to predict the path loss. It was found that the propose
approach improve the accuracy of path loss prediction compared to already published results. Sotiroudis et al. [94] used
synthetic and raw data for the modeling of path loss using RF before combining both the raw and the synthetic for the path
loss prediction. The datasets were collected through Ray Tracing simulations. The GBoost is used for the modeling of the
path loss using the ensemble of the data (raw + synthetic). Results indicated that the GBoost model perform better than
the RF with raw data and RF with synthetic data in predicting path loss. Table 9 shows a summary of the studies.

Hyperparameter settings

The setting of algorithm hyperparameters for the modeling of path loss influences the performance of the algorithm, thus
determining the accuracy of the path loss prediction. As such, makes the setting of the hyperparameters of the target algo-
rithm is critical regarding the prediction of path loss in wireless communication systems. The hyperparameters presented in
this section for different studies can guide researchers on the choice of hyperparameters for future modeling of path loss in
communication systems as there is no systematic way of getting the optimal hyperparameters.
Hyperparameters are parameters that, when appropriate values are provided, control the process of learning the model
and also determine the values of the model. With the recent development in machine learning, hyperparameter automation
and optimization have been constantly used. Sometimes, it is most efficient to know when and how to go about setting
the desired parameters for every given model or algorithm to function. Hyperparameter settings seek to answer questions
such as, given a model or an algorithm, what are its most important parameters, and what could be its required values
for proper processing (Rijn & Hutter, 2018). The prefix “hyper” is suggestive of the fact that these parameters are top-level
and they control the learning process of the model. Hyperparameter settings is an activity that is done by machine learning
engineers whereby they choose and set hyperparameter values for the learning algorithm to use even before training of
the model begins. Therefore, this process is an external activity that happens because the model will not be able to alter
its parameters during the training process (Rijn & Hutter, 2018). They are not part of the model itself but are used by the
learning algorithm; that is, these values remain the same at the end of the training procedure. Given the critical nature of
the hyperparameter settings in modeling path loss, we have extracted the parameters of the algorithm used by different
studies to predict path loss in wireless communication systems. Tables 10–13 present hyperparameters extracted from these

21
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 10
Hyperparameter settings for some deep learning.

Refs. Algorithm Hyperparameter

[99] ANN–CNN CNN: Activation function: ReLu; Kernel size:

(5,5, 3,3), (3,3), (3,3, 2.2, 2,2); Max pooling: 2;
padding: 2; stride: 1; Convolutions: 200, 100,
50, 25, 12, and 1.
ANN: layers: 200, 200; Activation function:
ReLu.
[4] VGG-16 Batch size: 6, learning rate: 0.0001,
VGG-16 is pre-trained using the 1000-class
[7] VGG-16 – ResNet-50 The batch size is set to 16 for VGG-16 and 32
for ResNet-50. The learning rate is set to
0.001 for both architectures
The weight (λ) of Sigmoid Cross-Entropy loss
is tested in the range [3, 10].
[92] XGBoost-CNN XGBoost: XGBoost: n_estimators=700,
learning_rate=0.1, colsample_bytree=0.9,
subsample=0.9, max_depth=8
convolutional_layers=4,
filter_numbers=16,32,64,128, filter_size=(3,3),
Pooling filter size=(2,2), Fully connected
layers=3, Nodes= 32, 8, 1.
[52] CNN The number of layers: 6, Learning Rate:
5 × 10−4
[101] CNN 3 convolutional layers, 1 fully connected layer,
Activation function: LeakyReLU, Filter Size: (5,
5) and 2, Number of Neurons in FC layers:
1281 and1
[82] CNN ResNet50V number of epochs: 100, Random
Test Split Sample: 10, Adam Optimizer
learning rate: 0.001, batch size of 12. The
number of layers in MLP: 249, 35, and 1.
[14] AE-CNN Learning Rate: 0.001, Batch Size: 64,
Minimum Training Epoch: 100
[13] CNN-Based Number of neurons: 20, number of hidden
layers: 1, Training Algorithm: Adam
Algorithm, Learning Rate: 0.001.
[59] AlexNet Number of convolutional layers: 4, number of
fully connected layers: 3.
[44] Deep NN Input length: 502, Output length: 1, Number
of hidden layers: 3, Number of Neurons: 502,
128, 8., Dropout rate: 0.18, Batch size: 128,
Optimizer: Adam, Loss function: MSE,
Activation: ReLU & Sigmoid.
[91] CNN-Image based Training set: 200 epochs, Adam Optimizer,
Learning rate: 0.001
[45] AE-GAN Convolution layer: 3, max pooling, fully
connected layer: 3, Deconvolution layer: 3,
activation function: ReLu and Sigmoid, fully
connected layer 1: 0.3 dropout, 512 neurons,
linear activation, fully connected layer 2: 0.3
dropout, 2500 neurons, sigmoid activation.
[49] CNN Convolutional layer: 5, pooling layer: 3 and
fully connected layer: 3.
[77] PPNet Max pooling, convolution, deconvolution and
fully connected layer.
[11] VGG-16 3 × 3 convolution layers, max-pooling layers,
3 fully connected layers. Sigmoid activation.

different studies of path loss modeling using machine learning algorithms. We have only presented the hyperparameters of
the studies that are available in the published work and only the hyperparameters reported in the paper.

Study environment for path loss prediction

When considering the design of wireless communication networks and determining propagation path loss in wireless
signal transmission, it is important to consider the environment in which the signal is to be transmitted. The nature of the
propagation environment determines the impact of propagation path loss experienced in the transmission process. Propagat-
ing signals are expected to travel from the source transmitter to the destination receiver which, under normal circumstances,
shouldn’t be impacted by any factor in the line of transmission. Radonjić & Gaćeša [78] tried to find out the effect of the

22
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 11
Hyperparameter settings for artificial neural network.

Refs. Algorithm Hyperparameter

[104] ANN One hidden layer activation function:

8(1) − 14(1) Activation function: tansig/purelin,
Two hidden layers: 10(1) , 5(2) Three Hidden
layers: 10(1) , 5(2) , 3(3) Epoch: 1000
[102] BPNN Activation function hidden layer: Sigmoid
Activation function neuron in output layer:
Sigmoid
[61] RBFNN and MLPNN MLPNN Layers: 3, input layers have 3 units,
the hidden layer has 20 units, output layer
has 1 unit. Hidden layer Activation function:
relu, Output layer activation function:
Sigmoid.
RBFNN Layer: 3, 3 input units, 30 hidden
units, and 1 output unit. 30 clusters centroids.
[43] ANN, Gaussian process, and PCA Learning rate: 0.001, Epoch: 1000, Tolerance
rate: 0.00001
[74] ANN Number of hidden neurons: 9, Activation
Functions: Hyperbolic Tangent Sigmoid
(Tansig)
[106] ANN Activation functions: Gaussian,
Multi-quadratic, inverse multi-quadratic,
Cauchy function.
Number of epochs: 500, Training parameter
goal 10−4
[89] ANN Input layer nodes: 10, output layer node: 1,
number of epochs: 80,
[62] ANN Input normalization range: −1 and 1,
Activation function f (. )
[74] ANN Activation function: HTAF and logistics
sigmoid, Hidden layers: 80, number of
neurons: between 10 and 100.
[89] ANN The number of hidden layers: 2 and 3,
Activation function: hyperbolic Tangent
function for hidden layers, Linear for output
layers.
[27] ANN The number of hidden layers: varied from 31
to 39 incremental steps of 2, Activation
functions: logsig, purelin, and tansig.
[63] ANN Activation Functions: Rectifier, hyperbolic
tangent, logistics sigmoid., Learning Rate:
0.001, Epoch: 1000, Tolerance rate: 0.00001
[72] ANN The number of hidden layers: 1, Number of
output layers: is 1.
[73] MLP-RBF RBF: Activation Function: Gaussian, Spread of
the Gaussian is 0.01, RBF.
[46] ANN Activation Function: hyperbolic tangent
sigmoid function (Tansig) in the hidden layer
and Purline for the output layer.

