Traditional path loss prediction models can be broadly divided into two classes: deterministic and statistical. Deterministic models, such as ray tracing, rely on physical propagation equations to make very accurate predictions. They require, however, extensive environmental data and significant computational resources, and they are difficult to scale (Hoomod et al., 2018; Akpaida et al., 2018).

Statistical models, including the COST 231-Hata and ECC-33 models, predict path loss trends from empirical data. While being of useful real-world relevance, they generally lack site-specific accuracy (Tarhuni and Ouni, 2022; Zhang et al., 2023). As established by Rappaport (2002) in his foundational work on wireless communications, these statistical models provide a practical compromise between accuracy and implementation complexity for large-scale deployment scenarios.

For 5.9 GHz city-wide V2I communications of critical significance to intelligent transportation systems (Boban et al., 2011), path loss models must be capable of addressing significant challenges presented by dynamic obstacles (e.g., cars, buildings, and foliage), road topology variations, and changing line-of-sight (LOS) conditions. While the classic Hata (1980) model established the foundation for urban path loss prediction, modern V2I systems operate at higher frequencies and in more complex environments, necessitating enhanced modeling approaches. Experimental data indicate that path loss exponents are likely to be approximately 2.0 or higher in LOS scenarios and considerably greater than 3.5 in NLOS situations in dense urban settings (Abbas et al., 2013). This observation aligns with the seminal work by Chang and Yang (1997), who established the fundamental relationship between urban morphology and path loss exponent variations. To better model LOS-to-NLOS transitions, hybrid approaches like two-slope models have been proposed (Fernández et al., 2024).

While deterministic devices such as 3D ray tracing are more effective than others in modeling complex wave interactions such as diffraction and reflection, their computational expense emphasizes the value of balanced techniques with optimized efficiency and accuracy in the design of urban V2I networks (Pätzold and De Nardis, 2020). Fraile et al. (2000) demonstrated that simplified ray tracing models can achieve reasonable accuracy while significantly reducing computational requirements, making them suitable for large-scale network planning applications.

ML has emerged as a powerful tool for improving path loss modeling in wireless communication, offering enhanced accuracy and adaptability over traditional empirical and deterministic models (Goodfellow et al., 2016). ML techniques have been widely applied in various domains, including image recognition (Krizhevsky et al., 2012), natural language processing (Devlin et al., 2019), and wireless communication systems (Zhang et al., 2019). Path loss prediction, a crucial component of wireless network planning, is inherently a regression problem, making it well-suited for supervised learning techniques such as support vector machines, artificial neural networks (ANN), random forests, and K-nearest neighbors (Murphy, 2012). Among these, ANN models have demonstrated superior accuracy in path loss estimation compared to empirical models (Bishop, 2006). For instance, Zhang et al. (2019) developed a real-time channel prediction model that estimates path loss (PL) and packet drop probability in Dedicated Short-Range Communications (DSRC) systems, highlighting the flexibility of ML approaches in dynamic vehicular environments.

The adaptability of ML in path loss modeling is attributed to its ability to learn directly from measured propagation data, allowing models to generalize better across varying environments (Oroza et al., 2017). This approach addresses the limitations of traditional channel models as identified by Molisch et al. (2011), who emphasized the need for context-aware propagation modeling in vehicular communications. Recent comprehensive studies by Huang et al. (2022a) have demonstrated that ML models can effectively capture the complex relationship between environmental features and path loss in V2I scenarios. The follow-up work by Huang et al. (2022b) further validated these findings through extensive field measurements across diverse urban environments. Unlike traditional models that rely on predefined propagation assumptions, ML-based approaches use extensive datasets collected from real-world conditions to train algorithms for accurate signal prediction. Furthermore, ML models incorporate key radio environment factors, including distance, frequency, antenna height, terrain type, and obstacles, to enhance predictive accuracy (Uccellari et al., 2016).

Deep learning, a subset of ML characterized by neural networks with multiple hidden layers, has shown promising results in path loss prediction. Deep neural networks (DNNs) can automatically extract features from raw data, eliminating the need for manual feature engineering (Tarhuni and Ouni, 2022). This capability is particularly valuable in complex urban environments where signal propagation is influenced by numerous factors. For example, Thrane et al. (2020) demonstrated that DNNs outperform traditional path loss models in urban scenarios by capturing intricate relationships between environmental variables and signal attenuation (Gho et al., 2019).

Recurrent Neural Networks and Long Short-Term Memory networks have been applied to path loss prediction in dynamic environments, leveraging their ability to model temporal dependencies in signal propagation (Pätzold and De Nardis, 2020). These approaches are especially relevant for V2I communications, where both the vehicle and the surrounding environment may change rapidly. Additionally, Convolutional Neural Networks have been utilized to process spatial information from environmental maps, enabling more accurate path loss predictions in urban settings (Filippi and Bazzi, 2021).

Despite their advantages, deep learning approaches face challenges such as the need for large training datasets, computational complexity, and potential overfitting (Gozalvez and Sepulcre, 2022). Wu et al. (2020) proposed a novel transfer learning approach to address these limitations by leveraging knowledge from source domains with abundant data to improve prediction accuracy in target domains with limited data. Hybrid models that combine deep learning with traditional path loss models have been proposed to address these limitations, offering a balance between accuracy and computational efficiency (Huang et al., 2022a).

