Machine-Learning-Based Adaptive Wireless Network Selection for Terrestrial and Non-Terrestrial Networks in 5G and Beyond

Abstract

Non-terrestrial networks (NTNs) have become increasingly crucial, particularly with the standardization of fifth-generation (5G) technology. In parallel, the rise of Internet of Things (IoT) technologies has amplified the need for human-centric solutions in 5G and beyond (5 GB) systems. To address diverse communication requirements from a human-centric perspective, leveraging the advantages of both terrestrial networks (TNs) and NTNs has emerged as a key focus for 5 GB communications. In this paper, a machine learning (ML)-based approach is proposed to facilitate decision making between TN and NTN networks within a multi-connectivity scenario, aiming to provide a human-centric solution. For this approach, a novel synthetic dataset is constructed using various sensing information, based on the assumption that numerous interconnected sensor systems will be available in smart city networks with sixth-generation (6G) technologies. The ML results are derived from this newly generated dataset. These simulation results demonstrate that the proposed approach, designed to meet the requirements of next-generation systems, can be effectively utilized with 6G.

1. Introduction

Fifth generation (5G) cellular communications systems have introduced three distinct application groups: enhanced mobile broadband (eMBB), ultra-reliable low-latency communications (URLLC), and massive machine-type communications (mMTC) [1]. When we examine the purpose of this categorization, it can be seen its relationship with consumer electronics and consumer technologies. The increasing number of Internet of Things (IoT) consumer electronics results in the more wireless communications requirement richness in time [2]. Thus, 5G cellular systems are designed to meet different types of communications requirements for three application groups synchronously.

The richness of requirements will increase in the next decades [3]. In the International Mobile Telecommunications 2030 (IMT-2030) [3], which is currently under development by the International Telecommunication Union (ITU), advanced and broader usage scenarios are discussed, replacing the eMBB, URLLC, and mMTC usage scenarios introduced in IMT-2020 [4] with new capabilities named immersive communication, hyper-reliable and low-latency communication, and massive communication, respectively. In addition to these scenarios, IMT-2030 introduces three new usage scenarios that were not supported in the previous generations: artificial intelligence (AI) and communication, integrated sensing and communication, and ubiquitous connectivity. In the ubiquitous connectivity usage scenario, the focus is on addressing areas that are currently out of coverage or nearly uncovered. Additionally, the AI and communication scenario discusses the inclusion of AI and information processing capabilities, including the collection, processing, and preparation of data from various sources, such as user equipment. Therefore, the utilization of non-terrestrial networks (NTN) and AI in 6th generation (6G) wireless systems is clearly emphasized with the main 6G scenarios of ITU in the IMT-2030 documentation [3].

To overcome new challenges with new requirements, there is a need for more flexible designs in 6G communications [2]. In other words, these flexible wireless designs will form the foundational building blocks of 6G communications. Indeed, 6G will give rise to consumer electronics applications that will require the next generation of communication systems. This cycle, visualized in Figure 1, is expected to continue. The development of flexibility-based wireless systems and new consumer electronics technologies will trigger each other’s advancements, serving as mutually reinforcing factors.

To leverage the advantages of non-terrestrial networks (NTNs) as part of next-generation flexible wireless communications, the integration of terrestrial networks (TNs) and NTN systems has become essential. NTN systems provide an increasing number of alternative solutions [5,6,7,8]. Compared to TNs, NTN platforms offer several benefits across different scenarios [9]. These systems enhance coverage and capacity while improving the line-of-sight (LOS). Additionally, they play critical roles in challenging environments, densely populated areas, and disaster situations.

The coordination units are needed for an interoperability between the TN and NTN networks. As shown in Figure 2, it is useful to determine the TN and NTN network options dynamically under changing environments. Different communications requirements can be met via both TN and NTN networks together. It is also possible to serve with multi-connectivity capabilities. Therefore, there is a need to design a method that decides between TN, NTN, and TN+NTN networks under different scenarios to take advantages of both TN and NTN systems. This can be also called as the user-network association if there are available different types of communications networks.

