SINMCEL- Connectivity Solution for Remote Environments. Case Study: Combu Island

Abstract

Modeling is the safest and most cost-effective way to test new and complex technologies. A model represents a system created by combining concepts that help to understand, simulate, or predict the behavior of the object or process being represented. Signal propagation modeling tools have become increasingly relevant as analytical resources for detailing the scenario of interest and predicting outcomes. This study uses a supervised machine learning technique to propose a simulation and analysis method for transmission systems in atypical and remote environments. Its main goal is to suggest choices that maximize efficiency regarding the propagation models used in the scenarios under analysis. For validation purposes, a case study was conducted in the Amazon region, specifically on Combú Island, in the state of Pará, Brazil, where aspects related to the system's reference parameters were analyzed using the proposed algorithm named SINMCEL V.2.

Index Terms
Transmission System; SINMCEL; Supervised Machine Learning; Telecommunications.

I. INTRODUCTION

Due to its geographic and environmental specificities, the Brazilian Amazon has always attracted international attention. This became even more evident after Brazil was chosen as the host country for the 30th United Nations (UN) Conference on Climate Change (COP30), scheduled for November 2025 in Belém, Pará. This global event represents an excellent opportunity to debate ideas and strategies that support feasible actions for the region’s technological development [¹]. The event will require significant investments from the telecommunications sector and increased engineering efforts to provide, guarantee, and ensure fast and secure connectivity to users at any time and any place-especially in remote environments without internet access, which will serve as venues for important activities during the event [¹].

This work aims to develop a radio propagation prediction tool that efficiently assists in optimizing the choice of propagation models, focusing primarily on atypical and remote environments.

Most closed-source prediction software presents its results through simulations of propagation models [²]. This approach is generally effective, but it is clear that the selection of the propagation model in the simulation is determined solely by the user, without suggestions from the tool itself. In other words, there is a direct relationship between the results obtained from the link and the accurate choice of the propagation model.

This work contributes by presenting a method implemented in the SINMCEL tool (Intelligent Leveling Tool that Maps and Controls Data between Links), which provides a differentiated approach to analyzing transmission systems [³]. The method uses artificial intelligence (AI), specifically supervised machine learning, to establish optimized topographic and morphological mapping criteria in atypical environments. Based on this information, the software suggests the most appropriate propagation model to the user, considering the specific characteristics of the environment where the link is located.

To validate the proposed model, a case study was conducted in an atypical and remote location in Belém, Pará. The selected area was Combú Island, which represents a complex environment that combines, in the same setting, wild and well-preserved nature alongside a sophisticated tourist environment attracting high user traffic. This location is expected to be one of the most visited during the COP30 event. An important aspect is that despite the island’s high level of economic activity, quality connectivity is still lacking.

The case study involved simulations and field measurements in the SHF spectrum. This frequency range was chosen because of the technical characteristics of the equipment used in the link.

This article is organized as follows: Section II presents the propagation models used in the study. Section III briefly discusses the application of AI methods through machine learning. Section IV presents the parameters used in the study. Section V presents and discusses the results obtained. Finally, Section VI offers the conclusions.

II. PROPAGATION MODELS

The main objective of radio transmission is to deliver to the receiver a signal that is as faithful as possible to the transmitted one. The signal is degraded during its journey due to several factors, such as free space attenuation, absorption and obstruction by obstacles, reflection, scattering, and absorption by atmospheric gases, among others. The operation of a transmission system follows criteria that ensure the reliability and integrity of the received signal, considering the end-to-end link process [⁴].

To ensure integrity, reliability, and robustness throughout this process, efficient signal modeling techniques must be employed to accurately represent the reality of propagation in the communication channel. In pursuit of this efficiency, several techniques have been developed and refined, including empirical, deterministic, statistical, and hybrid approaches [⁵].

In this study, five propagation models were selected-three empirical and two theoretical-each employing mathematical analysis methods with varying levels of complexity. These models were chosen to validate the proposed method based on their specific applicability to the objectives of this study: Longley-Rice, Stanford University Interim (SUI), ITU-R Recommendation P.833-7 (Attenuation due to Vegetation), Adjusted Ray-Tracing, and Ericsson.

• Longley-Rice (RC) model: This is a more accurate propagation model in terms of the medium's characteristics. It estimates average signal loss over irregular terrain using a mathematical method that accounts for both free space attenuation and various propagation mechanisms. The path attenuation reference, in dB, is defined by the sum of the free space loss and additional losses caused by propagation mechanisms, as shown in Table I [⁶].

TABLE I
ATTENUATION REGIONS OF THE LONGLEY-RICE MODEL

Region	Description
Unobstructed view - ALOS	Calculated as a direct function of the distance between stations.
Diffraction- ADIF	Function that characterizes the region of very rapid increase in attenuation.
Spreading- $A_{ES}$	Function that characterizes the effect of dispersive phenomena of the electromagnetic signal

Resulting:

(1) $$\begin{array}{c}{A}_{RC}=\left\{\begin{array}{c}{A}_{EL}+A_{LOS}+\left(Y_L+Y_S+Y_T\right),ford<d_{lb}\\ A_{EL}+A_{DIF}+\left(Y_L+Y_S+Y_T\right),for{d}_{lb}\le d\le d_{lm}\\ A_{EL}+A_{ES}+\left(Y_L+Y_S+Y_T\right),ford>d_{lm}\end{array}\right.\end{array}$$

Where: Y_L represents the percentage of time the link does not exceed a specific attenuation value; Y_Srepresents the percentage referring to the location of the paths where degradation occurs; Y_T represents the percentage of areas in similar regions that face loss relative to the propagating signal; d is the distance between the transmitter and receiver in [km]; d_lb is the line of sight of the transmitting antenna radio in [km] and d_lm represents the radio line of sight of the receiving antenna in [km].

