Data-Driven Fusion of Gaussian Process Regression and 3GPP Models for Urban Path Loss Prediction

Abstract

Accurate cell-specific signal strength maps are essential for coverage analysis and interference assessment in dense urban networks. However, drive-test measurements are typically spatially sparse, leaving many street segments unobserved. This paper proposes a physics-informed hybrid model for street-level pathloss prediction in dense urban environments. We employ a calibrated antenna and pathloss model as a physical mean function, while graph-based Gaussian Process Regression models residual shadowing along the street topology. Using measurements from 107 cells, the approach transforms unreliable planning data—starting with an initial RMSE of over 17 dB—into precise radio footprints. The proposed model achieves an overall MAE of 3.6 dB and an LOO-MAE of 3.2 dB. The results indicate that combining physical calibration with topology-aware residual learning produces plausible, generalized cell footprints even from sparse measurement data.