In the literature, there are several multitier optimization studies for resource allocation [10,11], machine learning (ML)-based resource allocation [12], AI-based network orchestration [13,14,15,16], and ML-based service management [17,18]. There are also available studies for the TN and satellite integration for spectrum sharing [19] and general concept [20]. In [21,22], multi/dual connectivity is studied for TN and NTN systems. Link scheduling algorithms are proposed under multi-connectivity scenarios in [23,24]. To the best of this author’s knowledge, dynamic decision systems for TN, NTN, and TN+NTN networks under different multi-connectivity scenarios have not been studied in the literature.

Cellular wireless communication systems have not yet extensively utilized other sensor signal sources. However, with 6G communications, it is anticipated that valuable sensing information will be exploited to enhance environmental awareness and support human-centric solutions [25]. To achieve this, integrating ambient systems into cellular networks is essential for sensing environmental conditions and addressing user requirements. This approach enables wireless channels to be estimated directly based on environmental and user-specific data, bypassing traditional channel estimation techniques. Furthermore, data science tools are preferable, as they facilitate the analysis of diverse information sources collectively. Establishing relationships between various data inputs is challenging with classical methods, making advanced tools like ML techniques more suitable.

In this paper, a machine learning (ML)-based approach is proposed to dynamically select between terrestrial networks (TNs) and non-terrestrial networks (NTNs) within a multi-connectivity scenario, providing a human-centric solution. Furthermore, the concept of useful sensing information is explored. For the proposed method, a novel synthetic dataset is created using various sensing inputs, based on the assumption that numerous interconnected sensor systems will be available in 6G-enabled smart city networks. Additionally, an example AI-driven approach is introduced for next-generation wireless systems. Simulation results confirm that the proposed approach, designed to address the requirements of diverse consumer electronics and technologies, can be effectively implemented in 6G wireless systems.

In the remaining part of the paper, Section 2 presents the fundamental topics for wireless communications channel models and environment awareness. The proposed multi-connectivity decision approach is discussed in Section 3. ML results are provided in Section 4. Finally, Section 5 concludes the paper, also including several open issues.

2. Wireless Communications Channel Models and Environment Awareness

In this section, wireless channels are discussed from the environment awareness perspective to form a basis for the dataset generation that is explained in the next section.

The behavior of wireless channels is highly dependent on the environment in which the communications takes place. Signal distortion effects under different wireless channels can be clearly observed as the environment changes. There are many different channel models considering the environment awareness in the literature [26,27]. In this paper, air-to-ground (A2G) and ground-to-ground (G2G) wireless channel models are taken into account to work with terrestrial and non-terrestrial networks together.

Properties of the wireless channel may vary based on the characteristics of different environments. In order to exemplify the environmental effects on the A2G channel, the series of studies from Matolak et al. is reviewed [28,29,30,31]. In their first study [28], the wireless channel is reviewed under such conditions where water surface is the operating environment. As a continuation of their study, the wireless channel is characterized in hilly/mountainous [29] and suburban/near-urban [30] environments. In the last study of the series [31], they examined air-frame shadowing loss. From the given A2G channel studies, it can be verified that the channel characteristics are highly depending on the environment conditions in terms channel properties of path loss, shadowing and multipath propagation.

The characteristics of wireless channel exhibits different behaviors also in G2G links. To exemplify, the varying time delay of G2G links is shown in [32] for different environments, such as homogeneously/inhomogeneously placed buildings, closely spaced buildings, linear streets, and stadiums.

For the below path loss equation [27], path loss exponent varies under different environments for A2G and G2G links.

$$\begin{array}{cc}\hfill P_{loss}\left(d\right)& =P_{\mathrm{loss}}\left(d_0\right)+10\eta log\left({\displaystyle \frac{d}{d_0}}\right)\hfill \\ \hfill d_0<d\end{array}$$

(1)

where d is the link range, $P_{loss}$$\left(d\right)$ is the path loss value, $d_0$ is the reference distance from the transmitter, which is determined based on measurements close to the transmitter, $P_{loss}$$\left(d_0\right)$ is the free space path loss, and $\eta$ is the parameter for path loss exponent that is incorporated based on the propagation environment in which the values of $\eta$ is given in

Table 1. Path loss exponent under different environments for air-to-ground (A2G) links [26].

