Beyond Homoscedastic Channel Models: An Uncertainty-Aware Stochastic Neural Network Approach for C-Band Total Path Loss Prediction in Vegetated Scenarios
Academic Article in Scopus
Achieving the multi-layered spectral objectives of sixth-generation (6G) networks requires high-fidelity channel characterization, particularly for C-band deployments in complex vegetated environments supporting cellular vehicle-to-everything (C-V2X) and non-terrestrial networks (NTN). This study presents a physics-aware, heteroscedasticity-driven channel modeling framework that utilizes Variational Gaussian Processes (VGP) to quantify both aleatory and epistemic uncertainties, thereby characterizing the non-stationary dynamics of foliage-induced shadowing. By employing an Exponentiated Quadratic (EQ) kernel, the proposed stochastic architecture captures spatial correlation from real-world measurements at 4.9, 5.85, and 6.1 GHz. The results demonstrate a significant enhancement in predictive accuracy, achieving average Root Mean Square Error (RMSE) reductions of 73%, 76%, 82%, and 75% over traditional Floating-Intercept (FI), Close-In (CI), ABG, and CIF empirical benchmarks, respectively. These improvements are further validated through coverage probability and spread ratio metrics. Our findings establish variance-adaptive probabilistic regression as a robust alternative to the homoscedastic linear assumption, advancing 6G propagation modeling in high-entropy, vegetated scenarios.