[43] ANN/ PCA and Gaussian Process Activation functions: rectifier, logistic sigmoid,
and hyperbolic tangent functions.
[103] BPNN, SVR, Random Forest, AdaBoost BPNN: one input layer, one hidden layer and
varies between 2 and 8 at an interval of one,
one output layer.
SVR: Kernel Function= Gaussian kernel,
insensitivity loss is 0.2.
[98] ANN, RT, and k-NN Hidden layers: 100, 150, 100
Activation function: ReLu,
[19] SVR and ANN SVR Training and testing set of 80% and 20%
respectively, ANN Training and testing set of
80% and 20% respectively, Number of
neurons=4
[26] ANN ANN number of sample=200, ANN has 30
hidden layers, with three activation functions
such as sigmoid
[57] ANN ANN: number of hidden layers = 3, 50 hidden
neurons, used non-linear activation function.
Levenberg–Marquardt’s backpropagation
algorithm was adopted for the learning
process.
(continued on next page)
23
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 11 (continued)

Refs. Algorithm Hyperparameter

[88] ANN, RF ANN: 2 hidden layers each with 32 neurons,

Activation function: rectified linear unit based
RF: number of nodes – 500 and 2
[71] SVM, ANN Dataset 2356 points, 2000 for training and
356 points for testing

Table 12
Hyperparameter settings for Fuzzy Neural Network.

Refs. Algorithm Hyperparameter

[2] ANFIS Hybrid learning algorithm that combines Least

Square Error and Gradient Descent method is
used for the training, a Single input layer, three
hidden layers, and one output layer
[12] K-Means fuzzy logic The fuzzy logic limit is defined using the
triangular membership function (trimf), and the
distance membership function is [0 400].
[16] Neuro-Fuzzy Activation function: hyperbolic tangent, 1 neuron
in the input layer, 95 neurons in the hidden layer,
1 neuron in the output layer.
[17] ANFIS Number of layers: 5, Membership function:
Gaussian distribution
[18] Fuzzy logic Membership functions: asymmetric triangular,
Other membership functions: Gaussian, Generic
and trapezoidal
[34] ANFIS Membership function of Sugeno-type,
Generalization: Complete data set has been
divided into 3 parts.500 samples for tuning, 500
samples for testing, and the remaining 500 for
checking,
Checking: Overfitting of training data is prevented
using separate data set of 500 samples
[35] Fuzzy-Logic Membership Function: triangular membership
functions, classify the fuzzy variable into 5 levels
as follows: NZ: Nearly zero, S: Small, M: Medium,
L: Large, and VL: Very large.
[40] Neuro-Fuzzy The number of simulations: 15, 1 to ratio 14 was
reserved for testing and training.
[56] BPSK Membership function: Triangular,
[69] Fuzzy Logic Membership Function: Triangular, Fuzzy variables
are classified into 5 levels.
[66] Fuzzy-Logic Membership Function: Triangular
[66] Fuzzy-Logic Membership Function: Triangular
[67] Fuzzy Logic Network Analyzer, the measurement point was
about 10 to 20 m

[65] Fuzzy Regression Triangular Fuzzy set, 3 Crisp parameters, fuzzy

number if determined by conventional regression
[81] Neuro-Fuzzy Model The number of layers: 5, first order Sugeno model
was used to describe the structure, membership
function: generalized bell membership function.
[87] Fuzzy Logic Membership function: RSPower and Distance
input has Gaussian membership function, path
loss member function is the Triangular
membership function. Defuzzification method:
Centroid method
[97] ANFIS Membership Function: Pi and Number of epochs:
300–1000,

environment on the propagation of wireless signals and asserted that the quality of the transmission of wireless signals
from source to destination depends on the environmental impact on the transmitted signals. Different environments can be
considered for the study of path loss propagation as explained in the following sub-sections.

Urban environment

The urban environment is the region that is surrounding a developed city. Characteristically, they are very developed
with a high level and concentration of structures such as commercial buildings, houses, cabins, overhead bridges, railways,

24
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 13
Hyperparameter settings for nature-inspired meta-heuristic algorithm.

Reference Algorithm Hyperparameter

[58] Optimized COST 231 Hata Model Population Size: 50, Number of iterations: 100,
Cross-over rate: 2, Mutation rate: 1
[42] PSO W = 0.5, C1 =C2 =1.494, best solution key, and
best particle gain.
[6] ABC-PLOSS Fitness function: COST 231 Hata model, Colony
Size: 3, Employed Bees limit trials: 5, maximum
forage cycle: 5, Parameter size: 3
[31] Optimized COST231 Optimization Parameter: k1=46.3, k2=44.9,
k3=13.82, k4=6.55, k5=33.9
[3] Optimized Okumura-Hata optimization variables used is 6, variable limits
between 0 and 3, The GA parameters: population
size of 12, the mutation rate of 2, selection of 0.5

Table 14
Breakdown of study environment for different studies used for path loss prediction.

Refs. Environment

[82] Multiple Environment

[4,27,58,72,73] Urban and Suburban
[101] Smart factory
[3,13,16,18,35,43] Suburban
[15,17,44,45,49,53,55,60,63,66,74,77,84,89,90,91,92,94,95,98,100,116]. Urban
[59] Open Square
[54] Rural and Urban
[62,57] Rural
[6,56,83] Urban, Suburban and Rural
[106] Mine Environment
[103] Cabin
[88] Campus
[65,67] Forest
[31,97] Rural and Suburban
[87] Metropolitan
[42] Agriculture environment

and so on. Propagation of wireless signals in this kind of environment requires a good understanding of the possibility of
propagation path loss and ways to avoid such. With the presence of structures, such as skyscrapers and tall buildings in the
urban environment, the signal loss could easily be experienced, hence the need for proactive surveillance [10].

Suburban environment

Suburban areas (Suburbs) or environments are areas that are lower population density environment, than urban, which
separate places where people reside and places where there is a high concentration of commercial activities. A Suburban
area is either part of an urban area or a rural area. It is not a super-crowded settlement and the houses in a suburban
environment are based on neighborhood. It can also be seen as an area on the outskirt of a city. Studies have shown
significant results in path loss determination of path loss in this kind of environment [30].

Rural environment

The rural environment is an open space of land that has very few buildings with a minimal percentage of humans living
there. The urban environment, which is rocked by so many developmental structures and infrastructural developments, is
the direct opposite of the rural environment. In rural areas, it is believed that agriculture is the primary industry. Businesses
are located far away because of the absence of people and buildings and, wildlife farming is mostly carried out in rural
areas. Here, wireless signal propagation might be affected by a few parameters, such as vegetation, when compared with
urban environments [64]. Different studies were conducted in different study environments, many of which typically report
the study environment where the data were collected for the research. The survey extracted the study environment for
different studies as presented in Table 14. Only studies that revealed the study environment in their published research
report are included.

System and environmental parameters for path loss prediction

It is important to note that proper data collection, choice of system variables, required parameter settings of the measure-
ment set-up, and choice of the environmental variables are desired in the development of an efficient path loss prediction

25
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 15
Breakdown of feature engineering for deep learning path loss modeling and radio equipment used for measurement.

Refs. Feature selection method Feature extraction method Radio measurement equipment

[99] N/A CNN Rohde & Schwarz (R&S) TSMW

[4] N/A CNN Not available
[7] N/A Deep learning techniques PlaceMaker plugin, SketchUp Software
[92,83] N/A CNN Not available
[52] N/A CNN Not available
[101] N/A CNN Not available
[82] N/A ResNet50V2 CNN Not available
[14] N/A AE-CNN Not available
[13] N/A CNN Not available
[55] N/A CNN Not available
[54] N/A CNN GPS Unit, SX1278 Transceiver, STM32LD MCU, Battery
[91] N/A CNN Not available
[95] N/A CNN RF planning software suite EDX
[45] N/A CNN Not available
[49] N/A CNN Rx antenna, Tx antennas and vehicle
[77] N/A No mention 3D-ray-tracing software
[11] N/A VGG-16 3D ray tracing software

model. With the required inputs, it is possible to choose and set desired parameters/variables for prediction using machine
learning algorithms [74].
Most importantly, the study of the environment, whether open, suburban, or urban areas, where the signal will be prop-
agated will give a clear instance of how the environmental variables will be chosen and, perhaps, set its parameters for the
prediction purpose. It is worth mentioning that most researchers based their research and set their environmental parame-
ters accordingly by observing the specific case study. For example, path loss prediction in open areas or build-up environ-
ments is entirely different from mines environment or vegetation environments. Similarly, the indoor scenario is different
from the outdoor, and so on.
According to Popoola et al. [74], the received signal strength is the basic idea behind predicting propagation path loss.
The transmitting antenna height and its direction of transmission cannot be left out in setting environmental variables and
parameters. To further set the system for prediction, the transmitting power must be determined before thinking of carrying
out prediction functions. To measure path loss, the general setup for this purpose comprises mainly the global positioning
system (GPS), spectrum analyzer or field strength meter, vector signal generator, and a personal computer [74]. The spectrum
analyzer is mainly used to measure the received signal strength while the GPS gives the spatial coordinates (in degrees) of
the measurement points while determining the elevation of the receiver’s location [75]. Thus, many variables are consid-
ered for measuring path loss. However, the general-purpose path loss data consists of the terrain profile, transmitter, and
receiver heights, transmission distance, climatic condition, transmitting frequency, and transmitter power level. Sometimes,
other variables such as antenna polarization, azimuth, directivity, antenna types, antenna gains, and path loss exponent,
among many others, are also considered. These variables, from the study environment, can be referred to as features. The
equipment, with specifications used by different studies for measurement in the different study environments, is reported in
Tables 15–18 for only the studies that make available the measuring equipment. Tables 15–18 present the different feature
engineering methods used for path loss prediction as adopted by different studies.

