Propagation-Informed Pathloss Radiomap Prediction with Cross-City and Multi-Band Generalization

Abstract

Radiomap prediction provides a high-resolution spatial representation of channel metrics that is essential for network planning, interference management, and resource allocation. Empirical and statistical models lack site specificity, whereas deterministic ray tracing (RT) is too computationally expensive for large-scale or real-time use. We present a deep-learning framework that predicts the pathloss radiomap of an urban region directly from geographic information system (GIS) data and base-station (BS) engineering parameters. The framework decouples the task into two stages: a propagation feature extraction module that converts raw topography, building layouts, and BS configurations into eight physics-informed feature maps, and a U-Net regression module that reconstructs the pathloss radiomap. The model is trained and evaluated on a large-scale RT-simulation dataset spanning seven Chinese cities and four carrier frequencies from 900 MHz to 3.5 GHz. On held-out test regions the model attains an outdoor root-mean-square error (RMSE) of 3.92 dB while predicting a full map in approximately 20 ms, several orders of magnitude faster than RT. Evaluated on three additional cities absent from training, the model retains an outdoor RMSE of 3.74 dB, demonstrating that the propagation-informed representation transfers across urban morphologies; a single jointly-trained model further remains accurate across all four bands. We quantify the contribution of each propagation feature through an ablation study.