Attenuation coefficients (A_LOS), (A_DIF) e (A_ES) are determined from equations (2), (3), (4), (5), and (6).

(2) $$A_{LOS}=20log\left(1+\frac{\left(d_{lb}+d_{lm}\right)\Delta h}{h_{eb}{h}_{em}}\right)$$

(3) $$\begin{array}{c}{A}_{DIF}=\left(1+0.045{({\Delta }_h/\lambda )}^{\frac{1}{2}}{(\alpha {\theta }_e+\left(d_{lb}+d_{lm}\right)/d)}^{\frac{1}{2}}\right)\end{array}$$

(4) $$\begin{array}{c}{A}_{ES}=H_0+10log\left(f{\theta }_e^4\right)-0.1\left(N_s-301\right)e^{\left(\frac{-\theta e^d}{40}\right)}\end{array}$$

(5) $$\begin{array}{c}\alpha =\frac{6370}{\left(1-0.04665e^{0,005577N_s}\right)}\end{array}$$

(6) $$\begin{array}{c}{H}_0=\frac{\left(h_{eb}^{-1}\times h_{em}^{-1}\right)}{{\theta }_e\times f\times 0.007-0.058{\theta }_e}\end{array}$$

Where: f is the operating frequency in [MHz]; λ represents the wavelength in meters [m]; h_eb and h_em are the heights of the transmitting and receiving antennas respectively [m]; θ_e is the angle of the crossing of the horizons of the link antennas [°]; ∆_h is the irregularity of the terrain [m] and N_s represents the surface refractivity index.

When a detailed terrain profile is available, the parameters for the propagation path can be defined more effectively. Typical values of terrain irregularity (∆h) for various types of terrain are shown in Table II.

TABLE II
ESTIMATED VALUES FOR LAND IRREGULARITY

Land Type	∆h
Water or very smooth plains	0 a 50
Plains	30
Hill	80 the 150
Mountains	150 the 300
Mountain ranges	300 the 700

• Sui Model (SU) ˗ This model applies to three types of environments: regions with rugged terrain and dense vegetation, flat landscapes with sparse vegetation, and intermediate scenarios [⁷]. The propagation loss in the Sui model is given by (7), (8), (9), (10), (11), and (12).

(7) $$\begin{array}{c}{A}_{SU}=A+10\gamma lo{g}_{10}\frac{d}{d_0}+\Delta L_f+\Delta L_h+Sparad>d_0\end{array}$$

(8) $$\begin{array}{c}A=20lo{g}_{10}\left(\frac{4\pi d_0}{\lambda }\right)\end{array}$$

(9) $$\begin{array}{c}\gamma =\alpha -b{h}_{em}+\left(c/h_{eb}\right)\end{array}$$

(10) $$\begin{array}{c}\Delta L_f=6lo{g}_{10}\left(f/2000\right)fortypeAandBforenvironments\end{array}$$

(11) $$\begin{array}{c}\Delta L_h=-10,8lo{g}_{10}\left(h_{em}/2\right)fortypeCenvironments\end{array}$$

(12) $$S=0,65{(\log \left(f\right))}^2-1,3+\log \left(f\right)+\alpha$$

Where: f is the operating frequency in [MHz]; d represents the total distance of the link [km]; α propagation loss [dBm]; b represents the types of soils; h_b effective height of transmitting antenna [m]; c represents the types of vegetation; d₀ represents the reference distance (100 meters); ΔL_f is the frequency correction factor [MHz]; ΔL_h represents the antenna height correction factor in the receiver [m] and h_eb and h_em are the heights of the antennas in [m].

• Ericsson Model 9999 (ERI) ˗ Model initially developed to operate in the frequency range from 150 MHz to 2 GHz. However, its use can be extended to higher ranges through adjustments in certain parameters [⁸] and [⁹]. The signal attenuation is determined by (13) and (14).

(13) $$A_{ERI}=a_0+a_{1}lo{g}_{10}\left(d\right)+a_{2}lo{g}_{10}\left(h_{em}\right)+a_{3}lo{g}_{10}\left(h_{eb}\right)lo{g}_{10}\left(d\right)-3,2{(lo{g}_{10}11,75h_{em})}^2$$

(14) $$\begin{array}{c}G\left(f\right)=44,9lo{g}_{10}\left(f\right)-4,78lo{g}_{10}\left(f\right)\end{array}$$

Where a₀, a₁, a₂ and a₃ are the constants that can be modified according to the scenario under analysis. The default values of these constants are shown in Table III.