Feature engineering for modeling path loss in communication systems

When the number of features (e.g., variables/inputs/parameters) is very large to the extent that it is not feasible to
use all the features, the irrelevant and unwanted ones are removed, typically by the dimension reduction technique, thus
leaving only the relevant features. The dimension reduction is conducted to improve accuracy and visualization and provide
a better understating of the acquired knowledge. Feature selection and feature extraction are the main feature engineering
techniques typically employed to solve the problems of redundant and irrelevant features. Feature selection is therefore
the process of eliminating irrelevant features while selecting the most relevant, whereas feature extraction is the process of
transforming large input features into smaller subsets of the features that preserve the highly relevant features (Khalid et al.,
2014). Therefore, feature selection and extraction are required for feature engineering to improve the prediction accuracy of
path loss.

Meta-data analysis and discussion

Machine learning as an alternative to traditional models

The study of path loss prediction seeks to generally investigate what happens between the point of signal transmission
(Transmitter) and the destination or the target (Receiver). Here, many research efforts have been put in place towards in-
vestigating this problem. Many traditional models for path loss prediction, such as basic models, supplementary models,

26
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Table 16
Feature engineering methods for artificial neural network and radio equipment used for measurement.

Refs. Feature selection method Feature extraction method Radio measurement equipment

[104] N/A Not available Not available

[43] N/A PCA Not available
[74] N/A Not available SEFRAM 7806, GARMIN GPS MAP 76CS, LAPTOP
[27] N/A Not available GPS, Laptop, inverter, TEMS Phones.
[60] N/A Not available VISAFONE CDMA2000, GPS, LAPTOP
[102] N/A Not available Not available
[106] N/A Not available Not available
[89] N/A Not available Not available
[62] N/A Not available Not available
[74] N/A Not available Not available
[89] N/A Not available Not available
[63] N/A Not available Not available
[72] N/A Not available Not available
[73] N/A Not available Not available
[46] N/A Not available Not available
[43] N/A PCA Not available
[98] N/A Not available Laptop, GPS, GPS receiver, ad HAFP
[19] N/A Not available Not available
[26] N/A Not available GPS, Camera, Geographical mapping system
[57] N/A PCA Not available
[88] Evaluating correlation Not available Not available
[71] N/A PCA Not available

Table 17
Feature engineering methods for Fuzzy Neural Network and radio equipment used for measurement.

Refs. Feature selection method Feature extraction method Radio measurement equipment

[2] N/A Not available 3 G Ericson (W995) TEMS Phone, GPS, Computer,
[12] N/A Not available Not available
[16] N/A Not available Samsung smartphone, model Galaxy J5 Duos
[17] N/A Not available TEMS (Test Mobile System)
[18] N/A Not available Not available
[34] N/A Not available XCAL, Laptop,
[35] N/A Not available Not available
[40] N/A Not available Not available
[56] N/A Not available Not available
[69] N/A Not available Not available
[66] N/A Not available Not available
[68] N/A Not available Vehicle or a handcart
[67] N/A Not available Microcomputer, network analyzer
[65] N/A Not available Car, Network Analyzer, Microcomputer,
[81] N/A Not available Not available
[87] N/A Not available Not available
[97] N/A Not available Spectrum Analyzer, GPS, Memory system,

Table 18
Feature engineering methods for nature-inspired meta-heuristic algorithms and radio equipment used for
measurement.

Res. Feature selection method Feature extraction method Radio measurement equipment

[58] N/A Not available Not available

[42] N/A Not available Drone, Router, Laptop, XBee S2C
[6] N/A Not available Not available
[31] N/A Not available Laptop, GPS, Network analyzer,
[3] N/A Not available GPS
[15] N/A Not available Not available
[100] N/A Not available Not available
[51] N/A Not available Not available

foundational models, many ray models, terrain models, measurement-based/corrected models, and stochastic models, were
the dominant methods for path loss prediction before now. The results obtained from these models limit the accuracy,
hence the motivation to investigate new methods of predicting path loss [70]. This, for example, necessitated the adoption
of machine learning in path loss prediction models. A lot of efforts have been invested in path loss prediction in wireless
communication signal propagation as communications technology rapidly advances. Machine learning has proven to be an
alternative approach to traditional models for the prediction of path loss. Machine learning techniques have shown to be

27
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

3.5

publicaons 2.5

1.5

0.5

0
2019 2020 2021 2022
Year
Fig 5. Trends of publication using hybrid deep learning architecture.

more effective, efficient, and robust than traditional models in the prediction of path loss in wireless communication sys-
tems. It is revealed that most of the studies indicate that the 5G technology is the dominant technology and is more recently
used by the research community when developing path loss prediction models.

Deep learning as an alternative to shallow machine learning methods

Deep learning algorithms have become the de-facto algorithms for developing modern path loss predictive systems.
While existing machine learning approaches for path loss modeling depend largely on hand engineering feature extrac-
tion and selection phases, deep learning architecture has become an alternative technology to address this problem. Deep
learning models are known for the automatic extraction of features or parameters once the architecture has been carefully
designed. Thus, it offers better and comparable performance to the model built with hand engineering features. In this sur-
vey, we have shown cases where deep learning models have outperformed the traditional methods and machine learning
methods that are based on hand engineering features.

Hybrid deep learning

Research findings show that propagation path loss prediction results from single models have no absolute value of ac-
curacy. This makes researchers hybridize two or more deep learning models to overcome the limitation of a single deep
learning algorithm and improve performance. Thus, hybridizing of deep learning algorithms to predict path loss is emerging
in recent times with superior prediction accuracy as compared to the single deep learning algorithms and traditional mod-
els. The frequency of the applications of deep learning algorithms for the prediction of path loss was collected and depicted
in Fig. 5. The trend indicates the growing interest in using hybrid deep learning to predict path loss.
At the time of carrying out this survey in 2021, there was an increase in the number of research efforts that use hybrid
deep learning architecture geared towards propagation path loss prediction in wireless communication networks.

Convolutional neural networks

CNN variants have been popular in solving prediction problems involving image processing because of the advantages
attributed to the CNN having an in-built automated feature extraction mechanism for feature extraction without the need for
extra effort. As a result, several research efforts have used CNN architecture in propagation path loss prediction in wireless
communication networks. Fig. 6 shows the distribution of the application of the CNN architectures in path loss prediction. At
the time of this research, seven papers in the literature predicted propagation path loss in wireless communication networks
using CNN. Findings have it that, though CNN has been in existence for a while now, the first category that gave birth to the
modern distribution of CNN is the AlexNet. Besides it, other distributions as submitted in the research included RadioUnet
and VGG-16.

Shallow machine learning algorithms

At the time of this research, the ANN has the highest number of publications in the literature. More than twenty re-
search publications used ANN to predict propagation path loss in wireless communication networks. Some of the studies
hybridized the ANN with deep learning architecture. Results obtained from using these architectures show an enhanced

28
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Fig. 6. Distribution of CNN variants for path loss prediction as published in the literature.

Fig. 7. Percentage distribution of the types of Fuzzy networks.

path loss prediction, thus advancing the cause for research in many ways. The ANN is still active in the prediction of path
loss in wireless communication systems. After the ANN, fuzzy ANN is the second most commonly used algorithm for the
prediction of path loss in wireless communication systems. An analysis of the distribution of various Fuzzy ANN models in
the literature is presented in Fig. 7.
It can be observed from Fig. 8 that, of all the architectures available in the literature that predicted path loss, ANN
had the highest number of publications, closely followed by the fuzzy neural network. On the other hand, deep learning is
showing growing interest.
Worthy of note is the nature-inspired meta-heuristic algorithms (e.g., GA, PSO) that are used for wireless network prop-
agation path loss prediction, as there exist a certain number of published works at the time of this study work.

