Video Generation Empowered Long-Term Radio Map Prediction in UAV-Assisted Communication

Radio maps provide rich information about the radio landscape that is useful to a myriad of wireless communication and sensing applications. In contrast to the scenario with fixed-position transmitters, the transmitter in the unmanned aerial vehicle (UAV)-assisted communication scenario can move along a predefined trajectory, resulting in a continuously evolving \emph{long-term radio map}. To address this long-term radio map prediction problem, existing deep learning-based methods that follow the image-to-image translation approach suffer from significant complexity and limited accuracy due to their independent frame-by-frame construction method, which neglects the temporal correlation of radio characteristics. In this work, we formulate the long-term radio map prediction problem in the UAV-assisted communication system as a \emph{conditioned video generation} task. In particular, we propose a \emph{lightweight} model that encodes the static environment only once, conditions radio map prediction on both the environment and the UAV trajectory, and refines the generated sequence in the temporal domain by exploiting double-sided dependencies. Experiments on a large-scale simulated dataset demonstrate that the proposed method consistently improves prediction accuracy and temporal consistency while reducing computational redundancy compared with both static baselines and sequential baselines. The source code is available at \href{https://github.com/LINZHU-PolyU/DynamicRadioMap}{GitHub repository}.