TABLE III
STANDARD VALUES ACCORDING TO THE STUDY SCENARIO

Environment	a0	a1	a2	a3
Urban	36.2	30.2	12	0.1
Suburban	43.20	68.93	12	0.1
Rural	45.95	100.6	12	0.1

• Recommendation ITU-R P.833-7 ˗ It shows, through different mathematical methods, the effects of vegetation on signal attenuation. These models are used for specific frequency bands and vegetation geometries, [¹⁰]. This study will use the model that considers the terrestrial path with a terminal in the forest, which is determined by (15).

(15) $$\begin{array}{c}{A}_{ev}=A_m\left[1-exp\left(-d\frac{\gamma }{A_m}\right)\right]\end{array}$$

Where: d path distance within woodland [m]; γ specific attenuation for very short vegetative paths [dB/m]; A_m maximum attenuation for one terminal within a specific type and depth of vegetation (dB)

• Ray-tracing ˗ Mathematical model based on the principle of geometric optics, describing signal propagation behavior by tracking its paths (reflected or diffracted signals). Given that the study scenario will be in remote areas characterized by dense vegetation, rivers, humidity, and temperature typical of the Amazon region, it was decided to use the Adjusted Ray-Tracing model developed by [¹¹], explicitly designed for such conditions. This model is recommended for use in mixed and atypical environments. The propagation loss for the Adjusted Ray Tracing model is calculated as follows:

(16) $$\begin{array}{c}{P}_L=P_t-{\left[E\left(d\right)\right]}^2\end{array}$$

Where: P_t transmitted power [dBm]; E(d)² represents the square of the amplitude of the electric field and is calculated by equation (17).

(17) $$\begin{array}{c}E\left(d\right)=E_{0m}{d}_{0m}\left(\left(\frac{exp\left(-j\frac{2\pi }{\lambda }{d}_1\right)}{d_1}+\left(\Gamma \frac{exp\left(-j\frac{2\pi }{\lambda }{d}_2\right)}{d_2}\right)\right)\right)\end{array}$$

Where: E₀ represents the electric field adjusted in free space; d₀ reference distance, used to normalize the amplitude of the electric field [m]; λ represents the wavelength of the electromagnetic wave; d₁ represents the distance traveled by the first ray, with direct sight; d₂ represents the distance traveled by the second ray, derived from reflections or diffractions; Γ represents the reflection coefficient for mixed environments calculated by the equations (18) and (19).

(18) $$\begin{array}{c}\Gamma =\frac{\left({\epsilon }_r-j\left(\frac{{18.10}^9\sigma }{f}\right)\right)sin\theta -\sqrt{\left({\epsilon }_r-j\left(\frac{{18.10}^9\sigma }{f}\right)-co{s}^2\theta \right)}}{\left({\epsilon }_r-j\left(\frac{{18.10}^9\sigma }{f}\right)\right)sin\theta +\sqrt{\left({\epsilon }_r-j\left(\frac{{18.10}^9\sigma }{f}\right)-co{s}^2\theta \right)}}\end{array}$$

(19) $$\begin{array}{c}\theta =ta{n}^{-1}\left(\frac{h_t{h}_r}{d}\right)\end{array}$$

Where: f is the operating frequency; θ represents the angle of incidence of the wave; h_t represents the height of the transmitting antenna; h_r represents the height of the receiving antenna; ɛ_r represents the relative permittivity of the material; σ represents the electrical conductivity of the material.

The adjustments made to the Ray-tracing model refer to the propagation components of the electric field, permittivity, and conductivity. In this work, the components that reflect on dry land (ɛr = 15 and σ = 0.000.1) and the components reflected in the river (ɛr = 80 and σ = 0.5) will be used.

III. MACHINE LEARNING DEFINITIONS

Recent advances in Artificial Intelligence (AI) are evident in the significant ways in which the behavior of various sectors of society has been transformed and innovated across multiple areas of human activity. This technology has become an integral part of everyday life, offering advanced solutions by developing systems that emulate human thought processes. As a result, machines are capable of effectively operating intelligent algorithms [¹²].

Given the broad scope of AI, this paper focuses specifically on the field of machine learning-a fundamental method in AI development. Machine learning is based on the principle that systems can learn and improve performance through experience. These algorithms are designed to process large datasets, identify underlying patterns, and make accurate decisions with minimal human intervention, showcasing the adaptive and autonomous capabilities of modern computing systems.

Machine learning methods are typically classified into three categories [¹³]:

• • Supervised learning - This method trains the algorithm by mapping input data to corresponding output data. The goal is to teach the system to predict and associate both existing and new inputs with their correct outputs. An example of its use is in customer service voice recognition systems employed by mobile network operators.

• • Unsupervised learning - In this approach, only input data is provided. The algorithm must identify the underlying structure in the input to infer appropriate outputs. An application example is identifying customer mobility patterns by analyzing usage across cell towers.

• • Reinforcement learning - This method is based on interaction with the environment. The algorithm learns by taking actions that result in rewards or penalties, depending on the outcomes. It is commonly used to optimize network parameters, such as reducing interference and increasing coverage in mobile communications.

Table IV presents examples of commonly used algorithms, categorized by learning type.