Hyperparameter settings

The performance level of the machine learning algorithms for predicting path loss is sensitive to hyperparameter settings.
The choice of hyperparameters used is therefore crucial to the accuracy of the path loss prediction. The hyperparameters are
the basic standard for tuning models to carry out predictions. Whenever the required optimal parameters are provided there
is an assurance that the model will produce the desired results. Parameter settings are mostly done before the learning pro-
cess of models which hitherto controls the outcome of the model or the algorithm as it performs its designed functions. Of
all the architectures used, parameter settings were presented for each paper found in the literature as highlighted in Tables
10–13. It can be deduced from the Tables that, up to date, there is no systematic way of getting uniform hyperparameters
for every algorithm because each algorithm has unique hyperparameter settings.

29
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

30
25
20
Frequency 15
10
5
0

Architecture
Fig. 8. Frequency distribution of published literature based on different architectures.

Fig. 9. Distribution of the study environment as found in the survey.

Feature engineering

This survey has considered feature engineering, bringing to bear all the processes by which models learn features from
the data fed into the model for better path loss prediction result delivery. It is found that the research community heav-
ily relies on in-built data extraction mechanisms in deep learning architecture for feature extraction in the process of path
loss prediction, though handcraft feature extraction is the predominant method. In deep learning, models can learn and ex-
tract features from 3D image data and, through a trained model, make informed predictions with a good level of accuracy.
However, many of the studies do not give preference to data feature extraction despite its significance in path loss predic-
tion, while others make the feature extraction method used in the study available in the published research work. Feature
selection is highly scarce in the domain of path loss prediction.

Study environment

The environment to collect data and conduct research on path loss prediction plays a crucial role. Since signal attenuation
is caused by a lot of factors in the environment as signals are being propagated, the concentration of certain factors in
the environment will determine the kind of network to be designed by the engineers, hence the need to determine the
nature of the environment. The studies conducted on path loss prediction were carried out in different environments as
discussed in Section 9. Most researchers always give the nature of the environment where the signal is propagated. It is
found that the research community mostly conducts research on path loss prediction in wireless communication systems in
the urban environment as shown in Fig. 9. This is likely because the urban area has the largest concentration of wireless
communication systems compared to the other environment like rural or suburban. Some of the researchers conducted the
research in multiple environments as indicated in Fig. 9, the trend is likely to gain interest in the future.

30
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Challenges and new point of view for solving the challenges in future research

The existing path loss models are faced with different challenges. Research in path loss prediction has therefore recently
taken a new dimension due to the development in AI technologies, robust deep learning architectures, machine learning
methods, and computational intelligence techniques., some of which are widely discussed in this section. We also discuss
some of these challenges and provide possible recommendations for future research in the domain of path loss prediction.

Deep learning modeling

Deep learning models are faced with several challenges that the existing studies for path loss need to critically address
them. As an example, deep learning architecture is time-consuming to train but, once trained, offline inferencing can be done
fast and in real-time. Nowadays, the availability of Graphics processing unit architecture and high-performance computing
has reduced the training time of deep learning models. However, these systems are expensive to purchase individually.
Leveraging the advantages of cloud-based infrastructure, researchers have been able to successfully train and deploy robust
deep learning solutions in many domains. Future research in path loss predictive modeling should explore this direction.
While this proposal is not new, several studies have been leveraging this advantage. For instance, Thrane et al. [99] pro-
posed the hybridization of traditional ANN and CNN (ANN-CNN) for the prediction of path loss in mobile communication
systems using satellite images as input. In the architecture, the CNN processes the satellite images while the ANN computes
the numerical features. It was established that this architecture improved the prediction accuracy of the path loss compared
to the stochastic models and ray-tracing methods. However, the research community in path loss predictive modeling is
still faced with the question of which deep learning architecture provides optimal performance for developing efficient path
loss predictive solutions. Furthermore, consideration should be geared towards the deployment of current deep learning
technologies in the development of accurate predictive path loss propagation models for emerging technologies.

Feature engineering

One of the important stages of developing a path loss predictive model is the identification of discriminative features
for developing the system. Features or parameters used for path loss modeling can be categorized into two: system and
environment features. The system-dependent parameters include carrier frequency, transmitter-receiver separation distance,
antenna gains, transmitter and receiver heights, operating frequency, and angle between the line of sight and horizontal
plane among others. These features are independent of the environment propagation and as such can be extracted sepa-
rately. The environment-dependent parameters include terrain, vegetation, and building condition, which are determined by
the geographical environment and condition of the weather. Invariably, the features or parameters are obtained from 3D
maps and topographic databases, as well as the database for the land cover. The parameters or features obtained from the
weather include humidity, temperature, precipitation rate, and so on.
Hand-engineering of features for building path loss models has been the major direction of the existing studies; how-
ever, this approach suffers from generalizability. The empirical propagation models rely mainly on measurement data from
specific environments of interest which limits their applicability. Ray-tracing-based solutions are usually employed to com-
plement this limitation; they are however computationally expensive. Furthermore, the non-availability of manual extraction
of features that could characterize the complex physical and geometric structures of the propagation environment hinders
the models’ generalizability ability. Numerous unwanted and irrelevant parameters have been employed in previous studies.
Future research must employ feature selection methods such as filter, wrapper, and embedded methods to select relevant
parameters that will focus on developing a path loss predictive model with generalization property.
There is therefore a need for an extensive comparison of models based on hand engineering features and those of the
models based on automated feature extraction using architecture such as a convolutional neural network.

Algorithms and hyperparameter settings

Correct selection of algorithms will play a significant role in building an efficient and accurate path loss predictive system.
In addition, tuning of hyperparameters of these algorithms is also important in future work. Existing studies to investigate
extensively the performance of the most widely used algorithms for developing path loss predictive model is lacking in the
literature. The use of factors, such as the accuracy and complexity of the algorithm, can be explored to select an appropriate
algorithm for the path loss predictive system. The hyperparameter values of the algorithms are set before the beginning of
the learning process. Different approaches exist for the selection of the optimum hyperparameters settings such as the ran-
dom search, grid search, and Bayesian optimization. The hyperparameters settings play a critical role. Although researchers
have explored diverse neural network architectures, system parameters, and learning algorithms to develop path loss models.
These are majorly based on deep learning models. Thus, there is a need to conduct an extensive comparative analysis study
that will provide more insightful knowledge in this direction to accommodate other computational intelligence techniques.

31
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

Performance metrics for model evaluation

Existing studies have employed different performance metrics to evaluate the efficacy of their path loss predictive mod-
els. Some of the performance metrics used include mean squared error, root mean square error, correlation factor, maximum
prediction error, mean absolute error, mean absolute percentage error, standard deviation error, regression coefficient, pre-
diction time, and accuracy. There are however no universally acceptable performance metrics for path loss predictive model
evaluation as different studies used different performance metrics. Each study reports one or more of these metrics, which
makes it difficult to accurately compare the performance of existing studies in the literature. Future work should address
this challenge by providing a comprehensive evaluation of existing studies using clusters of performance metrics. This will
provide more insight into which metrics or groups of metrics should be universally adopted to evaluate path loss predictive
systems.

Interference and signal attenuation

One of the major requirements for a communication system is the presence of the sender and receiver. During wireless
communication between sender and receiver, the signals from sender to receiver can get obstructed by many obstacles.
Buildings and terrain profiles affect signal coverage areas, which can lead to a poor communication system. Similarly, atten-
uation from competitive signals within the coverage areas is another challenge to building an effective telecommunication
system. Signals from an antenna, depending on the propagation environment, can pick a variety of paths, leading to the
receiver causing every path to communicate with its surroundings differently and, thus, causing a delay in arrival by a par-
ticular amount of the signals at the receiver. These delays may lead to constructive interference when they are in phase
with each other, or destructive, not in phase with each other. The attenuation from either of these interferences will result
in multipath fading, which may occur based on the type of objects that causes the obstruction. For instance, if attenuation
is caused by static large obstacles such as mountains and buildings, the resultant effect is slow-fading. However, If the at-
tenuation is caused by small obstacles with transient properties and with a variation of time, a fast-fading or small-scale
fading will occur. Whenever signals communicate with their surroundings they cause frequency delay, reflections, or shifted
frequency because of diffractions. A frequency shift may also occur for the mobile receiver because of Doppler spreading.
Conversely, small-scale fading occurs as a result of frequency shifts and delay. Therefore, when designing an effective path
loss system, different configurations should be examined to capture the underlying characteristics of the propagating signals.