Environment	Path Loss Exponent ($\mathit{\eta }$)
Suburban	2.54–3.04
Urban	2.20–2.60
Near airports	2.00–2.25
Hilly, mountainous	1.00–1.80
Over sea	0.14–2.46

and

Table 2. Path loss exponent under different environments for ground-to-ground (G2G) links [27].

Environment	Path Loss Exponent ($\mathit{\eta }$)
LOS in buildings	1.50–2.00
LOS free space	2.00
In factories	2.00–3.00
Urban area	2.70–3.50
Obstructed in buildings	4.00–6.00

under different environments for A2G and G2G links, respectively. Considering Equation (1), path loss parameters changes with different environmental conditions. As a consequence, varying path loss values are obtained.

As a different perspective, satellite images of an environment are employed to analyze the path loss characteristics with a ML-based technique in [33]. This study shows that the path loss is observed more severely when the building density is increased. Therefore, when the path loss calculations are operated, effects of the environment must be taken under consideration to represent a more accurate channel models.

The other wireless channel properties such as shadowing and multipath propagation also vary with environmental conditions. In this section, only path loss component of the channel is examined. Wireless channel and environment relation is exploited while generating a dataset in the next section.

3. Multi-Connectivity Decisions

In this section, the proposed methods for synthetic data generation, feature extraction, network handover control, and classification decision are presented under the subsections.

3.1. Synthetic Data Generation and Features

As stated in Section 1, several sensing information are used for the proposed solution in this paper. However, there is not any corresponding dataset that is related with the target USI. Moreover, it is a difficult task to form a real-world dataset based on parallel sensor measurements. Hence, a synthetic dataset is preferred to solve the given problem and introduce an example decision method. In addition, it should be noted that the synthetic dataset usage has some important advantages such as providing more data quality, scalability, and utilization simplicity compared to real-world datasets.

MATLAB 2023b platform is employed to generate a task-oriented synthetic dataset. The class labels are generated in balance at first step. Then, the features are generated with specific rules and randomnesses for each class label. These specific rules and randomnesses are defined through wireless channel relations that are summarized in

Table 3. Definition and analysis of the used features. eMBB: enhanced mobile broadband, URLCC: ultra-reliable low-latency communications, LoS: line of sight, TN-TP: terrestrial networks transmission point, NTN-TP: non-terrestrial networks transmission point, Info. Gain: Information Gain.

No	Definition	Domain	Type	Channel Relation	Range	Info. Gain
1	Geographic Characteristic	Region	Environmental	Extreme environments cause rich multipath.	Extreme (1)– Basic (10)	0.097
2	Residential Area Plan	Region	Environmental	High possibility of rich multipath environment for urban scenarios. Path loss characteristic changes remarkably.	Urban (1)– Rural (10)	0.068
3	User Density	Region	Environmental	Spectral efficiency is an important criteria.	Dense (1)– Scarce (10)	0.028
4	User Mobility	User	Environmental	High mobility may cause Doppler spread.	High (1)– Low (10)	0.094
5	eMBB Necessity	User	Requirement	To meet the eMBB requirements, low tolerance for the multipath effects. Needs an increased channel capacity.	High (1)– Low (10)	0.040
6	URLLC Necessity	User	Requirement	To meet the URLLC requirements, less delay spread and Doppler spread. Needs a balance between subcarrier spacing and symbol duration for the waveform.	High (1)– Low (10)	0.094
7	TN-TP Distance	User	Environmental	Path loss characteristic changes remarkably. High possibility of more interference.	High (1)– Low (10)	0.115
8	NTN-TP Distance	User	Environmental	Path loss characteristic changes remarkably. High possibility of more interference.	High (1)– Low (10)	0.110
9	TN-TP LoS Availability	User	Environmental	Path loss characteristic changes remarkably. High possibility of more interference.	High (1)– Low (10)	0.291

. Moreover, the dependencies between the features are also considered. It is assumed that the features are normalized between 1 and 10 according to the range definitions that are given in Table 3. Each feature in the dataset is graded on a scale from 1 to 10 to reflect realistic scenario conditions. For instance, Geographic Characteristics are mapped such that a value of 1 corresponds to dense urban canyons with severe multipath, while a value of 10 corresponds to open rural plains with minimal obstacles. As another example, User Density ranges from 1 (ultra-dense hotspot, >5000 users/km²) to 10 (sparse users, <50 users/km²). These quantification rules ensure that the abstract feature scores are reproducible in actual deployment scenarios.

