PhySNet: Physics-Inspired Spectral–Spatial Decoupled Attention for Radio Map Estimation

Abstract

Radio map estimation (RME) plays a critical role in spectrum management and network planning for 5G and beyond. In dense urban environments, however, the received signal strength (RSS) exhibits complex multi-scale characteristics, where smooth large-scale path loss coexists with sharp local variations caused by geometry-induced line-of-sight/non-line-of-sight (LOS/NLOS) transitions. Such heterogeneity poses a significant challenge for RME, especially under sparse sampling conditions. Existing interpolation- and learning-based methods often suffer from severe performance degradation, as they lack intrinsic physical propagation priors to effectively model these heterogeneous patterns. To address this challenge, we propose PhySNet, a physics-inspired framework that explicitly disentangles large-scale trends and local variations through FFT-based spectral decomposition. The resulting decoupled components are then utilized as physics-guided queries in a novel Spectral Cross-Window Attention (SCWA) mechanism, enabling efficient spatial feature aggregation. Subsequently, a spatially aware fusion network further refines the aggregated features by incorporating local environmental semantics. Experimental results on ray-tracing-based urban simulation datasets demonstrate that PhySNet consistently outperforms state-of-the-art methods in terms of RMSE and spatial consistency, offering a robust and scalable solution for rapid radio map construction in complex electromagnetic environments.