TABLE IV
MACHINE LEARNING ALGORITHMS AND LEARNING TYPES

ALGORITHMS	TYPES OF LEARNING	CLASS	APPLICATION
Artificial Neural Networks	Supervised Learning	Predict which category a new observation belongs to	Binary Inputs
K-means	Unsupervised Learning	Group data into clusters with similar characteristics	Clustering
Decision Trees	Supervised Learning	Predict which category a new observation belongs to	Binary Inputs
Deep Q-Learning (DQN)	Reinforcement Learning	Q-Learning extension that uses neural networks	Deep Neural Networks
Linear Regression	Supervised Learning	Predict a continuous value	Continuous Variables
Principal Component Analysis (PCA)	Unsupervised Learning	Reduce the number of variables to simplify visualization and modeling	Data Compression
Gaussian Mixture Models (GMM)	Unsupervised Learning	Clustering	Generative Models
Q-Learning	Reinforcement Learning	Explore the Environment	Operations Optimization

TABLE V
DATABASE SYNTHESIS

Type of Morphology	Alluvial Plains; Firm Ground; River Islands.
Topology Type	Floodplain Forest and Terra Firme and Dense Forest
Physical Geometry	Geographic Coordinates; Azimuth; Elevation Angle; Distance.
Power	Transmitted Received; Reception Threshold; Fade Margin; Fade Margin Interference Fading.
Noise	Thermal considering 100% and Operating time on Equipment, Echo, RSR, and Due to different types of polarization.
Frequencies of Operation	Plumbing; Power Limits Transmitted; Spurious Signal Limits; Distance Limits.
Features of Equipment	Equipment Model (TX and RX)
Specifications of Antennas	Height; Gain; Type of Polarization.
Propagation Models	Longley Rice, Sui, Ray- Tracing adjusted, Ericsson e Recommenders ITU-R P.833-7;
Machine Learning	Supervised with Tree Decision Algorithm

Since the goal is to present an efficient tool capable of providing accurate analyses in atypical scenarios using AI techniques, this work employs the supervised machine learning method in conjunction with the Decision Tree algorithm. This algorithm is suitable for scenarios involving both discrete classification (e.g., binary decisions such as "yes" or "no") and continuous prediction tasks, where numerical values are estimated [¹⁴,¹⁵]. The conceptual framework is illustrated in Fig. 1.

Fig. 1
Illustration of the decision tree algorithm.

The Decision Tree algorithm operates based on the concept of nodes and leaves, where decision nodes represent a test or condition applied to an attribute:

• • The root node contains the initial dataset (input data);

• • The leaf nodes represent the final output or class resulting from each split.

In this context, it is understood that the root node is one of the attributes of the database, and the leaf nodes are the class or value that will be generated as a response. In this work, the decision tree algorithm will portray problems related to information classification. To evaluate the quality of the information gain, Equation (20) is used:

(20) $$InformationGain\left(S,A\right)=Entropy\left(S\right)-\sum _{\vartheta \in Values\left(A\right)}\frac{\left|S_{\vartheta }\right|}{S}xEntropy\left(S_{\vartheta }\right)$$

Where: This equation defines Information Gain as the difference between the entropy of the original set SSS and the weighted average entropy of the subsets Values S_ϑ generated by splitting according to the values ϑ of attribute (A). Here, $\frac{\left|S_{\vartheta }\right|}{S}$ represents the proportion of examples in the subset corresponding to the value ϑ of (A). Entropy measures the impurity or uncertainty of the data; thus, Information Gain quantifies the reduction in uncertainty achieved by partitioning S based on attribute (A). The greater the Information Gain, the better the attribute is at classifying the data, making it more suitable for splitting at that node in the decision tree. Impurity measure of the original data set calculated by the equation (21).

(21) $$Entropy\left(S\right)=\sum _{i=1}^c{P}_{i}lo{g}_2\left(P_i\right)$$

Where: c represents the number of classes and P_i is the proportion of examples in the class i.

IV. METHODOLOGY

The premise for developing the methodology is based on exploring new intelligent techniques that assist in decision-making in scenarios with heterogeneous paths, where the signal path includes fractions passing through fresh water and dense vegetation; this requires an appropriate propagation model. The proposed method presents the following characteristics:

• • Calculates and presents the main operating parameters that constitute the transmission system.

• • Simulations are based solely on topographic and morphological mapping, as well as the specific characteristics of the links under analysis.

• • A decision support system is based on geoinformation and link characteristics. Using this information, it is possible to generate recommendations for propagation models for each scenario through supervised machine learning with a decision tree algorithm.

• • Provides warning alerts to the user about possible data input errors or information incompatible with the established system limits.

• • Combines system parameters aligned with technical recommendations and current standards, allowing users to work with technically reliable and legally compliant data.

• • Generates a report of the results obtained in PDF format.

The programming resources shown in Fig. 2 were utilized in the research.

Fig. 2
Resources used for the development of the methodology. Fonte: Author.

Fig. 2 illustrates the main steps for implementing resources in the SINMCEL system, which was developed in four stages, as outlined below:

• • The first stage involves creating the data structure that feeds the system. PostgreSQL, a relational database recommended for high-performance applications, was used for this purpose. This data structure must include all relevant information related to morphology types, topography, physical geometry, specific system parameters (such as power, noise, and frequency), and propagation models, as shown in Table V. Additionally, in this stage, artificial intelligence was implemented by training the database using the Decision Trees algorithm.

• Second Stage: Supabase is an open-source interface that provides a robust infrastructure of authentication, data storage, and real-time APIs. In this stage, Supabase is powered by the PostgreSQL database, [¹⁶].