The generated dataset contains the three types of classes listed below. The decision system selects one of these network options, considering the features to meet wireless communications requirements under different intelligent transportation scenarios.

–

Class 1: Terrestrial networks (TN)

–

Class 2: Non-Terrestrial networks (NTN)

–

Class 3: TN and NTN together (TN-NTN)

Nine different features in Table 3 include both environmental effects and some user requirements such as eMBB and/or URLLC necessities. The feature type (environmental or requirement) is given in the fourth column in the table. The common intersection point for the used features is the wireless communications channel. In other words, the given features commonly affect the channel characteristic. These features provide an environment-aware perspective rather than giving wireless channel information directly. The wireless communications channel varies depending on the environmental conditions, as explained in Section 2. Additionally, the feature domain (region based or user based) is provided in the third column of Table 3 to point out the sensor/information source.

As the first two features given in Table 3, the geographic characteristic and the residential area planning (e.g., urban and rural areas) of a region have effects on the wireless channel from the aspect of multipath environment level. As an example, there is a high possibility of a rich multipath environment under extreme geographical conditions and urban scenarios. Moreover, path loss characteristic changes remarkably under different residential area plannings. For both features, the NTN links are preferable under extreme geographical conditions and urban scenarios thanks to the behaviors of A2G wireless channel as discussed through Section 2 in detail.

Another feature provides the user density level of a region and it gives a clue for the spectral efficiency requirements. The more user density comes with the more spectral efficiency necessity. It is possible to take advantages of TN and NTN networks while meeting the high spectral efficiency requirement in the case of dense region scenarios.

For the user mobility feature, there is a critical channel relation that a high mobility of a user may cause Doppler spread. NTN links are more preferable under the high mobility scenarios since NTN platforms are in the sky and the relative mobility differences are low compared to the TN links.

eMBB necessity for a user is taken as another feature. To meet the eMBB requirements such as high throughput and data rates, there is a low tolerance for the multipath effects. Moreover, channel capacity needs to be increased. If the eMBB necessity is high, TN and NTN networks can be used together to complement each other.

For the URLLC necessity feature, the amount of delay spread and Doppler spread are critical. As an example, there is a need to balance subcarrier spacing and symbol duration for the waveform design considering the reliability and latency performance expectations. Additionally, URLLC service type is essential for the V2X applications. Generally, NTN links are preferable if the URLLC necessity is high.

The other three features are terrestrial networks transmission point (TN-TP) distance, non-terrestrial networks transmission point (NTN-TP), and TN-TP LoS availability. For all of these features, path loss characteristic changes remarkably under different scenarios. If TN-TP distance is high, NTN link is preferable; however, TN link should be used if NTN-TP distance is high and TN-TP LoS is available.

The discriminative power of each feature is quantified using the Information Gain given in Table 3, computed via an ID3 decision tree splitting criterion. For a dataset D with entropy $H\left(D\right)$, the Information Gain of feature X is defined as in Equation (2).

$$IG(D,X)=H\left(D\right)-\sum _{v\in Values\left(X\right)}{\displaystyle \frac{|D_v|}{\left|D\right|}}H\left(D_v\right)$$

(2)

where $D_v$ denotes the subset of D where $X=v$. This metric highlights features that reduce uncertainty the most during classification.

3.2. Network Handover Control

As shown in the block diagram of the proposed method (Figure 3), it is assumed that nine features are extracted from two types of raw data that include environmental information and wireless communications requirements. The features of geographic characteristic, residential area plan, user density, user mobility, TN-TP distance, NTN-TP distance, and TN-TP LoS availability are extracted from the environmental information. Additionally, eMBB necessity and URLLC necessity features are extracted from wireless communications requirements information.