Traditional methods

Traditional methods have been in existence and are widely used to develop path loss predictive systems when it is diffi-
cult to obtain the actual desired measurements. These methods make a prediction based on the availability of prior knowl-
edge of the underlying path loss predictive system. Theoretical models, for instance, are purely analytical models which are
derivative of idealized electromagnetic propagation theory. Additionally, traditional methods such as terrain models, supple-
mentary models, stochastic fading models, and many-ray models have also been widely used to develop path loss predictive
systems. While these traditional models are suitable for analytical scenarios, they require accurate and precise knowledge
of the environment of interest which may be difficult to model directly in most cases. As an example of such requirements,
many-ray models try to compute the path loss by taking the summation of the losses along distinct paths, rather than on
LOS path. This constraint can be lifted to develop a robust path loss system. In reality, modeling a path loss system that
considers robust parameters associated with the environment and the system may be difficult to achieve. However, future
research should take into consideration some of the useful environmental factors that could improve the performance of the
target path loss predictive systems.

Existing path loss models

No doubt, several models have been proposed in the literature which focuses majorly on developing efficient path loss
predictive system. Some of the well-known models for path loss prediction are Okumura-Hata, COST-231, Walfish-Ikegami,
Stanford University Interim, Ericsson, Egli, ECC-33, Standard Propagation Model, and ITU 452 Model, just to mention few.
These are the baseline models commonly used by researchers to compare the performance of their developed path loss
system. Therefore, it is almost impossible to know the performance trends of the newly developed path loss models in the
literature. It is therefore important that future research should consider conducting a robust comparative analysis study that
will provide more insight into the performance of the existing path loss models in the literature beyond the traditional or
common path loss models that are commonly used for performance comparison.

Machine learning methods

The development of machine learning technologies is attributed to the successful deployment of path loss predictive
models with improved accuracies. Machine learning models such as ANN, ANFIS, and computational intelligence methods
(e.g., GA, ABC, and PSO) have been at the forefront of path loss predictive modeling. For example, studies have observed

32
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

that PSO produces improved accuracy for path loss prediction when compared with GA. While the ANN has played a signif-
icant role in developing a path loss predictive model, studies have observed that increasing the number of hidden neurons
ordinarily, does not yield a better result as a tradeoff must be achieved to reach the best architecture. Additionally, the
expandability of machine learning algorithms to provide the underlying explanation and interpretation of their predictive
outputs has been considered this leads to techniques such as the SHapley Additive exPlanations techniques.
As machine learning methods have achieved significant performance for path loss predictive modeling, however, most of
the techniques explored in the literature are in the supervised learning paradigm of machine learning. This places a strin-
gent requirement on the need to obtain clear measurements that are free of environmental influence to build a realistic
path loss predictive system. Additionally, these measurements must be labeled so that a function that models the output
of interest (i.e. path loss) can be easily developed. Obtaining realistic measurements that are free of human influence and
critical environmental factors is one of the major challenges of building effective path loss predictive models using machine
learning methods. Therefore, future research should critically consider this challenge. In addition, other paradigms of ma-
chine learning such as unsupervised, semi-supervised, and reinforcement learning are under-explored for developing path
loss predictive systems.

Conclusions

This survey paper has presented a comprehensive survey on the adoption of shallow machine learning and deep learn-
ing algorithms to predict path loss in wireless communication systems. The survey conducted a large-scale survey on the
applications of machine learning algorithms such as ANN, ANFIS, GA, random forest, KNN, etc., and deep learning architec-
tures such as the ResNet, VGG, deep LSTM, AlexNet, etc., for the prediction of path loss. The survey has presented feature
engineering, hyperparameter settings, study environment, and publication trends. Synthesis and analysis of the published
works on path loss prediction using machine learning and deep learning models are presented. It is found in the survey
that deep learning architectures are fast gaining interest from the research community because it has proven to have better
performance than conventional machine learning algorithms like the ANN. The studies that applied deep learning for path
loss prediction heavily depend on the automated feature extraction mechanism in the deep learning architecture. Feature
selection is scarce in the domain of path loss prediction. It is found that most of the studies were conducted in urban areas,
suburban, rural, or multiple environments. The challenges facing the prediction of path loss using machine learning meth-
ods were outlined and discussed. Lastly, the directions for future research were pointed out for easy identification of the
direction to follow for solving the challenges in the future.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgment

This Work is Funded by the Federal Republic of Nigeria under the National Research Fund (NRF) of the Tertiary Education
Trust Fund (TETFund) Grant No. TETF/ES/DR..&D-CE/NRF2020/SETI/64/VOL.1 and the Nigeria Communications Commission
(NCC) under Grant No. NCC/R&D/RG/SLU/001.

References

[1] A. Abdulkarim, N. Faruk, E. Alozie, A.O. Sowande, I.Y. Olayinka, A.D. Usman, K.S. Adewole, A.A. Oloyede, H. Chiroma, S. Garba, A.L. Imoize, L.S. Taura,
Application of machine learning algorithms to path loss modeling: a review, in: Proceedings of the 5th IEEE ITED Conference, IEEE, 2022, pp. 1–5.
[2] Z. Adeyemo, T. Olawuyi, O. Oseni, S. Ojo, Development of a path-loss prediction model using adaptive neuro-fuzzy inference system, Int. J. Wirel.
Microw. Technol. 9 (6) (2019) 40–53.
[3] Z.K. Adeyemo, O.K. Ogunremi, A.O. Akande, Genetic algorithm based pathloss optimization for long term evolution in Lagos, Nigeria, Int. J. Appl. 6
(2) (2016) Available on https://www.ijastnet.com/journal/index/827.
[4] O. Ahmadien, H.F. Ates, T. Baykas, B.K. Gunturk, Predicting path loss distribution of an area from satellite images using deep learning, IEEE Access 8
(2020) 64982–64991.
[5] P.P. Angelov, Evolving Rule-Based models: a Tool For Design of Flexible Adaptive Systems, 92, Springer Science & Business Media, 2002.
[6] V.I.E. Anireh, E.N. Osegi, ABC-PLOSS: a software tool for path-loss minimisation in GSM telecom networks using artificial bee colony algorithm, Int. J.
Swarm Intell. 4 (1) (2019) 20–37.
[7] H.F. Ates, S.M. Hashir, T. Baykas, B.K. Gunturk, Path loss exponent and shadowing factor prediction from satellite images using deep learning, IEEE
Access 7 (2019) 101366–101375.
[8] M. Ayadi, A.B. Zineb, S. Tabbane, A UHF path loss model using learning machine for heterogeneous networks, IEEE Trans. Antennas Propag. 65 (7)
(2017) 3675–3683.
[9] M. Bkassiny, Y. Li, S.K. Jayaweera, A survey on machine-learning techniques in cognitive radios, IEEE Commun. Surv. Amp Tutor. 15 (3) (2013) 1136–
1159, doi:10.1109/surv.2012.10 0412.0 0 017.
[10] J. Baten, Creating firms for a new century: determinants of firm creation around 1900, Eur. Rev. Econ. Hist. 7 (3) (2003) 301–329.
[11] M. Bal, A. Marey, H.F. Ates, T. Baykas, B.K. Gunturk, Regression of large-scale path loss parameters using deep neural networks, IEEE Antennas Wirel.
Propag. Lett. 21 (8) (2022) 1562–1566.
[12] W. Bhupuak, S. Tooprakai, Minimizing path loss prediction error using k-means clustering and fuzzy logic, Turk. J. Electr. Eng. Comput. Sci. 26 (4)
(2018) 1989–2002.