• Third Stage: In this phase, the Stored Procedures tool represents the SQL code blocks responsible for all mathematical operations required by the system, [¹⁷]. In this way, all operations will be stored and executed directly on the database server.

• Fourth Step: The last step refers to the Postman tool, which performs HTTP requests and provides a graphical interface for analyzing, monitoring, and validating responses via APIs.

Fig. 3 summarizes, in a simplified way, the end-to-end design of the developed methodology. The flowchart shows the main steps used in implementing the proposed methodology. The procedures start from the database built for the SINMCEL prediction software. The functioning of this structure depends directly on the data required at the input interface.

Fig. 3
Summary of the methodology presented through a flowchart.

The tool provides users with two possible simulation formats. The first is the intelligent simulation format, in which the Decision Tree algorithm selects the most appropriate propagation model-an approach that optimizes system performance. The second is the free simulation format, which allows the user to manually select a propagation model. In this mode, the system loses part of its decision-making autonomy.

When the intelligent simulation option is selected, the software recommends the most suitable propagation model for the specific environment and then presents the simulation results. In contrast, the free simulation mode requires the user to choose one of the available models, after which the system runs the simulation and displays the results accordingly.

A. Description of attributes used in the decision tree algorithm

When the intelligent simulation mode is selected, the system uses the Decision Tree algorithm to determine the most appropriate propagation model. This process involves the following steps:

Upon entering the input data, the scenario analysis algorithm evaluates the following nine attributes:

• Morphology - Terrain Class (rough, urban, forested, or mixed);

• Topology - Particularities of the relief (altitude);

• Presence of direct line of sight - LoS

• Buildings - Constructions (high, medium, or none);

• Vegetation - Type (dense, balanced, or none);

• Freshwater stretches (present or none);

• Vegetation with stretches of fresh water (present or none);

• Vegetation with the presence of buildings (present or none);

• Signal attenuation as a function of link distance;

The data obtained from on-site measurements and the information available in free prediction tools were organized and processed, as shown in Table VI.

TABLE VI
CLASSIFICATION OF ATTRIBUTES THAT WILL BE USED IN THE DECISION TREE ALGORITHM

DESCRIPTION	ATTRIBUTES	DATA CLASSIFICATION	CHOICE OPTIONS	SUGGESTED MODEL
1) Morphology	Terrain type	-Categorical (user-defined)	Rugged environment	Important data that provides accuracy to the result
			Urban environment	Important data that provides accuracy to the result
			Forest environment	Important data that provides accuracy to the result
			Mixed environment	Important data that provides accuracy to the result
2)Topology	Altitude	Categorical	Defined by geographic coordinates	Important data that provides accuracy in the result
3) Direct line of sight (LoS)	LoS	Binary	Yes	Free space path loss Analyze the other options.
			Not
4)Buildings	Presence of medium-rise buildings	Binary	Yes	Sui
			Not	Analyze the other options.
	Presence of high-rise building	Binary	Yes	Longley Rice
			Not	Analyze the other options.
5)Vegetation	Presence of dense vegetation	Binary	Yes	Recommendation ITU-R P.833-7
	Presence of balanced vegetation	Binary	Not	Analyze the other options
			Yes	Ericsson
			Not	Analyze the other options.
6)Stretches of fresh water	Presence of freshwater stretches	Binary	Not	Analyze the other options
			Yes	Adjusted Ray-Tracing
7)Vegetation with buildings	Vegetation with buildings	Binary	Not	Analyze the other options
			Yes	Longley Rice
8)Signal attenuation	According to the model suggested by the system	Numeric	According to the model suggested	Suggested model response
9) Estimation of received power	According to the model suggested by the system	Numeric	According to the model suggested	Suggested model response

Based on the information in Table VI, it is understood that the operation of the Decision Tree algorithm relies on the description and classification of data-referred to as attributes-which are linked to the available propagation model options. In this study, nine attribute classes were considered for constructing the decision-making process.

The first and most critical attribute at this stage is the presence or absence of Line of Sight (LoS), which represents the root node of the tree. This determination is based on the geographic coordinates provided by the user. The decision path proceeds as follows:

• If LoS is present, the link tends to be simpler, and the suggested propagation model is Free Space Path Loss there is LOS, the link tends to be simpler, and the suggested propagation model would be Free Space Path Loss;

• If LoS is absent, the algorithm continues by evaluating other factors, such as terrain type, altitude, and environmental obstacles.

For the decision nodes-which structure the branches of the tree-the following rules apply:

• Dense vegetation (forest environments): The recommended model is ITU-R P.833-7.

• Moderate vegetation (forest environments): The system suggests the Ericsson model.

• Medium-sized buildings in urban areas: The system recommends the SUI model.

• Large buildings in urban or hilly terrain: The algorithm selects the Longley-Rice model.

• Presence of freshwater bodies in mixed environments: The model used is Adjusted Ray-Tracing.

• Large buildings with vegetation in urban or hilly environments: The Longley-Rice model is also recommended.

Which represent the final classification outcomes-there is:

• decision support tool that provides a precise and reliable recommendation for the most appropriate propagation model to use in the simulation.

