Channel Knowledge Map Construction With Radio Propagation Graph Representation Learning

This paper presents a novel framework for constructing channel knowledge maps with high accuracy and computational efficiency. The proposed radio propagation graph representation learning (RPGRL) framework integrates physicsbased modeling and data-driven learning through end-to-end optimization using only measured received power data. In RPGRL, both propagation probability and received power estimations are jointly optimized without requiring explicit path supervision, enabling adaptive graph-based representations that account for uncertainties in reflections caused by surface scattering and angular errors introduced by three-dimensional building map quantization. Experimental evaluations using real-world signal measurement data collected in an urban environment demonstrate that RPGRL significantly improves received power estimation accuracy while reducing computation time by more than 70% compared to conventional ray tracing. Furthermore, it outperforms existing deep learning-based methods, such as RadioUNet, by leveraging path-aware estimation grounded in radio propagation physics. These results highlight RPGRL as a promising and practical approach for environment-aware communication and intelligent wireless network design toward 6G.