Sionna 레이 트레이싱 기반 건물 마스크 조건부 CNN을 이용한 전파 환경 지도의 공간 초해상도

초록

Radio environment maps (REMs) are essential for beam management and coverage planning in sixth-generation networks, but dense measurements are costly, so an REM is often available only on a coarse grid. This paper proposes a spatial super-resolution method for reconstructing a high-resolution REM from such an observation. Besides the interpolated low-resolution REM, the proposed lightweight residual convolutional network takes a building mask, an analytic free-space power map, and a normalized transmitter-distance map as conditioning channels, all obtainable from a map database. It is trained with a validity-masked reconstruction loss and a gradient-matching loss on data generated by NVIDIA Sionna RT ray tracing over distinct urban scenes. For a down-sampling factor of four, the method lowers the root-mean-square error of bicubic interpolation by 13.9 % on seen scenes and 7.9 % on unseen ones, and raises the coverage intersection-over-union on unseen scenes from 0.72 to 0.79. An ablation study shows that the building mask accounts for most of the gain.

키워드