B. Methodology Validation Process

Field measurements were taken at 12 points distributed between 500 meters and about 6 km from the transmitter, covering a total distance of approximately 5.5 km. The spacing between points was not uniform, as we aimed to better capture signal variations caused by different terrain and environmental conditions. In a specific 1.2 km segment, points were placed more regularly, spaced approximately every 100 meters, to provide a more detailed analysis of that area.

The selection of measurement locations considered GPS coordinates and topographic data from the Shuttle Radar Topography Mission (SRTM), taking into account terrain relief, vegetation density, and proximity to water bodies, ensuring that the measurements accurately reflected real propagation conditions.

The receiver antenna was kept at 1.5 meters above the ground throughout all measurements, ensuring data consistency and reflecting real deployment conditions in open areas.

Measurements were conducted using a portable spectrum analyzer connected to a directional sector antenna via a low-loss coaxial cable, both configured to operate at 5.4 GHz with a 40 MHz bandwidth. The analyzer functioned as a passive device, solely recording the received signal power level without establishing an active connection with the fixed transmitter. The antenna was manually aligned at each point to aim the signal toward the transmitter, ensuring measurement accuracy. At each point, signal strength was recorded in dBm, with five consecutive samples taken to mitigate short-term fluctuations.

This setup allowed us to confidently assess the link’s performance under varying environmental conditions along the route.

These field measurements were carried out on Combu Island, a riverine region near Belém, to support a temporary connectivity solution. The selection of measurement locations was based on the route with the highest traffic flow on the island. The selected points-Mangal das Garças Park, Casa de Chocolate (Filha do Combu), and Restaurante Ribeirinho-are tourist attractions that will also host activities during the COP30 event. The route connecting these sites presents a heterogeneous scenario that includes dense urban segments, forested areas, and stretches over freshwater, as illustrated in Figs. 4, 5, and 6.

Fig. 4
Analysis scenario - route taken by heads of state inside the chocolate house - daughter of the combu. Source: author

Fig. 5
Analysis scenario - riverside restaurant. Source: Google.

Fig. 6
Analysis scenario - Mangal das herças park. Source: Google.

C. Carrying out Measurements

Four point-to-point devices, two server radios, and two client radios from the manufacturer Cambium Networks, model PTP450i, were installed to perform the measurements, as shown in Fig. 7 (a), Fig 7(b), Fig. 8 (a) and Fig. 8 (b).

Fig. 7
Cambium Radio Server, model PTP450i

Fig. 8
Cambium radio client, model PTP450i

The PTP 450i devices suit various environments and applications, especially long-distance and high-capacity links. The PTP 450i model has an integrated antenna system that can reach a gain of 23 dBi and cover a frequency range of 4900-5925 MHz.

V. RESULT

Measurements on three point-to-point links, with different distances, covering dense vegetation and freshwater environments will confirm the method. Table VII summarizes all geometric and physical parameters of the link.

TABLE VII
GEOMETRIC AND PHYSICAL PARAMETERS OF THE LINK

		Link 1 (CT_ CC)	Link 2 (CT_ RR)	Link 3 (CT_ PM)
Point A [°]	Latitude	-1.464650	-1.464650	-1.464650
	Longitude	-48.49842	-48.49842	-48.49842
Point B [°]	Latitude	-1.500082	-1.505321	-1.449889
	Longitude	-48.46088	-48.46285	-48.50030
Distance [km]	d	6	6,54	1,02
Azimuth in A and B [°]	B→A	313,30	318,83	87,64
	A→B	46,64	41,16	272,35
Elevation Angle [°]	AE	-0.006454	-0.006753	-0.00115
Antenna Gain [dBi]	G	17/17	17/17	13/15
Frequency [GHz]	f	5.4/5.8	5.4/5.8	5.4/5.8
Bandwidth [MHz]	RF	40	40	40
Transmitted Power [dBm]	PT	25	27	20
Reception Threshold [dBm]	LR	- 75 to 64 QAM	- 80 to 16 QAM	-70 to 256 QAM

Where CC - represents the physical station’s location - Chocolate House - Combu Island; CT - Represents the physical station’s location - CIOP Tower (Owner Prodepa); RR- Represents the physical station’s location of the at Riverside Restaurant; PM - represents the physical location of the station Mangal das Herças Park.

The links under analysis were then established, as shown in Fig. 9.

Fig. 9
Location of stations using the Google Earth tool.

The simulation was performed for Link 1 (CT-CC) at 5.4 GHz with a bandwidth of 40 MHz. The antenna heights were determined based on the geographic locations, according to the field study: CT at 50 m and CC at 44 m. The transmitted power at this frequency is limited to 27 dBm. Since this link crosses a forested region with segments over freshwater, a power level of 25 dBm was initially used, considering a reception threshold of -75 dBm. The algorithm processed the input data and recommended the Adjusted Ray-Tracing propagation model as the most appropriate option for the scenario under analysis, as shown in Fig. 10.

Fig. 10
Simulation result of the CT_CC link considering the Adjusted Ray-Tracing propagation model.

Fig. 10 illustrates the convergence between the field-measured data and the result obtained using the model suggested by the software. The signal attenuation measured in the field was 64 dB, while the value predicted by the Adjusted Ray-Tracing model was 66 dB. To validate these results, a statistical analysis was performed using the Pearson correlation coefficient to quantify the relationship between the measured data and the simulated values, as presented in Fig. 11.