33
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

[13] H. Cheng, S. Ma, H. Lee, CNN-based mmWave path loss modeling for fixed wireless access in suburban scenarios, IEEE Antennas Wirel. Propag. Lett.
19 (10) (2020) 1694–1698.
[14] H. Cheng, S. Ma, H. Lee, M. Cho, Millimeter wave path loss modeling for 5G communications using deep learning with dilated convolution and
attention, IEEE Access 9 (2021) 62867–62879.
[15] C.C. Chiu, Y.T. Cheng, C.W. Chang, Comparison of particle swarm optimization and genetic algorithm for the path loss reduction in an urban area, J.
Appl. Sci. Eng. 15 (4) (2012) 371–380.
[16] H. Cruz, R. Nascimento, J. Araujo, E. Pelaes, G. Cavalcante, Methodologies for path loss prediction in LTE-1.8GHz networks using neuro-fuzzy and ANN,
in: Proceedings of the Paper Presented at the 2017 SBMO/IEEE MTT-S International Microwave and Optoelectronics Conference (IMOC), 2017.
[17] T.E. Dalkiliç, B.Y. Hanci, A. Apaydin, Fuzzy adaptive neural network approach to path loss prediction in urban areas at GSM-900 band, Turk. J. Electr.
Eng. Comput. Sci. 18 (6) (2010) 1077–1094.
[18] A. Danladi, P. Vasira, Path loss modeling for next generation wireless network using fuzzy logic-spline interpolation technique, J. Eng. Res. Rep. (2018)
1–10.
[19] S. Duangsuwan, P. Juengkittikul, M.M. Myint, Path loss characterization using machine learning models for GS-to-UAV-enabled communication in
smart farming scenarios, Int. J. Antennas Propag. (2021) 2021.
[20] N. Faruk, Y.A. Adediran, A.A. Ayeni, Error bounds of empirical path loss models at VHF/UHF Bands in Kwara State, L. Karadzinov, T. Capuder,
K. Mokhtar (Eds.), Eurocon, Nigeria, 2013, doi:10.1109/EUROCON.2013.6625043.
[21] N. Faruk, Y.A. Adediran, A.A Ayeni, Optimization of Davidson model based On RF measurement conducted In UHF/VHF Bands, in: Proceedings of the
6th IEEE Conference on Information Technology, Jordan, IEEE ICIT, 2013 Available on.
[22] O.W. Faruk Bello, A.A. Oloyede, N.T. Surajudeen Bakinde, O. Obiyemi, L.A. Olawoyin, M. Ali, A.N. Jimoh, Clutter and terrain effects on path loss in the
VHF/UHF bands, IET Microw. Antennas Propag. 12 (1) (2018) 69–76 The Institution of Engineering and Technology, doi:10.1049/iet-map.2016.0809.
[23] S.I. Faruk Popoola, N.T. Surajudeen-Bakinde, A.A. Oloyede, A. Abdulkarim, L.A. Olawoyin, M. Ali, C.T. Calafate, A.A. Atayero, N, Path loss predictions in
the VHF and UHF bands within urban environments: experimental investigation of empirical, heuristics and geospatial models, IEEE Access 7 (2019),
doi:10.1109/ACCESS.2019.2921411.
[24] I.Y. Faruk Abdulrasheed, N.T. Surajudeen-Bakinde, E. Adetiba, A.A. Oloyede, A. Abdulkarim, O. Sowande, A.H. Ifijeh, A.A. Atayero, N, Large-scale radio
propagation path loss measurements and predictions in the VHF and UHF bands, Heliyon 7 (6) (2021) E07298, doi:10.1016/j.heliyon.2021.e07298.
[25] J. Fu, P. Juyal, E.J. Jorgensen, A. Zajić, Comparison of Statistical and Deep Learning Path Loss Model for Motherboard Desktop Environment, in: Pro-
ceedings of the 2022 16th European Conference on Antennas and Propagation (EuCAP), IEEE, 2022, pp. 1–5.
[26] Y. Egi, C.E. Otero, Machine-learning and 3D point-cloud based signal power path loss model for the deployment of wireless communication systems,
IEEE Access 7 (2019) 42507–42517.
[27] J.O. Eichie, O.D. Oyedum, M. Ajewole, A.M. Aibinu, Comparative analysis of basic models and artificial neural network based model for path loss
prediction, Prog. Electromagn. Res. 61 (2017) 133–146.
[28] K. Fukushima, A neural network model for selective attention in visual pattern recognition, Biol. Cybern. 55 (1) (1986) 5–15.
[29] K. Fukushima, S. Miyake, T. Ito, Neocognitron: a neural network model for a mechanism of visual pattern recognition, IEEE Trans. Syst. Man Cybern.
(5) (1983) 826–834.
[30] J.F. Gainsborough, Slow growth and urban sprawl: support for a new regional agenda? Urban Aff. Rev. 37 (5) (2002) 728–744.
[31] M. Garah, H. Oudira, L. Djouane, N. Hamdiken, Particle swarm optimization for the path loss reduction in suburban and rural area, Int. J. Electr.
Comput. Eng. 7 (4) (2017) 2125.
[32] Garg, V. (2007). Radio propagation and propagation path-loss models. Wireless Communications & Networking, 47–84.
[33] A. Ghosh, A. Sufian, F. Sultana, A. Chakrabarti, D. De, Fundamental concepts of convolutional neural network, in: Recent Trends and Advances in
Artificial Intelligence and Internet of Things, Springer, 2020, pp. 519–567.
[34] A. Gupta, K. Ghanshala, R. Joshi, Path loss predictions for fringe areas using adaptive neuro-fuzzy inference system, Int. J. Syst. Assur. Eng. Manag.
(2021) 1–14.
[35] V. Gupta, S. Sharma, Secure path loss prediction in fringe areas using fuzzy logic approach, in: Proceedings of the Paper Presented at the 2009
International Conference on Advances in Computing, Control, and Telecommunication Technologies, 2009.
[36] Halperin, D., Hu, W., Sheth, A., & Wetherall, D. (2010). Predictable 802.11 packet delivery from wireless channel measurements. ACM SIGCOMM
Computer Communication Review, 40(4), 159–170.
[37] M.H. Hassoun, Fundamentals of Artificial Neural Networks, MIT press, 1995.
[38] H.K. H Hoomod, I. Al-Mejibli, A.I. Jabboory, Analyzing study of path loss propagation models in wireless communications at 0.8GHz, J. Phys. Conf.
Ser. (2018) 012028 1003, doi:10.1088/1742-6596/1003/1/012028.
[39] D.H. Hubel, T.N. Wiesel, Receptive fields and functional architecture of monkey striate cortex, J. Physiol. 195 (1) (1968) 215–243.
[40] M.J. Jafri, J.J. Ely, L.L. Vahala, Interference path loss prediction in A319/320 Airplanes using modulated fuzzy logic and neural networks, in: Proceedings
of the Paper Presented at the NAFIPS 20 07-20 07 Annual Meeting of the North American Fuzzy Information Processing Society, 2007.
[41] J.S.R. Jang, C.T. Sun, E. Mizutani, Neuro-fuzzy and soft computing-a computational approach to learning and machine intelligence [Book Review], IEEE
Trans. Autom. Control 42 (10) (1997) 1482–1484.
[42] H.M. Jawad, A.M. Jawad, R. Nordin, S.K. Gharghan, N.F. Abdullah, M. Ismail, M.J. Abu-AlShaeer, Accurate empirical path-loss model based on particle
swarm optimization for wireless sensor networks in smart agriculture, IEEE Sens. J. 20 (1) (2019) 552–561.
[43] H.-.S. Jo, C. Park, E. Lee, H.K. Choi, J. Park, Path loss prediction based on machine learning techniques: principal component analysis, artificial neural
network, and Gaussian process, Sensors 20 (7) (2020) 1927.
[44] R.T. Juang, Explainable deep-learning-based path loss prediction from path profiles in urban environments, Appl. Sci. 11 (15) (2021) 6690.
[45] R.T. Juang, Deep learning-based path loss model in urban environments using image-to-image translation, IEEE Trans. Antennas Propag. (2022).
[46] M. Kalakh, N. Kandil, N. Hakem, Neural networks model of an UWB channel path loss in a mine environment, in: Proceedings of the Paper Presented
at the 2012 IEEE 75th Vehicular Technology Conference (VTC Spring), 2012.
[47] D. Karaboga, E. Kaya, Adaptive network based fuzzy inference system (ANFIS) training approaches: a comprehensive survey, Artif. Intell. Rev. 52 (4)
(2019) 2263–2293.
[48] F.A. Khan, A. Ali, H. Abbas, N.A.H. Haldar, A cloud-based healthcare framework for security and patients’ data privacy using wireless body area
networks, Procedia Comput. Sci. 34 (2014) 511–517.
[49] N. Kuno, M. Inomata, M. Sasaki, W. Yamada, Deep learning-based path loss prediction using side-view images in an UMa environment, in: Proceedings
of the 2022 16th European Conference on Antennas and Propagation (EuCAP), IEEE, 2022, pp. 1–5.
[50] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proc. IEEE 86 (11) (1998) 2278–2324.
[51] B.h. Lee, D. Ham, J. Choi, S.-.C. Kim, Y.-.H. Kim, Genetic algorithm for path loss model selection in signal strength-based indoor localization, IEEE Sens.
J. 21 (21) (2021) 24285–24296.
[52] J.-.Y. Lee, M.Y. Kang, S.C. Kim, Path loss exponent prediction for outdoor millimeter wave channels through deep learning, in: Proceedings of the
Paper Presented at the 2019 IEEE Wireless Communications and Networking Conference (WCNC), 2019.
[53] R. Levie, Ç. Yapar, G. Kutyniok, G. Caire, Pathloss prediction using deep learning with applications to cellular optimization and efficient D2D link
scheduling, in: Proceedings of the Paper Presented at the ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2020.
[54] L. Liu, Y. Yao, Z. Cao, M. Zhang, Deeplora: learning accurate path loss model for long distance links in lpwan, in: Proceedings of the Paper Presented
at the Proceedings of IEEE INFOCOM, 2021.
[55] S. Ma, H. Cheng, H. Lee, A practical approach for indoor path loss modeling based on deep learning, J. Comput. Sci. Eng. 15 (2) (2021) 84–95.