The other basic assumption is that environmental information is obtained through smart city networks. Smart cities in the near future will have all necessary information related to the environment. Moreover, geographic information systems (GIS) already have many valuable information that can be used in many different environmental-aware systems. The wireless communications requirements information already exists in the new-generation cellular networks to be benefited under different subsystems.

To decrease the ping-pong effects during the network handover between TN and NTN transmission points, a feature control algorithm is proposed. The key idea is to smooth sudden variations of the extracted features over time, avoiding unnecessary switching. Three equations are defined for this purpose:

$${\sigma }_k^2\left(t\right)=\left\{\begin{array}{cc}0,\hfill & \mathrm{if}t<T\hfill \\ {\widehat{\sigma }}_k^2\left(t\right),\hfill & \mathrm{else}\hfill \end{array}\right.$$

(3)

where ${\sigma }_k^2\left(t\right)$ denotes the variance of the k-th feature at time t. The parameter T represents the time window threshold that must be exceeded before a reliable variance estimate can be obtained. This prevents premature handovers caused by short-term fluctuations.

$${\sigma }_{k,b}^2\left(t\right)=\left\{\begin{array}{cc}0,\hfill & \mathrm{if}t<T\hfill \\ {\sigma }_k^2\left(t\right),\hfill & \mathrm{if}t=T\hfill \\ {\sigma }_{k,b}^2(t-1),\hfill & \mathrm{if}{\sigma }_k^2\left(t\right)=0\mathrm{or}{\sigma }_k^2\left(t\right)>\beta {\sigma }_{k,b}^2(t-1)\hfill \\ \alpha {\sigma }_{k,b}^2(t-1)+(1-\alpha ){\sigma }_k^2\left(t\right),\hfill & \mathrm{else}\hfill \end{array}\right.$$

(4)

In Equation (4), ${\sigma }_{k,b}^2\left(t\right)$ is the background variance of the k-th feature. The parameter $\alpha \in [0,1]$ controls the learning rate: when $\alpha$ is close to 1, the background variance changes slowly, making the system more stable; when $\alpha$ decreases, recent variance values have more impact. The parameter $\beta$ is a tolerance factor: it avoids updating the background variance when sudden spikes are detected, thereby preventing unnecessary handovers.

$$f_k\left(t\right)=\left\{\begin{array}{cc}{f}_k\left(t\right),\hfill & \mathrm{if}t<T\mathrm{or}{\sigma }_{k,b}^2(t-1)=0\hfill \\ f_k\left(t\right),\hfill & \mathrm{if}\left|{\displaystyle \frac{{\sigma }_{k,b}^2\left(t\right)}{{\sigma }_{k,b}^2(t-1)}}-1\right|>B\hfill \\ f_k(t-1),\hfill & \mathrm{else}\hfill \end{array}\right.$$

(5)

Finally, Equation (5) defines how the k-th feature $f_k\left(t\right)$ is updated. The threshold B determines the sensitivity of feature updates. A smaller B increases responsiveness (faster adaptation), while a larger B reduces the update frequency and avoids ping-pong effects.

The decision-making module receives smoothed inputs by stabilizing the feature dynamics, which reduces the probability of oscillatory network selection. In other words, the optimization problem of minimizing unnecessary handovers is indirectly addressed through controlled feature adaptation.

Algorithm 1 ensures that only stable and consistent feature updates are passed to the decision engine, thereby mitigating the ping-pong effect of frequent TN/NTN switching.

Algorithm 1: Proposed feature control algorithm for network handover

1. Initialize background variance ${\sigma }_{k,b}^2\left(0\right)=0$ for all features. 2. For each time instant t: (a)   Compute variance ${\sigma }_k^2\left(t\right)$ using Equation (3). (b)   Update background variance ${\sigma }_{k,b}^2\left(t\right)$ using Equation (4). (c)   Update feature value $f_k\left(t\right)$ using Equation (5). 3. Feed updated feature values into the ML-based decision module.

3.3. Classification Decision Method

A multiclass classification problem is studied since there are three classes on the dataset. Supervised learning techniques are applied with hyperparameter optimizations to increase the success rate performances. At the start, all users are connected to TN transmission point if possible. Then, the proposed approach is started to decide on the communications networks. The decisions are made using ML models as user-based despite some of the features are region-based. For the region-based features, a region can be defined in different way. A vehicle cluster can also be taken as a small region. Hence, it is possible to use both small clusters and the larger regions.