Fig. 11
Scatter diagram to analyze the correlation between CT_CC link variables.

The adjusted ray-tracing propagation model is highly accurate compared to the data measured in situ. Comparatively, the results obtained using the “free simulation” option showed considerable differences concerning the measurements, as shown in Figs. 12 and 13.

Fig. 12
Simulation result of the CT_CC link considering the Adjusted Ray-Tracing, Longley-Rice and SUI propagation models.

Fig. 13
Simulation result of the CT_CC link considering the Adjusted Ray-Tracing, Ericsson, LOS, and ITU-R P.833-7 propagation models.

According to Fig. 12, the link attenuation obtained using the Longley-Rice model was 50 dB, whereas the SUI model yielded a value of approximately 47 dB. In Fig. 13, the attenuation results obtained from the Ericsson, LOS, and ITU-R P.833-7 models were 48 dB, 36 dB, and 56 dB, respectively.

The models selected under the “free simulation” mode consistently produced lower attenuation values compared to the result obtained using the Adjusted Ray-Tracing model. These discrepancies can be attributed to the specific assumptions and propagation mechanisms considered in each model. For example, the Longley-Rice model incorporates diffraction and scattering effects in non-line-of-sight (NLOS) scenarios, while the SUI model accounts for terrain-induced shadowing to varying degrees.

Given these considerations, it is evident that the propagation model recommended by SINMCEL was the most appropriate for the analyzed link. Therefore, the received power result will be presented exclusively for the Adjusted Ray-Tracing model, as illustrated in Fig. 14.

Fig. 14
Received power level on the CT_CC link.

The result presented in Fig. 14 shows a received power of -56.29 dBm. In this case, the transmitted power was 25 dBm, with a reception threshold of -75 dBm.

For organizational purposes, since the result analysis for the CT_CC link has been fully detailed, it is unnecessary to repeat points already discussed throughout the article. The subsequent analyses and discussions for the other simulation scenarios will be addressed individually and concisely.

The second measurement was performed for Link 2 (CT_RR), also at 5.4 GHz and 40 MHz bandwidth. The antenna heights were 50 m at CT and 47 m at RR. In this case, since the link traverses a forested region (characterized by dense vegetation), the maximum permitted power of 27 dBm was used, with a reception threshold set at -80 dBm. After processing the input data, the SINMCEL software recommended the ITU-R P.833-7 model as the most suitable option for the scenario under analysis, as illustrated in Figs. 15, 16, and 17.

Fig. 15
Simulation result of the CT_RR link considering the ITU-R P.833-7 Recommendation.

Fig. 16
Scatter diagram to analyze the correlation between CT_RR link variables.

Fig. 17
Simulation result of the CT_CC link considering the Adjusted Ray-Tracing, Longley-Rice, LOS and SUI propagation models.

By analyzing Figs. 15, 16, and 17, it can be concluded that there is strong agreement between the measured and simulated results. For the CT_RR link, the field measurements indicated an attenuation of 70.23 dB, while the simulated value using the SINMCEL-suggested model was 72.16 dB. Given the specific characteristics of this link-such as dense vegetation and terrain conditions-the propagation model selected by SINMCEL proved to be the most appropriate for the analysis. The corresponding received power is presented in Fig. 18.

Fig. 18
Received power level on the CT_RR link.

As shown in Fig. 18, the received power was -52.30 dBm, considering a reception threshold of -80 dBm. Subsequently, measurements for link 3 (CT_PM) were carried out in a region characterized by dense vegetation and tall buildings. Due to the short-distance nature of this link, a transmission power of 20 dBm was used, with a reception threshold set at -70 dBm. The antenna heights were 50 m at CT and 40 m at PM, while the frequency and bandwidth remained the same as in previous analyses (5.4 GHz and 40 MHz, respectively). Based on this configuration, the SINMCEL software recommended the Longley-Rice model as the most suitable for the scenario. Figs. 19, 20, 21, and 22 present the results obtained.

Fig. 19
Simulation result of the CT_PM link considering the longley rice propagation model.

Fig. 20
Scatter diagram to analyze correlation between CT_PM link variables.

Fig. 21
Simulation result of the CT_CC link considering the Longley-Rice, Adjusted Ray-Tracing, LOS and SUI propagation models.

Fig. 22
Received power level on the CT_PM link.

The data presented in Figs. 19, 20, 21, and 22 allow us to conclude that the Longley-Rice model is appropriate for analyzing this link. The measured attenuation was 57.86 dB, corresponding to a received power level of -53.95 dBm.

VI. FINAL CONSIDERATIONS

In this article, we aimed to establish a consistent method that combines machine learning with propagation models through the implementation of a predictive tool. The objective was to provide users with an intelligent analysis approach tailored to atypical scenarios, though not limited exclusively to them. Field measurements were conducted in the city of Belém to validate the proposed method. Three point-to-point links with varying characteristics-including different distances, dense forest regions, and freshwater routes-were analyzed. Throughout the study, the advantages and limitations of the SINMCEL tool were discussed. The software employs a decision tree algorithm to intelligently recommend the most appropriate propagation model for each analysis scenario. Ultimately, this work presents an efficient method capable of offering a nuanced perspective on the analysis of transmission systems.