34
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

[56] S. Mathew, K. Shylaja, T. Jayasri, M. Hemalatha, Path loss prediction in wireless communication system using fuzzy logic, Indian J. Sci. Technol. 7 (5)
(2014) 642–647.
[57] N. Moraitis, L. Tsipi, D. Vouyioukas, A. Gkioni, S. Louvros, Performance evaluation of machine learning methods for path loss prediction in rural
environment at 3.7GHz, Wirel. Netw. 27 (6) (2021) 4169–4188.
[58] S. Nafea, E.K. Hamza, Path loss optimization in WIMAX network using genetic algorithm, Int. J. Comput. 20 (1) (2020) 24–30 Communications &
Control (IJCCC).
[59] K. Nobuaki, W. YAMADA, M. SASAKI, Y. TAKATORI, Convolutional neural network for prediction method of path loss characteristics considering diffrac-
tion and reflection in an open-square environment, in: Proceedings of the Paper Presented at the 2019 URSI Asia-Pacific Radio Science Conference
(AP-RASC), 2019.
[60] A. Obot, O. Simeon, J. Afolayan, Comparative analysis of path loss prediction models for urban macrocellular environments, Niger. J. Technol. 30 (3)
(2011) 50–59.
[61] S. Ojo, A. Imoize, D. Alienyi, Radial Basis Function Neural Network path loss prediction model for LTE networks in multitransmitter signal propagation
environments, Int. J. Commun. Syst. 34 (3) (2021) e4680.
[62] E. Ostlin, H.-.J. Zepernick, H. Suzuki, Macrocell path-loss prediction using artificial neural networks, IEEE Trans. Veh. Technol. 59 (6) (2010) 2735–2747.
[63] C. Park, D.K. Tettey, H.-S. Jo, 2019. arXiv preprint arXiv: 1904.02383.
[64] T. Pateman, Rural and urban areas: comparing lives using rural/urban classifications, Reg. Trends 43 (1) (2011) 11–86.
[65] S. Phaiboon, P. Phokharatkul, An empirical based path loss model with tree density effects for 1.8GHz mobile communications using fuzzy regression,
WSEAS Trans. Commun. 5 (3) (2006) 564–571.
[66] S. Phaiboon, P. Phokharatkul, S. Somkuarnpanit, 2 to 16GHz microwave line-of-sight path loss prediction on urban streets by fuzzy logic models, in:
Proceedings of the Paper Presented at the TENCON 20 05-20 05 IEEE Region 10 Conference, 2005.
[67] S. Phaiboon, P. Phokharatkul, S. Somkuarnpanit, Muti-layer fuzzy logic sets for mobile path loss in forests, in: Proceedings of the Paper Presented at
the TENCON 20 06-20 06 IEEE Region 10 Conference, 2006.
[68] S. Phaiboon, P. Phokharatkul, S. Somkuarnpanit, S. Boonpiyathud, Microwave line-of-sight path loss prediction on urban street by fuzzy logic model,
in: Proceedings of the Paper Presented at the 2005 Asia-Pacific Microwave Conference, 2005.
[69] S. Phaiboon, P. Phokharutkul, S. Somkurnpanich, Mobile propagation path loss models for suburban areas using fuzzy logic approximation, in: Pro-
ceedings of the Paper Presented at the 2004 IEEE Region 10 Conference TENCON 2004, 2004.
[70] C. Phillips, D. Sicker, D. Grunwald, A survey of wireless path loss prediction and coverage mapping methods, IEEE Commun. Surv. Tutor. 15 (1) (2012)
255–270.
[71] M. Piacentini, F. Rinaldi, Path loss prediction in urban environment using learning machines and dimensionality reduction techniques, Comput.
Manag. Sci. 8 (4) (2011) 371–385.
[72] I. Popescu, I. Nafornita, P. Constantinou, Comparison of neural network models for path loss prediction, in: Proceedings of the Paper Presented at the
WiMob’2005), IEEE International Conference on Wireless And Mobile Computing, Networking And Communications, 2005 2005.
[73] I. Popescu, D. Nikitopoulos, P. Constantinou, I. Nafornita, Comparison of ANN based models for path loss prediction in indoor environment, in:
Proceedings of the Paper Presented at the IEEE Vehicular Technology Conference, 2006.
[74] S. Popoola, N. Faruk, A. Oloyede, A. Atayero, N. Surajudeen-Bakinde, L. Olawoyin, Characterization of path loss in the VHF band using neural network
modeling technique, in: Proceedings of the Paper Presented at the 2019 19th International Conference on Computational Science and Its Applications
(ICCSA), 2019.
[75] S.I. Popoola, A. Jefia, A.A. Atayero, O. Kingsley, N. Faruk, O.F. Oseni, R.O. Abolade, Determination of neural network parameters for path loss prediction
in very high frequency wireless channel, IEEE Access 7 (2019) 150462–150483.
[76] S.I. Popoola, S. Misra, A.A. Atayero, Outdoor path loss predictions based on extreme learning machine, Wirel. Pers. Commun. 99 (1) (2018) 441–460.
[77] K. Qiu, S. Bakirtzis, H. Song, J. Zhang, I. Wassell, Pseudo ray-tracing: deep leaning assisted outdoor mm-wave path loss prediction, IEEE Wirel. Com-
mun. Lett. (2022).
[78] V.M. Radonjić, N.N. Gaćeša, Effect of environment on the propagation of electromagnetic waves in GRC 408E digital radiorelay devices, Vojnotehn.
Glas. 59 (1) (2011) 40–61.
[79] Rydell, J.B., Otterlind, O., & Sjöö, A. (2021). A review of wireless positioning from past and current to emerging technologies. Decision Support Systems
and Industrial IoT in Smart Grid, Factories, and Cities, 39–61.
[80] M.A. Saeed, M.U. Saeed, M.A.S. Hassan, T. Javed, Impact of propagation path loss by varying BTS height and frequency for combining multiple path
loss approaches in macro-femto environment, Arab. J. Sci. Eng. 47 (2) (2022) 1227–1238.
[81] M.A. Salman, S.I. Popoola, N. Faruk, N. Surajudeen-Bakinde, A.A. Oloyede, L.A. Olawoyin, Adaptive Neuro-Fuzzy model for path loss prediction in the
VHF band, in: Proceedings of the Paper Presented at the 2017 International Conference on Computing Networking and Informatics (ICCNI), 2017.
[82] U.S. Sani, D.T.C. Lai, O.A. Malik, A hybrid combination of a convolutional neural network with a regression model for path loss prediction using tiles
of 2D satellite images, in: Proceedings of the Paper Presented at the 2020 8th International Conference on Intelligent and Advanced Systems (ICIAS),
2021.
[83] U.S. Sani, O.A. Malik, D.T.C. Lai, Dynamic regressor/ensemble selection for a multi-frequency and multi-environment path loss prediction, Information
13 (11) (2022) 519.
[84] M. Sasaki, N. Kuno, T. Nakahira, M. Inomata, W. Yamada, T. Moriyama, Deep learning based channel prediction at 2–26GHz band using long short-term
memory network, in: Proceedings of the Paper Presented at the 2021 15th European Conference on Antennas and Propagation (EuCAP), 2021.
[85] P.K. Sharma, R. Singh, Comparative analysis of propagation path loss models with field measured data, Int. J. Eng. Sci. Technol. 2 (6) (2010) 2008–2013.
[86] A. Seretis, C.D. Sarris, An overview of machine learning techniques for radiowave propagation modeling, IEEE Trans. Antennas Propag. (2021).
[87] O.O. Shoewu, M.A. Adedoyin, L.A. Akinyemi, L.I. Oborkhale, Fuzzy-logic based path loss models for Metropolitan environment, in: Proceedings of the
Paper Presented at the 2018 IEEE International Workshop on Signal Processing Systems (SiPS), 2018.
[88] H. Singh, S. Gupta, C. Dhawan, A. Mishra, Path loss prediction in smart campus environment: machine learning-based approaches, in: Proceedings of
the Paper Presented at the 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring), 2020.
[89] S.P. Sotiroudis, S.K. Goudos, K.A. Gotsis, K. Siakavara, J.N. Sahalos, Application of a composite differential evolution algorithm in optimal neural net-
work design for propagation path-loss prediction in mobile communication systems, IEEE Antennas Wirel. Propag. Lett. 12 (2013) 364–367.
[90] S.P. Sotiroudis, S.K. Goudos, K. Siakavara, Neural networks and random forests: a comparison regarding prediction of propagation path loss for NB-IoT
networks, in: Proceedings of the Paper Presented at the 2019 8th International Conference on Modern Circuits and Systems Technologies (MOCAST),
2019.
[91] S.P. Sotiroudis, S.K. Goudos, K. Siakavara, Deep learning for radio propagation: using image-driven regression to estimate path loss in urban areas, ICT
Express 6 (3) (2020) 160–165.
[92] S.P. Sotiroudis, P. Sarigiannidis, S.K. Goudos, K. Siakavara, Fusing diverse input modalities for path loss prediction: a deep learning approach, IEEE
Access 9 (2021) 30441–30451.
[93] S.P. Sotiroudis, K. Siakavara, G.P. Koudouridis, P. Sarigiannidis, S.K. Goudos, Enhancing Machine Learning Models for Path Loss Prediction Using Image
Texture Techniques, IEEE Antennas Wirel. Propag. Lett. 20 (8) (2021) 1443–1447.
[94] S.P. Sotiroudis, G. Athanasiadou, G.V. Tsoulos, C. Christodoulou, S.K. Goudos, Ensemble learning for 5G flying base station path loss modelling, in:
Proceedings of the 2022 16th European Conference on Antennas and Propagation (EuCAP), IEEE, 2022, pp. 1–4.
[95] S.P. Sotiroudis, S.K. Goudos, C. Christodoulou, A hybrid CNN-NGBoost model for probabilistic image-driven path loss prediction, in: Proceedings of the
2022 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting (AP-S/URSI), IEEE, 2022, pp. 898–899.