Since the proposed handover control scheme relies on an ML-based decision module, the inference delay of the employed models must be considered to ensure real-time applicability from the latency considerations perspective. In our implementation, lightweight classifiers such as Random Forests and Gradient Boosting algorithms are preferred due to their relatively low computational complexity compared to deep neural networks. The average inference time was measured on a workstation equipped with an AMD Ryzen 5800X processor, NVIDIA RTX 3070Ti graphic card, and 32 GB DDR4 memory. The results show that a single inference operation requires approximately 0.8 ms for Random Forests and 1.5 ms for Gradient Boosted Trees, which is negligible compared to the typical handover preparation time in 5G/NTN systems (20–50 ms). Thus, the inference latency does not constitute a bottleneck for the proposed approach.

4. Machine Learning Results

Throughout this study, the Orange Data Mining program, which is an open source data analysis tool, is employed to get ML results on the generated dataset (https://drive.google.com/drive/folders/1iQv7Dooy4hAwSpucu7WCvXvA6Fof4A88, accessed on 30 September 2025). The specifications of the hardware that is used in the simulations are given as follows: AMD Ryzen 5800X processor, NVIDIA RTX 3070Ti graphics card, and 32 GB DDR4 memory. During the simulations, a dataset that contains $\mathrm{10,000}$ samples is generated on the MATLAB platform. Then, by using these samples, an ultimate dataset that contains $\mathrm{30,000}$ samples is obtained via a generative adversarial networks (GAN)-based deep learning data synthesizer named the conditional tabular generative adversarial network (CTGAN) that is specifically designed to be used for tabular datasets [34]. The CTGAN model is trained for 300 epochs with a batch size of 500, learning rate $2\times {10}^{-4}$, and latent dimension of 128. The generator and discriminator are optimized alternately using the Adam optimizer. This setup provides stable training convergence and ensures high-fidelity synthetic samples aligned with real-world feature distributions. In the ultimate dataset, each data sample has nine different features and a class label. The distribution plot of the features on three class labels for the ultimate dataset is shown in Figure 4. Feature selection and reduction methods are not applied on the dataset. Nine features are directly employed while training supervised ML models. To verify the absence of multicollinearity, we calculated the Pearson correlation coefficients between all features. No strong correlations ($\left|\rho \right|>0.8$) were observed, confirming that the features contribute complementary information.

For the proposed algorithm, several ML models are employed such as Gradient Boosting, Neural Networks (NN), Random Forest, and k-Nearest Neighbors (kNN). To ensure robustness, we applied 5-fold cross-validation on the training dataset. The optimized hyperparameters of each model along with their classification accuracy and F1 scores are given in

Table 4. The best hyperparameters and corresponding results for different ML models.

Model	The Optimized Hyperparameters	Accuracy	F1 Score
Gradient Boosting	Learning rate: 0.2 Max depth: 3 Min samples split: 2 Number of estimators: 200	0.821	0.82
Neural Networks (NN)	Activation: ReLu Solver: Adam Alpha: 0.0001 Hidden Layers: 90 Number of iterations: 500	0.777	0.775
Random Forest	Min samples split: 5 Number of estimators: 19	0.774	0.773
k-Nearest Neighbors (kNN)	Metric: Euclidean Number of neighbors: 20	0.712	0.707

. Moreover, the corresponding confusion matrices are shown in Figure 5. The models of Gradient Boosting, NN, Random Forest, and kNN achieve 82.1%, 77.7%, 77.4%, and 71.2% classification accuracies on the ultimate dataset with CTGAN, respectively. The low deviations confirm the stability of the models across folds.

The obtained ML results indicate that the Gradient Boosting model outperforms other models in terms of classification accuracy and F1 score, primarily due to the inherent challenges faced by other models when dealing with tabular data [35]. These challenges result in the classical models performing less effectively compared to the Gradient Boosting model. Furthermore, the total training and optimization time of the classical models is longer, as the Gradient Boosting model requires fewer iterations and, consequently, a shorter runtime.