The software used the decision tree algorithm to intelligently recommend the appropriate propagation model for each analysis scenario. The objective of the work was to present an efficient method capable of providing a differentiated perspective in the analysis of transmission systems.

REFERENCES

• ^[1] GE Martins and R. S. da Silva. "Desafios da Participação Popular nos Projetos Urbanos da COP30 em Belém do Pará”. Challenges of popular participation in cop30 urban projects in Belém do Pará, 2024.
• ^[2] K konstantinou, S Kang, T Brown, C Tzaras. “Measurement and Modelling of the Propagation Channel Between Low-Heigth Terminals Microwaves”. Antenas pace Sciences. Springer link, vol. 5, pp. 412-418, 2019.
• ^[3] T Otobo and H Tertuliano. “An Innovative Graphical Viewer Analysis Applied in a Multipoint Transmission System”. IEEE Access, vol. 7, pp. 82473 -82490, 2019.
• ^[4] TK Sarkar, J Zhong and K Kyungjung. Medouri and M. Salazar-Palma, "A Survey of Various Propagation Models for Mobile Communication," IEEE Antennas and Propagation Magazine, vol.5, pp. 597-615 2006, Doi: 10.1109/MAP.2003.1232163.
  » https://doi.org/10.1109/MAP.2003.1232163.
• ^[5] M Vasudevan and M Yuksel, "Machine Learning for Radio Propagation Modeling: A Comprehensive Survey," IEEE Open Journal of the Communications Society, vol. 5, pp. 5123-5153, 2024, Doi: 10.1109/OJCOMS.2024.3446457.
  » https://doi.org/10.1109/OJCOMS.2024.3446457.
• ^[6] R Shumate. “Longley-Rice and ITU-P.1546 Combined: A New International Terrain-Specific Propagation Model”. IEEE 72nd Vehicular Technology Conference - Fall Ottawa. k, pp. 1-5, vol 1, 2010, Doi: 10.1109/VETECF.2010.5594342.
  » https://doi.org/10.1109/VETECF.2010.5594342.
• ^[7] B Castro, C Ribeiro and CJ Cavalcante. “COST231-Hata and SUI Models performance using a LMS tuning algorithm on 5.8GHz in Amazon Region cities”. Proceedings of the Fourth European Conference on Antennas and Propagation. Barcelona, pp. 1-3, vol. 5, 2010.
• ^[8] A Saeed, H Rehman and M Masood. “Performance Analysis and Comparison of Radio Propagation Models for Outdoor Environment in 4G LTE Network”. Publication blekinge Institute of Technology, Dissertation, pp. 1-63, 2013.
• ^[9] Z El Khaled, A Wessam and M Hamid. "An accurate empirical path loss model for heterogeneous fixed wireless networks below 5.8 GHz frequencies." IEEE Access, pp. 182755-182775, vol.8, 2020.
• ^[10] B Mohammad and M MACGREGOR. “Modelling vegetation effects on RF propagation”. 39th Annual IEEE Conference on Local Computer Networks. IEEE, pp. 450-453, vol. 39, 2014, Doi: 10.1109/LCN.2014.6925814.
  » https://doi.org/10.1109/LCN.2014.6925814.
• ^[11] Z Yun and M F Iskander "Ray Tracing for Radio Propagation Modeling: Principles and Applications," IEEE Access, vol. 3, pp. 1089-1100, 2015, Doi: 10.1109/ACCESS.2015.2453991.
  » https://doi.org/10.1109/ACCESS.2015.2453991.
• ^[12] C Huang, et al. "Artificial Intelligence Enabled Radio Propagation for Communications-Part I: Channel Characterization and Antenna-Channel Optimization," IEEE Transactions on Antennas and Propagation, vol.70, pp. 3939-3954, 2022, Doi: 10.1109/TAP.2022.3149663.
  » https://doi.org/10.1109/TAP.2022.3149663.
• ^[13] A Seretis and CD Sarris, "An Overview of Machine Learning Techniques for Radiowave Propagation Modeling," IEEE Transactions on Antennas and Propagation, vol. 70, pp. 3970-3985, 2022, Doi: 10.1109/TAP.2021.3098616.
  » https://doi.org/10.1109/TAP.2021.3098616.
• ^[14] V. Shats. “Error-Free Training via Information Structuring in the Classification Problem”. Journal of Intelligent Learning Systems and Applications Vol. 9, No. 3, pp. 9-11, 2017.
• ^[15] S. Reginald. “Machine Learning Mapping of Soil Apparent Electrical Conductivity on a Research Farm in Mississippi”. Open Journal of Statistics Vol. 7, No. 4, pp. 4-8, 2017.
• ^[16] MS Abdul, et al. "Database replication: A survey of open source and commercial tools." International Journal of Computer Applications, vol.13, pp. 1-8, 2011.
• ^[17] MMW Vanhoof, and H Hainaut. "Relational symbolic execution of SQL code for unit testing of database programs" Science of Computer Programming, vol. 105, pp. 44-72, 2015.

Publication Dates

• Publication in this collection
  03 Nov 2025
• Date of issue
  2025

History

• Received
  04 Dec 2024
• Reviewed
  12 Mar 2025
• Accepted
  16 July 2025