Physics-Informed Transfer Learning for Data-Efficient Radio Environment Maps in High-Speed Railway Communication Networks

Radio environment map (REM) enables rapid access to channel state information through precise electromagnetic signal propagation modeling. However, REM estimation in high-speed railway (HSR) scenes faces challenges of sparse training data and insufficient cross-scene generalization capability. To address these challenges, we propose an efficient cross-scene REM estimation framework that integrates physical information with transfer learning (TL), which effectively leverages inherent commonalities across HSR scenes to achieve knowledge transfer. First, we propose a scene classification algorithm based on density clustering of scene features, achieving rapid scene segmentation by fusing multi-level geographic features. We then combine the Okumura-Hata model with distance scaling factors and the spatial loss field theory to capture universal macroscopic fading patterns. Additionally, we employ a residual prediction network that learns nonlinear heterogeneous propagation patterns from multi-source features. Finally, we construct a progressive TL mechanism that defines knowledge transfer as the transmission of universal physical information. The mechanism pre-trains on data-rich scenes to learn scene-invariant common laws, then efficiently adapts to data-sparse environments by fine-tuning with minimal target samples to focus on unique environmental layouts and fading patterns. Experimental results based on real HSR measurement data demonstrate that the proposed REM estimation framework significantly outperforms benchmark methods in both prediction accuracy and model generalization.