Since the proposed scheme relies on an ML-based decision module, the inference delay of the employed models must be considered to ensure real-time applicability. In our implementation, four classifiers are evaluated. The average inference latency is measured and the results show that a single-sample inference requires approximately 0.9 ms for Random Forest, 1.5 ms for Gradient Boosting, 3.2 ms for kNN, and 6.7 ms for the NN model. These values remain well below the typical handover preparation time in 5G/NTN systems (20–50 ms). Therefore, the inference latency of the employed ML models does not constitute a bottleneck for the proposed approach as shown in

Table 5. Inference latency of the ML models used in this study (single-sample prediction). Reported values are indicative measurements (average over many runs) for a typical workstation.

Model	Average Inference Latency (ms)	Std. Dev. (ms)
Gradient Boosting	1.50	0.20
Neural Networks (NN)	6.70	0.50
Random Forest	0.90	0.12
k-Nearest Neighbors (kNN)	3.20	0.40

.

Parameter Sensitivity Analysis

In order to further analyze the robustness of the proposed handover control algorithm, a sensitivity study with respect to the control parameters $\alpha$ and $\beta$ in Equations (4) and (5) has been conducted. These parameters play a critical role in determining the responsiveness and stability of the feature update mechanism.

The parameter $\alpha$ controls the balance between the new variance and the historical background variance. Smaller values of $\alpha$ lead to faster adaptation but increase the risk of unnecessary handovers (ping-pong effect). Larger $\alpha$ values, on the other hand, provide stability but may delay the adaptation to sudden environmental changes.

Similarly, the parameter $\beta$ determines the tolerance against sudden variance spikes. A low $\beta$ value makes the system more sensitive to minor fluctuations, potentially triggering excessive handovers. A high $\beta$ value smooths out these effects but may cause the algorithm to overlook significant changes.

The parameters T, $\alpha$, $\beta$, and B were selected based on extensive simulations and prior domain knowledge. Specifically, T (observation window length) was set to 50 ms to align with typical 5G handover preparation intervals. $\alpha$ (variance update weight) was optimized within [0.6, 0.8] to balance responsiveness and stability. $\beta$ (variance tolerance factor) was chosen within [1.3, 1.6] to suppress noise while capturing genuine mobility-induced changes. B (handover decision threshold) was determined empirically to minimize ping-pong events without degrading accuracy. A parameter sensitivity analysis (see Figure 6 and Figure 7) illustrates the performance impact of these parameters. The proposed method achieves the best trade-off when $\alpha$ is set between $0.6$ and $0.8$, and $\beta$ is set between $1.3$ and $1.6$. These parameter ranges minimize the ping-pong effect while preserving high classification accuracy, confirming the effectiveness of the handover control scheme under different operating conditions.

To demonstrate the practical value of the proposed scheme, we consider a high-density vehicular-to-everything (V2X) scenario in an urban environment. In this setting, user density exceeds 3000 vehicles/km², channel conditions are highly dynamic due to mobility and multipath, and both terrestrial and non-terrestrial links must be efficiently managed. Our handover control algorithm successfully reduced the ping-pong ratio by 5% compared to a baseline SINR-threshold method, while maintaining service continuity and low latency. This case confirms the algorithm’s applicability to demanding real-world NTN deployments.

5. Conclusions

AI-empowered IoT systems, such as the one proposed in this study, hold significant promise for 6G communication systems, given the continuous growth in the size and diversity of data generated by IoT applications. In this paper, several human-centric features of smart city IoT networks are considered while proposing an ML-based network decision approach. Simulation results demonstrate that the ML-based approach can effectively facilitate decision-making between TN and NTN networks in 5G and beyond. The proposed dynamic network handover approach addresses the requirements of immersive communication, hyper-reliable and low-latency communication, massive communication, AI-integrated communication, combined sensing and communication, and ubiquitous connectivity in new 6G scenarios.

For future studies, more diverse and informative sensing data inputs can be incorporated into ML models. Subsequently, feature selection and dimensionality reduction techniques may be applied to the dataset to enhance model efficiency. Additionally, a reinforcement learning framework can be developed in a simulated environment to better model wireless channels and communication networks.