T-MAP: A Motion-Aware Dual-Stream Framework for High-Dynamic Directional Spectrum Map Prediction
Abstract
Spectrum map prediction plays a critical role in reliable vehicular perception for intelligent transportation systems (ITS). However, existing methods rely on an isotropic energy diffusion assumption, which fails to capture the severe nonstationarity and blurring caused by the dual-dynamic nature of V2X scenarios: high-speed carrier mobility and rapid directional beam scanning. To address this, we propose T-MAP, a motionaware dual-stream prediction framework designed to resolve the inherent conflict between morphological appearance maintenance and kinematic trajectory tracking. Unlike black-box regression models, T-MAP explicitly decouples spatial features and employs a motion-aware matrix to transform the prediction task into the inference of electromagnetic energy transition probabilities. By integrating a parallelized temporal attention unit (TAU), the framework bypasses the inductive bias of traditional recurrent structures, enabling precise capture of long-range beam rotation trends. Experimental results on a high-fidelity V2X simulation dataset demonstrate that T-MAP significantly outperforms stateof-the-art baselines, achieving superior structural fidelity and robustness in tracking high-dynamic directional signals.