35
H. Chiroma, P. Nickolas, N. Faruk et al. Scientific African 19 (2023) e01550

[96] H. Steendam, M. Moeneclaey, Analysis and optimization of the performance of OFDM on frequency-selective time-selective fading channels, IEEE
Trans. Commun. 47 (12) (1999) 1811–1819.
[97] N.T. Surajudeen-Bakinde, N. Faruk, S.I. Popoola, M.A. Salman, A.A. Oloyede, L.A. Olawoyin, C.T. Calafate, Path loss predictions for multi-transmitter
radio propagation in VHF bands using Adaptive Neuro-Fuzzy Inference System, Eng. Sci. Technol. Int. J. 21 (4) (2018) 679–691.
[98] A. Tahat, T. Edwan, H. Al-Sawwaf, J. Al-Baw, M. Amayreh, Simplistic machine learning-based air-to-ground path loss modeling in an urban environ-
ment, in: Proceedings of the Paper Presented at the 2020 Fifth International Conference on Fog and Mobile Edge Computing (FMEC), 2020.
[99] J. Thrane, D. Zibar, H.L. Christiansen, Model-aided deep learning method for path loss prediction in mobile communication systems at 2.6GHz, IEEE
Access 8 (2020) 7925–7936.
[100] T.C. Tu, C.C. Chiu, Path loss reduction in an urban area by genetic algorithms, J. Electromagn. Waves Appl 20 (3) (2006) 319–330.
[101] P. Wang, H. Lee, Indoor path loss modeling for 5G Communications in smart factory scenarios based on meta-learning, in: Proceedings of the Paper
Presented at the 2021 Twelfth International Conference on Ubiquitous and Future Networks (ICUFN), 2021.
[102] Y. Wang, M.Y. Liang, J. Hu, T. Song, A new method for radio wave propagation prediction based on BP-neural network and path loss model, in:
Proceedings of the Paper Presented at the 2020 12th International Conference on Knowledge and Smart Technology (KST), 2020.
[103] J. Wen, Y. Zhang, G. Yang, Z. He, W. Zhang, Path loss prediction based on machine learning methods for aircraft cabin environments, IEEE Access 7
(2019) 159251–159261.
[104] L. Wu, D. He, B. Ai, J. Wang, H. Qi, K. Guan, Z. Zhong, Artificial neural network based path loss prediction for wireless communication network, IEEE
Access 8 (2020) 199523–199538.
[105] G. Yang, Y. Zhang, Z. He, J. Wen, Z. Ji, Y. Li, Machine-learning-based prediction methods for path loss and delay spread in air-to-ground millime-
tre-wave channels, IET Microw. Antennas Propag. 13 (8) (2019) 1113–1121.
[106] N. Zaarour, S. Affes, N. Kandil, N. Hakem, Comparative study on a 60GHz path loss channel modeling in a mine environment using neural networks,
in: Proceedings of the Paper Presented at the 2015 IEEE International Conference on Ubiquitous Wireless Broadband (ICUWB), 2015.
[107] Y. Zhang, J. Wen, G. Yang, Z. He, J. Wang, Path loss prediction based on machine learning: principle, method, and data expansion, Appl. Sci. 9 (9)
(2019) 1908.
[108] G.B. Afolayan, N. Faruk, O.W. Bello, A.A. Ayeni, M.H. Ali, M.O. Olufemi, Error Analysis for Empirical Path-Loss Models In GSM and WCDMA Bands
in the Urban Area, Journal of Science, Technology and Mathematics Education (JOSTMED) 11 (2) (2015) 29–35 Published by Federal University of
Technology, Minna.
[109] A.A. Oloyede, N Faruk, O.W. Bello, Variation of clutter height and its impact on path loss in the VHF/UHF band, Advances in Wireless and Optical
Communications (RTUWO) (2016) 129–132 Riga, 2016, doi:10.1109/RTUWO.2016.7821870.
[110] A.A. Jimoh, N.T. Surajudeen-Bakinde, N. Faruk, O.W. Bello, A.A. Ayeni, Clutter Height Variation Effects on Frequency Dependent Path Loss Models at
UHF Band in Build-Up Areas, Science, Technology & Arts Research (STAR) Journal 4 (4) (2016) 138–147 Published by Wollega University, Ethiopia.
Available online https://www.ajol.info/index.php/star/article/view/145846.
[111] E. Greenberg, E. Klodzh, Comparison of deterministic, empirical and physical propagation models in urban environments, in: 2015 IEEE International
Conference on Microwaves, Communications, Antennas and Electronic Systems (COMCAS), IEEE, 2015, pp. 1–5.
[112] S.I. Popoola, A.A. Atayero, N. Faruk, C.T. Calafate, L.A. Olawoyin, V.O. Matthews, et al., Standard Propagation Model Tuning for Path Loss Predictions
in Built-Up Environments, in: O. Gervasi, et al. (Eds.), Computational Science and Its Applications – ICCSA 2017, 10409, Lecture Notes in Computer
Science, 2017, pp. 363–375.
[113] S.I. Popoola, N. Faruk, N.T. Surajudeen-Bakinde, A.A. Atayero, S. Misra, Artificial Neural Network Model for Path Loss Predictions in the VHF Band, in:
Conference Proceedings of ICDLAIR2019, Springer International Publishing, 2021, pp. 161–169.
[114] N. Surajudeen-Bakinde, N. Faruk, A. Oloyede, A. Abdulkarim, L. Olawoyin, S. Popoola, E. Adetiba, Effect of membership functions and data size on the
performance of ANFIS-based model for predicting path losses in the VHF and UHF bands, J. Eng. Res. (2021) 1–37, doi:10.36909/jer.10457.
[115] Q. Ramon Adebowale, N. Faruk, K.S. Adewole, A. Abdulkarim, L.A. Olawoyin, A.A. Oloyede, H. Chiroma, A.D. Usman, C.T. Calafate, Application of
computational intelligence algorithms in radio propagation: A systematic review and metadata analysis, Mob. Inf. Syst. (2021) (2021) 20 Article ID
6619364, doi:10.1155/2021/6619364.
[116] A.A. Budalal, I. Shayea, M.R. Islam, M.H. Azmi, H. Mohamad, S.A. Saad, Y.I Daradkeh, Millimetre-wave propagation channel based on NYUSIM channel
model with consideration of rain fade in tropical climates, IEEE Access 10 (2021) 1990–2005.

36