We utilize an open-access dataset (30), which was gathered under a V2I measurement campaign in Bologna. Eight RSU positions as transmitters (Tx) are included in this dataset, which function in between 5.9 GHz. The transmitters were mounted at two different heights: 6.5 and 10.5 m. The receiver (Rx) was mounted on the roof of a car at a height of 1.75 m. The dataset includes RSSI measurements, which were converted to path loss using the formula shown in Equation 1:

where P Tx is the transmission power, and P Rx is the received power, which in this context is the measured RSSI value, L cable represents the cable loss, and G Tx and G Rx are the Tx and Rx antenna gains, respectively.

The dataset includes measurements from various urban environments, including open areas, narrow streets, and areas with dense buildings. GPS coordinates for both transmitters and receivers were recorded, allowing for accurate distance calculations and environmental classification.

To ensure data quality and reliability prior to model training, several preprocessing techniques were applied to the raw V2I communication dataset. These techniques were designed to handle noise, normalize feature scales, and preserve the statistical distribution of environmental classes across training, validation, and testing sets.

Feature importance analysis was conducted to identify the most significant factors affecting path loss prediction accuracy. Figure 2 shows the relative importance of different features in the XGBoost model.

The analysis reveals that distance, environmental class, and transmitter height are the most significant factors affecting path loss prediction accuracy. While receiver latitude and longitude also show high importance, this is primarily because they implicitly capture the road network geometry and building distribution patterns in our study areas.

Regarding the use of latitude and longitude as features, we acknowledge that an intrinsic coordinate system (such as a polar coordinate system with radius = distance and angle with origin at the RSU might be theoretically more elegant. However, we found that latitude and longitude provided practical advantages in our specific dataset, as they inherently encode spatial relationships between the transmitter and receiver within the urban landscape. Future work could explore alternative coordinate systems to potentially improve model performance further. The most critical hyperparameters include in Table 1.

We classify the urban environments into three categories based on building density, street layout, and vegetation coverage:

This classification system provides a structured approach to understanding how different urban characteristics affect signal propagation and path loss prediction.

This section presents a comprehensive overview of the four path loss prediction models evaluated in this study, each representing a different class of modeling approaches. As shown in Figure 3, we begin with the dual slope model, an empirical formulation known for its simplicity and effectiveness in capturing distinct propagation characteristics at varying distance intervals. This is followed by the 3GPP model, a standardized model widely adopted in cellular communication research. The last two models, XGBoost and MLP, belong to the machine learning domain and leverage data-driven training to enhance prediction accuracy. Together, these models provide a diverse analytical foundation for evaluating path loss behavior in V2I communication scenarios.

Figure 4 illustrates the path loss behavior in an open area with minimal obstructions and compares the performance of machine learning models with traditional approaches. XGBoost performed exceptionally well, achieving an RMSE of just 0.25 dB and an R² of 0.94, outperforming all other models. The MLP model performed well, accurately identifying subtle propagation effects, including ground reflections, with an RMSE of 0.43 dB and an R² of 0.83. In contrast, the traditional models lagged considerably: the Dual Slope model achieved an RMSE of 4.68 dB and an R² of 0.61, whereas the 3GPP TR 38.901 model performed much worse, with an RMSE of 16.18 dB and an R² of −3.47, indicating performance worse than a naive mean-value prediction. These results highlight the shortcomings of static, rule-based models for use in open settings where complicated and multidimensional signal propagation dynamics are the main concern

Figure 5 illustrates the path loss patterns in a suburban environment with moderate building density and partial obstructions. XGBoost demonstrated the highest accuracy, achieving an RMSE of 0.29 dB and an R² of 0.92, while the MLP model also performed well, with an RMSE of 0.41 dB and an R² of 0.83. In contrast, traditional models showed significantly lower accuracy: the Dual Slope model had an RMSE of 5.55 dB and an R² of 0.52, and the 3GPP TR 38.901 model performed poorly, with an RMSE of 12.29 dB and a negative R² of −1.32. These results highlight the superior adaptability of machine learning models in handling the variability of suburban environments—such as the presence of trees and mid-rise buildings—compared to the static nature of traditional, formula-based models.

Figure 6 illustrates path loss behavior in dense urban areas characterized by tall buildings and pronounced multipath effects. XGBoost delivered outstanding performance, achieving an RMSE of 0.22 dB and an R² of 0.95, while the MLP model also performed well, with an RMSE of 0.37 dB and an R² of 0.86. In contrast, traditional models struggled to represent the complexity of the environment: the Dual Slope model recorded an RMSE of 6.79 dB and an R² of 0.64, while the 3GPP TR 38.901 model fared the worst, with an RMSE of 14.84 dB and a negative R² of −0.72. These results underscore XGBoost’s ability to capture nonlinear propagation effects such as shadowing and diffraction and clearly reveal the limitations of static, formula-based models like 3GPP in accurately modeling signal behavior in complex urban settings.

Our analysis indicates that environmental factors significantly influence path loss characteristics. Table 3 summarizes the critical variables affecting signal reliability in different urban environments and proposes specific actions for improving connectivity. Gozalvez et al. (2012) similarly observed that environmental context awareness is crucial for optimizing V2I communication systems, particularly in heterogeneous urban settings. In open urban environments, transmission power optimization can significantly improve connectivity. However, in dense urban environments with sharp road curves, increasing transmission power provides minimal benefits, and strategic RSU placement becomes more critical.