Non-Stationary UAV A2G Channel Characterization: From Urban Measurements to Expert-Assisted Neural Modeling

Unmanned aerial vehicles (UAVs) have become indispensable in modern applications ranging from emergency response to wireless relaying; however, characterizing their air-to-ground (A2G) channels in urban environments remains challenging due to complex propagation dynamics such as multipath fading, Doppler effects, and non-stationary conditions. This study addresses these limitations by proposing a comprehensive framework that bridges empirical measurement, statistical analysis, and intelligent modeling. Our core contributions are threefold: First, we conducted an extensive measurement campaign at 70 MHz across representative urban topographies, establishing a high-fidelity empirical dataset that captures time-varying channel responses. Second, we performed a rigorous statistical characterization of both macroscopic and microscopic fading phenomena to enhance conventional path loss models with elevation angle corrections and developed a Markov-chain-based Tapped Delay Line (TDL) model for temporal multipath evolution. Third, we designed an expert-assisted Attention Network (EAANET) platform that synergizes deep learning with physical channel laws, achieving superior real-time path loss prediction accuracy compared to traditional baselines. Experimental validations demonstrate the proposed framework's superior accuracy and operational stability, providing a complete toolkit for urban A2G channel analysis that advances the design and optimization of reliable UAV communication systems.