RadioMapMotion: A Dataset and Benchmark for Proactive Spatio-Temporal Radio Environment Prediction

Radio maps (RMs), which provide location-dependent pathloss estimations, serve as a foundational enabler for proactive and environment-aware communications in 6G networks. However, most existing deep learning-based RM construction methods treat dynamic environments as a series of independent static snapshots, thereby omitting the temporal continuity inherent in signal propagation variations induced by the movement of dynamic entities. To overcome this limitation, we investigate the task of spatio-temporal RM prediction, which entails forecasting a sequence of future maps based on historical observations. A major challenge in this predictive paradigm lies in the absence of datasets that capture the continuous evolution of environmental dynamics. To fill this gap, we introduce RadioMapMotion, the first large-scale public dataset comprising continuous RM sequences generated from physically consistent vehicular trajectories. As a benchmark for this task, we further propose RadioMotionNet, a UNet architecture incorporating Convolutional Long Short-Term Memory (ConvLSTM) modules, tailored for multi-step sequence forecasting. Experimental evaluations show that RadioMotionNet achieves superior prediction accuracy and structural fidelity compared to representative baselines, while maintaining low inference latency, indicating its potential for real-time network deployment.